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.
|
2011-10-16 06:41:05 +00:00
|
|
|
* Copyright 2011 Nexenta Systems, Inc. All rights reserved.
|
2016-06-09 18:18:16 +00:00
|
|
|
* Copyright (c) 2011, 2015 by Delphix. All rights reserved.
|
2015-04-01 13:07:48 +00:00
|
|
|
* Copyright (c) 2014, Joyent, Inc. All rights reserved.
|
2016-01-06 21:22:48 +00:00
|
|
|
* Copyright 2014 HybridCluster. All rights reserved.
|
2016-06-09 18:46:42 +00:00
|
|
|
* Copyright 2016 RackTop Systems.
|
2014-03-22 09:07:14 +00:00
|
|
|
* Copyright (c) 2016 Actifio, Inc. All rights reserved.
|
2011-10-16 06:41:05 +00:00
|
|
|
*/
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
#include <sys/dmu.h>
|
|
|
|
#include <sys/dmu_impl.h>
|
|
|
|
#include <sys/dmu_tx.h>
|
|
|
|
#include <sys/dbuf.h>
|
|
|
|
#include <sys/dnode.h>
|
|
|
|
#include <sys/zfs_context.h>
|
|
|
|
#include <sys/dmu_objset.h>
|
|
|
|
#include <sys/dmu_traverse.h>
|
|
|
|
#include <sys/dsl_dataset.h>
|
|
|
|
#include <sys/dsl_dir.h>
|
2010-05-28 20:45:14 +00:00
|
|
|
#include <sys/dsl_prop.h>
|
2008-11-20 20:01:55 +00:00
|
|
|
#include <sys/dsl_pool.h>
|
|
|
|
#include <sys/dsl_synctask.h>
|
2013-05-03 21:17:21 +00:00
|
|
|
#include <sys/spa_impl.h>
|
2008-11-20 20:01:55 +00:00
|
|
|
#include <sys/zfs_ioctl.h>
|
|
|
|
#include <sys/zap.h>
|
|
|
|
#include <sys/zio_checksum.h>
|
2010-05-28 20:45:14 +00:00
|
|
|
#include <sys/zfs_znode.h>
|
|
|
|
#include <zfs_fletcher.h>
|
|
|
|
#include <sys/avl.h>
|
|
|
|
#include <sys/ddt.h>
|
2010-08-26 21:24:34 +00:00
|
|
|
#include <sys/zfs_onexit.h>
|
2013-09-04 12:00:57 +00:00
|
|
|
#include <sys/dmu_send.h>
|
|
|
|
#include <sys/dsl_destroy.h>
|
2014-06-05 21:19:08 +00:00
|
|
|
#include <sys/blkptr.h>
|
2013-12-11 22:33:41 +00:00
|
|
|
#include <sys/dsl_bookmark.h>
|
2014-06-05 21:19:08 +00:00
|
|
|
#include <sys/zfeature.h>
|
2015-12-22 01:31:57 +00:00
|
|
|
#include <sys/bqueue.h>
|
2014-03-22 09:07:14 +00:00
|
|
|
#include <sys/zvol.h>
|
2016-06-07 16:16:52 +00:00
|
|
|
#include <sys/policy.h>
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2011-11-17 18:14:36 +00:00
|
|
|
/* Set this tunable to TRUE to replace corrupt data with 0x2f5baddb10c */
|
|
|
|
int zfs_send_corrupt_data = B_FALSE;
|
2015-12-22 01:31:57 +00:00
|
|
|
int zfs_send_queue_length = 16 * 1024 * 1024;
|
|
|
|
int zfs_recv_queue_length = 16 * 1024 * 1024;
|
2016-06-09 18:46:42 +00:00
|
|
|
/* Set this tunable to FALSE to disable setting of DRR_FLAG_FREERECORDS */
|
|
|
|
int zfs_send_set_freerecords_bit = B_TRUE;
|
2011-11-17 18:14:36 +00:00
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
static char *dmu_recv_tag = "dmu_recv_tag";
|
2016-01-06 21:22:48 +00:00
|
|
|
const char *recv_clone_name = "%recv";
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
#define BP_SPAN(datablkszsec, indblkshift, level) \
|
|
|
|
(((uint64_t)datablkszsec) << (SPA_MINBLOCKSHIFT + \
|
|
|
|
(level) * (indblkshift - SPA_BLKPTRSHIFT)))
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
static void byteswap_record(dmu_replay_record_t *drr);
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
struct send_thread_arg {
|
|
|
|
bqueue_t q;
|
|
|
|
dsl_dataset_t *ds; /* Dataset to traverse */
|
|
|
|
uint64_t fromtxg; /* Traverse from this txg */
|
|
|
|
int flags; /* flags to pass to traverse_dataset */
|
|
|
|
int error_code;
|
|
|
|
boolean_t cancel;
|
2016-01-06 21:22:48 +00:00
|
|
|
zbookmark_phys_t resume;
|
2015-12-22 01:31:57 +00:00
|
|
|
};
|
|
|
|
|
|
|
|
struct send_block_record {
|
|
|
|
boolean_t eos_marker; /* Marks the end of the stream */
|
|
|
|
blkptr_t bp;
|
|
|
|
zbookmark_phys_t zb;
|
|
|
|
uint8_t indblkshift;
|
|
|
|
uint16_t datablkszsec;
|
|
|
|
bqueue_node_t ln;
|
|
|
|
};
|
|
|
|
|
2013-05-03 21:17:21 +00:00
|
|
|
typedef struct dump_bytes_io {
|
|
|
|
dmu_sendarg_t *dbi_dsp;
|
|
|
|
void *dbi_buf;
|
|
|
|
int dbi_len;
|
|
|
|
} dump_bytes_io_t;
|
|
|
|
|
|
|
|
static void
|
2015-12-02 19:53:37 +00:00
|
|
|
dump_bytes_cb(void *arg)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
2013-05-03 21:17:21 +00:00
|
|
|
dump_bytes_io_t *dbi = (dump_bytes_io_t *)arg;
|
|
|
|
dmu_sendarg_t *dsp = dbi->dbi_dsp;
|
2016-01-06 21:22:48 +00:00
|
|
|
dsl_dataset_t *ds = dmu_objset_ds(dsp->dsa_os);
|
2008-11-20 20:01:55 +00:00
|
|
|
ssize_t resid; /* have to get resid to get detailed errno */
|
2016-06-09 19:07:01 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* The code does not rely on this (len being a multiple of 8). We keep
|
|
|
|
* this assertion because of the corresponding assertion in
|
|
|
|
* receive_read(). Keeping this assertion ensures that we do not
|
|
|
|
* inadvertently break backwards compatibility (causing the assertion
|
|
|
|
* in receive_read() to trigger on old software).
|
|
|
|
*
|
|
|
|
* Removing the assertions could be rolled into a new feature that uses
|
|
|
|
* data that isn't 8-byte aligned; if the assertions were removed, a
|
|
|
|
* feature flag would have to be added.
|
|
|
|
*/
|
|
|
|
|
2013-05-10 21:17:03 +00:00
|
|
|
ASSERT0(dbi->dbi_len % 8);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2012-05-09 22:05:14 +00:00
|
|
|
dsp->dsa_err = vn_rdwr(UIO_WRITE, dsp->dsa_vp,
|
2013-05-03 21:17:21 +00:00
|
|
|
(caddr_t)dbi->dbi_buf, dbi->dbi_len,
|
2008-11-20 20:01:55 +00:00
|
|
|
0, UIO_SYSSPACE, FAPPEND, RLIM64_INFINITY, CRED(), &resid);
|
2012-05-09 22:05:14 +00:00
|
|
|
|
|
|
|
mutex_enter(&ds->ds_sendstream_lock);
|
2013-05-03 21:17:21 +00:00
|
|
|
*dsp->dsa_off += dbi->dbi_len;
|
2012-05-09 22:05:14 +00:00
|
|
|
mutex_exit(&ds->ds_sendstream_lock);
|
2013-05-03 21:17:21 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static int
|
|
|
|
dump_bytes(dmu_sendarg_t *dsp, void *buf, int len)
|
|
|
|
{
|
|
|
|
dump_bytes_io_t dbi;
|
|
|
|
|
|
|
|
dbi.dbi_dsp = dsp;
|
|
|
|
dbi.dbi_buf = buf;
|
|
|
|
dbi.dbi_len = len;
|
|
|
|
|
2015-12-02 19:53:37 +00:00
|
|
|
#if defined(HAVE_LARGE_STACKS)
|
|
|
|
dump_bytes_cb(&dbi);
|
|
|
|
#else
|
2013-05-03 21:17:21 +00:00
|
|
|
/*
|
|
|
|
* The vn_rdwr() call is performed in a taskq to ensure that there is
|
|
|
|
* always enough stack space to write safely to the target filesystem.
