2008-11-20 20:01:55 +00:00
|
|
|
/*
|
|
|
|
* CDDL HEADER START
|
|
|
|
*
|
|
|
|
* The contents of this file are subject to the terms of the
|
|
|
|
* Common Development and Distribution License (the "License").
|
|
|
|
* You may not use this file except in compliance with the License.
|
|
|
|
*
|
|
|
|
* You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
|
|
|
|
* or http://www.opensolaris.org/os/licensing.
|
|
|
|
* See the License for the specific language governing permissions
|
|
|
|
* and limitations under the License.
|
|
|
|
*
|
|
|
|
* When distributing Covered Code, include this CDDL HEADER in each
|
|
|
|
* file and include the License file at usr/src/OPENSOLARIS.LICENSE.
|
|
|
|
* If applicable, add the following below this CDDL HEADER, with the
|
|
|
|
* fields enclosed by brackets "[]" replaced with your own identifying
|
|
|
|
* information: Portions Copyright [yyyy] [name of copyright owner]
|
|
|
|
*
|
|
|
|
* CDDL HEADER END
|
|
|
|
*/
|
|
|
|
/*
|
2010-05-28 20:45:14 +00:00
|
|
|
* Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
|
2015-07-02 16:23:20 +00:00
|
|
|
* Copyright (c) 2012, 2015 by Delphix. All rights reserved.
|
2015-04-02 03:44:32 +00:00
|
|
|
* Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
|
2008-11-20 20:01:55 +00:00
|
|
|
*/
|
|
|
|
|
|
|
|
#include <sys/zfs_context.h>
|
|
|
|
#include <sys/dbuf.h>
|
|
|
|
#include <sys/dnode.h>
|
|
|
|
#include <sys/dmu.h>
|
|
|
|
#include <sys/dmu_impl.h>
|
|
|
|
#include <sys/dmu_tx.h>
|
|
|
|
#include <sys/dmu_objset.h>
|
|
|
|
#include <sys/dsl_dir.h>
|
|
|
|
#include <sys/dsl_dataset.h>
|
|
|
|
#include <sys/spa.h>
|
|
|
|
#include <sys/zio.h>
|
|
|
|
#include <sys/dmu_zfetch.h>
|
2014-04-16 03:40:22 +00:00
|
|
|
#include <sys/range_tree.h>
|
2014-12-13 02:07:39 +00:00
|
|
|
#include <sys/trace_dnode.h>
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
static kmem_cache_t *dnode_cache;
|
2010-08-26 21:24:34 +00:00
|
|
|
/*
|
|
|
|
* Define DNODE_STATS to turn on statistic gathering. By default, it is only
|
|
|
|
* turned on when DEBUG is also defined.
|
|
|
|
*/
|
|
|
|
#ifdef DEBUG
|
|
|
|
#define DNODE_STATS
|
|
|
|
#endif /* DEBUG */
|
|
|
|
|
|
|
|
#ifdef DNODE_STATS
|
|
|
|
#define DNODE_STAT_ADD(stat) ((stat)++)
|
|
|
|
#else
|
|
|
|
#define DNODE_STAT_ADD(stat) /* nothing */
|
|
|
|
#endif /* DNODE_STATS */
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2010-08-26 16:53:00 +00:00
|
|
|
ASSERTV(static dnode_phys_t dnode_phys_zero);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
int zfs_default_bs = SPA_MINBLOCKSHIFT;
|
|
|
|
int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
|
|
|
|
|
2010-08-27 17:55:07 +00:00
|
|
|
#ifdef _KERNEL
|
2010-08-26 21:24:34 +00:00
|
|
|
static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
|
2010-08-27 17:55:07 +00:00
|
|
|
#endif /* _KERNEL */
|
2010-08-26 21:24:34 +00:00
|
|
|
|
2015-04-03 03:14:28 +00:00
|
|
|
static int
|
|
|
|
dbuf_compare(const void *x1, const void *x2)
|
|
|
|
{
|
|
|
|
const dmu_buf_impl_t *d1 = x1;
|
|
|
|
const dmu_buf_impl_t *d2 = x2;
|
|
|
|
|
2016-08-27 18:12:53 +00:00
|
|
|
int cmp = AVL_CMP(d1->db_level, d2->db_level);
|
|
|
|
if (likely(cmp))
|
|
|
|
return (cmp);
|
2015-04-03 03:14:28 +00:00
|
|
|
|
2016-08-27 18:12:53 +00:00
|
|
|
cmp = AVL_CMP(d1->db_blkid, d2->db_blkid);
|
|
|
|
if (likely(cmp))
|
|
|
|
return (cmp);
|
2015-04-03 03:14:28 +00:00
|
|
|
|
2015-05-06 17:08:25 +00:00
|
|
|
if (d1->db_state == DB_SEARCH) {
|
|
|
|
ASSERT3S(d2->db_state, !=, DB_SEARCH);
|
2015-04-03 03:14:28 +00:00
|
|
|
return (-1);
|
2015-05-06 17:08:25 +00:00
|
|
|
} else if (d2->db_state == DB_SEARCH) {
|
|
|
|
ASSERT3S(d1->db_state, !=, DB_SEARCH);
|
2015-04-01 15:10:58 +00:00
|
|
|
return (1);
|
|
|
|
}
|
|
|
|
|
2016-08-27 18:12:53 +00:00
|
|
|
return (AVL_PCMP(d1, d2));
|
2015-04-03 03:14:28 +00:00
|
|
|
}
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
/* ARGSUSED */
|
|
|
|
static int
|
|
|
|
dnode_cons(void *arg, void *unused, int kmflag)
|
|
|
|
{
|
|
|
|
dnode_t *dn = arg;
|
2010-08-26 21:24:34 +00:00
|
|
|
int i;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
Identify locks flagged by lockdep
When running a kernel with CONFIG_LOCKDEP=y, lockdep reports possible
recursive locking in some cases and possible circular locking dependency
in others, within the SPL and ZFS modules.
This patch uses a mutex type defined in SPL, MUTEX_NOLOCKDEP, to mark
such mutexes when they are initialized. This mutex type causes
attempts to take or release those locks to be wrapped in lockdep_off()
and lockdep_on() calls to silence the dependency checker and allow the
use of lock_stats to examine contention.
For RW locks, it uses an analogous lock type, RW_NOLOCKDEP.
The goal is that these locks are ultimately changed back to type
MUTEX_DEFAULT or RW_DEFAULT, after the locks are annotated to reflect
their relationship (e.g. z_name_lock below) or any real problem with the
lock dependencies are fixed.
Some of the affected locks are:
tc_open_lock:
=============
This is an array of locks, all with same name, which txg_quiesce must
take all of in order to move txg to next state. All default to the same
lockdep class, and so to lockdep appears recursive.
zp->z_name_lock:
================
In zfs_rmdir,
dzp = znode for the directory (input to zfs_dirent_lock)
zp = znode for the entry being removed (output of zfs_dirent_lock)
zfs_rmdir()->zfs_dirent_lock() takes z_name_lock in dzp
zfs_rmdir() takes z_name_lock in zp
Since both dzp and zp are type znode_t, the locks have the same default
class, and lockdep considers it a possible recursive lock attempt.
l->l_rwlock:
============
zap_expand_leaf() sometimes creates two new zap leaf structures, via
these call paths:
zap_deref_leaf()->zap_get_leaf_byblk()->zap_leaf_open()
zap_expand_leaf()->zap_create_leaf()->zap_expand_leaf()->zap_create_leaf()
Because both zap_leaf_open() and zap_create_leaf() initialize
l->l_rwlock in their (separate) leaf structures, the lockdep class is
the same, and the linux kernel believes these might both be the same
lock, and emits a possible recursive lock warning.
Signed-off-by: Olaf Faaland <faaland1@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3895
2015-10-15 20:08:27 +00:00
|
|
|
rw_init(&dn->dn_struct_rwlock, NULL, RW_NOLOCKDEP, NULL);
|
2008-11-20 20:01:55 +00:00
|
|
|
mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
|
|
|
|
mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
|
2009-01-15 21:59:39 +00:00
|
|
|
cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
|
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
/*
|
|
|
|
* Every dbuf has a reference, and dropping a tracked reference is
|
|
|
|
* O(number of references), so don't track dn_holds.
|
|
|
|
*/
|
|
|
|
refcount_create_untracked(&dn->dn_holds);
|
2008-11-20 20:01:55 +00:00
|
|
|
refcount_create(&dn->dn_tx_holds);
|
2010-08-26 21:24:34 +00:00
|
|
|
list_link_init(&dn->dn_link);
|
|
|
|
|
|
|
|
bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
|
|
|
|
bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
|
|
|
|
bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
|
|
|
|
bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
|
|
|
|
bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
|
|
|
|
bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
|
|
|
|
bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
for (i = 0; i < TXG_SIZE; i++) {
|
2010-08-26 21:24:34 +00:00
|
|
|
list_link_init(&dn->dn_dirty_link[i]);
|
2014-04-16 03:40:22 +00:00
|
|
|
dn->dn_free_ranges[i] = NULL;
|
2008-11-20 20:01:55 +00:00
|
|
|
list_create(&dn->dn_dirty_records[i],
|
|
|
|
sizeof (dbuf_dirty_record_t),
|
|
|
|
offsetof(dbuf_dirty_record_t, dr_dirty_node));
|
|
|
|
}
|
|
|
|
|
2010-08-26 21:24:34 +00:00
|
|
|
dn->dn_allocated_txg = 0;
|
|
|
|
dn->dn_free_txg = 0;
|
|
|
|
dn->dn_assigned_txg = 0;
|
|
|
|
dn->dn_dirtyctx = 0;
|
|
|
|
dn->dn_dirtyctx_firstset = NULL;
|
|
|
|
dn->dn_bonus = NULL;
|
|
|
|
dn->dn_have_spill = B_FALSE;
|
|
|
|
dn->dn_zio = NULL;
|
|
|
|
dn->dn_oldused = 0;
|
|
|
|
dn->dn_oldflags = 0;
|
|
|
|
dn->dn_olduid = 0;
|
|
|
|
dn->dn_oldgid = 0;
|
|
|
|
dn->dn_newuid = 0;
|
|
|
|
dn->dn_newgid = 0;
|
|
|
|
dn->dn_id_flags = 0;
|
|
|
|
|
|
|
|
dn->dn_dbufs_count = 0;
|
2013-08-21 04:11:52 +00:00
|
|
|
dn->dn_unlisted_l0_blkid = 0;
|
2015-04-03 03:14:28 +00:00
|
|
|
avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
|
2008-11-20 20:01:55 +00:00
|
|
|
offsetof(dmu_buf_impl_t, db_link));
|
|
|
|
|
2010-08-26 21:24:34 +00:00
|
|
|
dn->dn_moved = 0;
|
2008-11-20 20:01:55 +00:00
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* ARGSUSED */
|
|
|
|
static void
|
|
|
|
dnode_dest(void *arg, void *unused)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
dnode_t *dn = arg;
|
|
|
|
|
|
|
|
rw_destroy(&dn->dn_struct_rwlock);
|
|
|
|
mutex_destroy(&dn->dn_mtx);
|
|
|
|
mutex_destroy(&dn->dn_dbufs_mtx);
|
2009-01-15 21:59:39 +00:00
|
|
|
cv_destroy(&dn->dn_notxholds);
|
2008-11-20 20:01:55 +00:00
|
|
|
refcount_destroy(&dn->dn_holds);
|
|
|
|
refcount_destroy(&dn->dn_tx_holds);
|
2010-08-26 21:24:34 +00:00
|
|
|
ASSERT(!list_link_active(&dn->dn_link));
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
for (i = 0; i < TXG_SIZE; i++) {
|
2010-08-26 21:24:34 +00:00
|
|
|
ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
|
2014-04-16 03:40:22 +00:00
|
|
|
ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
|
2008-11-20 20:01:55 +00:00
|
|
|
list_destroy(&dn->dn_dirty_records[i]);
|
2013-05-10 21:17:03 +00:00
|
|
|
ASSERT0(dn->dn_next_nblkptr[i]);
|
|
|
|
ASSERT0(dn->dn_next_nlevels[i]);
|
|
|
|
ASSERT0(dn->dn_next_indblkshift[i]);
|
|
|
|
ASSERT0(dn->dn_next_bonustype[i]);
|
|
|
|
ASSERT0(dn->dn_rm_spillblk[i]);
|
|
|
|
ASSERT0(dn->dn_next_bonuslen[i]);
|
|
|
|
ASSERT0(dn->dn_next_blksz[i]);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
2013-05-10 21:17:03 +00:00
|
|
|
ASSERT0(dn->dn_allocated_txg);
|
|
|
|
ASSERT0(dn->dn_free_txg);
|
|
|
|
ASSERT0(dn->dn_assigned_txg);
|
|
|
|
ASSERT0(dn->dn_dirtyctx);
|
2010-08-26 21:24:34 +00:00
|
|
|
ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
|
|
|
|
ASSERT3P(dn->dn_bonus, ==, NULL);
|
|
|
|
ASSERT(!dn->dn_have_spill);
|
|
|
|
ASSERT3P(dn->dn_zio, ==, NULL);
|
2013-05-10 21:17:03 +00:00
|
|
|
ASSERT0(dn->dn_oldused);
|
|
|
|
ASSERT0(dn->dn_oldflags);
|
|
|
|
ASSERT0(dn->dn_olduid);
|
|
|
|
ASSERT0(dn->dn_oldgid);
|
|
|
|
ASSERT0(dn->dn_newuid);
|
|
|
|
ASSERT0(dn->dn_newgid);
|
|
|
|
ASSERT0(dn->dn_id_flags);
|
|
|
|
|
|
|
|
ASSERT0(dn->dn_dbufs_count);
|
2013-08-21 04:11:52 +00:00
|
|
|
ASSERT0(dn->dn_unlisted_l0_blkid);
|
2015-04-03 03:14:28 +00:00
|
|
|
avl_destroy(&dn->dn_dbufs);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
void
|
|
|
|
dnode_init(void)
|
|
|
|
{
|
2010-08-26 21:24:34 +00:00
|
|
|
ASSERT(dnode_cache == NULL);
|
2011-11-01 23:56:48 +00:00
|
|
|
dnode_cache = kmem_cache_create("dnode_t", sizeof (dnode_t),
|
2014-05-15 01:17:39 +00:00
|
|
|
0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
|
2010-08-26 21:24:34 +00:00
|
|
|
kmem_cache_set_move(dnode_cache, dnode_move);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
void
|
|
|
|
dnode_fini(void)
|
|
|
|
{
|
|
|
|
kmem_cache_destroy(dnode_cache);
|
2010-08-26 21:24:34 +00:00
|
|
|
dnode_cache = NULL;
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
#ifdef ZFS_DEBUG
|
|
|
|
void
|
|
|
|
dnode_verify(dnode_t *dn)
|
|
|
|
{
|
|
|
|
int drop_struct_lock = FALSE;
|
|
|
|
|
|
|
|
ASSERT(dn->dn_phys);
|
|
|
|
ASSERT(dn->dn_objset);
|
2010-08-26 21:24:34 +00:00
|
|
|
ASSERT(dn->dn_handle->dnh_dnode == dn);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2012-12-13 23:24:15 +00:00
|
|
|
ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
|
|
|
|
return;
|
|
|
|
|
|
|
|
if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
|
|
|
|
rw_enter(&dn->dn_struct_rwlock, RW_READER);
|
|
|
|
drop_struct_lock = TRUE;
|
|
|
|
}
|
|
|
|
if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
|
|
|
|
int i;
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
int max_bonuslen = DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots);
|
2008-11-20 20:01:55 +00:00
|
|
|
ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
|
|
|
|
if (dn->dn_datablkshift) {
|
|
|
|
ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
|
|
|
|
ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
|
|
|
|
ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
|
|
|
|
}
|
|
|
|
ASSERT3U(dn->dn_nlevels, <=, 30);
|
2012-12-13 23:24:15 +00:00
|
|
|
ASSERT(DMU_OT_IS_VALID(dn->dn_type));
|
2008-11-20 20:01:55 +00:00
|
|
|
ASSERT3U(dn->dn_nblkptr, >=, 1);
|
|
|
|
ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
ASSERT3U(dn->dn_bonuslen, <=, max_bonuslen);
|
2008-11-20 20:01:55 +00:00
|
|
|
ASSERT3U(dn->dn_datablksz, ==,
|
|
|
|
dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
|
|
|
|
ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
|
|
|
|
ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
dn->dn_bonuslen, <=, max_bonuslen);
|
2008-11-20 20:01:55 +00:00
|
|
|
for (i = 0; i < TXG_SIZE; i++) {
|
|
|
|
ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
if (dn->dn_phys->dn_type != DMU_OT_NONE)
|
|
|
|
ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
|
2009-07-02 22:44:48 +00:00
|
|
|
ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
|
2008-11-20 20:01:55 +00:00
|
|
|
if (dn->dn_dbuf != NULL) {
|
|
|
|
ASSERT3P(dn->dn_phys, ==,
|
|
|
|
(dnode_phys_t *)dn->dn_dbuf->db.db_data +
|
|
|
|
(dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
|
|
|
|
}
|
|
|
|
if (drop_struct_lock)
|
|
|
|
rw_exit(&dn->dn_struct_rwlock);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
|
|
|
void
|
|
|
|
dnode_byteswap(dnode_phys_t *dnp)
|
|
|
|
{
|
|
|
|
uint64_t *buf64 = (void*)&dnp->dn_blkptr;
|
|
|
|
int i;
|
|
|
|
|
|
|
|
if (dnp->dn_type == DMU_OT_NONE) {
|
|
|
|
bzero(dnp, sizeof (dnode_phys_t));
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
|
|
|
dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
|
|
|
|
dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
dnp->dn_extra_slots = BSWAP_8(dnp->dn_extra_slots);
|
2008-11-20 20:01:55 +00:00
|
|
|
dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
|
|
|
|
dnp->dn_used = BSWAP_64(dnp->dn_used);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* dn_nblkptr is only one byte, so it's OK to read it in either
|
|
|
|
* byte order. We can't read dn_bouslen.
|
|
|
|
*/
|
|
|
|
ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
|
|
|
|
ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
|
|
|
|
for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
|
|
|
|
buf64[i] = BSWAP_64(buf64[i]);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* OK to check dn_bonuslen for zero, because it won't matter if
|
|
|
|
* we have the wrong byte order. This is necessary because the
|
|
|
|
* dnode dnode is smaller than a regular dnode.
|
|
|
|
*/
|
|
|
|
if (dnp->dn_bonuslen != 0) {
|
|
|
|
/*
|
|
|
|
* Note that the bonus length calculated here may be
|
|
|
|
* longer than the actual bonus buffer. This is because
|
|
|
|
* we always put the bonus buffer after the last block
|
|
|
|
* pointer (instead of packing it against the end of the
|
|
|
|
* dnode buffer).