|
|
|
|
* The ZIO_TYPE_FREE threads are used because there can be a lot of
|
|
|
|
* them and they are used in vdev_file.c for a similar purpose.
|
|
|
|
*/
|
|
|
|
spa_taskq_dispatch_sync(dmu_objset_spa(dsp->dsa_os), ZIO_TYPE_FREE,
|
2015-12-02 19:53:37 +00:00
|
|
|
ZIO_TASKQ_ISSUE, dump_bytes_cb, &dbi, TQ_SLEEP);
|
|
|
|
#endif /* HAVE_LARGE_STACKS */
|
2012-05-09 22:05:14 +00:00
|
|
|
|
|
|
|
return (dsp->dsa_err);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
2015-07-06 03:20:31 +00:00
|
|
|
/*
|
|
|
|
* For all record types except BEGIN, fill in the checksum (overlaid in
|
|
|
|
* drr_u.drr_checksum.drr_checksum). The checksum verifies everything
|
|
|
|
* up to the start of the checksum itself.
|
|
|
|
*/
|
|
|
|
static int
|
|
|
|
dump_record(dmu_sendarg_t *dsp, void *payload, int payload_len)
|
|
|
|
{
|
|
|
|
ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
|
|
|
|
==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
|
2016-07-22 15:52:49 +00:00
|
|
|
(void) fletcher_4_incremental_native(dsp->dsa_drr,
|
2015-07-06 03:20:31 +00:00
|
|
|
offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
|
|
|
|
&dsp->dsa_zc);
|
2016-09-22 23:01:19 +00:00
|
|
|
if (dsp->dsa_drr->drr_type == DRR_BEGIN) {
|
|
|
|
dsp->dsa_sent_begin = B_TRUE;
|
|
|
|
} else {
|
2015-07-06 03:20:31 +00:00
|
|
|
ASSERT(ZIO_CHECKSUM_IS_ZERO(&dsp->dsa_drr->drr_u.
|
|
|
|
drr_checksum.drr_checksum));
|
|
|
|
dsp->dsa_drr->drr_u.drr_checksum.drr_checksum = dsp->dsa_zc;
|
|
|
|
}
|
2016-09-22 23:01:19 +00:00
|
|
|
if (dsp->dsa_drr->drr_type == DRR_END) {
|
|
|
|
dsp->dsa_sent_end = B_TRUE;
|
|
|
|
}
|
2016-07-22 15:52:49 +00:00
|
|
|
(void) fletcher_4_incremental_native(&dsp->dsa_drr->
|
2015-07-06 03:20:31 +00:00
|
|
|
drr_u.drr_checksum.drr_checksum,
|
|
|
|
sizeof (zio_cksum_t), &dsp->dsa_zc);
|
|
|
|
if (dump_bytes(dsp, dsp->dsa_drr, sizeof (dmu_replay_record_t)) != 0)
|
|
|
|
return (SET_ERROR(EINTR));
|
|
|
|
if (payload_len != 0) {
|
2016-07-22 15:52:49 +00:00
|
|
|
(void) fletcher_4_incremental_native(payload, payload_len,
|
2015-07-06 03:20:31 +00:00
|
|
|
&dsp->dsa_zc);
|
|
|
|
if (dump_bytes(dsp, payload, payload_len) != 0)
|
|
|
|
return (SET_ERROR(EINTR));
|
|
|
|
}
|
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
2016-06-09 18:18:16 +00:00
|
|
|
/*
|
|
|
|
* Fill in the drr_free struct, or perform aggregation if the previous record is
|
|
|
|
* also a free record, and the two are adjacent.
|
|
|
|
*
|
|
|
|
* Note that we send free records even for a full send, because we want to be
|
|
|
|
* able to receive a full send as a clone, which requires a list of all the free
|
|
|
|
* and freeobject records that were generated on the source.
|
|
|
|
*/
|
2008-11-20 20:01:55 +00:00
|
|
|
static int
|
2012-05-09 22:05:14 +00:00
|
|
|
dump_free(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
|
2008-11-20 20:01:55 +00:00
|
|
|
uint64_t length)
|
|
|
|
{
|
2012-05-09 22:05:14 +00:00
|
|
|
struct drr_free *drrf = &(dsp->dsa_drr->drr_u.drr_free);
|
2010-05-28 20:45:14 +00:00
|
|
|
|
2013-07-29 18:58:53 +00:00
|
|
|
/*
|
|
|
|
* When we receive a free record, dbuf_free_range() assumes
|
|
|
|
* that the receiving system doesn't have any dbufs in the range
|
|
|
|
* being freed. This is always true because there is a one-record
|
|
|
|
* constraint: we only send one WRITE record for any given
|
2016-01-06 21:22:48 +00:00
|
|
|
* object,offset. We know that the one-record constraint is
|
2013-07-29 18:58:53 +00:00
|
|
|
* true because we always send data in increasing order by
|
|
|
|
* object,offset.
|
|
|
|
*
|
|
|
|
* If the increasing-order constraint ever changes, we should find
|
|
|
|
* another way to assert that the one-record constraint is still
|
|
|
|
* satisfied.
|
|
|
|
*/
|
|
|
|
ASSERT(object > dsp->dsa_last_data_object ||
|
|
|
|
(object == dsp->dsa_last_data_object &&
|
|
|
|
offset > dsp->dsa_last_data_offset));
|
|
|
|
|
2012-06-22 14:17:50 +00:00
|
|
|
if (length != -1ULL && offset + length < offset)
|
|
|
|
length = -1ULL;
|
|
|
|
|
2010-05-28 20:45:14 +00:00
|
|
|
/*
|
|
|
|
* If there is a pending op, but it's not PENDING_FREE, push it out,
|
|
|
|
* since free block aggregation can only be done for blocks of the
|
|
|
|
* same type (i.e., DRR_FREE records can only be aggregated with
|
|
|
|
* other DRR_FREE records. DRR_FREEOBJECTS records can only be
|
|
|
|
* aggregated with other DRR_FREEOBJECTS records.
|
|
|
|
*/
|
2012-05-09 22:05:14 +00:00
|
|
|
if (dsp->dsa_pending_op != PENDING_NONE &&
|
|
|
|
dsp->dsa_pending_op != PENDING_FREE) {
|
2015-07-06 03:20:31 +00:00
|
|
|
if (dump_record(dsp, NULL, 0) != 0)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINTR));
|
2012-05-09 22:05:14 +00:00
|
|
|
dsp->dsa_pending_op = PENDING_NONE;
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
|
|
|
|
2012-05-09 22:05:14 +00:00
|
|
|
if (dsp->dsa_pending_op == PENDING_FREE) {
|
2010-05-28 20:45:14 +00:00
|
|
|
/*
|
|
|
|
* There should never be a PENDING_FREE if length is -1
|
|
|
|
* (because dump_dnode is the only place where this
|
|
|
|
* function is called with a -1, and only after flushing
|
|
|
|
* any pending record).
|
|
|
|
*/
|
|
|
|
ASSERT(length != -1ULL);
|
|
|
|
/*
|
|
|
|
* Check to see whether this free block can be aggregated
|
|
|
|
* with pending one.
|
|
|
|
*/
|
|
|
|
if (drrf->drr_object == object && drrf->drr_offset +
|
|
|
|
drrf->drr_length == offset) {
|
|
|
|
drrf->drr_length += length;
|
|
|
|
return (0);
|
|
|
|
} else {
|
|
|
|
/* not a continuation. Push out pending record */
|
2015-07-06 03:20:31 +00:00
|
|
|
if (dump_record(dsp, NULL, 0) != 0)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINTR));
|
2012-05-09 22:05:14 +00:00
|
|
|
dsp->dsa_pending_op = PENDING_NONE;
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
/* create a FREE record and make it pending */
|
2012-05-09 22:05:14 +00:00
|
|
|
bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
|
|
|
|
dsp->dsa_drr->drr_type = DRR_FREE;
|
2010-05-28 20:45:14 +00:00
|
|
|
drrf->drr_object = object;
|
|
|
|
drrf->drr_offset = offset;
|
|
|
|
drrf->drr_length = length;
|
2012-05-09 22:05:14 +00:00
|
|
|
drrf->drr_toguid = dsp->dsa_toguid;
|
2010-05-28 20:45:14 +00:00
|
|
|
if (length == -1ULL) {
|
2015-07-06 03:20:31 +00:00
|
|
|
if (dump_record(dsp, NULL, 0) != 0)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINTR));
|
2010-05-28 20:45:14 +00:00
|
|
|
} else {
|
2012-05-09 22:05:14 +00:00
|
|
|
dsp->dsa_pending_op = PENDING_FREE;
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
|
|
|
static int
|
2014-06-05 21:19:08 +00:00
|
|
|
dump_write(dmu_sendarg_t *dsp, dmu_object_type_t type,
|
2016-07-11 17:45:52 +00:00
|
|
|
uint64_t object, uint64_t offset, int lsize, int psize, const blkptr_t *bp,
|
|
|
|
void *data)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
2016-07-11 17:45:52 +00:00
|
|
|
uint64_t payload_size;
|
2012-05-09 22:05:14 +00:00
|
|
|
struct drr_write *drrw = &(dsp->dsa_drr->drr_u.drr_write);
|
2010-05-28 20:45:14 +00:00
|
|
|
|
2013-07-29 18:58:53 +00:00
|
|
|
/*
|
|
|
|
* We send data in increasing object, offset order.
|
|
|
|
* See comment in dump_free() for details.
|
|
|
|
*/
|
|
|
|
ASSERT(object > dsp->dsa_last_data_object ||
|
|
|
|
(object == dsp->dsa_last_data_object &&
|
|
|
|
offset > dsp->dsa_last_data_offset));
|
|
|
|
dsp->dsa_last_data_object = object;
|
2016-07-11 17:45:52 +00:00
|
|
|
dsp->dsa_last_data_offset = offset + lsize - 1;
|
2010-05-28 20:45:14 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* If there is any kind of pending aggregation (currently either
|
|
|
|
* a grouping of free objects or free blocks), push it out to
|
|
|
|
* the stream, since aggregation can't be done across operations
|
|
|
|
* of different types.
|
|
|
|
*/
|
2012-05-09 22:05:14 +00:00
|
|
|
if (dsp->dsa_pending_op != PENDING_NONE) {
|
2015-07-06 03:20:31 +00:00
|
|
|
if (dump_record(dsp, NULL, 0) != 0)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINTR));
|
2012-05-09 22:05:14 +00:00
|
|
|
dsp->dsa_pending_op = PENDING_NONE;
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
2015-07-06 03:20:31 +00:00
|
|
|
/* write a WRITE record */
|
2012-05-09 22:05:14 +00:00
|
|
|
bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
|
|
|
|
dsp->dsa_drr->drr_type = DRR_WRITE;
|
2010-05-28 20:45:14 +00:00
|
|
|
drrw->drr_object = object;
|
|
|
|
drrw->drr_type = type;
|
|
|
|
drrw->drr_offset = offset;
|
2012-05-09 22:05:14 +00:00
|
|
|
drrw->drr_toguid = dsp->dsa_toguid;
|
2016-07-11 17:45:52 +00:00
|
|
|
drrw->drr_logical_size = lsize;
|
|
|
|
|
|
|
|
/* only set the compression fields if the buf is compressed */
|
|
|
|
if (lsize != psize) {
|
|
|
|
ASSERT(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_COMPRESSED);
|
|
|
|
ASSERT(!BP_IS_EMBEDDED(bp));
|
|
|
|
ASSERT(!BP_SHOULD_BYTESWAP(bp));
|
|
|
|
ASSERT(!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)));
|
|
|
|
ASSERT3U(BP_GET_COMPRESS(bp), !=, ZIO_COMPRESS_OFF);
|
|
|
|
ASSERT3S(psize, >, 0);
|
|
|
|
ASSERT3S(lsize, >=, psize);
|
|
|
|
|
|
|
|
drrw->drr_compressiontype = BP_GET_COMPRESS(bp);
|
|
|
|
drrw->drr_compressed_size = psize;
|
|
|
|
payload_size = drrw->drr_compressed_size;
|
|
|
|
} else {
|
|
|
|
payload_size = drrw->drr_logical_size;
|
|
|
|
}
|
|
|
|
|
2014-11-03 20:15:08 +00:00
|
|
|
if (bp == NULL || BP_IS_EMBEDDED(bp)) {
|
2014-06-05 21:19:08 +00:00
|
|
|
/*
|
2014-11-03 20:15:08 +00:00
|
|
|
* There's no pre-computed checksum for partial-block
|
|
|
|
* writes or embedded BP's, so (like
|
|
|
|
* fletcher4-checkummed blocks) userland will have to
|
|
|
|
* compute a dedup-capable checksum itself.
|
2014-06-05 21:19:08 +00:00
|
|
|
*/
|
|
|
|
drrw->drr_checksumtype = ZIO_CHECKSUM_OFF;
|
|
|
|
} else {
|
|
|
|
drrw->drr_checksumtype = BP_GET_CHECKSUM(bp);
|
2016-06-15 22:47:05 +00:00
|
|
|
if (zio_checksum_table[drrw->drr_checksumtype].ci_flags &
|
|
|
|
ZCHECKSUM_FLAG_DEDUP)
|
2014-06-05 21:19:08 +00:00
|
|
|
drrw->drr_checksumflags |= DRR_CHECKSUM_DEDUP;
|
|
|
|
DDK_SET_LSIZE(&drrw->drr_key, BP_GET_LSIZE(bp));
|
|
|
|
DDK_SET_PSIZE(&drrw->drr_key, BP_GET_PSIZE(bp));
|
|
|
|
DDK_SET_COMPRESS(&drrw->drr_key, BP_GET_COMPRESS(bp));
|
|
|
|
drrw->drr_key.ddk_cksum = bp->blk_cksum;
|
|
|
|
}
|
2010-05-28 20:45:14 +00:00
|
|
|
|
2016-07-11 17:45:52 +00:00
|
|
|
if (dump_record(dsp, data, payload_size) != 0)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINTR));
|
2010-05-28 20:45:14 +00:00
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
2014-06-05 21:19:08 +00:00
|
|
|
static int
|
|
|
|
dump_write_embedded(dmu_sendarg_t *dsp, uint64_t object, uint64_t offset,
|
|
|
|
int blksz, const blkptr_t *bp)
|
|
|
|
{
|
|
|
|
char buf[BPE_PAYLOAD_SIZE];
|
|
|
|
struct drr_write_embedded *drrw =
|
|
|
|
&(dsp->dsa_drr->drr_u.drr_write_embedded);
|
|
|
|
|
|
|
|
if (dsp->dsa_pending_op != PENDING_NONE) {
|
2015-07-06 03:20:31 +00:00
|
|
|
if (dump_record(dsp, NULL, 0) != 0)
|
2014-06-05 21:19:08 +00:00
|
|
|
return (EINTR);
|
|
|
|
dsp->dsa_pending_op = PENDING_NONE;
|
|
|
|
}
|
|
|
|
|
|
|
|
ASSERT(BP_IS_EMBEDDED(bp));
|
|
|
|
|
|
|
|
bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
|
|
|
|
dsp->dsa_drr->drr_type = DRR_WRITE_EMBEDDED;
|
|
|
|
drrw->drr_object = object;
|
|
|
|
drrw->drr_offset = offset;
|
|
|
|
drrw->drr_length = blksz;
|
|
|
|
drrw->drr_toguid = dsp->dsa_toguid;
|
|
|
|
drrw->drr_compression = BP_GET_COMPRESS(bp);
|
|
|
|
drrw->drr_etype = BPE_GET_ETYPE(bp);
|
|
|
|
drrw->drr_lsize = BPE_GET_LSIZE(bp);
|
|
|
|
drrw->drr_psize = BPE_GET_PSIZE(bp);
|
|
|
|
|
|
|
|
decode_embedded_bp_compressed(bp, buf);
|
|
|
|
|
2015-07-06 03:20:31 +00:00
|
|
|
if (dump_record(dsp, buf, P2ROUNDUP(drrw->drr_psize, 8)) != 0)
|
2014-06-05 21:19:08 +00:00
|
|
|
return (EINTR);
|
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
2010-05-28 20:45:14 +00:00
|
|
|
static int
|
2012-05-09 22:05:14 +00:00
|
|
|
dump_spill(dmu_sendarg_t *dsp, uint64_t object, int blksz, void *data)
|
2010-05-28 20:45:14 +00:00
|
|
|
{
|
2012-05-09 22:05:14 +00:00
|
|
|
struct drr_spill *drrs = &(dsp->dsa_drr->drr_u.drr_spill);
|
2010-05-28 20:45:14 +00:00
|
|
|
|
2012-05-09 22:05:14 +00:00
|
|
|
if (dsp->dsa_pending_op != PENDING_NONE) {
|
2015-07-06 03:20:31 +00:00
|
|
|
if (dump_record(dsp, NULL, 0) != 0)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINTR));
|
2012-05-09 22:05:14 +00:00
|
|
|
dsp->dsa_pending_op = PENDING_NONE;
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/* write a SPILL record */
|
2012-05-09 22:05:14 +00:00
|
|
|
bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
|
|
|
|
dsp->dsa_drr->drr_type = DRR_SPILL;
|
2010-05-28 20:45:14 +00:00
|
|
|
drrs->drr_object = object;
|
|
|
|
drrs->drr_length = blksz;
|
2012-05-09 22:05:14 +00:00
|
|
|
drrs->drr_toguid = dsp->dsa_toguid;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2015-07-06 03:20:31 +00:00
|
|
|
if (dump_record(dsp, data, blksz) != 0)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINTR));
|
2008-11-20 20:01:55 +00:00
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
|
|
|
static int
|
2012-05-09 22:05:14 +00:00
|
|
|
dump_freeobjects(dmu_sendarg_t *dsp, uint64_t firstobj, uint64_t numobjs)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
2012-05-09 22:05:14 +00:00
|
|
|
struct drr_freeobjects *drrfo = &(dsp->dsa_drr->drr_u.drr_freeobjects);
|
2010-05-28 20:45:14 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* If there is a pending op, but it's not PENDING_FREEOBJECTS,
|
|
|
|
* push it out, since free block aggregation can only be done for
|
|
|
|
* blocks of the same type (i.e., DRR_FREE records can only be
|
|
|
|
* aggregated with other DRR_FREE records. DRR_FREEOBJECTS records
|
|
|
|
* can only be aggregated with other DRR_FREEOBJECTS records.