|
|
|
|
*/
|
|
|
|
int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
int slots = dnp->dn_extra_slots + 1;
|
|
|
|
size_t len = DN_SLOTS_TO_BONUSLEN(slots) - off;
|
2012-12-13 23:24:15 +00:00
|
|
|
dmu_object_byteswap_t byteswap;
|
|
|
|
ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
|
|
|
|
byteswap = DMU_OT_BYTESWAP(dnp->dn_bonustype);
|
|
|
|
dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
2010-05-28 20:45:14 +00:00
|
|
|
|
|
|
|
/* Swap SPILL block if we have one */
|
|
|
|
if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
byteswap_uint64_array(DN_SPILL_BLKPTR(dnp), sizeof (blkptr_t));
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
void
|
|
|
|
dnode_buf_byteswap(void *vbuf, size_t size)
|
|
|
|
{
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
int i = 0;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
|
|
|
|
ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
|
|
|
|
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
while (i < size) {
|
|
|
|
dnode_phys_t *dnp = vbuf + i;
|
|
|
|
dnode_byteswap(dnp);
|
|
|
|
|
|
|
|
i += DNODE_MIN_SIZE;
|
|
|
|
if (dnp->dn_type != DMU_OT_NONE)
|
|
|
|
i += dnp->dn_extra_slots * DNODE_MIN_SIZE;
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
void
|
|
|
|
dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
|
|
|
|
{
|
|
|
|
ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
|
|
|
|
|
|
|
|
dnode_setdirty(dn, tx);
|
|
|
|
rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
ASSERT3U(newsize, <=, DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
|
2008-11-20 20:01:55 +00:00
|
|
|
(dn->dn_nblkptr-1) * sizeof (blkptr_t));
|
|
|
|
dn->dn_bonuslen = newsize;
|
|
|
|
if (newsize == 0)
|
|
|
|
dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
|
|
|
|
else
|
|
|
|
dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
|
|
|
|
rw_exit(&dn->dn_struct_rwlock);
|
|
|
|
}
|
|
|
|
|
2010-05-28 20:45:14 +00:00
|
|
|
void
|
|
|
|
dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
|
|
|
|
{
|
|
|
|
ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
|
|
|
|
dnode_setdirty(dn, tx);
|
|
|
|
rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
|
|
|
|
dn->dn_bonustype = newtype;
|
|
|
|
dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
|
|
|
|
rw_exit(&dn->dn_struct_rwlock);
|
|
|
|
}
|
|
|
|
|
|
|
|
void
|
|
|
|
dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
|
|
|
|
{
|
|
|
|
ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
|
|
|
|
ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
|
|
|
|
dnode_setdirty(dn, tx);
|
|
|
|
dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
|
|
|
|
dn->dn_have_spill = B_FALSE;
|
|
|
|
}
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
static void
|
|
|
|
dnode_setdblksz(dnode_t *dn, int size)
|
|
|
|
{
|
2013-05-10 21:17:03 +00:00
|
|
|
ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
|
2008-11-20 20:01:55 +00:00
|
|
|
ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
|
|
|
|
ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
|
|
|
|
ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
|
|
|
|
1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
|
|
|
|
dn->dn_datablksz = size;
|
|
|
|
dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
|
2014-04-16 03:40:22 +00:00
|
|
|
dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static dnode_t *
|
2010-05-28 20:45:14 +00:00
|
|
|
dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
|
2010-08-26 21:24:34 +00:00
|
|
|
uint64_t object, dnode_handle_t *dnh)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
2015-04-02 03:44:32 +00:00
|
|
|
dnode_t *dn;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2015-04-02 03:44:32 +00:00
|
|
|
dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
|
2010-08-26 21:24:34 +00:00
|
|
|
ASSERT(!POINTER_IS_VALID(dn->dn_objset));
|
|
|
|
dn->dn_moved = 0;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Defer setting dn_objset until the dnode is ready to be a candidate
|
|
|
|
* for the dnode_move() callback.
|
|
|
|
*/
|
2008-11-20 20:01:55 +00:00
|
|
|
dn->dn_object = object;
|
|
|
|
dn->dn_dbuf = db;
|
2010-08-26 21:24:34 +00:00
|
|
|
dn->dn_handle = dnh;
|
2008-11-20 20:01:55 +00:00
|
|
|
dn->dn_phys = dnp;
|
|
|
|
|
2010-08-26 21:24:34 +00:00
|
|
|
if (dnp->dn_datablkszsec) {
|
2008-11-20 20:01:55 +00:00
|
|
|
dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
|
2010-08-26 21:24:34 +00:00
|
|
|
} else {
|
|
|
|
dn->dn_datablksz = 0;
|
|
|
|
dn->dn_datablkszsec = 0;
|
|
|
|
dn->dn_datablkshift = 0;
|
|
|
|
}
|
2008-11-20 20:01:55 +00:00
|
|
|
dn->dn_indblkshift = dnp->dn_indblkshift;
|
|
|
|
dn->dn_nlevels = dnp->dn_nlevels;
|
|
|
|
dn->dn_type = dnp->dn_type;
|
|
|
|
dn->dn_nblkptr = dnp->dn_nblkptr;
|
|
|
|
dn->dn_checksum = dnp->dn_checksum;
|
|
|
|
dn->dn_compress = dnp->dn_compress;
|
|
|
|
dn->dn_bonustype = dnp->dn_bonustype;
|
|
|
|
dn->dn_bonuslen = dnp->dn_bonuslen;
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
dn->dn_num_slots = dnp->dn_extra_slots + 1;
|
2008-11-20 20:01:55 +00:00
|
|
|
dn->dn_maxblkid = dnp->dn_maxblkid;
|
2010-05-28 20:45:14 +00:00
|
|
|
dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
|
|
|
|
dn->dn_id_flags = 0;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
dmu_zfetch_init(&dn->dn_zfetch, dn);
|
|
|
|
|
2012-12-13 23:24:15 +00:00
|
|
|
ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
|
2010-08-26 21:24:34 +00:00
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
mutex_enter(&os->os_lock);
|
2015-04-02 03:44:32 +00:00
|
|
|
if (dnh->dnh_dnode != NULL) {
|
|
|
|
/* Lost the allocation race. */
|
|
|
|
mutex_exit(&os->os_lock);
|
|
|
|
kmem_cache_free(dnode_cache, dn);
|
|
|
|
return (dnh->dnh_dnode);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Exclude special dnodes from os_dnodes so an empty os_dnodes
|
|
|
|
* signifies that the special dnodes have no references from
|
|
|
|
* their children (the entries in os_dnodes). This allows
|
|
|
|
* dnode_destroy() to easily determine if the last child has
|
|
|
|
* been removed and then complete eviction of the objset.
|
|
|
|
*/
|
|
|
|
if (!DMU_OBJECT_IS_SPECIAL(object))
|
|
|
|
list_insert_head(&os->os_dnodes, dn);
|
2010-08-26 21:24:34 +00:00
|
|
|
membar_producer();
|
2015-04-02 03:44:32 +00:00
|
|
|
|
2010-08-26 21:24:34 +00:00
|
|
|
/*
|
2015-04-02 03:44:32 +00:00
|
|
|
* Everything else must be valid before assigning dn_objset
|
|
|
|
* makes the dnode eligible for dnode_move().
|
2010-08-26 21:24:34 +00:00
|
|
|
*/
|
|
|
|
dn->dn_objset = os;
|
2015-04-02 03:44:32 +00:00
|
|
|
|
|
|
|
dnh->dnh_dnode = dn;
|
2008-11-20 20:01:55 +00:00
|
|
|
mutex_exit(&os->os_lock);
|
|
|
|
|
2016-07-13 12:42:40 +00:00
|
|
|
arc_space_consume(sizeof (dnode_t), ARC_SPACE_DNODE);
|
2008-11-20 20:01:55 +00:00
|
|
|
return (dn);
|
|
|
|
}
|
|
|
|
|
2010-08-26 21:24:34 +00:00
|
|
|
/*
|
|
|
|
* Caller must be holding the dnode handle, which is released upon return.
|
|
|
|
*/
|
2008-11-20 20:01:55 +00:00
|
|
|
static void
|
|
|
|
dnode_destroy(dnode_t *dn)
|
|
|
|
{
|
2010-05-28 20:45:14 +00:00
|
|
|
objset_t *os = dn->dn_objset;
|
2015-04-02 03:44:32 +00:00
|
|
|
boolean_t complete_os_eviction = B_FALSE;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2010-05-28 20:45:14 +00:00
|
|
|
ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
mutex_enter(&os->os_lock);
|
2010-08-26 21:24:34 +00:00
|
|
|
POINTER_INVALIDATE(&dn->dn_objset);
|
2015-04-02 03:44:32 +00:00
|
|
|
if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
|
|
|
|
list_remove(&os->os_dnodes, dn);
|
|
|
|
complete_os_eviction =
|
|
|
|
list_is_empty(&os->os_dnodes) &&
|
|
|
|
list_link_active(&os->os_evicting_node);
|
|
|
|
}
|
2008-11-20 20:01:55 +00:00
|
|
|
mutex_exit(&os->os_lock);
|
|
|
|
|
2010-08-26 21:24:34 +00:00
|
|
|
/* the dnode can no longer move, so we can release the handle */
|
|
|
|
zrl_remove(&dn->dn_handle->dnh_zrlock);
|
|
|
|
|
|
|
|
dn->dn_allocated_txg = 0;
|
|
|
|
dn->dn_free_txg = 0;
|
|
|
|
dn->dn_assigned_txg = 0;
|
|
|
|
|
|
|
|
dn->dn_dirtyctx = 0;
|
|
|
|
if (dn->dn_dirtyctx_firstset != NULL) {
|
2008-11-20 20:01:55 +00:00
|
|
|
kmem_free(dn->dn_dirtyctx_firstset, 1);
|
|
|
|
dn->dn_dirtyctx_firstset = NULL;
|
|
|
|
}
|
2010-08-26 21:24:34 +00:00
|
|
|
if (dn->dn_bonus != NULL) {
|
2008-11-20 20:01:55 +00:00
|
|
|
mutex_enter(&dn->dn_bonus->db_mtx);
|
2016-06-02 04:04:53 +00:00
|
|
|
dbuf_destroy(dn->dn_bonus);
|
2008-11-20 20:01:55 +00:00
|
|
|
dn->dn_bonus = NULL;
|
|
|
|
}
|
2010-08-26 21:24:34 +00:00
|
|
|
dn->dn_zio = NULL;
|
|
|
|
|
|
|
|
dn->dn_have_spill = B_FALSE;
|
|
|
|
dn->dn_oldused = 0;
|
|
|
|
dn->dn_oldflags = 0;
|
|
|
|
dn->dn_olduid = 0;
|
|
|
|
dn->dn_oldgid = 0;
|
|
|
|
dn->dn_newuid = 0;
|
|
|
|
dn->dn_newgid = 0;
|
|
|
|
dn->dn_id_flags = 0;
|
2013-08-21 04:11:52 +00:00
|
|
|
dn->dn_unlisted_l0_blkid = 0;
|
2010-08-26 21:24:34 +00:00
|
|
|
|
2015-12-26 21:10:31 +00:00
|
|
|
dmu_zfetch_fini(&dn->dn_zfetch);
|
2008-11-20 20:01:55 +00:00
|
|
|
kmem_cache_free(dnode_cache, dn);
|
2016-07-13 12:42:40 +00:00
|
|
|
arc_space_return(sizeof (dnode_t), ARC_SPACE_DNODE);
|
2015-04-02 03:44:32 +00:00
|
|
|
|
|
|
|
if (complete_os_eviction)
|
|
|
|
dmu_objset_evict_done(os);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
void
|
|
|
|
dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
|
|
|
int i;
|
|
|
|
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
ASSERT3U(dn_slots, >, 0);
|
|
|
|
ASSERT3U(dn_slots << DNODE_SHIFT, <=,
|
|
|
|
spa_maxdnodesize(dmu_objset_spa(dn->dn_objset)));
|
2014-11-03 20:15:08 +00:00
|
|
|
ASSERT3U(blocksize, <=,
|
|
|
|
spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
|
2008-11-20 20:01:55 +00:00
|
|
|
if (blocksize == 0)
|
|
|
|
blocksize = 1 << zfs_default_bs;
|
|
|
|
else
|
|
|
|
blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
|
|
|
|
|
|
|
|
if (ibs == 0)
|
|
|
|
ibs = zfs_default_ibs;
|
|
|
|
|
|
|
|
ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
|
|
|
|
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d dn_slots=%d\n",
|
|
|
|
dn->dn_objset, dn->dn_object, tx->tx_txg, blocksize, ibs, dn_slots);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
ASSERT(dn->dn_type == DMU_OT_NONE);
|
|
|
|
ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
|
|
|
|
ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
|
|
|
|
ASSERT(ot != DMU_OT_NONE);
|
2012-12-13 23:24:15 +00:00
|
|
|
ASSERT(DMU_OT_IS_VALID(ot));
|
2008-11-20 20:01:55 +00:00
|
|
|
ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
|
2010-05-28 20:45:14 +00:00
|
|
|
(bonustype == DMU_OT_SA && bonuslen == 0) ||
|
2008-11-20 20:01:55 +00:00
|
|
|
(bonustype != DMU_OT_NONE && bonuslen != 0));
|
2012-12-13 23:24:15 +00:00
|
|
|
ASSERT(DMU_OT_IS_VALID(bonustype));
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
ASSERT3U(bonuslen, <=, DN_SLOTS_TO_BONUSLEN(dn_slots));
|
2008-11-20 20:01:55 +00:00
|
|
|
ASSERT(dn->dn_type == DMU_OT_NONE);
|
2013-05-10 21:17:03 +00:00
|
|
|
ASSERT0(dn->dn_maxblkid);
|
|
|
|
ASSERT0(dn->dn_allocated_txg);
|
|
|
|
ASSERT0(dn->dn_assigned_txg);
|
2008-11-20 20:01:55 +00:00
|
|
|
ASSERT(refcount_is_zero(&dn->dn_tx_holds));
|
|
|
|
ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
|
2015-04-03 03:14:28 +00:00
|
|
|
ASSERT(avl_is_empty(&dn->dn_dbufs));
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
for (i = 0; i < TXG_SIZE; i++) {
|
2013-05-10 21:17:03 +00:00
|
|
|
ASSERT0(dn->dn_next_nblkptr[i]);
|
|
|
|
ASSERT0(dn->dn_next_nlevels[i]);
|
|
|
|
ASSERT0(dn->dn_next_indblkshift[i]);
|
|
|
|
ASSERT0(dn->dn_next_bonuslen[i]);
|
|
|
|
ASSERT0(dn->dn_next_bonustype[i]);
|
|
|
|
ASSERT0(dn->dn_rm_spillblk[i]);
|
|
|
|
ASSERT0(dn->dn_next_blksz[i]);
|
2008-11-20 20:01:55 +00:00
|
|
|
ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
|
|
|
|
ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
|
2014-04-16 03:40:22 +00:00
|
|
|
ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
dn->dn_type = ot;
|
|
|
|
dnode_setdblksz(dn, blocksize);
|
|
|
|
dn->dn_indblkshift = ibs;
|
|
|
|
dn->dn_nlevels = 1;
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
dn->dn_num_slots = dn_slots;
|
2010-05-28 20:45:14 +00:00
|
|
|
if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
|
|
|
|
dn->dn_nblkptr = 1;
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
else {
|
|
|
|
dn->dn_nblkptr = MIN(DN_MAX_NBLKPTR,
|
|
|
|
1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
|
|
|
|
SPA_BLKPTRSHIFT));
|
|
|
|
}
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
dn->dn_bonustype = bonustype;
|
|
|
|
dn->dn_bonuslen = bonuslen;
|
|
|
|
dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
|
|
|
|
dn->dn_compress = ZIO_COMPRESS_INHERIT;
|
|
|
|
dn->dn_dirtyctx = 0;
|
|
|
|
|
|
|
|
dn->dn_free_txg = 0;
|
|
|
|
if (dn->dn_dirtyctx_firstset) {
|
|
|
|
kmem_free(dn->dn_dirtyctx_firstset, 1);
|
|
|
|
dn->dn_dirtyctx_firstset = NULL;
|
|
|
|
}
|
|
|
|
|
|
|
|
dn->dn_allocated_txg = tx->tx_txg;
|
2010-05-28 20:45:14 +00:00
|
|
|
dn->dn_id_flags = 0;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
dnode_setdirty(dn, tx);
|
|
|
|
dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
|
|
|
|
dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
|
2010-05-28 20:45:14 +00:00
|
|
|
dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
|
2008-11-20 20:01:55 +00:00
|
|
|
dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
|
|
|
|
}
|
|
|
|
|
|
|
|
void
|
|
|
|
dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
dmu_object_type_t bonustype, int bonuslen, int dn_slots, dmu_tx_t *tx)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
2009-07-02 22:44:48 +00:00
|
|
|
int nblkptr;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
|
2014-11-03 20:15:08 +00:00
|
|
|
ASSERT3U(blocksize, <=,
|
|
|
|
spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
|
2013-05-10 21:17:03 +00:00
|
|
|
ASSERT0(blocksize % SPA_MINBLOCKSIZE);
|
2008-11-20 20:01:55 +00:00
|
|
|
ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
|
|
|
|
ASSERT(tx->tx_txg != 0);
|
|
|
|
ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
|
2010-05-28 20:45:14 +00:00
|
|
|
(bonustype != DMU_OT_NONE && bonuslen != 0) ||
|
|
|
|
(bonustype == DMU_OT_SA && bonuslen == 0));
|
2012-12-13 23:24:15 +00:00
|
|
|
ASSERT(DMU_OT_IS_VALID(bonustype));
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
ASSERT3U(bonuslen, <=,
|
|
|
|
DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(dn->dn_objset))));
|
|
|
|
|
|
|
|
dn_slots = dn_slots > 0 ? dn_slots : DNODE_MIN_SLOTS;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
/* clean up any unreferenced dbufs */
|
|
|
|
dnode_evict_dbufs(dn);
|
2009-02-18 20:51:31 +00:00
|
|
|
|
2010-05-28 20:45:14 +00:00
|
|
|
dn->dn_id_flags = 0;
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
|
|
|
|
dnode_setdirty(dn, tx);
|
2009-07-02 22:44:48 +00:00
|
|
|
if (dn->dn_datablksz != blocksize) {
|
|
|
|
/* change blocksize */
|
|
|
|
ASSERT(dn->dn_maxblkid == 0 &&
|
|
|
|
(BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
|
|
|
|
dnode_block_freed(dn, 0)));
|
|
|
|
dnode_setdblksz(dn, blocksize);
|
|
|
|
dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
|
|
|
|
}
|
|
|
|
if (dn->dn_bonuslen != bonuslen)
|
|
|
|
dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
|
2010-05-28 20:45:14 +00:00
|
|
|
|
|
|
|
if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
|
|
|
|
nblkptr = 1;
|
|
|
|
else
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
nblkptr = MIN(DN_MAX_NBLKPTR,
|
|
|
|
1 + ((DN_SLOTS_TO_BONUSLEN(dn_slots) - bonuslen) >>
|
|
|
|
SPA_BLKPTRSHIFT));
|
2010-05-28 20:45:14 +00:00
|
|
|
if (dn->dn_bonustype != bonustype)
|
|
|
|
dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
|
2009-02-18 20:51:31 +00:00