|
|
|
|
*/
|
2012-05-09 22:05:14 +00:00
|
|
|
if (dsp->dsa_pending_op != PENDING_NONE &&
|
|
|
|
dsp->dsa_pending_op != PENDING_FREEOBJECTS) {
|
2015-07-06 03:20:31 +00:00
|
|
|
if (dump_record(dsp, NULL, 0) != 0)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINTR));
|
2012-05-09 22:05:14 +00:00
|
|
|
dsp->dsa_pending_op = PENDING_NONE;
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
2012-05-09 22:05:14 +00:00
|
|
|
if (dsp->dsa_pending_op == PENDING_FREEOBJECTS) {
|
2010-05-28 20:45:14 +00:00
|
|
|
/*
|
|
|
|
* See whether this free object array can be aggregated
|
|
|
|
* with pending one
|
|
|
|
*/
|
|
|
|
if (drrfo->drr_firstobj + drrfo->drr_numobjs == firstobj) {
|
|
|
|
drrfo->drr_numobjs += numobjs;
|
|
|
|
return (0);
|
|
|
|
} else {
|
|
|
|
/* can't be aggregated. Push out pending record */
|
2015-07-06 03:20:31 +00:00
|
|
|
if (dump_record(dsp, NULL, 0) != 0)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINTR));
|
2012-05-09 22:05:14 +00:00
|
|
|
dsp->dsa_pending_op = PENDING_NONE;
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
/* write a FREEOBJECTS record */
|
2012-05-09 22:05:14 +00:00
|
|
|
bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
|
|
|
|
dsp->dsa_drr->drr_type = DRR_FREEOBJECTS;
|
2010-05-28 20:45:14 +00:00
|
|
|
drrfo->drr_firstobj = firstobj;
|
|
|
|
drrfo->drr_numobjs = numobjs;
|
2012-05-09 22:05:14 +00:00
|
|
|
drrfo->drr_toguid = dsp->dsa_toguid;
|
2010-05-28 20:45:14 +00:00
|
|
|
|
2012-05-09 22:05:14 +00:00
|
|
|
dsp->dsa_pending_op = PENDING_FREEOBJECTS;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
|
|
|
static int
|
2012-05-09 22:05:14 +00:00
|
|
|
dump_dnode(dmu_sendarg_t *dsp, uint64_t object, dnode_phys_t *dnp)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
2012-05-09 22:05:14 +00:00
|
|
|
struct drr_object *drro = &(dsp->dsa_drr->drr_u.drr_object);
|
2010-05-28 20:45:14 +00:00
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
if (object < dsp->dsa_resume_object) {
|
|
|
|
/*
|
|
|
|
* Note: when resuming, we will visit all the dnodes in
|
|
|
|
* the block of dnodes that we are resuming from. In
|
|
|
|
* this case it's unnecessary to send the dnodes prior to
|
|
|
|
* the one we are resuming from. We should be at most one
|
|
|
|
* block's worth of dnodes behind the resume point.
|
|
|
|
*/
|
|
|
|
ASSERT3U(dsp->dsa_resume_object - object, <,
|
|
|
|
1 << (DNODE_BLOCK_SHIFT - DNODE_SHIFT));
|
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
if (dnp == NULL || dnp->dn_type == DMU_OT_NONE)
|
2012-05-09 22:05:14 +00:00
|
|
|
return (dump_freeobjects(dsp, object, 1));
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2012-05-09 22:05:14 +00:00
|
|
|
if (dsp->dsa_pending_op != PENDING_NONE) {
|
2015-07-06 03:20:31 +00:00
|
|
|
if (dump_record(dsp, NULL, 0) != 0)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINTR));
|
2012-05-09 22:05:14 +00:00
|
|
|
dsp->dsa_pending_op = PENDING_NONE;
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
/* write an OBJECT record */
|
2012-05-09 22:05:14 +00:00
|
|
|
bzero(dsp->dsa_drr, sizeof (dmu_replay_record_t));
|
|
|
|
dsp->dsa_drr->drr_type = DRR_OBJECT;
|
2010-05-28 20:45:14 +00:00
|
|
|
drro->drr_object = object;
|
|
|
|
drro->drr_type = dnp->dn_type;
|
|
|
|
drro->drr_bonustype = dnp->dn_bonustype;
|
|
|
|
drro->drr_blksz = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
|
|
|
|
drro->drr_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
|
|
|
drro->drr_dn_slots = dnp->dn_extra_slots + 1;
|
2010-05-28 20:45:14 +00:00
|
|
|
drro->drr_checksumtype = dnp->dn_checksum;
|
|
|
|
drro->drr_compress = dnp->dn_compress;
|
2012-05-09 22:05:14 +00:00
|
|
|
drro->drr_toguid = dsp->dsa_toguid;
|
2010-05-28 20:45:14 +00:00
|
|
|
|
2014-11-03 20:15:08 +00:00
|
|
|
if (!(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
|
|
|
|
drro->drr_blksz > SPA_OLD_MAXBLOCKSIZE)
|
|
|
|
drro->drr_blksz = SPA_OLD_MAXBLOCKSIZE;
|
|
|
|
|
2015-07-06 03:20:31 +00:00
|
|
|
if (dump_record(dsp, DN_BONUS(dnp),
|
|
|
|
P2ROUNDUP(dnp->dn_bonuslen, 8)) != 0) {
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINTR));
|
2015-07-06 03:20:31 +00:00
|
|
|
}
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2013-07-29 18:58:53 +00:00
|
|
|
/* Free anything past the end of the file. */
|
2012-05-09 22:05:14 +00:00
|
|
|
if (dump_free(dsp, object, (dnp->dn_maxblkid + 1) *
|
2013-07-29 18:58:53 +00:00
|
|
|
(dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT), -1ULL) != 0)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINTR));
|
2013-09-04 12:00:57 +00:00
|
|
|
if (dsp->dsa_err != 0)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINTR));
|
2008-11-20 20:01:55 +00:00
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
2014-06-05 21:19:08 +00:00
|
|
|
static boolean_t
|
|
|
|
backup_do_embed(dmu_sendarg_t *dsp, const blkptr_t *bp)
|
|
|
|
{
|
|
|
|
if (!BP_IS_EMBEDDED(bp))
|
|
|
|
return (B_FALSE);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Compression function must be legacy, or explicitly enabled.
|
|
|
|
*/
|
|
|
|
if ((BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_LEGACY_FUNCTIONS &&
|
2016-07-11 17:45:52 +00:00
|
|
|
!(dsp->dsa_featureflags & DMU_BACKUP_FEATURE_LZ4)))
|
2014-06-05 21:19:08 +00:00
|
|
|
return (B_FALSE);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Embed type must be explicitly enabled.
|
|
|
|
*/
|
|
|
|
switch (BPE_GET_ETYPE(bp)) {
|
|
|
|
case BP_EMBEDDED_TYPE_DATA:
|
|
|
|
if (dsp->dsa_featureflags & DMU_BACKUP_FEATURE_EMBED_DATA)
|
|
|
|
return (B_TRUE);
|
|
|
|
break;
|
|
|
|
default:
|
|
|
|
return (B_FALSE);
|
|
|
|
}
|
|
|
|
return (B_FALSE);
|
|
|
|
}
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
/*
|
|
|
|
* This is the callback function to traverse_dataset that acts as the worker
|
|
|
|
* thread for dmu_send_impl.
|
|
|
|
*/
|
|
|
|
/*ARGSUSED*/
|
|
|
|
static int
|
|
|
|
send_cb(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
|
|
|
|
const zbookmark_phys_t *zb, const struct dnode_phys *dnp, void *arg)
|
|
|
|
{
|
|
|
|
struct send_thread_arg *sta = arg;
|
|
|
|
struct send_block_record *record;
|
|
|
|
uint64_t record_size;
|
|
|
|
int err = 0;
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
|
|
|
|
zb->zb_object >= sta->resume.zb_object);
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
if (sta->cancel)
|
|
|
|
return (SET_ERROR(EINTR));
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
if (bp == NULL) {
|
|
|
|
ASSERT3U(zb->zb_level, ==, ZB_DNODE_LEVEL);
|
|
|
|
return (0);
|
|
|
|
} else if (zb->zb_level < 0) {
|
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
|
|
|
record = kmem_zalloc(sizeof (struct send_block_record), KM_SLEEP);
|
|
|
|
record->eos_marker = B_FALSE;
|
|
|
|
record->bp = *bp;
|
|
|
|
record->zb = *zb;
|
|
|
|
record->indblkshift = dnp->dn_indblkshift;
|
|
|
|
record->datablkszsec = dnp->dn_datablkszsec;
|
|
|
|
record_size = dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT;
|
|
|
|
bqueue_enqueue(&sta->q, record, record_size);
|
|
|
|
|
|
|
|
return (err);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This function kicks off the traverse_dataset. It also handles setting the
|
|
|
|
* error code of the thread in case something goes wrong, and pushes the End of
|
|
|
|
* Stream record when the traverse_dataset call has finished. If there is no
|
|
|
|
* dataset to traverse, the thread immediately pushes End of Stream marker.
|
|
|
|
*/
|
|
|
|
static void
|
|
|
|
send_traverse_thread(void *arg)
|
|
|
|
{
|
|
|
|
struct send_thread_arg *st_arg = arg;
|
|
|
|
int err;
|
|
|
|
struct send_block_record *data;
|
2016-08-21 13:22:32 +00:00
|
|
|
fstrans_cookie_t cookie = spl_fstrans_mark();
|
2015-12-22 01:31:57 +00:00
|
|
|
|
|
|
|
if (st_arg->ds != NULL) {
|
2016-01-06 21:22:48 +00:00
|
|
|
err = traverse_dataset_resume(st_arg->ds,
|
|
|
|
st_arg->fromtxg, &st_arg->resume,
|
|
|
|
st_arg->flags, send_cb, st_arg);
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
if (err != EINTR)
|
|
|
|
st_arg->error_code = err;
|
|
|
|
}
|
|
|
|
data = kmem_zalloc(sizeof (*data), KM_SLEEP);
|
|
|
|
data->eos_marker = B_TRUE;
|
|
|
|
bqueue_enqueue(&st_arg->q, data, 1);
|
2016-08-21 13:22:32 +00:00
|
|
|
spl_fstrans_unmark(cookie);
|
2017-01-18 23:10:35 +00:00
|
|
|
thread_exit();
|
2015-12-22 01:31:57 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This function actually handles figuring out what kind of record needs to be
|
|
|
|
* dumped, reading the data (which has hopefully been prefetched), and calling
|
|
|
|
* the appropriate helper function.
|
|
|
|
*/
|
2008-11-20 20:01:55 +00:00
|
|
|
static int
|
2015-12-22 01:31:57 +00:00
|
|
|
do_dump(dmu_sendarg_t *dsa, struct send_block_record *data)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
2015-12-22 01:31:57 +00:00
|
|
|
dsl_dataset_t *ds = dmu_objset_ds(dsa->dsa_os);
|
|
|
|
const blkptr_t *bp = &data->bp;
|
|
|
|
const zbookmark_phys_t *zb = &data->zb;
|
|
|
|
uint8_t indblkshift = data->indblkshift;
|
|
|
|
uint16_t dblkszsec = data->datablkszsec;
|
|
|
|
spa_t *spa = ds->ds_dir->dd_pool->dp_spa;
|
2008-11-20 20:01:55 +00:00
|
|
|
dmu_object_type_t type = bp ? BP_GET_TYPE(bp) : DMU_OT_NONE;
|
|
|
|
int err = 0;
|
2015-12-22 01:31:57 +00:00
|
|
|
uint64_t dnobj;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
ASSERT3U(zb->zb_level, >=, 0);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
ASSERT(zb->zb_object == DMU_META_DNODE_OBJECT ||
|
|
|
|
zb->zb_object >= dsa->dsa_resume_object);
|
|
|
|
|
2010-05-28 20:45:14 +00:00
|
|
|
if (zb->zb_object != DMU_META_DNODE_OBJECT &&
|
|
|
|
DMU_OBJECT_IS_SPECIAL(zb->zb_object)) {
|
2009-07-02 22:44:48 +00:00
|
|
|
return (0);
|
2013-12-09 18:37:51 +00:00
|
|
|
} else if (BP_IS_HOLE(bp) &&
|
|
|
|
zb->zb_object == DMU_META_DNODE_OBJECT) {
|
2015-12-22 01:31:57 +00:00
|
|
|
uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level);
|
2008-12-03 20:09:06 +00:00
|
|
|
uint64_t dnobj = (zb->zb_blkid * span) >> DNODE_SHIFT;
|
2015-12-22 01:31:57 +00:00
|
|
|
err = dump_freeobjects(dsa, dnobj, span >> DNODE_SHIFT);
|
2013-12-09 18:37:51 +00:00
|
|
|
} else if (BP_IS_HOLE(bp)) {
|
2015-12-22 01:31:57 +00:00
|
|
|
uint64_t span = BP_SPAN(dblkszsec, indblkshift, zb->zb_level);
|
|
|
|
uint64_t offset = zb->zb_blkid * span;
|
|
|
|
err = dump_free(dsa, zb->zb_object, offset, span);
|
2008-12-03 20:09:06 +00:00
|
|
|
} else if (zb->zb_level > 0 || type == DMU_OT_OBJSET) {
|
|
|
|
return (0);
|
|
|
|
} else if (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
|
|
|
dnode_phys_t *blk;
|
|
|
|
int epb = BP_GET_LSIZE(bp) >> DNODE_SHIFT;
|
2014-12-06 17:24:32 +00:00
|
|
|
arc_flags_t aflags = ARC_FLAG_WAIT;
|
2008-12-03 20:09:06 +00:00
|
|
|
arc_buf_t *abuf;
|
2015-12-22 01:31:57 +00:00
|
|
|
int i;
|
|
|
|
|
|
|
|
ASSERT0(zb->zb_level);
|
2008-12-03 20:09:06 +00:00
|
|
|
|
2013-07-02 20:26:24 +00:00
|
|
|
if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
|
|
|
|
ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
|
|
|
|
&aflags, zb) != 0)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EIO));
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2008-12-03 20:09:06 +00:00
|
|
|
blk = abuf->b_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
|
|
|
dnobj = zb->zb_blkid * epb;
|
|
|
|
for (i = 0; i < epb; i += blk[i].dn_extra_slots + 1) {
|
2015-12-22 01:31:57 +00:00
|
|
|
err = dump_dnode(dsa, dnobj + i, blk + i);
|
2013-09-04 12:00:57 +00:00
|
|
|
if (err != 0)
|
2008-11-20 20:01:55 +00:00
|
|
|
break;
|
|
|
|
}
|
2016-06-02 04:04:53 +00:00
|
|
|
arc_buf_destroy(abuf, &abuf);
|
2010-05-28 20:45:14 +00:00
|
|
|
} else if (type == DMU_OT_SA) {
|
2014-12-06 17:24:32 +00:00
|
|
|
arc_flags_t aflags = ARC_FLAG_WAIT;
|
2008-12-03 20:09:06 +00:00
|
|
|
arc_buf_t *abuf;
|
2008-11-20 20:01:55 +00:00
|
|
|
int blksz = BP_GET_LSIZE(bp);
|
2008-12-03 20:09:06 +00:00
|
|
|
|
2013-07-02 20:26:24 +00:00
|
|
|
if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
|
|
|
|
ZIO_PRIORITY_ASYNC_READ, ZIO_FLAG_CANFAIL,
|
|
|
|
&aflags, zb) != 0)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EIO));
|
2008-12-03 20:09:06 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
err = dump_spill(dsa, zb->zb_object, blksz, abuf->b_data);
|
2016-06-02 04:04:53 +00:00
|
|
|
arc_buf_destroy(abuf, &abuf);
|
2015-12-22 01:31:57 +00:00
|
|
|
} else if (backup_do_embed(dsa, bp)) {
|
2014-06-05 21:19:08 +00:00
|
|
|
/* it's an embedded level-0 block of a regular object */
|
2015-12-22 01:31:57 +00:00
|
|
|
int blksz = dblkszsec << SPA_MINBLOCKSHIFT;
|
|
|
|
ASSERT0(zb->zb_level);
|
|
|
|
err = dump_write_embedded(dsa, zb->zb_object,
|
2014-06-05 21:19:08 +00:00
|
|
|
zb->zb_blkid * blksz, blksz, bp);
|
2015-12-22 01:31:57 +00:00
|
|
|
} else {
|
|
|
|
/* it's a level-0 block of a regular object */
|
2014-12-06 17:24:32 +00:00
|
|
|
arc_flags_t aflags = ARC_FLAG_WAIT;
|
2010-05-28 20:45:14 +00:00
|
|
|
arc_buf_t *abuf;
|
2015-12-22 01:31:57 +00:00
|
|
|
int blksz = dblkszsec << SPA_MINBLOCKSHIFT;
|
|
|
|
uint64_t offset;
|
2016-07-11 17:45:52 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* If we have large blocks stored on disk but the send flags
|
|
|
|
* don't allow us to send large blocks, we split the data from
|
|
|
|
* the arc buf into chunks.