|
|
|
if (dn->dn_nblkptr != nblkptr)
|
|
|
|
dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
|
2010-05-28 20:45:14 +00:00
|
|
|
if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
|
|
|
|
dbuf_rm_spill(dn, tx);
|
|
|
|
dnode_rm_spill(dn, tx);
|
|
|
|
}
|
2008-11-20 20:01:55 +00:00
|
|
|
rw_exit(&dn->dn_struct_rwlock);
|
|
|
|
|
|
|
|
/* change type */
|
|
|
|
dn->dn_type = ot;
|
|
|
|
|
|
|
|
/* change bonus size and type */
|
|
|
|
mutex_enter(&dn->dn_mtx);
|
|
|
|
dn->dn_bonustype = bonustype;
|
|
|
|
dn->dn_bonuslen = bonuslen;
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
dn->dn_num_slots = dn_slots;
|
2009-02-18 20:51:31 +00:00
|
|
|
dn->dn_nblkptr = nblkptr;
|
2008-11-20 20:01:55 +00:00
|
|
|
dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
|
|
|
|
dn->dn_compress = ZIO_COMPRESS_INHERIT;
|
|
|
|
ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
|
|
|
|
|
2009-02-18 20:51:31 +00:00
|
|
|
/* fix up the bonus db_size */
|
|
|
|
if (dn->dn_bonus) {
|
2008-11-20 20:01:55 +00:00
|
|
|
dn->dn_bonus->db.db_size =
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
DN_SLOTS_TO_BONUSLEN(dn->dn_num_slots) -
|
|
|
|
(dn->dn_nblkptr-1) * sizeof (blkptr_t);
|
2008-11-20 20:01:55 +00:00
|
|
|
ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
|
|
|
|
}
|
|
|
|
|
|
|
|
dn->dn_allocated_txg = tx->tx_txg;
|
|
|
|
mutex_exit(&dn->dn_mtx);
|
|
|
|
}
|
|
|
|
|
2010-08-27 17:55:07 +00:00
|
|
|
#ifdef _KERNEL
|
2010-08-26 21:24:34 +00:00
|
|
|
#ifdef DNODE_STATS
|
|
|
|
static struct {
|
|
|
|
uint64_t dms_dnode_invalid;
|
|
|
|
uint64_t dms_dnode_recheck1;
|
|
|
|
uint64_t dms_dnode_recheck2;
|
|
|
|
uint64_t dms_dnode_special;
|
|
|
|
uint64_t dms_dnode_handle;
|
|
|
|
uint64_t dms_dnode_rwlock;
|
|
|
|
uint64_t dms_dnode_active;
|
|
|
|
} dnode_move_stats;
|
|
|
|
#endif /* DNODE_STATS */
|
|
|
|
|
|
|
|
static void
|
|
|
|
dnode_move_impl(dnode_t *odn, dnode_t *ndn)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
|
|
|
|
ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
|
|
|
|
ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
|
|
|
|
ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
|
|
|
|
ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
|
|
|
|
|
|
|
|
/* Copy fields. */
|
|
|
|
ndn->dn_objset = odn->dn_objset;
|
|
|
|
ndn->dn_object = odn->dn_object;
|
|
|
|
ndn->dn_dbuf = odn->dn_dbuf;
|
|
|
|
ndn->dn_handle = odn->dn_handle;
|
|
|
|
ndn->dn_phys = odn->dn_phys;
|
|
|
|
ndn->dn_type = odn->dn_type;
|
|
|
|
ndn->dn_bonuslen = odn->dn_bonuslen;
|
|
|
|
ndn->dn_bonustype = odn->dn_bonustype;
|
|
|
|
ndn->dn_nblkptr = odn->dn_nblkptr;
|
|
|
|
ndn->dn_checksum = odn->dn_checksum;
|
|
|
|
ndn->dn_compress = odn->dn_compress;
|
|
|
|
ndn->dn_nlevels = odn->dn_nlevels;
|
|
|
|
ndn->dn_indblkshift = odn->dn_indblkshift;
|
|
|
|
ndn->dn_datablkshift = odn->dn_datablkshift;
|
|
|
|
ndn->dn_datablkszsec = odn->dn_datablkszsec;
|
|
|
|
ndn->dn_datablksz = odn->dn_datablksz;
|
|
|
|
ndn->dn_maxblkid = odn->dn_maxblkid;
|
|
|
|
bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
|
|
|
|
sizeof (odn->dn_next_nblkptr));
|
|
|
|
bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
|
|
|
|
sizeof (odn->dn_next_nlevels));
|
|
|
|
bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
|
|
|
|
sizeof (odn->dn_next_indblkshift));
|
|
|
|
bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
|
|
|
|
sizeof (odn->dn_next_bonustype));
|
|
|
|
bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
|
|
|
|
sizeof (odn->dn_rm_spillblk));
|
|
|
|
bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
|
|
|
|
sizeof (odn->dn_next_bonuslen));
|
|
|
|
bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
|
|
|
|
sizeof (odn->dn_next_blksz));
|
|
|
|
for (i = 0; i < TXG_SIZE; i++) {
|
|
|
|
list_move_tail(&ndn->dn_dirty_records[i],
|
|
|
|
&odn->dn_dirty_records[i]);
|
|
|
|
}
|
2014-04-16 03:40:22 +00:00
|
|
|
bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
|
|
|
|
sizeof (odn->dn_free_ranges));
|
2010-08-26 21:24:34 +00:00
|
|
|
ndn->dn_allocated_txg = odn->dn_allocated_txg;
|
|
|
|
ndn->dn_free_txg = odn->dn_free_txg;
|
|
|
|
ndn->dn_assigned_txg = odn->dn_assigned_txg;
|
|
|
|
ndn->dn_dirtyctx = odn->dn_dirtyctx;
|
|
|
|
ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
|
|
|
|
ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
|
|
|
|
refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
|
2015-04-03 03:14:28 +00:00
|
|
|
ASSERT(avl_is_empty(&ndn->dn_dbufs));
|
|
|
|
avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
|
2010-08-26 21:24:34 +00:00
|
|
|
ndn->dn_dbufs_count = odn->dn_dbufs_count;
|
2013-08-21 04:11:52 +00:00
|
|
|
ndn->dn_unlisted_l0_blkid = odn->dn_unlisted_l0_blkid;
|
2010-08-26 21:24:34 +00:00
|
|
|
ndn->dn_bonus = odn->dn_bonus;
|
|
|
|
ndn->dn_have_spill = odn->dn_have_spill;
|
|
|
|
ndn->dn_zio = odn->dn_zio;
|
|
|
|
ndn->dn_oldused = odn->dn_oldused;
|
|
|
|
ndn->dn_oldflags = odn->dn_oldflags;
|
|
|
|
ndn->dn_olduid = odn->dn_olduid;
|
|
|
|
ndn->dn_oldgid = odn->dn_oldgid;
|
|
|
|
ndn->dn_newuid = odn->dn_newuid;
|
|
|
|
ndn->dn_newgid = odn->dn_newgid;
|
|
|
|
ndn->dn_id_flags = odn->dn_id_flags;
|
|
|
|
dmu_zfetch_init(&ndn->dn_zfetch, NULL);
|
|
|
|
list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
|
|
|
|
ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Update back pointers. Updating the handle fixes the back pointer of
|
|
|
|
* every descendant dbuf as well as the bonus dbuf.
|
|
|
|
*/
|
|
|
|
ASSERT(ndn->dn_handle->dnh_dnode == odn);
|
|
|
|
ndn->dn_handle->dnh_dnode = ndn;
|
|
|
|
if (ndn->dn_zfetch.zf_dnode == odn) {
|
|
|
|
ndn->dn_zfetch.zf_dnode = ndn;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Invalidate the original dnode by clearing all of its back pointers.
|
|
|
|
*/
|
|
|
|
odn->dn_dbuf = NULL;
|
|
|
|
odn->dn_handle = NULL;
|
2015-04-03 03:14:28 +00:00
|
|
|
avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
|
2010-08-26 21:24:34 +00:00
|
|
|
offsetof(dmu_buf_impl_t, db_link));
|
|
|
|
odn->dn_dbufs_count = 0;
|
2013-08-21 04:11:52 +00:00
|
|
|
odn->dn_unlisted_l0_blkid = 0;
|
2010-08-26 21:24:34 +00:00
|
|
|
odn->dn_bonus = NULL;
|
|
|
|
odn->dn_zfetch.zf_dnode = NULL;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Set the low bit of the objset pointer to ensure that dnode_move()
|
|
|
|
* recognizes the dnode as invalid in any subsequent callback.
|
|
|
|
*/
|
|
|
|
POINTER_INVALIDATE(&odn->dn_objset);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Satisfy the destructor.
|
|
|
|
*/
|
|
|
|
for (i = 0; i < TXG_SIZE; i++) {
|
|
|
|
list_create(&odn->dn_dirty_records[i],
|
|
|
|
sizeof (dbuf_dirty_record_t),
|
|
|
|
offsetof(dbuf_dirty_record_t, dr_dirty_node));
|
2014-04-16 03:40:22 +00:00
|
|
|
odn->dn_free_ranges[i] = NULL;
|
2010-08-26 21:24:34 +00:00
|
|
|
odn->dn_next_nlevels[i] = 0;
|
|
|
|
odn->dn_next_indblkshift[i] = 0;
|
|
|
|
odn->dn_next_bonustype[i] = 0;
|
|
|
|
odn->dn_rm_spillblk[i] = 0;
|
|
|
|
odn->dn_next_bonuslen[i] = 0;
|
|
|
|
odn->dn_next_blksz[i] = 0;
|
|
|
|
}
|
|
|
|
odn->dn_allocated_txg = 0;
|
|
|
|
odn->dn_free_txg = 0;
|
|
|
|
odn->dn_assigned_txg = 0;
|
|
|
|
odn->dn_dirtyctx = 0;
|
|
|
|
odn->dn_dirtyctx_firstset = NULL;
|
|
|
|
odn->dn_have_spill = B_FALSE;
|
|
|
|
odn->dn_zio = NULL;
|
|
|
|
odn->dn_oldused = 0;
|
|
|
|
odn->dn_oldflags = 0;
|
|
|
|
odn->dn_olduid = 0;
|
|
|
|
odn->dn_oldgid = 0;
|
|
|
|
odn->dn_newuid = 0;
|
|
|
|
odn->dn_newgid = 0;
|
|
|
|
odn->dn_id_flags = 0;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Mark the dnode.
|
|
|
|
*/
|
|
|
|
ndn->dn_moved = 1;
|
|
|
|
odn->dn_moved = (uint8_t)-1;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*ARGSUSED*/
|
|
|
|
static kmem_cbrc_t
|
|
|
|
dnode_move(void *buf, void *newbuf, size_t size, void *arg)
|
|
|
|
{
|
|
|
|
dnode_t *odn = buf, *ndn = newbuf;
|
|
|
|
objset_t *os;
|
|
|
|
int64_t refcount;
|
|
|
|
uint32_t dbufs;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The dnode is on the objset's list of known dnodes if the objset
|
|
|
|
* pointer is valid. We set the low bit of the objset pointer when
|
|
|
|
* freeing the dnode to invalidate it, and the memory patterns written
|
|
|
|
* by kmem (baddcafe and deadbeef) set at least one of the two low bits.
|
|
|
|
* A newly created dnode sets the objset pointer last of all to indicate
|
|
|
|
* that the dnode is known and in a valid state to be moved by this
|
|
|
|
* function.
|
|
|
|
*/
|
|
|
|
os = odn->dn_objset;
|
|
|
|
if (!POINTER_IS_VALID(os)) {
|
|
|
|
DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
|
|
|
|
return (KMEM_CBRC_DONT_KNOW);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Ensure that the objset does not go away during the move.
|
|
|
|
*/
|
|
|
|
rw_enter(&os_lock, RW_WRITER);
|
|
|
|
if (os != odn->dn_objset) {
|
|
|
|
rw_exit(&os_lock);
|
|
|
|
DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
|
|
|
|
return (KMEM_CBRC_DONT_KNOW);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If the dnode is still valid, then so is the objset. We know that no
|
|
|
|
* valid objset can be freed while we hold os_lock, so we can safely
|
|
|
|
* ensure that the objset remains in use.
|
|
|
|
*/
|
|
|
|
mutex_enter(&os->os_lock);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Recheck the objset pointer in case the dnode was removed just before
|
|
|
|
* acquiring the lock.
|
|
|
|
*/
|
|
|
|
if (os != odn->dn_objset) {
|
|
|
|
mutex_exit(&os->os_lock);
|
|
|
|
rw_exit(&os_lock);
|
|
|
|
DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
|
|
|
|
return (KMEM_CBRC_DONT_KNOW);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* At this point we know that as long as we hold os->os_lock, the dnode
|
|
|
|
* cannot be freed and fields within the dnode can be safely accessed.
|
|
|
|
* The objset listing this dnode cannot go away as long as this dnode is
|
|
|
|
* on its list.
|
|
|
|
*/
|
|
|
|
rw_exit(&os_lock);
|
|
|
|
if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
|
|
|
|
mutex_exit(&os->os_lock);
|
|
|
|
DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
|
|
|
|
return (KMEM_CBRC_NO);
|
|
|
|
}
|
|
|
|
ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Lock the dnode handle to prevent the dnode from obtaining any new
|
|
|
|
* holds. This also prevents the descendant dbufs and the bonus dbuf
|
|
|
|
* from accessing the dnode, so that we can discount their holds. The
|
|
|
|
* handle is safe to access because we know that while the dnode cannot
|
|
|
|
* go away, neither can its handle. Once we hold dnh_zrlock, we can
|
|
|
|
* safely move any dnode referenced only by dbufs.
|
|
|
|
*/
|
|
|
|
if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
|
|
|
|
mutex_exit(&os->os_lock);
|
|
|
|
DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
|
|
|
|
return (KMEM_CBRC_LATER);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Ensure a consistent view of the dnode's holds and the dnode's dbufs.
|
|
|
|
* We need to guarantee that there is a hold for every dbuf in order to
|
|
|
|
* determine whether the dnode is actively referenced. Falsely matching
|
|
|
|
* a dbuf to an active hold would lead to an unsafe move. It's possible
|
|
|
|
* that a thread already having an active dnode hold is about to add a
|
|
|
|
* dbuf, and we can't compare hold and dbuf counts while the add is in
|
|
|
|
* progress.
|
|
|
|
*/
|
|
|
|
if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
|
|
|
|
zrl_exit(&odn->dn_handle->dnh_zrlock);
|
|
|
|
mutex_exit(&os->os_lock);
|
|
|
|
DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
|
|
|
|
return (KMEM_CBRC_LATER);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* A dbuf may be removed (evicted) without an active dnode hold. In that
|
|
|
|
* case, the dbuf count is decremented under the handle lock before the
|
|
|
|
* dbuf's hold is released. This order ensures that if we count the hold
|
|
|
|
* after the dbuf is removed but before its hold is released, we will
|
|
|
|
* treat the unmatched hold as active and exit safely. If we count the
|
|
|
|
* hold before the dbuf is removed, the hold is discounted, and the
|
|
|
|
* removal is blocked until the move completes.
|
|
|
|
*/
|
|
|
|
refcount = refcount_count(&odn->dn_holds);
|
|
|
|
ASSERT(refcount >= 0);
|
|
|
|
dbufs = odn->dn_dbufs_count;
|
|
|
|
|
|
|
|
/* We can't have more dbufs than dnode holds. */
|
|
|
|
ASSERT3U(dbufs, <=, refcount);
|
|
|
|
DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
|
|
|
|
uint32_t, dbufs);
|
|
|
|
|
|
|
|
if (refcount > dbufs) {
|
|
|
|
rw_exit(&odn->dn_struct_rwlock);
|
|
|
|
zrl_exit(&odn->dn_handle->dnh_zrlock);
|
|
|
|
mutex_exit(&os->os_lock);
|
|
|
|
DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
|
|
|
|
return (KMEM_CBRC_LATER);
|
|
|
|
}
|
|
|
|
|
|
|
|
rw_exit(&odn->dn_struct_rwlock);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* At this point we know that anyone with a hold on the dnode is not
|
|
|
|
* actively referencing it. The dnode is known and in a valid state to
|
|
|
|
* move. We're holding the locks needed to execute the critical section.
|
|
|
|
*/
|
|
|
|
dnode_move_impl(odn, ndn);
|
|
|
|
|
|
|
|
list_link_replace(&odn->dn_link, &ndn->dn_link);
|
|
|
|
/* If the dnode was safe to move, the refcount cannot have changed. */
|
|
|
|
ASSERT(refcount == refcount_count(&ndn->dn_holds));
|
|
|
|
ASSERT(dbufs == ndn->dn_dbufs_count);
|
|
|
|
zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
|
|
|
|
mutex_exit(&os->os_lock);
|
|
|
|
|
|
|
|
return (KMEM_CBRC_YES);
|
|
|
|
}
|
|
|
|
#endif /* _KERNEL */
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
void
|
2010-08-26 21:24:34 +00:00
|
|
|
dnode_special_close(dnode_handle_t *dnh)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
2010-08-26 21:24:34 +00:00
|
|
|
dnode_t *dn = dnh->dnh_dnode;
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
/*
|
|
|
|
* Wait for final references to the dnode to clear. This can
|
|
|
|
* only happen if the arc is asyncronously evicting state that
|
|
|
|
* has a hold on this dnode while we are trying to evict this
|
|
|
|
* dnode.
|
|
|
|
*/
|
|
|
|
while (refcount_count(&dn->dn_holds) > 0)
|
|
|
|
delay(1);
|
2015-04-02 03:44:32 +00:00
|
|
|
ASSERT(dn->dn_dbuf == NULL ||
|
|
|
|
dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
|
2010-08-26 21:24:34 +00:00
|
|
|
zrl_add(&dnh->dnh_zrlock);
|
|
|
|
dnode_destroy(dn); /* implicit zrl_remove() */
|
|
|
|
zrl_destroy(&dnh->dnh_zrlock);
|
|
|
|
dnh->dnh_dnode = NULL;
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
2015-04-02 03:44:32 +00:00
|
|
|
void
|
2010-08-26 21:24:34 +00:00
|
|
|
dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
|
|
|
|
dnode_handle_t *dnh)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
2015-04-02 03:44:32 +00:00
|
|
|
dnode_t *dn;
|
|
|
|
|
|
|
|
dn = dnode_create(os, dnp, NULL, object, dnh);
|
2010-08-26 21:24:34 +00:00
|
|
|
zrl_init(&dnh->dnh_zrlock);
|
2008-11-20 20:01:55 +00:00
|
|
|
DNODE_VERIFY(dn);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void
|
2015-04-02 03:44:32 +00:00
|
|
|
dnode_buf_pageout(void *dbu)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
2015-04-02 03:44:32 +00:00
|
|
|
dnode_children_t *children_dnodes = dbu;
|
2008-11-20 20:01:55 +00:00
|
|
|
int i;
|
|
|
|
|
2015-04-02 03:44:32 +00:00
|
|
|
for (i = 0; i < children_dnodes->dnc_count; i++) {
|
2010-08-26 21:24:34 +00:00
|
|
|
dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
|
|
|
|
dnode_t *dn;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2010-08-26 21:24:34 +00:00
|
|
|
/*
|
|
|
|
* The dnode handle lock guards against the dnode moving to
|
|
|
|
* another valid address, so there is no need here to guard
|
|
|
|
* against changes to or from NULL.