|
|
|
|
*/
|
2017-04-11 21:56:54 +00:00
|
|
|
boolean_t split_large_blocks = blksz > SPA_OLD_MAXBLOCKSIZE &&
|
2016-07-11 17:45:52 +00:00
|
|
|
!(dsa->dsa_featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS);
|
|
|
|
/*
|
|
|
|
* We should only request compressed data from the ARC if all
|
|
|
|
* the following are true:
|
|
|
|
* - stream compression was requested
|
|
|
|
* - we aren't splitting large blocks into smaller chunks
|
|
|
|
* - the data won't need to be byteswapped before sending
|
|
|
|
* - this isn't an embedded block
|
|
|
|
* - this isn't metadata (if receiving on a different endian
|
|
|
|
* system it can be byteswapped more easily)
|
|
|
|
*/
|
|
|
|
boolean_t request_compressed =
|
|
|
|
(dsa->dsa_featureflags & DMU_BACKUP_FEATURE_COMPRESSED) &&
|
|
|
|
!split_large_blocks && !BP_SHOULD_BYTESWAP(bp) &&
|
|
|
|
!BP_IS_EMBEDDED(bp) && !DMU_OT_IS_METADATA(BP_GET_TYPE(bp));
|
2010-05-28 20:45:14 +00:00
|
|
|
|
2013-12-11 22:33:41 +00:00
|
|
|
ASSERT0(zb->zb_level);
|
2016-01-06 21:22:48 +00:00
|
|
|
ASSERT(zb->zb_object > dsa->dsa_resume_object ||
|
|
|
|
(zb->zb_object == dsa->dsa_resume_object &&
|
|
|
|
zb->zb_blkid * blksz >= dsa->dsa_resume_offset));
|
|
|
|
|
2017-04-11 21:56:54 +00:00
|
|
|
ASSERT3U(blksz, ==, BP_GET_LSIZE(bp));
|
|
|
|
|
|
|
|
enum zio_flag zioflags = ZIO_FLAG_CANFAIL;
|
2016-07-11 17:45:52 +00:00
|
|
|
if (request_compressed)
|
|
|
|
zioflags |= ZIO_FLAG_RAW;
|
|
|
|
|
2013-07-02 20:26:24 +00:00
|
|
|
if (arc_read(NULL, spa, bp, arc_getbuf_func, &abuf,
|
2017-04-11 21:56:54 +00:00
|
|
|
ZIO_PRIORITY_ASYNC_READ, zioflags, &aflags, zb) != 0) {
|
2011-11-17 18:14:36 +00:00
|
|
|
if (zfs_send_corrupt_data) {
|
|
|
|
/* Send a block filled with 0x"zfs badd bloc" */
|
2016-07-11 17:45:52 +00:00
|
|
|
abuf = arc_alloc_buf(spa, &abuf, ARC_BUFC_DATA,
|
|
|
|
blksz);
|
2017-04-11 21:56:54 +00:00
|
|
|
uint64_t *ptr;
|
2011-11-17 18:14:36 +00:00
|
|
|
for (ptr = abuf->b_data;
|
|
|
|
(char *)ptr < (char *)abuf->b_data + blksz;
|
|
|
|
ptr++)
|
2013-02-05 00:35:54 +00:00
|
|
|
*ptr = 0x2f5baddb10cULL;
|
2011-11-17 18:14:36 +00:00
|
|
|
} else {
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EIO));
|
2011-11-17 18:14:36 +00:00
|
|
|
}
|
|
|
|
}
|
2010-05-28 20:45:14 +00:00
|
|
|
|
2014-11-03 20:15:08 +00:00
|
|
|
offset = zb->zb_blkid * blksz;
|
|
|
|
|
2016-07-11 17:45:52 +00:00
|
|
|
if (split_large_blocks) {
|
|
|
|
ASSERT3U(arc_get_compression(abuf), ==,
|
|
|
|
ZIO_COMPRESS_OFF);
|
2017-04-11 21:56:54 +00:00
|
|
|
char *buf = abuf->b_data;
|
2014-11-03 20:15:08 +00:00
|
|
|
while (blksz > 0 && err == 0) {
|
|
|
|
int n = MIN(blksz, SPA_OLD_MAXBLOCKSIZE);
|
2015-12-22 01:31:57 +00:00
|
|
|
err = dump_write(dsa, type, zb->zb_object,
|
2016-07-11 17:45:52 +00:00
|
|
|
offset, n, n, NULL, buf);
|
2014-11-03 20:15:08 +00:00
|
|
|
offset += n;
|
|
|
|
buf += n;
|
|
|
|
blksz -= n;
|
|
|
|
}
|
|
|
|
} else {
|
2016-07-11 17:45:52 +00:00
|
|
|
err = dump_write(dsa, type, zb->zb_object, offset,
|
|
|
|
blksz, arc_buf_size(abuf), bp,
|
|
|
|
abuf->b_data);
|
2014-11-03 20:15:08 +00:00
|
|
|
}
|
2016-06-02 04:04:53 +00:00
|
|
|
arc_buf_destroy(abuf, &abuf);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
ASSERT(err == 0 || err == EINTR);
|
|
|
|
return (err);
|
|
|
|
}
|
|
|
|
|
2013-08-28 11:45:09 +00:00
|
|
|
/*
|
2015-12-22 01:31:57 +00:00
|
|
|
* Pop the new data off the queue, and free the old data.
|
|
|
|
*/
|
|
|
|
static struct send_block_record *
|
|
|
|
get_next_record(bqueue_t *bq, struct send_block_record *data)
|
|
|
|
{
|
|
|
|
struct send_block_record *tmp = bqueue_dequeue(bq);
|
|
|
|
kmem_free(data, sizeof (*data));
|
|
|
|
return (tmp);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Actually do the bulk of the work in a zfs send.
|
|
|
|
*
|
|
|
|
* Note: Releases dp using the specified tag.