|
|
|
|
*/
|
|
|
|
if (dnh->dnh_dnode == NULL) {
|
|
|
|
zrl_destroy(&dnh->dnh_zrlock);
|
2008-11-20 20:01:55 +00:00
|
|
|
continue;
|
2010-08-26 21:24:34 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
zrl_add(&dnh->dnh_zrlock);
|
|
|
|
dn = dnh->dnh_dnode;
|
2008-11-20 20:01:55 +00:00
|
|
|
/*
|
|
|
|
* If there are holds on this dnode, then there should
|
|
|
|
* be holds on the dnode's containing dbuf as well; thus
|
2010-08-26 21:24:34 +00:00
|
|
|
* it wouldn't be eligible for eviction and this function
|
2008-11-20 20:01:55 +00:00
|
|
|
* would not have been called.
|
|
|
|
*/
|
|
|
|
ASSERT(refcount_is_zero(&dn->dn_holds));
|
|
|
|
ASSERT(refcount_is_zero(&dn->dn_tx_holds));
|
|
|
|
|
2010-08-26 21:24:34 +00:00
|
|
|
dnode_destroy(dn); /* implicit zrl_remove() */
|
|
|
|
zrl_destroy(&dnh->dnh_zrlock);
|
|
|
|
dnh->dnh_dnode = NULL;
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
2010-08-26 21:24:34 +00:00
|
|
|
kmem_free(children_dnodes, sizeof (dnode_children_t) +
|
2015-04-02 03:44:32 +00:00
|
|
|
children_dnodes->dnc_count * sizeof (dnode_handle_t));
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
/*
|
|
|
|
* Return true if the given index is interior to a dnode already
|
|
|
|
* allocated in the block. That is, the index is neither free nor
|
|
|
|
* allocated, but is consumed by a large dnode.
|
|
|
|
*
|
|
|
|
* The dnode_phys_t buffer may not be in sync with the in-core dnode
|
|
|
|
* structure, so we try to check the dnode structure first and fall back
|
|
|
|
* to the dnode_phys_t buffer it doesn't exist.
|
|
|
|
*/
|
|
|
|
static boolean_t
|
|
|
|
dnode_is_consumed(dmu_buf_impl_t *db, int idx)
|
|
|
|
{
|
|
|
|
dnode_handle_t *dnh;
|
|
|
|
dmu_object_type_t ot;
|
|
|
|
dnode_children_t *children_dnodes;
|
|
|
|
dnode_phys_t *dn_block;
|
|
|
|
int skip;
|
|
|
|
int i;
|
|
|
|
|
|
|
|
children_dnodes = dmu_buf_get_user(&db->db);
|
|
|
|
dn_block = (dnode_phys_t *)db->db.db_data;
|
|
|
|
|
|
|
|
for (i = 0; i < idx; i += skip) {
|
|
|
|
dnh = &children_dnodes->dnc_children[i];
|
|
|
|
|
|
|
|
zrl_add(&dnh->dnh_zrlock);
|
|
|
|
if (dnh->dnh_dnode != NULL) {
|
|
|
|
ot = dnh->dnh_dnode->dn_type;
|
|
|
|
skip = dnh->dnh_dnode->dn_num_slots;
|
|
|
|
} else {
|
|
|
|
ot = dn_block[i].dn_type;
|
|
|
|
skip = dn_block[i].dn_extra_slots + 1;
|
|
|
|
}
|
|
|
|
zrl_remove(&dnh->dnh_zrlock);
|
|
|
|
|
|
|
|
if (ot == DMU_OT_NONE)
|
|
|
|
skip = 1;
|
|
|
|
}
|
|
|
|
|
|
|
|
return (i > idx);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Return true if the given index in the dnode block is a valid
|
|
|
|
* allocated dnode. That is, the index is not consumed by a large
|
|
|
|
* dnode and is not free.
|
|
|
|
*
|
|
|
|
* The dnode_phys_t buffer may not be in sync with the in-core dnode
|
|
|
|
* structure, so we try to check the dnode structure first and fall back
|
|
|
|
* to the dnode_phys_t buffer it doesn't exist.
|
|
|
|
*/
|
|
|
|
static boolean_t
|
|
|
|
dnode_is_allocated(dmu_buf_impl_t *db, int idx)
|
|
|
|
{
|
|
|
|
dnode_handle_t *dnh;
|
|
|
|
dmu_object_type_t ot;
|
|
|
|
dnode_children_t *children_dnodes;
|
|
|
|
dnode_phys_t *dn_block;
|
|
|
|
|
|
|
|
if (dnode_is_consumed(db, idx))
|
|
|
|
return (B_FALSE);
|
|
|
|
|
|
|
|
children_dnodes = dmu_buf_get_user(&db->db);
|
|
|
|
dn_block = (dnode_phys_t *)db->db.db_data;
|
|
|
|
|
|
|
|
dnh = &children_dnodes->dnc_children[idx];
|
|
|
|
|
|
|
|
zrl_add(&dnh->dnh_zrlock);
|
|
|
|
if (dnh->dnh_dnode != NULL)
|
|
|
|
ot = dnh->dnh_dnode->dn_type;
|
|
|
|
else
|
|
|
|
ot = dn_block[idx].dn_type;
|
|
|
|
zrl_remove(&dnh->dnh_zrlock);
|
|
|
|
|
|
|
|
return (ot != DMU_OT_NONE);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Return true if the given range of indices in the dnode block are
|
|
|
|
* free. That is, the starting index is not consumed by a large dnode
|
|
|
|
* and none of the indices are allocated.
|
|
|
|
*
|
|
|
|
* The dnode_phys_t buffer may not be in sync with the in-core dnode
|
|
|
|
* structure, so we try to check the dnode structure first and fall back
|
|
|
|
* to the dnode_phys_t buffer it doesn't exist.
|
|
|
|
*/
|
|
|
|
static boolean_t
|
|
|
|
dnode_is_free(dmu_buf_impl_t *db, int idx, int slots)
|
|
|
|
{
|
|
|
|
dnode_handle_t *dnh;
|
|
|
|
dmu_object_type_t ot;
|
|
|
|
dnode_children_t *children_dnodes;
|
|
|
|
dnode_phys_t *dn_block;
|
|
|
|
int i;
|
|
|
|
|
|
|
|
if (idx + slots > DNODES_PER_BLOCK)
|
|
|
|
return (B_FALSE);
|
|
|
|
|
|
|
|
children_dnodes = dmu_buf_get_user(&db->db);
|
|
|
|
dn_block = (dnode_phys_t *)db->db.db_data;
|
|
|
|
|
|
|
|
if (dnode_is_consumed(db, idx))
|
|
|
|
return (B_FALSE);
|
|
|
|
|
|
|
|
for (i = idx; i < idx + slots; i++) {
|
|
|
|
dnh = &children_dnodes->dnc_children[i];
|
|
|
|
|
|
|
|
zrl_add(&dnh->dnh_zrlock);
|
|
|
|
if (dnh->dnh_dnode != NULL)
|
|
|
|
ot = dnh->dnh_dnode->dn_type;
|
|
|
|
else
|
|
|
|
ot = dn_block[i].dn_type;
|
|
|
|
zrl_remove(&dnh->dnh_zrlock);
|
|
|
|
|
|
|
|
if (ot != DMU_OT_NONE)
|
|
|
|
return (B_FALSE);
|
|
|
|
}
|
|
|
|
|
|
|
|
return (B_TRUE);
|
|
|
|
}
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
/*
|
|
|
|
* errors:
|
|
|
|
* EINVAL - invalid object number.
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
* ENOSPC - hole too small to fulfill "slots" request
|
2008-11-20 20:01:55 +00:00
|
|
|
* EIO - i/o error.
|
|
|
|
* succeeds even for free dnodes.
|
|
|
|
*/
|
|
|
|
int
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
dnode_hold_impl(objset_t *os, uint64_t object, int flag, int slots,
|
2008-11-20 20:01:55 +00:00
|
|
|
void *tag, dnode_t **dnp)
|
|
|
|
{
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
int epb, idx, err, i;
|
2008-11-20 20:01:55 +00:00
|
|
|
int drop_struct_lock = FALSE;
|
|
|
|
int type;
|
|
|
|
uint64_t blk;
|
|
|
|
dnode_t *mdn, *dn;
|
|
|
|
dmu_buf_impl_t *db;
|
2010-08-26 21:24:34 +00:00
|
|
|
dnode_children_t *children_dnodes;
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
dnode_phys_t *dn_block_begin;
|
2010-08-26 21:24:34 +00:00
|
|
|
dnode_handle_t *dnh;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
ASSERT(!(flag & DNODE_MUST_BE_ALLOCATED) || (slots == 0));
|
|
|
|
ASSERT(!(flag & DNODE_MUST_BE_FREE) || (slots > 0));
|
|
|
|
|
2008-12-03 20:09:06 +00:00
|
|
|
/*
|
|
|
|
* If you are holding the spa config lock as writer, you shouldn't
|
2010-05-28 20:45:14 +00:00
|
|
|
* be asking the DMU to do *anything* unless it's the root pool
|
|
|
|
* which may require us to read from the root filesystem while
|
|
|
|
* holding some (not all) of the locks as writer.
|
2008-12-03 20:09:06 +00:00
|
|
|
*/
|
2010-05-28 20:45:14 +00:00
|
|
|
ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
|
|
|
|
(spa_is_root(os->os_spa) &&
|
2010-08-26 21:24:34 +00:00
|
|
|
spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
|
2008-12-03 20:09:06 +00:00
|
|
|
|
2009-07-02 22:44:48 +00:00
|
|
|
if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
|
|
|
|
dn = (object == DMU_USERUSED_OBJECT) ?
|
2010-08-26 21:24:34 +00:00
|
|
|
DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
|
2009-07-02 22:44:48 +00:00
|
|
|
if (dn == NULL)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(ENOENT));
|
2009-07-02 22:44:48 +00:00
|
|
|
type = dn->dn_type;
|
|
|
|
if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(ENOENT));
|
2009-07-02 22:44:48 +00:00
|
|
|
if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EEXIST));
|
2009-07-02 22:44:48 +00:00
|
|
|
DNODE_VERIFY(dn);
|
|
|
|
(void) refcount_add(&dn->dn_holds, tag);
|
|
|
|
*dnp = dn;
|
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
if (object == 0 || object >= DN_MAX_OBJECT)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2010-08-26 21:24:34 +00:00
|
|
|
mdn = DMU_META_DNODE(os);
|
|
|
|
ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
DNODE_VERIFY(mdn);
|
|
|
|
|
|
|
|
if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
|
|
|
|
rw_enter(&mdn->dn_struct_rwlock, RW_READER);
|
|
|
|
drop_struct_lock = TRUE;
|
|
|
|
}
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
db = dbuf_hold(mdn, blk, FTAG);
|
|
|
|
if (drop_struct_lock)
|
|
|
|
rw_exit(&mdn->dn_struct_rwlock);
|
|
|
|
if (db == NULL)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EIO));
|
2008-11-20 20:01:55 +00:00
|
|
|
err = dbuf_read(db, NULL, DB_RF_CANFAIL);
|
|
|
|
if (err) {
|
|
|
|
dbuf_rele(db, FTAG);
|
|
|
|
return (err);
|
|
|
|
}
|
|
|
|
|
|
|
|
ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
|
|
|
|
epb = db->db.db_size >> DNODE_SHIFT;
|
|
|
|
|
2010-08-26 21:24:34 +00:00
|
|
|
ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
|
2008-11-20 20:01:55 +00:00
|
|
|
children_dnodes = dmu_buf_get_user(&db->db);
|
|
|
|
if (children_dnodes == NULL) {
|
2010-08-26 21:24:34 +00:00
|
|
|
dnode_children_t *winner;
|
2015-04-02 03:44:32 +00:00
|
|
|
children_dnodes = kmem_zalloc(sizeof (dnode_children_t) +
|
2015-04-01 15:09:20 +00:00
|
|
|
epb * sizeof (dnode_handle_t), KM_SLEEP);
|
2010-08-26 21:24:34 +00:00
|
|
|
children_dnodes->dnc_count = epb;
|
|
|
|
dnh = &children_dnodes->dnc_children[0];
|
|
|
|
for (i = 0; i < epb; i++) {
|
|
|
|
zrl_init(&dnh[i].dnh_zrlock);
|
|
|
|
}
|
2015-04-02 03:44:32 +00:00
|
|
|
dmu_buf_init_user(&children_dnodes->dnc_dbu,
|
|
|
|
dnode_buf_pageout, NULL);
|
|
|
|
winner = dmu_buf_set_user(&db->db, &children_dnodes->dnc_dbu);
|
|
|
|
if (winner != NULL) {
|
2015-04-01 13:49:14 +00:00
|
|
|
|
|
|
|
for (i = 0; i < epb; i++) {
|
|
|
|
zrl_destroy(&dnh[i].dnh_zrlock);
|
|
|
|
}
|
|
|
|
|
2010-08-26 21:24:34 +00:00
|
|
|
kmem_free(children_dnodes, sizeof (dnode_children_t) +
|
2015-04-01 15:09:20 +00:00
|
|
|
epb * sizeof (dnode_handle_t));
|
2008-11-20 20:01:55 +00:00
|
|
|
children_dnodes = winner;
|
|
|
|
}
|
|
|
|
}
|
2010-08-26 21:24:34 +00:00
|
|
|
ASSERT(children_dnodes->dnc_count == epb);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
idx = object & (epb - 1);
|
|
|
|
dn_block_begin = (dnode_phys_t *)db->db.db_data;
|
|
|
|
|
|
|
|
if ((flag & DNODE_MUST_BE_FREE) && !dnode_is_free(db, idx, slots)) {
|
|
|
|
dbuf_rele(db, FTAG);
|
|
|
|
return (ENOSPC);
|
|
|
|
} else if ((flag & DNODE_MUST_BE_ALLOCATED) &&
|
|
|
|
!dnode_is_allocated(db, idx)) {
|
|
|
|
dbuf_rele(db, FTAG);
|
|
|
|
return (ENOENT);
|
|
|
|
}
|
|
|
|
|
2010-08-26 21:24:34 +00:00
|
|
|
dnh = &children_dnodes->dnc_children[idx];
|
|
|
|
zrl_add(&dnh->dnh_zrlock);
|
2015-04-02 03:44:32 +00:00
|
|
|
dn = dnh->dnh_dnode;
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
if (dn == NULL)
|
|
|
|
dn = dnode_create(os, dn_block_begin + idx, db, object, dnh);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
mutex_enter(&dn->dn_mtx);
|
|
|
|
type = dn->dn_type;
|
|
|
|
if (dn->dn_free_txg ||
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
((flag & DNODE_MUST_BE_FREE) && !refcount_is_zero(&dn->dn_holds))) {
|
2008-11-20 20:01:55 +00:00
|
|
|
mutex_exit(&dn->dn_mtx);
|
2010-08-26 21:24:34 +00:00
|
|
|
zrl_remove(&dnh->dnh_zrlock);
|
2008-11-20 20:01:55 +00:00
|
|
|
dbuf_rele(db, FTAG);
|
|
|
|
return (type == DMU_OT_NONE ? ENOENT : EEXIST);
|
|
|
|
}
|
|
|
|
if (refcount_add(&dn->dn_holds, tag) == 1)
|
2010-08-26 21:24:34 +00:00
|
|
|
dbuf_add_ref(db, dnh);
|
2015-04-02 03:44:32 +00:00
|
|
|
mutex_exit(&dn->dn_mtx);
|
|
|
|
|
2010-08-26 21:24:34 +00:00
|
|
|
/* Now we can rely on the hold to prevent the dnode from moving. */
|
|
|
|
zrl_remove(&dnh->dnh_zrlock);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
DNODE_VERIFY(dn);
|
|
|
|
ASSERT3P(dn->dn_dbuf, ==, db);
|
|
|
|
ASSERT3U(dn->dn_object, ==, object);
|
|
|
|
dbuf_rele(db, FTAG);
|
|
|
|
|
|
|
|
*dnp = dn;
|
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Return held dnode if the object is allocated, NULL if not.
|
|
|
|
*/
|
|
|
|
int
|
2010-05-28 20:45:14 +00:00
|
|
|
dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, 0, tag,
|
|
|
|
dnp));
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Can only add a reference if there is already at least one
|
|
|
|
* reference on the dnode. Returns FALSE if unable to add a
|
|
|
|
* new reference.
|
|
|
|
*/
|
|
|
|
boolean_t
|
|
|
|
dnode_add_ref(dnode_t *dn, void *tag)
|
|
|
|
{
|
|
|
|
mutex_enter(&dn->dn_mtx);
|
|
|
|
if (refcount_is_zero(&dn->dn_holds)) {
|
|
|
|
mutex_exit(&dn->dn_mtx);
|
|
|
|
return (FALSE);
|
|
|
|
}
|
|
|
|
VERIFY(1 < refcount_add(&dn->dn_holds, tag));
|
|
|
|
mutex_exit(&dn->dn_mtx);
|
|
|
|
return (TRUE);
|
|
|
|
}
|
|
|
|
|
|
|
|
void
|
|
|
|
dnode_rele(dnode_t *dn, void *tag)
|
2015-03-12 00:10:35 +00:00
|
|
|
{
|
|
|
|
mutex_enter(&dn->dn_mtx);
|
|
|
|
dnode_rele_and_unlock(dn, tag);
|
|
|
|
}
|
|
|
|
|
|
|
|
void
|
|
|
|
dnode_rele_and_unlock(dnode_t *dn, void *tag)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
|
|
|
uint64_t refs;
|
2010-08-26 21:24:34 +00:00
|
|
|
/* Get while the hold prevents the dnode from moving. */
|
|
|
|
dmu_buf_impl_t *db = dn->dn_dbuf;
|
|
|
|
dnode_handle_t *dnh = dn->dn_handle;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
refs = refcount_remove(&dn->dn_holds, tag);
|
|
|
|
mutex_exit(&dn->dn_mtx);
|
2010-08-26 21:24:34 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* It's unsafe to release the last hold on a dnode by dnode_rele() or
|
|
|
|
* indirectly by dbuf_rele() while relying on the dnode handle to
|
|
|
|
* prevent the dnode from moving, since releasing the last hold could
|
|
|
|
* result in the dnode's parent dbuf evicting its dnode handles. For
|
|
|
|
* that reason anyone calling dnode_rele() or dbuf_rele() without some
|
|
|
|
* other direct or indirect hold on the dnode must first drop the dnode
|
|
|
|
* handle.
|
|
|
|
*/
|
|
|
|
ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
/* NOTE: the DNODE_DNODE does not have a dn_dbuf */
|
2010-08-26 21:24:34 +00:00
|
|
|
if (refs == 0 && db != NULL) {
|
|
|
|
/*
|
|
|
|
* Another thread could add a hold to the dnode handle in
|
|
|
|
* dnode_hold_impl() while holding the parent dbuf. Since the
|
|
|
|
* hold on the parent dbuf prevents the handle from being
|
|
|
|
* destroyed, the hold on the handle is OK. We can't yet assert
|
|
|
|
* that the handle has zero references, but that will be
|
|
|
|
* asserted anyway when the handle gets destroyed.