|
2013-08-28 11:45:09 +00:00
|
|
|
*/
|
2013-09-04 12:00:57 +00:00
|
|
|
static int
|
2015-12-22 01:31:57 +00:00
|
|
|
dmu_send_impl(void *tag, dsl_pool_t *dp, dsl_dataset_t *to_ds,
|
2016-07-11 17:45:52 +00:00
|
|
|
zfs_bookmark_phys_t *ancestor_zb, boolean_t is_clone,
|
|
|
|
boolean_t embedok, boolean_t large_block_ok, boolean_t compressok,
|
|
|
|
int outfd, uint64_t resumeobj, uint64_t resumeoff,
|
2016-01-06 21:22:48 +00:00
|
|
|
vnode_t *vp, offset_t *off)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
2013-09-04 12:00:57 +00:00
|
|
|
objset_t *os;
|
2008-11-20 20:01:55 +00:00
|
|
|
dmu_replay_record_t *drr;
|
2012-05-09 22:05:14 +00:00
|
|
|
dmu_sendarg_t *dsp;
|
2008-11-20 20:01:55 +00:00
|
|
|
int err;
|
|
|
|
uint64_t fromtxg = 0;
|
2014-06-05 21:19:08 +00:00
|
|
|
uint64_t featureflags = 0;
|
2015-12-22 01:31:57 +00:00
|
|
|
struct send_thread_arg to_arg;
|
2016-01-06 21:22:48 +00:00
|
|
|
void *payload = NULL;
|
|
|
|
size_t payload_len = 0;
|
2015-12-22 01:31:57 +00:00
|
|
|
struct send_block_record *to_data;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
err = dmu_objset_from_ds(to_ds, &os);
|
2013-09-04 12:00:57 +00:00
|
|
|
if (err != 0) {
|
|
|
|
dsl_pool_rele(dp, tag);
|
|
|
|
return (err);
|
|
|
|
}
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
drr = kmem_zalloc(sizeof (dmu_replay_record_t), KM_SLEEP);
|
|
|
|
drr->drr_type = DRR_BEGIN;
|
|
|
|
drr->drr_u.drr_begin.drr_magic = DMU_BACKUP_MAGIC;
|
2010-05-28 20:45:14 +00:00
|
|
|
DMU_SET_STREAM_HDRTYPE(drr->drr_u.drr_begin.drr_versioninfo,
|
|
|
|
DMU_SUBSTREAM);
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
bzero(&to_arg, sizeof (to_arg));
|
|
|
|
|
2010-05-28 20:45:14 +00:00
|
|
|
#ifdef _KERNEL
|
2013-09-04 12:00:57 +00:00
|
|
|
if (dmu_objset_type(os) == DMU_OST_ZFS) {
|
2010-05-28 20:45:14 +00:00
|
|
|
uint64_t version;
|
2013-09-04 12:00:57 +00:00
|
|
|
if (zfs_get_zplprop(os, ZFS_PROP_VERSION, &version) != 0) {
|
2012-05-09 22:05:14 +00:00
|
|
|
kmem_free(drr, sizeof (dmu_replay_record_t));
|
2013-09-04 12:00:57 +00:00
|
|
|
dsl_pool_rele(dp, tag);
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2012-05-09 22:05:14 +00:00
|
|
|
}
|
2013-09-04 12:00:57 +00:00
|
|
|
if (version >= ZPL_VERSION_SA) {
|
2014-06-05 21:19:08 +00:00
|
|
|
featureflags |= DMU_BACKUP_FEATURE_SA_SPILL;
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
if (large_block_ok && to_ds->ds_feature_inuse[SPA_FEATURE_LARGE_BLOCKS])
|
2014-11-03 20:15:08 +00:00
|
|
|
featureflags |= DMU_BACKUP_FEATURE_LARGE_BLOCKS;
|
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 (to_ds->ds_feature_inuse[SPA_FEATURE_LARGE_DNODE])
|
|
|
|
featureflags |= DMU_BACKUP_FEATURE_LARGE_DNODE;
|
2014-06-05 21:19:08 +00:00
|
|
|
if (embedok &&
|
|
|
|
spa_feature_is_active(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA)) {
|
|
|
|
featureflags |= DMU_BACKUP_FEATURE_EMBED_DATA;
|
2016-07-11 17:45:52 +00:00
|
|
|
}
|
|
|
|
if (compressok) {
|
|
|
|
featureflags |= DMU_BACKUP_FEATURE_COMPRESSED;
|
|
|
|
}
|
|
|
|
if ((featureflags &
|
|
|
|
(DMU_BACKUP_FEATURE_EMBED_DATA | DMU_BACKUP_FEATURE_COMPRESSED)) !=
|
|
|
|
0 && spa_feature_is_active(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS)) {
|
|
|
|
featureflags |= DMU_BACKUP_FEATURE_LZ4;
|
2014-06-05 21:19:08 +00:00
|
|
|
}
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
if (resumeobj != 0 || resumeoff != 0) {
|
|
|
|
featureflags |= DMU_BACKUP_FEATURE_RESUMING;
|
|
|
|
}
|
|
|
|
|
2014-06-05 21:19:08 +00:00
|
|
|
DMU_SET_FEATUREFLAGS(drr->drr_u.drr_begin.drr_versioninfo,
|
|
|
|
featureflags);
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
drr->drr_u.drr_begin.drr_creation_time =
|
2015-12-22 01:31:57 +00:00
|
|
|
dsl_dataset_phys(to_ds)->ds_creation_time;
|
2013-09-04 12:00:57 +00:00
|
|
|
drr->drr_u.drr_begin.drr_type = dmu_objset_type(os);
|
2013-12-11 22:33:41 +00:00
|
|
|
if (is_clone)
|
2008-11-20 20:01:55 +00:00
|
|
|
drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CLONE;
|
2015-12-22 01:31:57 +00:00
|
|
|
drr->drr_u.drr_begin.drr_toguid = dsl_dataset_phys(to_ds)->ds_guid;
|
|
|
|
if (dsl_dataset_phys(to_ds)->ds_flags & DS_FLAG_CI_DATASET)
|
2008-11-20 20:01:55 +00:00
|
|
|
drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_CI_DATA;
|
2016-06-09 18:46:42 +00:00
|
|
|
if (zfs_send_set_freerecords_bit)
|
|
|
|
drr->drr_u.drr_begin.drr_flags |= DRR_FLAG_FREERECORDS;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
if (ancestor_zb != NULL) {
|
|
|
|
drr->drr_u.drr_begin.drr_fromguid =
|
|
|
|
ancestor_zb->zbm_guid;
|
|
|
|
fromtxg = ancestor_zb->zbm_creation_txg;
|
2013-12-11 22:33:41 +00:00
|
|
|
}
|
2015-12-22 01:31:57 +00:00
|
|
|
dsl_dataset_name(to_ds, drr->drr_u.drr_begin.drr_toname);
|
|
|
|
if (!to_ds->ds_is_snapshot) {
|
2013-12-11 22:33:41 +00:00
|
|
|
(void) strlcat(drr->drr_u.drr_begin.drr_toname, "@--head--",
|
|
|
|
sizeof (drr->drr_u.drr_begin.drr_toname));
|
2013-09-04 12:00:57 +00:00
|
|
|
}
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2012-05-09 22:05:14 +00:00
|
|
|
dsp = kmem_zalloc(sizeof (dmu_sendarg_t), KM_SLEEP);
|
|
|
|
|
|
|
|
dsp->dsa_drr = drr;
|
|
|
|
dsp->dsa_vp = vp;
|
|
|
|
dsp->dsa_outfd = outfd;
|
|
|
|
dsp->dsa_proc = curproc;
|
2013-09-04 12:00:57 +00:00
|
|
|
dsp->dsa_os = os;
|
2012-05-09 22:05:14 +00:00
|
|
|
dsp->dsa_off = off;
|
2015-12-22 01:31:57 +00:00
|
|
|
dsp->dsa_toguid = dsl_dataset_phys(to_ds)->ds_guid;
|
2012-05-09 22:05:14 +00:00
|
|
|
dsp->dsa_pending_op = PENDING_NONE;
|
2014-06-05 21:19:08 +00:00
|
|
|
dsp->dsa_featureflags = featureflags;
|
2016-01-06 21:22:48 +00:00
|
|
|
dsp->dsa_resume_object = resumeobj;
|
|
|
|
dsp->dsa_resume_offset = resumeoff;
|
2012-05-09 22:05:14 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
mutex_enter(&to_ds->ds_sendstream_lock);
|
|
|
|
list_insert_head(&to_ds->ds_sendstreams, dsp);
|
|
|
|
mutex_exit(&to_ds->ds_sendstream_lock);
|
2012-05-09 22:05:14 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
dsl_dataset_long_hold(to_ds, FTAG);
|
2013-04-10 21:54:56 +00:00
|
|
|
dsl_pool_rele(dp, tag);
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