|
|
|
|
*/
|
|
|
|
dbuf_rele(db, dnh);
|
|
|
|
}
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
void
|
|
|
|
dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
|
|
|
|
{
|
2010-05-28 20:45:14 +00:00
|
|
|
objset_t *os = dn->dn_objset;
|
2008-11-20 20:01:55 +00:00
|
|
|
uint64_t txg = tx->tx_txg;
|
|
|
|
|
2009-07-02 22:44:48 +00:00
|
|
|
if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
|
|
|
|
dsl_dataset_dirty(os->os_dsl_dataset, tx);
|
2008-11-20 20:01:55 +00:00
|
|
|
return;
|
2009-07-02 22:44:48 +00:00
|
|
|
}
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
DNODE_VERIFY(dn);
|
|
|
|
|
|
|
|
#ifdef ZFS_DEBUG
|
|
|
|
mutex_enter(&dn->dn_mtx);
|
|
|
|
ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
|
2010-08-26 21:24:34 +00:00
|
|
|
ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
|
2008-11-20 20:01:55 +00:00
|
|
|
mutex_exit(&dn->dn_mtx);
|
|
|
|
#endif
|
|
|
|
|
2010-05-28 20:45:14 +00:00
|
|
|
/*
|
|
|
|
* Determine old uid/gid when necessary
|
|
|
|
*/
|
|
|
|
dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
mutex_enter(&os->os_lock);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If we are already marked dirty, we're done.
|
|
|
|
*/
|
|
|
|
if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
|
|
|
|
mutex_exit(&os->os_lock);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
|
2015-04-03 03:14:28 +00:00
|
|
|
ASSERT(!refcount_is_zero(&dn->dn_holds) ||
|
|
|
|
!avl_is_empty(&dn->dn_dbufs));
|
2008-11-20 20:01:55 +00:00
|
|
|
ASSERT(dn->dn_datablksz != 0);
|
2013-05-10 21:17:03 +00:00
|
|
|
ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
|
|
|
|
ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
|
|
|
|
ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
|
|
|
|
dn->dn_object, txg);
|
|
|
|
|
|
|
|
if (dn->dn_free_txg > 0 && dn->dn_free_txg <= txg) {
|
|
|
|
list_insert_tail(&os->os_free_dnodes[txg&TXG_MASK], dn);
|
|
|
|
} else {
|
|
|
|
list_insert_tail(&os->os_dirty_dnodes[txg&TXG_MASK], dn);
|
|
|
|
}
|
|
|
|
|
|
|
|
mutex_exit(&os->os_lock);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The dnode maintains a hold on its containing dbuf as
|
|
|
|
* long as there are holds on it. Each instantiated child
|
2010-08-26 21:24:34 +00:00
|
|
|
* dbuf maintains a hold on the dnode. When the last child
|
2008-11-20 20:01:55 +00:00
|
|
|
* drops its hold, the dnode will drop its hold on the
|
|
|
|
* containing dbuf. We add a "dirty hold" here so that the
|
|
|
|
* dnode will hang around after we finish processing its
|
|
|
|
* children.
|
|
|
|
*/
|
|
|
|
VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
|
|
|
|
|
|
|
|
(void) dbuf_dirty(dn->dn_dbuf, tx);
|
|
|
|
|
|
|
|
dsl_dataset_dirty(os->os_dsl_dataset, tx);
|
|
|
|
}
|
|
|
|
|
|
|
|
void
|
|
|
|
dnode_free(dnode_t *dn, dmu_tx_t *tx)
|
|
|
|
{
|
|
|
|
int txgoff = tx->tx_txg & TXG_MASK;
|
|
|
|
|
|
|
|
dprintf("dn=%p txg=%llu\n", dn, tx->tx_txg);
|
|
|
|
|
|
|
|
/* we should be the only holder... hopefully */
|
|
|
|
/* ASSERT3U(refcount_count(&dn->dn_holds), ==, 1); */
|
|
|
|
|
|
|
|
mutex_enter(&dn->dn_mtx);
|
|
|
|
if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
|
|
|
|
mutex_exit(&dn->dn_mtx);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
dn->dn_free_txg = tx->tx_txg;
|
|
|
|
mutex_exit(&dn->dn_mtx);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If the dnode is already dirty, it needs to be moved from
|
|
|
|
* the dirty list to the free list.
|
|
|
|
*/
|
|
|
|
mutex_enter(&dn->dn_objset->os_lock);
|
|
|
|
if (list_link_active(&dn->dn_dirty_link[txgoff])) {
|
|
|
|
list_remove(&dn->dn_objset->os_dirty_dnodes[txgoff], dn);
|
|
|
|
list_insert_tail(&dn->dn_objset->os_free_dnodes[txgoff], dn);
|
|
|
|
mutex_exit(&dn->dn_objset->os_lock);
|
|
|
|
} else {
|
|
|
|
mutex_exit(&dn->dn_objset->os_lock);
|
|
|
|
dnode_setdirty(dn, tx);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Try to change the block size for the indicated dnode. This can only
|
|
|
|
* succeed if there are no blocks allocated or dirty beyond first block
|
|
|
|
*/
|
|
|
|
int
|
|
|
|
dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
|
|
|
|
{
|
2015-04-03 03:14:28 +00:00
|
|
|
dmu_buf_impl_t *db;
|
2008-12-03 20:09:06 +00:00
|
|
|
int err;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2014-11-03 20:15:08 +00:00
|
|
|
ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
|
2008-11-20 20:01:55 +00:00
|
|
|
if (size == 0)
|
|
|
|
size = SPA_MINBLOCKSIZE;
|
|
|
|
else
|
|
|
|
size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
|
|
|
|
|
|
|
|
if (ibs == dn->dn_indblkshift)
|
|
|
|
ibs = 0;
|
|
|
|
|
|
|
|
if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
|
|
|
|
return (0);
|
|
|
|
|
|
|
|
rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
|
|
|
|
|
|
|
|
/* Check for any allocated blocks beyond the first */
|
Illumos #4101, #4102, #4103, #4105, #4106
4101 metaslab_debug should allow for fine-grained control
4102 space_maps should store more information about themselves
4103 space map object blocksize should be increased
4105 removing a mirrored log device results in a leaked object
4106 asynchronously load metaslab
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Adam Leventhal <ahl@delphix.com>
Reviewed by: Sebastien Roy <seb@delphix.com>
Approved by: Garrett D'Amore <garrett@damore.org>
Prior to this patch, space_maps were preferred solely based on the
amount of free space left in each. Unfortunately, this heuristic didn't
contain any information about the make-up of that free space, which
meant we could keep preferring and loading a highly fragmented space map
that wouldn't actually have enough contiguous space to satisfy the
allocation; then unloading that space_map and repeating the process.
This change modifies the space_map's to store additional information
about the contiguous space in the space_map, so that we can use this
information to make a better decision about which space_map to load.
This requires reallocating all space_map objects to increase their
bonus buffer size sizes enough to fit the new metadata.
The above feature can be enabled via a new feature flag introduced by
this change: com.delphix:spacemap_histogram
In addition to the above, this patch allows the space_map block size to
be increase. Currently the block size is set to be 4K in size, which has
certain implications including the following:
* 4K sector devices will not see any compression benefit
* large space_maps require more metadata on-disk
* large space_maps require more time to load (typically random reads)
Now the space_map block size can adjust as needed up to the maximum size
set via the space_map_max_blksz variable.
A bug was fixed which resulted in potentially leaking an object when
removing a mirrored log device. The previous logic for vdev_remove() did
not deal with removing top-level vdevs that are interior vdevs (i.e.
mirror) correctly. The problem would occur when removing a mirrored log
device, and result in the DTL space map object being leaked; because
top-level vdevs don't have DTL space map objects associated with them.
References:
https://www.illumos.org/issues/4101
https://www.illumos.org/issues/4102
https://www.illumos.org/issues/4103
https://www.illumos.org/issues/4105
https://www.illumos.org/issues/4106
https://github.com/illumos/illumos-gate/commit/0713e23
Porting notes:
A handful of kmem_alloc() calls were converted to kmem_zalloc(). Also,
the KM_PUSHPAGE and TQ_PUSHPAGE flags were used as necessary.
Ported-by: Tim Chase <tim@chase2k.com>
Signed-off-by: Prakash Surya <surya1@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #2488
2013-10-01 21:25:53 +00:00
|
|
|
if (dn->dn_maxblkid != 0)
|
2008-11-20 20:01:55 +00:00
|
|
|
goto fail;
|
|
|
|
|
|
|
|
mutex_enter(&dn->dn_dbufs_mtx);
|
2015-04-03 03:14:28 +00:00
|
|
|
for (db = avl_first(&dn->dn_dbufs); db != NULL;
|
|
|
|
db = AVL_NEXT(&dn->dn_dbufs, db)) {
|
2010-05-28 20:45:14 +00:00
|
|
|
if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
|
|
|
|
db->db_blkid != DMU_SPILL_BLKID) {
|
2008-11-20 20:01:55 +00:00
|
|
|
mutex_exit(&dn->dn_dbufs_mtx);
|
|
|
|
goto fail;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
mutex_exit(&dn->dn_dbufs_mtx);
|
|
|
|
|
|
|
|
if (ibs && dn->dn_nlevels != 1)
|
|
|
|
goto fail;
|
|
|
|
|
2008-12-03 20:09:06 +00:00
|
|
|
/* resize the old block */
|
2015-12-22 01:31:57 +00:00
|
|
|
err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
|
2008-12-03 20:09:06 +00:00
|
|
|
if (err == 0)
|
2008-11-20 20:01:55 +00:00
|
|
|
dbuf_new_size(db, size, tx);
|
2008-12-03 20:09:06 +00:00
|
|
|
else if (err != ENOENT)
|
|
|
|
goto fail;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
dnode_setdblksz(dn, size);
|
|
|
|
dnode_setdirty(dn, tx);
|
|
|
|
dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
|
|
|
|
if (ibs) {
|
|
|
|
dn->dn_indblkshift = ibs;
|
|
|
|
dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
|
|
|
|
}
|
2008-12-03 20:09:06 +00:00
|
|
|
/* rele after we have fixed the blocksize in the dnode */
|
2008-11-20 20:01:55 +00:00
|
|
|
if (db)
|
|
|
|
dbuf_rele(db, FTAG);
|
|
|
|
|
|
|
|
rw_exit(&dn->dn_struct_rwlock);
|
|
|
|
return (0);
|
|
|
|
|
|
|
|
fail:
|
|
|
|
rw_exit(&dn->dn_struct_rwlock);
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(ENOTSUP));
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
2008-12-03 20:09:06 +00:00
|
|
|
/* read-holding callers must not rely on the lock being continuously held */
|
2008-11-20 20:01:55 +00:00
|
|
|
void
|
2008-12-03 20:09:06 +00:00
|
|
|
dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
|
|
|
uint64_t txgoff = tx->tx_txg & TXG_MASK;
|
|
|
|
int epbs, new_nlevels;
|
|
|
|
uint64_t sz;
|
|
|
|
|
2010-05-28 20:45:14 +00:00
|
|
|
ASSERT(blkid != DMU_BONUS_BLKID);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2008-12-03 20:09:06 +00:00
|
|
|
ASSERT(have_read ?
|
|
|
|
RW_READ_HELD(&dn->dn_struct_rwlock) :
|
|
|
|
RW_WRITE_HELD(&dn->dn_struct_rwlock));
|
|
|
|
|
|
|
|
/*
|
|
|
|
* if we have a read-lock, check to see if we need to do any work
|
|
|
|
* before upgrading to a write-lock.
|
|
|
|
*/
|
|
|
|
if (have_read) {
|
|
|
|
if (blkid <= dn->dn_maxblkid)
|
|
|
|
return;
|
|
|
|
|
|
|
|
if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
|
|
|
|
rw_exit(&dn->dn_struct_rwlock);
|
|
|
|
rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
|
|
|
|
}
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
if (blkid <= dn->dn_maxblkid)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
dn->dn_maxblkid = blkid;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Compute the number of levels necessary to support the new maxblkid.
|
|
|
|
*/
|
|
|
|
new_nlevels = 1;
|
|
|
|
epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
|
|
|
|
for (sz = dn->dn_nblkptr;
|
|
|
|
sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
|
|
|
|
new_nlevels++;
|
|
|
|
|
2016-08-31 08:12:08 +00:00
|
|
|
ASSERT3U(new_nlevels, <=, DN_MAX_LEVELS);
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
if (new_nlevels > dn->dn_nlevels) {
|
|
|
|
int old_nlevels = dn->dn_nlevels;
|
|
|
|
dmu_buf_impl_t *db;
|
|
|
|
list_t *list;
|
|
|
|
dbuf_dirty_record_t *new, *dr, *dr_next;
|
|
|
|
|
|
|
|
dn->dn_nlevels = new_nlevels;
|
|
|
|
|
|
|
|
ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
|
|
|
|
dn->dn_next_nlevels[txgoff] = new_nlevels;
|
|
|
|
|
|
|
|
/* dirty the left indirects */
|
|
|
|
db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
|
2010-05-28 20:45:14 +00:00
|
|
|
ASSERT(db != NULL);
|
2008-11-20 20:01:55 +00:00
|
|
|
new = dbuf_dirty(db, tx);
|
|
|
|
dbuf_rele(db, FTAG);
|
|
|
|
|
|
|
|
/* transfer the dirty records to the new indirect */
|
|
|
|
mutex_enter(&dn->dn_mtx);
|
|
|
|
mutex_enter(&new->dt.di.dr_mtx);
|
|
|
|
list = &dn->dn_dirty_records[txgoff];
|
|
|
|
for (dr = list_head(list); dr; dr = dr_next) {
|
|
|
|
dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
|
|
|
|
if (dr->dr_dbuf->db_level != new_nlevels-1 &&
|
2010-05-28 20:45:14 +00:00
|
|
|
dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
|
|
|
|
dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
|
2008-11-20 20:01:55 +00:00
|
|
|
ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
|
|
|
|
list_remove(&dn->dn_dirty_records[txgoff], dr);
|
|
|
|
list_insert_tail(&new->dt.di.dr_children, dr);
|
|
|
|
dr->dr_parent = new;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
mutex_exit(&new->dt.di.dr_mtx);
|
|
|
|
mutex_exit(&dn->dn_mtx);
|
|
|
|
}
|
|
|
|
|
|
|
|
out:
|
2008-12-03 20:09:06 +00:00
|
|
|
if (have_read)
|
|
|
|
rw_downgrade(&dn->dn_struct_rwlock);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
2015-07-02 16:23:20 +00:00
|
|
|
static void
|
|
|
|
dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
|
|
|
|
{
|
|
|
|
dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
|
|
|
|
if (db != NULL) {
|
|
|
|
dmu_buf_will_dirty(&db->db, tx);
|
|
|
|
dbuf_rele(db, FTAG);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
void
|
|
|
|
dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
|
|
|
|
{
|
|
|
|
dmu_buf_impl_t *db;
|
|
|
|
uint64_t blkoff, blkid, nblks;
|
2008-12-03 20:09:06 +00:00
|
|
|
int blksz, blkshift, head, tail;
|
2008-11-20 20:01:55 +00:00
|
|
|
int trunc = FALSE;
|
2008-12-03 20:09:06 +00:00
|
|
|
int epbs;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
|
|
|
|
blksz = dn->dn_datablksz;
|
2008-12-03 20:09:06 +00:00
|
|
|
blkshift = dn->dn_datablkshift;
|
|
|
|
epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2013-08-21 04:11:52 +00:00
|
|
|
if (len == DMU_OBJECT_END) {
|
2008-11-20 20:01:55 +00:00
|
|
|
len = UINT64_MAX - off;
|
|
|
|
trunc = TRUE;
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* First, block align the region to free:
|
|
|
|
*/
|
|
|
|
if (ISP2(blksz)) {
|
|
|
|
head = P2NPHASE(off, blksz);
|
|
|
|
blkoff = P2PHASE(off, blksz);
|
2008-12-03 20:09:06 +00:00
|
|
|
if ((off >> blkshift) > dn->dn_maxblkid)
|
|
|
|
goto out;
|
2008-11-20 20:01:55 +00:00
|
|
|
} else {
|
|
|
|
ASSERT(dn->dn_maxblkid == 0);
|
|
|
|
if (off == 0 && len >= blksz) {
|
2013-12-09 18:37:51 +00:00
|
|
|
/*
|
|
|
|
* Freeing the whole block; fast-track this request.
|
|
|
|
* Note that we won't dirty any indirect blocks,
|
|
|
|
* which is fine because we will be freeing the entire
|
|
|
|
* file and thus all indirect blocks will be freed
|
|
|
|
* by free_children().
|
|
|
|
*/
|
2008-12-03 20:09:06 +00:00
|
|
|
blkid = 0;
|
|
|
|
nblks = 1;
|
|
|
|
goto done;
|
|
|
|
} else if (off >= blksz) {
|
|
|
|
/* Freeing past end-of-data */
|
|
|
|
goto out;
|
2008-11-20 20:01:55 +00:00
|
|
|
} else {
|
|
|
|
/* Freeing part of the block. */
|
|
|
|
head = blksz - off;
|
|
|
|
ASSERT3U(head, >, 0);
|
|
|
|
}
|
|
|
|
blkoff = off;
|
|
|
|
}
|
|
|
|
/* zero out any partial block data at the start of the range */
|
|
|
|
if (head) {
|
|
|
|
ASSERT3U(blkoff + head, ==, blksz);
|
|
|
|
if (len < head)
|
|
|
|
head = len;
|
2015-12-22 01:31:57 +00:00
|
|
|
if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
|
|
|
|
TRUE, FALSE, FTAG, &db) == 0) {
|
2008-11-20 20:01:55 +00:00
|
|
|
caddr_t data;
|
|
|
|
|
|
|
|
/* don't dirty if it isn't on disk and isn't dirty */
|
|
|
|
if (db->db_last_dirty ||
|
|
|
|
(db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
|
|
|
|
rw_exit(&dn->dn_struct_rwlock);
|
2013-12-09 18:37:51 +00:00
|
|
|
dmu_buf_will_dirty(&db->db, tx);
|
2008-11-20 20:01:55 +00:00
|
|
|
rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
|
|
|
|
data = db->db.db_data;
|
|
|
|
bzero(data + blkoff, head);
|
|
|
|
}
|
|
|
|
dbuf_rele(db, FTAG);
|
|
|
|
}
|
|
|
|
off += head;
|
|
|
|
len -= head;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* If the range was less than one block, we're done */
|
2008-12-03 20:09:06 +00:00
|
|
|
if (len == 0)
|
2008-11-20 20:01:55 +00:00
|
|
|
goto out;
|
|
|
|
|
2008-12-03 20:09:06 +00:00
|
|
|
/* If the remaining range is past end of file, we're done */
|
|
|
|
if ((off >> blkshift) > dn->dn_maxblkid)
|
|
|
|
goto out;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2008-12-03 20:09:06 +00:00
|
|
|
ASSERT(ISP2(blksz));
|
|
|
|
if (trunc)
|
|
|
|
tail = 0;
|
|
|
|
else
|
|
|
|
tail = P2PHASE(len, blksz);
|
|
|
|
|
2013-05-10 21:17:03 +00:00
|
|
|
ASSERT0(P2PHASE(off, blksz));
|
2008-12-03 20:09:06 +00:00
|
|
|
/* zero out any partial block data at the end of the range */
|
|
|
|
if (tail) {
|
|
|
|
if (len < tail)
|
|
|
|
tail = len;
|
2015-12-22 01:31:57 +00:00
|
|
|
if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
|
|
|
|
TRUE, FALSE, FTAG, &db) == 0) {
|
2008-12-03 20:09:06 +00:00
|
|
|
/* don't dirty if not on disk and not dirty */
|
|
|
|
if (db->db_last_dirty ||
|
|
|
|
(db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
|
|
|
|
rw_exit(&dn->dn_struct_rwlock);
|
2013-12-09 18:37:51 +00:00
|
|
|
dmu_buf_will_dirty(&db->db, tx);
|
2008-12-03 20:09:06 +00:00
|
|
|
rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
|
|
|
|
bzero(db->db.db_data, tail);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
2008-12-03 20:09:06 +00:00
|
|
|
dbuf_rele(db, FTAG);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
2008-12-03 20:09:06 +00:00
|
|
|
len -= tail;
|
|
|
|
}
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2008-12-03 20:09:06 +00:00
|
|
|
/* If the range did not include a full block, we are done */
|
|
|
|
if (len == 0)
|
|
|
|
goto out;
|
|
|
|
|
|
|
|
ASSERT(IS_P2ALIGNED(off, blksz));
|
|
|
|
ASSERT(trunc || IS_P2ALIGNED(len, blksz));
|
|
|
|
blkid = off >> blkshift;
|
|
|
|
nblks = len >> blkshift;
|
|
|
|
if (trunc)
|
|
|
|
nblks += 1;
|
|
|
|
|
|
|
|
/*
|
2015-07-02 16:23:20 +00:00
|
|
|
* Dirty all the indirect blocks in this range. Note that only
|
|
|
|
* the first and last indirect blocks can actually be written
|
|
|
|
* (if they were partially freed) -- they must be dirtied, even if
|
|
|
|
* they do not exist on disk yet. The interior blocks will
|
|
|
|
* be freed by free_children(), so they will not actually be written.
|
|
|
|
* Even though these interior blocks will not be written, we
|
|
|
|
* dirty them for two reasons:
|
|
|
|
*
|
|
|
|
* - It ensures that the indirect blocks remain in memory until
|
|
|
|
* syncing context. (They have already been prefetched by
|
|
|
|
* dmu_tx_hold_free(), so we don't have to worry about reading
|
|
|
|
* them serially here.)
|
|
|
|
*
|
|
|
|
* - The dirty space accounting will put pressure on the txg sync
|
|
|
|
* mechanism to begin syncing, and to delay transactions if there
|
|
|
|
* is a large amount of freeing. Even though these indirect
|
|
|
|
* blocks will not be written, we could need to write the same
|
|
|
|
* amount of space if we copy the freed BPs into deadlists.
|
2008-12-03 20:09:06 +00:00
|
|
|
*/
|
|
|
|
if (dn->dn_nlevels > 1) {
|
2015-07-02 16:23:20 +00:00
|
|
|
uint64_t first, last, i, ibyte;
|
|
|
|
int shift, err;
|
2008-12-03 20:09:06 +00:00
|
|
|
|
|
|
|
first = blkid >> epbs;
|
2015-07-02 16:23:20 +00:00
|
|
|
dnode_dirty_l1(dn, first, tx);
|
2008-12-03 20:09:06 +00:00
|
|
|
if (trunc)
|
|
|
|
last = dn->dn_maxblkid >> epbs;
|
|
|
|
else
|
|
|
|
last = (blkid + nblks - 1) >> epbs;
|
2015-07-02 16:23:20 +00:00
|
|
|
if (last != first)
|
|
|
|
dnode_dirty_l1(dn, last, tx);
|
|
|
|
|
|
|
|
shift = dn->dn_datablkshift + dn->dn_indblkshift -
|
|
|
|
SPA_BLKPTRSHIFT;
|
|
|
|
for (i = first + 1; i < last; i++) {
|
|
|
|
/*
|
|
|
|
* Set i to the blockid of the next non-hole
|
|
|
|
* level-1 indirect block at or after i. Note
|
|
|
|
* that dnode_next_offset() operates in terms of
|
|
|
|
* level-0-equivalent bytes.
|
|
|
|
*/
|
|
|
|
ibyte = i << shift;
|
|
|
|
err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
|
|
|
|
&ibyte, 2, 1, 0);
|
|
|
|
i = ibyte >> shift;
|
|
|
|
if (i >= last)
|
|
|
|
break;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Normally we should not see an error, either
|
|
|
|
* from dnode_next_offset() or dbuf_hold_level()
|
|
|
|
* (except for ESRCH from dnode_next_offset).
|
|
|
|
* If there is an i/o error, then when we read
|
|
|
|
* this block in syncing context, it will use
|
|
|
|
* ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
|
|
|
|
* to the "failmode" property. dnode_next_offset()
|
|
|
|
* doesn't have a flag to indicate MUSTSUCCEED.
|
|
|
|
*/
|
|
|
|
if (err != 0)
|
|
|
|
break;
|
|
|
|
|
|
|
|
dnode_dirty_l1(dn, i, tx);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
}
|
2013-12-09 18:37:51 +00:00
|
|
|
|
2008-12-03 20:09:06 +00:00
|
|
|
done:
|
|
|
|
/*
|
|
|
|
* Add this range to the dnode range list.
|
|
|
|
* We will finish up this free operation in the syncing phase.
|
|
|
|
*/
|
2008-11-20 20:01:55 +00:00
|
|
|
mutex_enter(&dn->dn_mtx);
|
|
|
|
{
|
2014-04-16 03:40:22 +00:00
|
|
|
int txgoff = tx->tx_txg & TXG_MASK;
|
|
|
|
if (dn->dn_free_ranges[txgoff] == NULL) {
|
|
|
|
dn->dn_free_ranges[txgoff] =
|
|
|
|
range_tree_create(NULL, NULL, &dn->dn_mtx);
|
|
|
|
}
|
|
|
|
range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
|
|
|
|
range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
2014-04-16 03:40:22 +00:00
|
|
|
dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
|
|
|
|
blkid, nblks, tx->tx_txg);
|
2008-11-20 20:01:55 +00:00
|
|
|
mutex_exit(&dn->dn_mtx);
|
|
|
|
|
2008-12-03 20:09:06 +00:00
|
|
|
dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
|
2008-11-20 20:01:55 +00:00
|
|
|
dnode_setdirty(dn, tx);
|
|
|
|
out:
|
2008-12-03 20:09:06 +00:00
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
rw_exit(&dn->dn_struct_rwlock);
|
|
|
|
}
|
|
|
|
|
2010-05-28 20:45:14 +00:00
|
|
|
static boolean_t
|
|
|
|
dnode_spill_freed(dnode_t *dn)
|
|
|
|
{
|
|
|
|
int i;
|
|
|
|
|
|
|
|
mutex_enter(&dn->dn_mtx);
|
|
|
|
for (i = 0; i < TXG_SIZE; i++) {
|
|
|
|
if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
mutex_exit(&dn->dn_mtx);
|
|
|
|
return (i < TXG_SIZE);
|
|
|
|
}
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
/* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
|
|
|
|
uint64_t
|
|
|
|
dnode_block_freed(dnode_t *dn, uint64_t blkid)
|
|
|
|
{
|
|
|
|
void *dp = spa_get_dsl(dn->dn_objset->os_spa);
|
|
|
|
int i;
|
|
|
|
|
2010-05-28 20:45:14 +00:00
|
|
|
if (blkid == DMU_BONUS_BLKID)
|
2008-11-20 20:01:55 +00:00
|
|
|
return (FALSE);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If we're in the process of opening the pool, dp will not be
|
|
|
|
* set yet, but there shouldn't be anything dirty.
|
|
|
|
*/
|
|
|
|
if (dp == NULL)
|
|
|
|
return (FALSE);
|
|
|
|
|
|
|
|
if (dn->dn_free_txg)
|
|
|
|
return (TRUE);
|
|
|
|
|
2010-05-28 20:45:14 +00:00
|
|
|
if (blkid == DMU_SPILL_BLKID)
|
|
|
|
return (dnode_spill_freed(dn));
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
mutex_enter(&dn->dn_mtx);
|
|
|
|
for (i = 0; i < TXG_SIZE; i++) {
|
2014-04-16 03:40:22 +00:00
|
|
|
if (dn->dn_free_ranges[i] != NULL &&
|
|
|
|
range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
|
2008-11-20 20:01:55 +00:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
mutex_exit(&dn->dn_mtx);
|
|
|
|
return (i < TXG_SIZE);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* call from syncing context when we actually write/free space for this dnode */
|
|
|
|
void
|
|
|
|
dnode_diduse_space(dnode_t *dn, int64_t delta)
|
|
|
|
{
|
|
|
|
uint64_t space;
|
|
|
|
dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
|
|
|
|
dn, dn->dn_phys,
|
|
|
|
(u_longlong_t)dn->dn_phys->dn_used,
|
|
|
|
(longlong_t)delta);
|
|
|
|
|
|
|
|
mutex_enter(&dn->dn_mtx);
|
|
|
|
space = DN_USED_BYTES(dn->dn_phys);
|
|
|
|
if (delta > 0) {
|
|
|
|
ASSERT3U(space + delta, >=, space); /* no overflow */
|
|
|
|
} else {
|
|
|
|
ASSERT3U(space, >=, -delta); /* no underflow */
|
|
|
|
}
|
|
|
|
space += delta;
|
|
|
|
if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
|
|
|
|
ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
|
2013-05-10 21:17:03 +00:00
|
|
|
ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
|
2008-11-20 20:01:55 +00:00
|
|
|
dn->dn_phys->dn_used = space >> DEV_BSHIFT;
|
|
|
|
} else {
|
|
|
|
dn->dn_phys->dn_used = space;
|
|
|
|
dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
|
|
|
|
}
|
|
|
|
mutex_exit(&dn->dn_mtx);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
Illumos #4045 write throttle & i/o scheduler performance work
4045 zfs write throttle & i/o scheduler performance work
1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync
read, sync write, async read, async write, and scrub/resilver. The scheduler
issues a number of concurrent i/os from each class to the device. Once a class
has been selected, an i/o is selected from this class using either an elevator
algorithem (async, scrub classes) or FIFO (sync classes). The number of
concurrent async write i/os is tuned dynamically based on i/o load, to achieve
good sync i/o latency when there is not a high load of writes, and good write
throughput when there is. See the block comment in vdev_queue.c (reproduced
below) for more details.
2. The write throttle (dsl_pool_tempreserve_space() and
txg_constrain_throughput()) is rewritten to produce much more consistent delays
when under constant load. The new write throttle is based on the amount of
dirty data, rather than guesses about future performance of the system. When
there is a lot of dirty data, each transaction (e.g. write() syscall) will be
delayed by the same small amount. This eliminates the "brick wall of wait"
that the old write throttle could hit, causing all transactions to wait several
seconds until the next txg opens. One of the keys to the new write throttle is
decrementing the amount of dirty data as i/o completes, rather than at the end
of spa_sync(). Note that the write throttle is only applied once the i/o
scheduler is issuing the maximum number of outstanding async writes. See the
block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for
more details.
This diff has several other effects, including:
* the commonly-tuned global variable zfs_vdev_max_pending has been removed;
use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead.
* the size of each txg (meaning the amount of dirty data written, and thus the
time it takes to write out) is now controlled differently. There is no longer
an explicit time goal; the primary determinant is amount of dirty data.
Systems that are under light or medium load will now often see that a txg is
always syncing, but the impact to performance (e.g. read latency) is minimal.
Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this.
* zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression,
checksum, etc. This improves latency by not allowing these CPU-intensive tasks
to consume all CPU (on machines with at least 4 CPU's; the percentage is
rounded up).
--matt
APPENDIX: problems with the current i/o scheduler
The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem
with this is that if there are always i/os pending, then certain classes of
i/os can see very long delays.
For example, if there are always synchronous reads outstanding, then no async
writes will be serviced until they become "past due". One symptom of this
situation is that each pass of the txg sync takes at least several seconds
(typically 3 seconds).
If many i/os become "past due" (their deadline is in the past), then we must
service all of these overdue i/os before any new i/os. This happens when we
enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in
the future. If we can't complete all the i/os in 2.5 seconds (e.g. because
there were always reads pending), then these i/os will become past due. Now we
must service all the "async" writes (which could be hundreds of megabytes)
before we service any reads, introducing considerable latency to synchronous
i/os (reads or ZIL writes).
Notes on porting to ZFS on Linux:
- zio_t gained new members io_physdone and io_phys_children. Because
object caches in the Linux port call the constructor only once at
allocation time, objects may contain residual data when retrieved
from the cache. Therefore zio_create() was updated to zero out the two
new fields.
- vdev_mirror_pending() relied on the depth of the per-vdev pending queue
(vq->vq_pending_tree) to select the least-busy leaf vdev to read from.
This tree has been replaced by vq->vq_active_tree which is now used
for the same purpose.
- vdev_queue_init() used the value of zfs_vdev_max_pending to determine
the number of vdev I/O buffers to pre-allocate. That global no longer
exists, so we instead use the sum of the *_max_active values for each of
the five I/O classes described above.
- The Illumos implementation of dmu_tx_delay() delays a transaction by
sleeping in condition variable embedded in the thread
(curthread->t_delay_cv). We do not have an equivalent CV to use in
Linux, so this change replaced the delay logic with a wrapper called
zfs_sleep_until(). This wrapper could be adopted upstream and in other
downstream ports to abstract away operating system-specific delay logic.
- These tunables are added as module parameters, and descriptions added
to the zfs-module-parameters.5 man page.
spa_asize_inflation
zfs_deadman_synctime_ms
zfs_vdev_max_active
zfs_vdev_async_write_active_min_dirty_percent
zfs_vdev_async_write_active_max_dirty_percent
zfs_vdev_async_read_max_active
zfs_vdev_async_read_min_active
zfs_vdev_async_write_max_active
zfs_vdev_async_write_min_active
zfs_vdev_scrub_max_active
zfs_vdev_scrub_min_active
zfs_vdev_sync_read_max_active
zfs_vdev_sync_read_min_active
zfs_vdev_sync_write_max_active
zfs_vdev_sync_write_min_active
zfs_dirty_data_max_percent
zfs_delay_min_dirty_percent
zfs_dirty_data_max_max_percent
zfs_dirty_data_max
zfs_dirty_data_max_max
zfs_dirty_data_sync
zfs_delay_scale
The latter four have type unsigned long, whereas they are uint64_t in
Illumos. This accommodates Linux's module_param() supported types, but
means they may overflow on 32-bit architectures.
The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most
likely to overflow on 32-bit systems, since they express physical RAM
sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to
2^32 which does overflow. To resolve that, this port instead initializes
it in arc_init() to 25% of physical RAM, and adds the tunable
zfs_dirty_data_max_max_percent to override that percentage. While this
solution doesn't completely avoid the overflow issue, it should be a
reasonable default for most systems, and the minority of affected
systems can work around the issue by overriding the defaults.
- Fixed reversed logic in comment above zfs_delay_scale declaration.
- Clarified comments in vdev_queue.c regarding when per-queue minimums take
effect.
- Replaced dmu_tx_write_limit in the dmu_tx kstat file
with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts
how many times a transaction has been delayed because the pool dirty
data has exceeded zfs_delay_min_dirty_percent. The latter counts how
many times the pool dirty data has exceeded zfs_dirty_data_max (which
we expect to never happen).
- The original patch would have regressed the bug fixed in
zfsonlinux/zfs@c418410, which prevented users from setting the
zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE.
A similar fix is added to vdev_queue_aggregate().
- In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the
heap instead of the stack. In Linux we can't afford such large
structures on the stack.
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Adam Leventhal <ahl@delphix.com>
Reviewed by: Christopher Siden <christopher.siden@delphix.com>
Reviewed by: Ned Bass <bass6@llnl.gov>
Reviewed by: Brendan Gregg <brendan.gregg@joyent.com>
Approved by: Robert Mustacchi <rm@joyent.com>
References:
http://www.illumos.org/issues/4045
illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e
Ported-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #1913
2013-08-29 03:01:20 +00:00
|
|
|
* Call when we think we're going to write/free space in open context to track
|
|
|
|
* the amount of memory in use by the currently open txg.
|
2008-11-20 20:01:55 +00:00
|
|
|
*/
|
|
|
|
void
|
|
|
|
dnode_willuse_space(dnode_t *dn, int64_t space, dmu_tx_t *tx)
|
|
|
|
{
|
2010-05-28 20:45:14 +00:00
|
|
|
objset_t *os = dn->dn_objset;
|
2008-11-20 20:01:55 +00:00
|
|
|
dsl_dataset_t *ds = os->os_dsl_dataset;
|
Illumos #4045 write throttle & i/o scheduler performance work
4045 zfs write throttle & i/o scheduler performance work
1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync
read, sync write, async read, async write, and scrub/resilver. The scheduler
issues a number of concurrent i/os from each class to the device. Once a class
has been selected, an i/o is selected from this class using either an elevator
algorithem (async, scrub classes) or FIFO (sync classes). The number of
concurrent async write i/os is tuned dynamically based on i/o load, to achieve
good sync i/o latency when there is not a high load of writes, and good write
throughput when there is. See the block comment in vdev_queue.c (reproduced
below) for more details.
2. The write throttle (dsl_pool_tempreserve_space() and
txg_constrain_throughput()) is rewritten to produce much more consistent delays
when under constant load. The new write throttle is based on the amount of
dirty data, rather than guesses about future performance of the system. When
there is a lot of dirty data, each transaction (e.g. write() syscall) will be
delayed by the same small amount. This eliminates the "brick wall of wait"
that the old write throttle could hit, causing all transactions to wait several
seconds until the next txg opens. One of the keys to the new write throttle is
decrementing the amount of dirty data as i/o completes, rather than at the end
of spa_sync(). Note that the write throttle is only applied once the i/o
scheduler is issuing the maximum number of outstanding async writes. See the
block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for
more details.
This diff has several other effects, including:
* the commonly-tuned global variable zfs_vdev_max_pending has been removed;
use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead.
* the size of each txg (meaning the amount of dirty data written, and thus the
time it takes to write out) is now controlled differently. There is no longer
an explicit time goal; the primary determinant is amount of dirty data.
Systems that are under light or medium load will now often see that a txg is
always syncing, but the impact to performance (e.g. read latency) is minimal.
Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this.
* zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression,
checksum, etc. This improves latency by not allowing these CPU-intensive tasks
to consume all CPU (on machines with at least 4 CPU's; the percentage is
rounded up).
--matt
APPENDIX: problems with the current i/o scheduler
The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem
with this is that if there are always i/os pending, then certain classes of
i/os can see very long delays.
For example, if there are always synchronous reads outstanding, then no async
writes will be serviced until they become "past due". One symptom of this
situation is that each pass of the txg sync takes at least several seconds
(typically 3 seconds).
If many i/os become "past due" (their deadline is in the past), then we must
service all of these overdue i/os before any new i/os. This happens when we
enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in
the future. If we can't complete all the i/os in 2.5 seconds (e.g. because
there were always reads pending), then these i/os will become past due. Now we
must service all the "async" writes (which could be hundreds of megabytes)
before we service any reads, introducing considerable latency to synchronous
i/os (reads or ZIL writes).
Notes on porting to ZFS on Linux:
- zio_t gained new members io_physdone and io_phys_children. Because
object caches in the Linux port call the constructor only once at
allocation time, objects may contain residual data when retrieved
from the cache. Therefore zio_create() was updated to zero out the two
new fields.
- vdev_mirror_pending() relied on the depth of the per-vdev pending queue
(vq->vq_pending_tree) to select the least-busy leaf vdev to read from.
This tree has been replaced by vq->vq_active_tree which is now used
for the same purpose.
- vdev_queue_init() used the value of zfs_vdev_max_pending to determine
the number of vdev I/O buffers to pre-allocate. That global no longer
exists, so we instead use the sum of the *_max_active values for each of
the five I/O classes described above.
- The Illumos implementation of dmu_tx_delay() delays a transaction by
sleeping in condition variable embedded in the thread
(curthread->t_delay_cv). We do not have an equivalent CV to use in
Linux, so this change replaced the delay logic with a wrapper called
zfs_sleep_until(). This wrapper could be adopted upstream and in other
downstream ports to abstract away operating system-specific delay logic.
- These tunables are added as module parameters, and descriptions added
to the zfs-module-parameters.5 man page.
spa_asize_inflation
zfs_deadman_synctime_ms
zfs_vdev_max_active
zfs_vdev_async_write_active_min_dirty_percent
zfs_vdev_async_write_active_max_dirty_percent
zfs_vdev_async_read_max_active
zfs_vdev_async_read_min_active
zfs_vdev_async_write_max_active
zfs_vdev_async_write_min_active
zfs_vdev_scrub_max_active
zfs_vdev_scrub_min_active
zfs_vdev_sync_read_max_active
zfs_vdev_sync_read_min_active
zfs_vdev_sync_write_max_active
zfs_vdev_sync_write_min_active
zfs_dirty_data_max_percent
zfs_delay_min_dirty_percent
zfs_dirty_data_max_max_percent
zfs_dirty_data_max
zfs_dirty_data_max_max
zfs_dirty_data_sync
zfs_delay_scale
The latter four have type unsigned long, whereas they are uint64_t in
Illumos. This accommodates Linux's module_param() supported types, but
means they may overflow on 32-bit architectures.
The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most
likely to overflow on 32-bit systems, since they express physical RAM
sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to
2^32 which does overflow. To resolve that, this port instead initializes
it in arc_init() to 25% of physical RAM, and adds the tunable
zfs_dirty_data_max_max_percent to override that percentage. While this
solution doesn't completely avoid the overflow issue, it should be a
reasonable default for most systems, and the minority of affected
systems can work around the issue by overriding the defaults.
- Fixed reversed logic in comment above zfs_delay_scale declaration.
- Clarified comments in vdev_queue.c regarding when per-queue minimums take
effect.
- Replaced dmu_tx_write_limit in the dmu_tx kstat file
with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts
how many times a transaction has been delayed because the pool dirty
data has exceeded zfs_delay_min_dirty_percent. The latter counts how
many times the pool dirty data has exceeded zfs_dirty_data_max (which
we expect to never happen).
- The original patch would have regressed the bug fixed in
zfsonlinux/zfs@c418410, which prevented users from setting the
zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE.
A similar fix is added to vdev_queue_aggregate().
- In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the
heap instead of the stack. In Linux we can't afford such large
structures on the stack.
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Adam Leventhal <ahl@delphix.com>
Reviewed by: Christopher Siden <christopher.siden@delphix.com>
Reviewed by: Ned Bass <bass6@llnl.gov>
Reviewed by: Brendan Gregg <brendan.gregg@joyent.com>
Approved by: Robert Mustacchi <rm@joyent.com>
References:
http://www.illumos.org/issues/4045
illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e
Ported-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #1913
2013-08-29 03:01:20 +00:00
|
|
|
int64_t aspace = spa_get_asize(os->os_spa, space);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
Illumos #4045 write throttle & i/o scheduler performance work
4045 zfs write throttle & i/o scheduler performance work
1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync
read, sync write, async read, async write, and scrub/resilver. The scheduler
issues a number of concurrent i/os from each class to the device. Once a class
has been selected, an i/o is selected from this class using either an elevator
algorithem (async, scrub classes) or FIFO (sync classes). The number of
concurrent async write i/os is tuned dynamically based on i/o load, to achieve
good sync i/o latency when there is not a high load of writes, and good write
throughput when there is. See the block comment in vdev_queue.c (reproduced
below) for more details.
2. The write throttle (dsl_pool_tempreserve_space() and
txg_constrain_throughput()) is rewritten to produce much more consistent delays
when under constant load. The new write throttle is based on the amount of
dirty data, rather than guesses about future performance of the system. When
there is a lot of dirty data, each transaction (e.g. write() syscall) will be
delayed by the same small amount. This eliminates the "brick wall of wait"
that the old write throttle could hit, causing all transactions to wait several
seconds until the next txg opens. One of the keys to the new write throttle is
decrementing the amount of dirty data as i/o completes, rather than at the end
of spa_sync(). Note that the write throttle is only applied once the i/o
scheduler is issuing the maximum number of outstanding async writes. See the
block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for
more details.
This diff has several other effects, including:
* the commonly-tuned global variable zfs_vdev_max_pending has been removed;
use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead.
* the size of each txg (meaning the amount of dirty data written, and thus the
time it takes to write out) is now controlled differently. There is no longer
an explicit time goal; the primary determinant is amount of dirty data.
Systems that are under light or medium load will now often see that a txg is
always syncing, but the impact to performance (e.g. read latency) is minimal.
Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this.
* zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression,
checksum, etc. This improves latency by not allowing these CPU-intensive tasks
to consume all CPU (on machines with at least 4 CPU's; the percentage is
rounded up).
--matt
APPENDIX: problems with the current i/o scheduler
The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem
with this is that if there are always i/os pending, then certain classes of
i/os can see very long delays.
For example, if there are always synchronous reads outstanding, then no async
writes will be serviced until they become "past due". One symptom of this
situation is that each pass of the txg sync takes at least several seconds
(typically 3 seconds).
If many i/os become "past due" (their deadline is in the past), then we must
service all of these overdue i/os before any new i/os. This happens when we
enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in
the future. If we can't complete all the i/os in 2.5 seconds (e.g. because
there were always reads pending), then these i/os will become past due. Now we
must service all the "async" writes (which could be hundreds of megabytes)
before we service any reads, introducing considerable latency to synchronous
i/os (reads or ZIL writes).
Notes on porting to ZFS on Linux:
- zio_t gained new members io_physdone and io_phys_children. Because
object caches in the Linux port call the constructor only once at
allocation time, objects may contain residual data when retrieved
from the cache. Therefore zio_create() was updated to zero out the two
new fields.
- vdev_mirror_pending() relied on the depth of the per-vdev pending queue
(vq->vq_pending_tree) to select the least-busy leaf vdev to read from.
This tree has been replaced by vq->vq_active_tree which is now used
for the same purpose.
- vdev_queue_init() used the value of zfs_vdev_max_pending to determine
the number of vdev I/O buffers to pre-allocate. That global no longer
exists, so we instead use the sum of the *_max_active values for each of
the five I/O classes described above.
- The Illumos implementation of dmu_tx_delay() delays a transaction by
sleeping in condition variable embedded in the thread
(curthread->t_delay_cv). We do not have an equivalent CV to use in
Linux, so this change replaced the delay logic with a wrapper called
zfs_sleep_until(). This wrapper could be adopted upstream and in other
downstream ports to abstract away operating system-specific delay logic.
- These tunables are added as module parameters, and descriptions added
to the zfs-module-parameters.5 man page.
spa_asize_inflation
zfs_deadman_synctime_ms
zfs_vdev_max_active
zfs_vdev_async_write_active_min_dirty_percent
zfs_vdev_async_write_active_max_dirty_percent
zfs_vdev_async_read_max_active
zfs_vdev_async_read_min_active
zfs_vdev_async_write_max_active
zfs_vdev_async_write_min_active
zfs_vdev_scrub_max_active
zfs_vdev_scrub_min_active
zfs_vdev_sync_read_max_active
zfs_vdev_sync_read_min_active
zfs_vdev_sync_write_max_active
zfs_vdev_sync_write_min_active
zfs_dirty_data_max_percent
zfs_delay_min_dirty_percent
zfs_dirty_data_max_max_percent
zfs_dirty_data_max
zfs_dirty_data_max_max
zfs_dirty_data_sync
zfs_delay_scale
The latter four have type unsigned long, whereas they are uint64_t in
Illumos. This accommodates Linux's module_param() supported types, but
means they may overflow on 32-bit architectures.
The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most
likely to overflow on 32-bit systems, since they express physical RAM
sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to
2^32 which does overflow. To resolve that, this port instead initializes
it in arc_init() to 25% of physical RAM, and adds the tunable
zfs_dirty_data_max_max_percent to override that percentage. While this
solution doesn't completely avoid the overflow issue, it should be a
reasonable default for most systems, and the minority of affected
systems can work around the issue by overriding the defaults.
- Fixed reversed logic in comment above zfs_delay_scale declaration.
- Clarified comments in vdev_queue.c regarding when per-queue minimums take
effect.
- Replaced dmu_tx_write_limit in the dmu_tx kstat file
with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts
how many times a transaction has been delayed because the pool dirty
data has exceeded zfs_delay_min_dirty_percent. The latter counts how
many times the pool dirty data has exceeded zfs_dirty_data_max (which
we expect to never happen).
- The original patch would have regressed the bug fixed in
zfsonlinux/zfs@c418410, which prevented users from setting the
zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE.
A similar fix is added to vdev_queue_aggregate().
- In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the
heap instead of the stack. In Linux we can't afford such large
structures on the stack.
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Adam Leventhal <ahl@delphix.com>
Reviewed by: Christopher Siden <christopher.siden@delphix.com>
Reviewed by: Ned Bass <bass6@llnl.gov>
Reviewed by: Brendan Gregg <brendan.gregg@joyent.com>
Approved by: Robert Mustacchi <rm@joyent.com>
References:
http://www.illumos.org/issues/4045
illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e
Ported-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #1913
2013-08-29 03:01:20 +00:00
|
|
|
if (ds != NULL) {
|
|
|
|
dsl_dir_willuse_space(ds->ds_dir, aspace, tx);
|
|
|
|
dsl_pool_dirty_space(dmu_tx_pool(tx), space, tx);
|
|
|
|
}
|
2008-11-20 20:01:55 +00:00
|
|
|
|
Illumos #4045 write throttle & i/o scheduler performance work
4045 zfs write throttle & i/o scheduler performance work
1. The ZFS i/o scheduler (vdev_queue.c) now divides i/os into 5 classes: sync
read, sync write, async read, async write, and scrub/resilver. The scheduler
issues a number of concurrent i/os from each class to the device. Once a class
has been selected, an i/o is selected from this class using either an elevator
algorithem (async, scrub classes) or FIFO (sync classes). The number of
concurrent async write i/os is tuned dynamically based on i/o load, to achieve
good sync i/o latency when there is not a high load of writes, and good write
throughput when there is. See the block comment in vdev_queue.c (reproduced
below) for more details.
2. The write throttle (dsl_pool_tempreserve_space() and
txg_constrain_throughput()) is rewritten to produce much more consistent delays
when under constant load. The new write throttle is based on the amount of
dirty data, rather than guesses about future performance of the system. When
there is a lot of dirty data, each transaction (e.g. write() syscall) will be
delayed by the same small amount. This eliminates the "brick wall of wait"
that the old write throttle could hit, causing all transactions to wait several
seconds until the next txg opens. One of the keys to the new write throttle is
decrementing the amount of dirty data as i/o completes, rather than at the end
of spa_sync(). Note that the write throttle is only applied once the i/o
scheduler is issuing the maximum number of outstanding async writes. See the
block comments in dsl_pool.c and above dmu_tx_delay() (reproduced below) for
more details.
This diff has several other effects, including:
* the commonly-tuned global variable zfs_vdev_max_pending has been removed;
use per-class zfs_vdev_*_max_active values or zfs_vdev_max_active instead.
* the size of each txg (meaning the amount of dirty data written, and thus the
time it takes to write out) is now controlled differently. There is no longer
an explicit time goal; the primary determinant is amount of dirty data.
Systems that are under light or medium load will now often see that a txg is
always syncing, but the impact to performance (e.g. read latency) is minimal.
Tune zfs_dirty_data_max and zfs_dirty_data_sync to control this.
* zio_taskq_batch_pct = 75 -- Only use 75% of all CPUs for compression,
checksum, etc. This improves latency by not allowing these CPU-intensive tasks
to consume all CPU (on machines with at least 4 CPU's; the percentage is
rounded up).
--matt
APPENDIX: problems with the current i/o scheduler
The current ZFS i/o scheduler (vdev_queue.c) is deadline based. The problem
with this is that if there are always i/os pending, then certain classes of
i/os can see very long delays.
For example, if there are always synchronous reads outstanding, then no async
writes will be serviced until they become "past due". One symptom of this
situation is that each pass of the txg sync takes at least several seconds
(typically 3 seconds).
If many i/os become "past due" (their deadline is in the past), then we must
service all of these overdue i/os before any new i/os. This happens when we
enqueue a batch of async writes for the txg sync, with deadlines 2.5 seconds in
the future. If we can't complete all the i/os in 2.5 seconds (e.g. because
there were always reads pending), then these i/os will become past due. Now we
must service all the "async" writes (which could be hundreds of megabytes)
before we service any reads, introducing considerable latency to synchronous
i/os (reads or ZIL writes).
Notes on porting to ZFS on Linux:
- zio_t gained new members io_physdone and io_phys_children. Because
object caches in the Linux port call the constructor only once at
allocation time, objects may contain residual data when retrieved
from the cache. Therefore zio_create() was updated to zero out the two
new fields.
- vdev_mirror_pending() relied on the depth of the per-vdev pending queue
(vq->vq_pending_tree) to select the least-busy leaf vdev to read from.
This tree has been replaced by vq->vq_active_tree which is now used
for the same purpose.
- vdev_queue_init() used the value of zfs_vdev_max_pending to determine
the number of vdev I/O buffers to pre-allocate. That global no longer
exists, so we instead use the sum of the *_max_active values for each of
the five I/O classes described above.
- The Illumos implementation of dmu_tx_delay() delays a transaction by
sleeping in condition variable embedded in the thread
(curthread->t_delay_cv). We do not have an equivalent CV to use in
Linux, so this change replaced the delay logic with a wrapper called
zfs_sleep_until(). This wrapper could be adopted upstream and in other
downstream ports to abstract away operating system-specific delay logic.
- These tunables are added as module parameters, and descriptions added
to the zfs-module-parameters.5 man page.
spa_asize_inflation
zfs_deadman_synctime_ms
zfs_vdev_max_active
zfs_vdev_async_write_active_min_dirty_percent
zfs_vdev_async_write_active_max_dirty_percent
zfs_vdev_async_read_max_active
zfs_vdev_async_read_min_active
zfs_vdev_async_write_max_active
zfs_vdev_async_write_min_active
zfs_vdev_scrub_max_active
zfs_vdev_scrub_min_active
zfs_vdev_sync_read_max_active
zfs_vdev_sync_read_min_active
zfs_vdev_sync_write_max_active
zfs_vdev_sync_write_min_active
zfs_dirty_data_max_percent
zfs_delay_min_dirty_percent
zfs_dirty_data_max_max_percent
zfs_dirty_data_max
zfs_dirty_data_max_max
zfs_dirty_data_sync
zfs_delay_scale
The latter four have type unsigned long, whereas they are uint64_t in
Illumos. This accommodates Linux's module_param() supported types, but
means they may overflow on 32-bit architectures.
The values zfs_dirty_data_max and zfs_dirty_data_max_max are the most
likely to overflow on 32-bit systems, since they express physical RAM
sizes in bytes. In fact, Illumos initializes zfs_dirty_data_max_max to
2^32 which does overflow. To resolve that, this port instead initializes
it in arc_init() to 25% of physical RAM, and adds the tunable
zfs_dirty_data_max_max_percent to override that percentage. While this
solution doesn't completely avoid the overflow issue, it should be a
reasonable default for most systems, and the minority of affected
systems can work around the issue by overriding the defaults.
- Fixed reversed logic in comment above zfs_delay_scale declaration.
- Clarified comments in vdev_queue.c regarding when per-queue minimums take
effect.
- Replaced dmu_tx_write_limit in the dmu_tx kstat file
with dmu_tx_dirty_delay and dmu_tx_dirty_over_max. The first counts
how many times a transaction has been delayed because the pool dirty
data has exceeded zfs_delay_min_dirty_percent. The latter counts how
many times the pool dirty data has exceeded zfs_dirty_data_max (which
we expect to never happen).
- The original patch would have regressed the bug fixed in
zfsonlinux/zfs@c418410, which prevented users from setting the
zfs_vdev_aggregation_limit tuning larger than SPA_MAXBLOCKSIZE.
A similar fix is added to vdev_queue_aggregate().
- In vdev_queue_io_to_issue(), dynamically allocate 'zio_t search' on the
heap instead of the stack. In Linux we can't afford such large
structures on the stack.
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Adam Leventhal <ahl@delphix.com>
Reviewed by: Christopher Siden <christopher.siden@delphix.com>
Reviewed by: Ned Bass <bass6@llnl.gov>
Reviewed by: Brendan Gregg <brendan.gregg@joyent.com>
Approved by: Robert Mustacchi <rm@joyent.com>
References:
http://www.illumos.org/issues/4045
illumos/illumos-gate@69962b5647e4a8b9b14998733b765925381b727e
Ported-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #1913
2013-08-29 03:01:20 +00:00
|
|
|
dmu_tx_willuse_space(tx, aspace);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
2009-08-18 18:43:27 +00:00
|
|
|
/*
|
2013-06-11 17:12:34 +00:00
|
|
|
* Scans a block at the indicated "level" looking for a hole or data,
|
|
|
|
* depending on 'flags'.
|
|
|
|
*
|
|
|
|
* If level > 0, then we are scanning an indirect block looking at its
|
|
|
|
* pointers. If level == 0, then we are looking at a block of dnodes.
|
|
|
|
*
|
|
|
|
* If we don't find what we are looking for in the block, we return ESRCH.
|
|
|
|
* Otherwise, return with *offset pointing to the beginning (if searching
|
|
|
|
* forwards) or end (if searching backwards) of the range covered by the
|
|
|
|
* block pointer we matched on (or dnode).
|
2009-08-18 18:43:27 +00:00
|
|
|
*
|
|
|
|
* The basic search algorithm used below by dnode_next_offset() is to
|
|
|
|
* use this function to search up the block tree (widen the search) until
|
|
|
|
* we find something (i.e., we don't return ESRCH) and then search back
|
|
|
|
* down the tree (narrow the search) until we reach our original search
|
|
|
|
* level.
|
|
|
|
*/
|
2008-11-20 20:01:55 +00:00
|
|
|
static int
|
2008-12-03 20:09:06 +00:00
|
|
|
dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
|
2015-12-22 01:31:57 +00:00
|
|
|
int lvl, uint64_t blkfill, uint64_t txg)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
|
|
|
dmu_buf_impl_t *db = NULL;
|
|
|
|
void *data = NULL;
|
|
|
|
uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
|
|
|
|
uint64_t epb = 1ULL << epbs;
|
|
|
|
uint64_t minfill, maxfill;
|
2008-12-03 20:09:06 +00:00
|
|
|
boolean_t hole;
|
|
|
|
int i, inc, error, span;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2009-07-02 22:44:48 +00:00
|
|
|
hole = ((flags & DNODE_FIND_HOLE) != 0);
|
2008-12-03 20:09:06 +00:00
|
|
|
inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
|
|
|
|
ASSERT(txg == 0 || !hole);
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
if (lvl == dn->dn_phys->dn_nlevels) {
|
|
|
|
error = 0;
|
|
|
|
epb = dn->dn_phys->dn_nblkptr;
|
|
|
|
data = dn->dn_phys->dn_blkptr;
|
|
|
|
} else {
|
2015-12-22 01:31:57 +00:00
|
|
|
uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
|
|
|
|
error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
|
2008-11-20 20:01:55 +00:00
|
|
|
if (error) {
|
2008-12-03 20:09:06 +00:00
|
|
|
if (error != ENOENT)
|
|
|
|
return (error);
|
|
|
|
if (hole)
|
|
|
|
return (0);
|
|
|
|
/*
|
|
|
|
* This can only happen when we are searching up
|
|
|
|
* the block tree for data. We don't really need to
|
|
|
|
* adjust the offset, as we will just end up looking
|
|
|
|
* at the pointer to this block in its parent, and its
|
|
|
|
* going to be unallocated, so we will skip over it.
|
|
|
|
*/
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(ESRCH));
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
|
|
|
|
if (error) {
|
|
|
|
dbuf_rele(db, FTAG);
|
|
|
|
return (error);
|
|
|
|
}
|
|
|
|
data = db->db.db_data;
|
|
|
|
}
|
|
|
|
|
2013-12-09 18:37:51 +00:00
|
|
|
|
|
|
|
if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
|
|
|
|
db->db_blkptr->blk_birth <= txg ||
|
|
|
|
BP_IS_HOLE(db->db_blkptr))) {
|
2008-12-03 20:09:06 +00:00
|
|
|
/*
|
|
|
|
* This can only happen when we are searching up the tree
|
|
|
|
* and these conditions mean that we need to keep climbing.
|
|
|
|
*/
|
2013-03-08 18:41:28 +00:00
|
|
|
error = SET_ERROR(ESRCH);
|
2008-11-20 20:01:55 +00:00
|
|
|
} else if (lvl == 0) {
|
|
|
|
dnode_phys_t *dnp = data;
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
ASSERT(dn->dn_type == DMU_OT_DNODE);
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
ASSERT(!(flags & DNODE_FIND_BACKWARDS));
|
2008-11-20 20:01:55 +00:00
|
|
|
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
for (i = (*offset >> DNODE_SHIFT) & (blkfill - 1);
|
|
|
|
i < blkfill; i += dnp[i].dn_extra_slots + 1) {
|
2009-07-02 22:44:48 +00:00
|
|
|
if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
|
2008-11-20 20:01:55 +00:00
|
|
|
break;
|
|
|
|
}
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
|
|
|
|
if (i == blkfill)
|
2013-03-08 18:41:28 +00:00
|
|
|
error = SET_ERROR(ESRCH);
|
Implement large_dnode pool feature
Justification
-------------
This feature adds support for variable length dnodes. Our motivation is
to eliminate the overhead associated with using spill blocks. Spill
blocks are used to store system attribute data (i.e. file metadata) that
does not fit in the dnode's bonus buffer. By allowing a larger bonus
buffer area the use of a spill block can be avoided. Spill blocks
potentially incur an additional read I/O for every dnode in a dnode
block. As a worst case example, reading 32 dnodes from a 16k dnode block
and all of the spill blocks could issue 33 separate reads. Now suppose
those dnodes have size 1024 and therefore don't need spill blocks. Then
the worst case number of blocks read is reduced to from 33 to two--one
per dnode block. In practice spill blocks may tend to be co-located on
disk with the dnode blocks so the reduction in I/O would not be this
drastic. In a badly fragmented pool, however, the improvement could be
significant.
ZFS-on-Linux systems that make heavy use of extended attributes would
benefit from this feature. In particular, ZFS-on-Linux supports the
xattr=sa dataset property which allows file extended attribute data
to be stored in the dnode bonus buffer as an alternative to the
traditional directory-based format. Workloads such as SELinux and the
Lustre distributed filesystem often store enough xattr data to force
spill bocks when xattr=sa is in effect. Large dnodes may therefore
provide a performance benefit to such systems.
Other use cases that may benefit from this feature include files with
large ACLs and symbolic links with long target names. Furthermore,
this feature may be desirable on other platforms in case future
applications or features are developed that could make use of a
larger bonus buffer area.
Implementation
--------------
The size of a dnode may be a multiple of 512 bytes up to the size of
a dnode block (currently 16384 bytes). A dn_extra_slots field was
added to the current on-disk dnode_phys_t structure to describe the
size of the physical dnode on disk. The 8 bits for this field were
taken from the zero filled dn_pad2 field. The field represents how
many "extra" dnode_phys_t slots a dnode consumes in its dnode block.
This convention results in a value of 0 for 512 byte dnodes which
preserves on-disk format compatibility with older software.
Similarly, the in-memory dnode_t structure has a new dn_num_slots field
to represent the total number of dnode_phys_t slots consumed on disk.
Thus dn->dn_num_slots is 1 greater than the corresponding
dnp->dn_extra_slots. This difference in convention was adopted
because, unlike on-disk structures, backward compatibility is not a
concern for in-memory objects, so we used a more natural way to
represent size for a dnode_t.
The default size for newly created dnodes is determined by the value of
a new "dnodesize" dataset property. By default the property is set to
"legacy" which is compatible with older software. Setting the property
to "auto" will allow the filesystem to choose the most suitable dnode
size. Currently this just sets the default dnode size to 1k, but future
code improvements could dynamically choose a size based on observed
workload patterns. Dnodes of varying sizes can coexist within the same
dataset and even within the same dnode block. For example, to enable
automatically-sized dnodes, run
# zfs set dnodesize=auto tank/fish
The user can also specify literal values for the dnodesize property.
These are currently limited to powers of two from 1k to 16k. The
power-of-2 limitation is only for simplicity of the user interface.
Internally the implementation can handle any multiple of 512 up to 16k,
and consumers of the DMU API can specify any legal dnode value.
The size of a new dnode is determined at object allocation time and
stored as a new field in the znode in-memory structure. New DMU
interfaces are added to allow the consumer to specify the dnode size
that a newly allocated object should use. Existing interfaces are
unchanged to avoid having to update every call site and to preserve
compatibility with external consumers such as Lustre. The new
interfaces names are given below. The versions of these functions that
don't take a dnodesize parameter now just call the _dnsize() versions
with a dnodesize of 0, which means use the legacy dnode size.
New DMU interfaces:
dmu_object_alloc_dnsize()
dmu_object_claim_dnsize()
dmu_object_reclaim_dnsize()
New ZAP interfaces:
zap_create_dnsize()
zap_create_norm_dnsize()
zap_create_flags_dnsize()
zap_create_claim_norm_dnsize()
zap_create_link_dnsize()
The constant DN_MAX_BONUSLEN is renamed to DN_OLD_MAX_BONUSLEN. The
spa_maxdnodesize() function should be used to determine the maximum
bonus length for a pool.
These are a few noteworthy changes to key functions:
* The prototype for dnode_hold_impl() now takes a "slots" parameter.
When the DNODE_MUST_BE_FREE flag is set, this parameter is used to
ensure the hole at the specified object offset is large enough to
hold the dnode being created. The slots parameter is also used
to ensure a dnode does not span multiple dnode blocks. In both of
these cases, if a failure occurs, ENOSPC is returned. Keep in mind,
these failure cases are only possible when using DNODE_MUST_BE_FREE.
If the DNODE_MUST_BE_ALLOCATED flag is set, "slots" must be 0.
dnode_hold_impl() will check if the requested dnode is already
consumed as an extra dnode slot by an large dnode, in which case
it returns ENOENT.
* The function dmu_object_alloc() advances to the next dnode block
if dnode_hold_impl() returns an error for a requested object.
This is because the beginning of the next dnode block is the only
location it can safely assume to either be a hole or a valid
starting point for a dnode.
* dnode_next_offset_level() and other functions that iterate
through dnode blocks may no longer use a simple array indexing
scheme. These now use the current dnode's dn_num_slots field to
advance to the next dnode in the block. This is to ensure we
properly skip the current dnode's bonus area and don't interpret it
as a valid dnode.
zdb
---
The zdb command was updated to display a dnode's size under the
"dnsize" column when the object is dumped.
For ZIL create log records, zdb will now display the slot count for
the object.
ztest
-----
Ztest chooses a random dnodesize for every newly created object. The
random distribution is more heavily weighted toward small dnodes to
better simulate real-world datasets.
Unused bonus buffer space is filled with non-zero values computed from
the object number, dataset id, offset, and generation number. This
helps ensure that the dnode traversal code properly skips the interior
regions of large dnodes, and that these interior regions are not
overwritten by data belonging to other dnodes. A new test visits each
object in a dataset. It verifies that the actual dnode size matches what
was stored in the ztest block tag when it was created. It also verifies
that the unused bonus buffer space is filled with the expected data
patterns.
ZFS Test Suite
--------------
Added six new large dnode-specific tests, and integrated the dnodesize
property into existing tests for zfs allow and send/recv.
Send/Receive
------------
ZFS send streams for datasets containing large dnodes cannot be received
on pools that don't support the large_dnode feature. A send stream with
large dnodes sets a DMU_BACKUP_FEATURE_LARGE_DNODE flag which will be
unrecognized by an incompatible receiving pool so that the zfs receive
will fail gracefully.
While not implemented here, it may be possible to generate a
backward-compatible send stream from a dataset containing large
dnodes. The implementation may be tricky, however, because the send
object record for a large dnode would need to be resized to a 512
byte dnode, possibly kicking in a spill block in the process. This
means we would need to construct a new SA layout and possibly
register it in the SA layout object. The SA layout is normally just
sent as an ordinary object record. But if we are constructing new
layouts while generating the send stream we'd have to build the SA
layout object dynamically and send it at the end of the stream.
For sending and receiving between pools that do support large dnodes,
the drr_object send record type is extended with a new field to store
the dnode slot count. This field was repurposed from unused padding
in the structure.
ZIL Replay
----------
The dnode slot count is stored in the uppermost 8 bits of the lr_foid
field. The bits were unused as the object id is currently capped at
48 bits.
Resizing Dnodes
---------------
It should be possible to resize a dnode when it is dirtied if the
current dnodesize dataset property differs from the dnode's size, but
this functionality is not currently implemented. Clearly a dnode can
only grow if there are sufficient contiguous unused slots in the
dnode block, but it should always be possible to shrink a dnode.
Growing dnodes may be useful to reduce fragmentation in a pool with
many spill blocks in use. Shrinking dnodes may be useful to allow
sending a dataset to a pool that doesn't support the large_dnode
feature.
Feature Reference Counting
--------------------------
The reference count for the large_dnode pool feature tracks the
number of datasets that have ever contained a dnode of size larger
than 512 bytes. The first time a large dnode is created in a dataset
the dataset is converted to an extensible dataset. This is a one-way
operation and the only way to decrement the feature count is to
destroy the dataset, even if the dataset no longer contains any large
dnodes. The complexity of reference counting on a per-dnode basis was
too high, so we chose to track it on a per-dataset basis similarly to
the large_block feature.
Signed-off-by: Ned Bass <bass6@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #3542
2016-03-17 01:25:34 +00:00
|
|
|
|
|
|
|
*offset = (*offset & ~(DNODE_BLOCK_SIZE - 1)) +
|
|
|
|
(i << DNODE_SHIFT);
|
2008-11-20 20:01:55 +00:00
|
|
|
} else {
|
|
|
|
blkptr_t *bp = data;
|
2009-08-18 18:43:27 +00:00
|
|
|
uint64_t start = *offset;
|
2008-11-20 20:01:55 +00:00
|
|
|
span = (lvl - 1) * epbs + dn->dn_datablkshift;
|
|
|
|
minfill = 0;
|
|
|
|
maxfill = blkfill << ((lvl - 1) * epbs);
|
|
|
|
|
|
|
|
if (hole)
|
|
|
|
maxfill--;
|
|
|
|
else
|
|
|
|
minfill++;
|
|
|
|
|
2016-08-31 08:12:08 +00:00
|
|
|
if (span >= 8 * sizeof (*offset)) {
|
|
|
|
/* This only happens on the highest indirection level */
|
|
|
|
ASSERT3U((lvl - 1), ==, dn->dn_phys->dn_nlevels - 1);
|
|
|
|
*offset = 0;
|
|
|
|
} else {
|
|
|
|
*offset = *offset >> span;
|
|
|
|
}
|
|
|
|
|
2009-08-18 18:43:27 +00:00
|
|
|
for (i = BF64_GET(*offset, 0, epbs);
|
2008-12-03 20:09:06 +00:00
|
|
|
i >= 0 && i < epb; i += inc) {
|
2014-06-05 21:19:08 +00:00
|
|
|
if (BP_GET_FILL(&bp[i]) >= minfill &&
|
|
|
|
BP_GET_FILL(&bp[i]) <= maxfill &&
|
2008-12-03 20:09:06 +00:00
|
|
|
(hole || bp[i].blk_birth > txg))
|
2008-11-20 20:01:55 +00:00
|
|
|
break;
|
2009-08-18 18:43:27 +00:00
|
|
|
if (inc > 0 || *offset > 0)
|
|
|
|
*offset += inc;
|
|
|
|
}
|
2016-08-31 08:12:08 +00:00
|
|
|
|
|
|
|
if (span >= 8 * sizeof (*offset)) {
|
|
|
|
*offset = start;
|
|
|
|
} else {
|
|
|
|
*offset = *offset << span;
|
|
|
|
}
|
|
|
|
|
2009-08-18 18:43:27 +00:00
|
|
|
if (inc < 0) {
|
|
|
|
/* traversing backwards; position offset at the end */
|
|
|
|
ASSERT3U(*offset, <=, start);
|
|
|
|
*offset = MIN(*offset + (1ULL << span) - 1, start);
|
|
|
|
} else if (*offset < start) {
|
|
|
|
*offset = start;
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
2009-08-18 18:43:27 +00:00
|
|
|
if (i < 0 || i >= epb)
|
2013-03-08 18:41:28 +00:00
|
|
|
error = SET_ERROR(ESRCH);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
if (db)
|
|
|
|
dbuf_rele(db, FTAG);
|
|
|
|
|
|
|
|
return (error);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Find the next hole, data, or sparse region at or after *offset.
|
|
|
|
* The value 'blkfill' tells us how many items we expect to find
|
|
|
|
* in an L0 data block; this value is 1 for normal objects,
|
|
|
|
* DNODES_PER_BLOCK for the meta dnode, and some fraction of
|
|
|
|
* DNODES_PER_BLOCK when searching for sparse regions thereof.
|
|
|
|
*
|
|
|
|
* Examples:
|
|
|
|
*
|
2008-12-03 20:09:06 +00:00
|
|
|
* dnode_next_offset(dn, flags, offset, 1, 1, 0);
|
|
|
|
* Finds the next/previous hole/data in a file.
|
2008-11-20 20:01:55 +00:00
|
|
|
* Used in dmu_offset_next().
|
|
|
|
*
|
2008-12-03 20:09:06 +00:00
|
|
|
* dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
|
2008-11-20 20:01:55 +00:00
|
|
|
* Finds the next free/allocated dnode an objset's meta-dnode.
|
|
|
|
* Only finds objects that have new contents since txg (ie.
|
|
|
|
* bonus buffer changes and content removal are ignored).
|
|
|
|
* Used in dmu_object_next().
|
|
|
|
*
|
2008-12-03 20:09:06 +00:00
|
|
|
* dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
|
2008-11-20 20:01:55 +00:00
|
|
|
* Finds the next L2 meta-dnode bp that's at most 1/4 full.
|
|
|
|
* Used in dmu_object_alloc().
|
|
|
|
*/
|
|
|
|
int
|
2008-12-03 20:09:06 +00:00
|
|
|
dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
|
2008-11-20 20:01:55 +00:00
|
|
|
int minlvl, uint64_t blkfill, uint64_t txg)
|
|
|
|
{
|
2008-12-03 20:09:06 +00:00
|
|
|
uint64_t initial_offset = *offset;
|
2008-11-20 20:01:55 +00:00
|
|
|
int lvl, maxlvl;
|
|
|
|
int error = 0;
|
|
|
|
|
2008-12-03 20:09:06 +00:00
|
|
|
if (!(flags & DNODE_FIND_HAVELOCK))
|
|
|
|
rw_enter(&dn->dn_struct_rwlock, RW_READER);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
if (dn->dn_phys->dn_nlevels == 0) {
|
2013-03-08 18:41:28 +00:00
|
|
|
error = SET_ERROR(ESRCH);
|
2008-12-03 20:09:06 +00:00
|
|
|
goto out;
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
if (dn->dn_datablkshift == 0) {
|
|
|
|
if (*offset < dn->dn_datablksz) {
|
2008-12-03 20:09:06 +00:00
|
|
|
if (flags & DNODE_FIND_HOLE)
|
2008-11-20 20:01:55 +00:00
|
|
|
*offset = dn->dn_datablksz;
|
|
|
|
} else {
|
2013-03-08 18:41:28 +00:00
|
|
|
error = SET_ERROR(ESRCH);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
2008-12-03 20:09:06 +00:00
|
|
|
goto out;
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
maxlvl = dn->dn_phys->dn_nlevels;
|
|
|
|
|
|
|
|
for (lvl = minlvl; lvl <= maxlvl; lvl++) {
|
|
|
|
error = dnode_next_offset_level(dn,
|
2008-12-03 20:09:06 +00:00
|
|
|
flags, offset, lvl, blkfill, txg);
|
2008-11-20 20:01:55 +00:00
|
|
|
if (error != ESRCH)
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
|
2008-12-03 20:09:06 +00:00
|
|
|
while (error == 0 && --lvl >= minlvl) {
|
2008-11-20 20:01:55 +00:00
|
|
|
error = dnode_next_offset_level(dn,
|
2008-12-03 20:09:06 +00:00
|
|
|
flags, offset, lvl, blkfill, txg);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
2014-09-17 15:25:10 +00:00
|
|
|
/*
|
|
|
|
* There's always a "virtual hole" at the end of the object, even
|
|
|
|
* if all BP's which physically exist are non-holes.
|
|
|
|
*/
|
|
|
|
if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
|
|
|
|
minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
|
|
|
|
error = 0;
|
|
|
|
}
|
|
|
|
|
2008-12-03 20:09:06 +00:00
|
|
|
if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
|
|
|
|
initial_offset < *offset : initial_offset > *offset))
|
2013-03-08 18:41:28 +00:00
|
|
|
error = SET_ERROR(ESRCH);
|
2008-12-03 20:09:06 +00:00
|
|
|
out:
|
|
|
|
if (!(flags & DNODE_FIND_HAVELOCK))
|
|
|
|
rw_exit(&dn->dn_struct_rwlock);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
return (error);
|
|
|
|
}
|