if (resumeobj != 0 || resumeoff != 0) {
|
|
|
|
dmu_object_info_t to_doi;
|
|
|
|
nvlist_t *nvl;
|
|
|
|
err = dmu_object_info(os, resumeobj, &to_doi);
|
|
|
|
if (err != 0)
|
|
|
|
goto out;
|
|
|
|
SET_BOOKMARK(&to_arg.resume, to_ds->ds_object, resumeobj, 0,
|
|
|
|
resumeoff / to_doi.doi_data_block_size);
|
|
|
|
|
|
|
|
nvl = fnvlist_alloc();
|
|
|
|
fnvlist_add_uint64(nvl, "resume_object", resumeobj);
|
|
|
|
fnvlist_add_uint64(nvl, "resume_offset", resumeoff);
|
|
|
|
payload = fnvlist_pack(nvl, &payload_len);
|
|
|
|
drr->drr_payloadlen = payload_len;
|
|
|
|
fnvlist_free(nvl);
|
|
|
|
}
|
|
|
|
|
|
|
|
err = dump_record(dsp, payload, payload_len);
|
|
|
|
fnvlist_pack_free(payload, payload_len);
|
|
|
|
if (err != 0) {
|
2012-05-09 22:05:14 +00:00
|
|
|
err = dsp->dsa_err;
|
|
|
|
goto out;
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
err = bqueue_init(&to_arg.q, zfs_send_queue_length,
|
|
|
|
offsetof(struct send_block_record, ln));
|
|
|
|
to_arg.error_code = 0;
|
|
|
|
to_arg.cancel = B_FALSE;
|
|
|
|
to_arg.ds = to_ds;
|
|
|
|
to_arg.fromtxg = fromtxg;
|
|
|
|
to_arg.flags = TRAVERSE_PRE | TRAVERSE_PREFETCH;
|
|
|
|
(void) thread_create(NULL, 0, send_traverse_thread, &to_arg, 0, curproc,
|
|
|
|
TS_RUN, minclsyspri);
|
|
|
|
|
|
|
|
to_data = bqueue_dequeue(&to_arg.q);
|
|
|
|
|
|
|
|
while (!to_data->eos_marker && err == 0) {
|
|
|
|
err = do_dump(dsp, to_data);
|
|
|
|
to_data = get_next_record(&to_arg.q, to_data);
|
|
|
|
if (issig(JUSTLOOKING) && issig(FORREAL))
|
|
|
|
err = EINTR;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (err != 0) {
|
|
|
|
to_arg.cancel = B_TRUE;
|
|
|
|
while (!to_data->eos_marker) {
|
|
|
|
to_data = get_next_record(&to_arg.q, to_data);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
kmem_free(to_data, sizeof (*to_data));
|
|
|
|
|
|
|
|
bqueue_destroy(&to_arg.q);
|
|
|
|
|
|
|
|
if (err == 0 && to_arg.error_code != 0)
|
|
|
|
err = to_arg.error_code;
|
|
|
|
|
|
|
|
if (err != 0)
|
|
|
|
goto out;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2012-05-09 22:05:14 +00:00
|
|
|
if (dsp->dsa_pending_op != PENDING_NONE)
|
2015-07-06 03:20:31 +00:00
|
|
|
if (dump_record(dsp, NULL, 0) != 0)
|
2013-03-08 18:41:28 +00:00
|
|
|
err = SET_ERROR(EINTR);
|
2010-05-28 20:45:14 +00:00
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
if (err != 0) {
|
|
|
|
if (err == EINTR && dsp->dsa_err != 0)
|
2012-05-09 22:05:14 +00:00
|
|
|
err = dsp->dsa_err;
|
|
|
|
goto out;
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
bzero(drr, sizeof (dmu_replay_record_t));
|
|
|
|
drr->drr_type = DRR_END;
|
2012-05-09 22:05:14 +00:00
|
|
|
drr->drr_u.drr_end.drr_checksum = dsp->dsa_zc;
|
|
|
|
drr->drr_u.drr_end.drr_toguid = dsp->dsa_toguid;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
if (dump_record(dsp, NULL, 0) != 0)
|
2012-05-09 22:05:14 +00:00
|
|
|
err = dsp->dsa_err;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2012-05-09 22:05:14 +00:00
|
|
|
out:
|
2015-12-22 01:31:57 +00:00
|
|
|
mutex_enter(&to_ds->ds_sendstream_lock);
|
|
|
|
list_remove(&to_ds->ds_sendstreams, dsp);
|
|
|
|
mutex_exit(&to_ds->ds_sendstream_lock);
|
2012-05-09 22:05:14 +00:00
|
|
|
|
2016-09-22 23:01:19 +00:00
|
|
|
VERIFY(err != 0 || (dsp->dsa_sent_begin && dsp->dsa_sent_end));
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
kmem_free(drr, sizeof (dmu_replay_record_t));
|
2012-05-09 22:05:14 +00:00
|
|
|
kmem_free(dsp, sizeof (dmu_sendarg_t));
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
dsl_dataset_long_rele(to_ds, FTAG);
|
2013-09-04 12:00:57 +00:00
|
|
|
|
2012-05-09 22:05:14 +00:00
|
|
|
return (err);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
2011-11-17 18:14:36 +00:00
|
|
|
int
|
2013-09-04 12:00:57 +00:00
|
|
|
dmu_send_obj(const char *pool, uint64_t tosnap, uint64_t fromsnap,
|
2016-07-11 17:45:52 +00:00
|
|
|
boolean_t embedok, boolean_t large_block_ok, boolean_t compressok,
|
2014-11-03 20:15:08 +00:00
|
|
|
int outfd, vnode_t *vp, offset_t *off)
|
2013-09-04 12:00:57 +00:00
|
|
|
{
|
|
|
|
dsl_pool_t *dp;
|
|
|
|
dsl_dataset_t *ds;
|
|
|
|
dsl_dataset_t *fromds = NULL;
|
|
|
|
int err;
|
|
|
|
|
|
|
|
err = dsl_pool_hold(pool, FTAG, &dp);
|
|
|
|
if (err != 0)
|
|
|
|
return (err);
|
|
|
|
|
|
|
|
err = dsl_dataset_hold_obj(dp, tosnap, FTAG, &ds);
|
|
|
|
if (err != 0) {
|
|
|
|
dsl_pool_rele(dp, FTAG);
|
|
|
|
return (err);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (fromsnap != 0) {
|
2013-12-11 22:33:41 +00:00
|
|
|
zfs_bookmark_phys_t zb;
|
|
|
|
boolean_t is_clone;
|
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
err = dsl_dataset_hold_obj(dp, fromsnap, FTAG, &fromds);
|
|
|
|
if (err != 0) {
|
|
|
|
dsl_dataset_rele(ds, FTAG);
|
|
|
|
dsl_pool_rele(dp, FTAG);
|
|
|
|
return (err);
|
|
|
|
}
|
2013-12-11 22:33:41 +00:00
|
|
|
if (!dsl_dataset_is_before(ds, fromds, 0))
|
|
|
|
err = SET_ERROR(EXDEV);
|
2015-04-01 15:14:34 +00:00
|
|
|
zb.zbm_creation_time =
|
|
|
|
dsl_dataset_phys(fromds)->ds_creation_time;
|
|
|
|
zb.zbm_creation_txg = dsl_dataset_phys(fromds)->ds_creation_txg;
|
|
|
|
zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
|
2013-12-11 22:33:41 +00:00
|
|
|
is_clone = (fromds->ds_dir != ds->ds_dir);
|
|
|
|
dsl_dataset_rele(fromds, FTAG);
|
2014-11-03 20:15:08 +00:00
|
|
|
err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
|
2016-07-11 17:45:52 +00:00
|
|
|
embedok, large_block_ok, compressok, outfd, 0, 0, vp, off);
|
2013-12-11 22:33:41 +00:00
|
|
|
} else {
|
2014-11-03 20:15:08 +00:00
|
|
|
err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
|
2016-07-11 17:45:52 +00:00
|
|
|
embedok, large_block_ok, compressok, outfd, 0, 0, vp, off);
|
2013-09-04 12:00:57 +00:00
|
|
|
}
|
2013-12-11 22:33:41 +00:00
|
|
|
dsl_dataset_rele(ds, FTAG);
|
|
|
|
return (err);
|
2013-09-04 12:00:57 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
int
|
2016-01-06 21:22:48 +00:00
|
|
|
dmu_send(const char *tosnap, const char *fromsnap, boolean_t embedok,
|
2016-07-11 17:45:52 +00:00
|
|
|
boolean_t large_block_ok, boolean_t compressok, int outfd,
|
|
|
|
uint64_t resumeobj, uint64_t resumeoff,
|
2016-01-06 21:22:48 +00:00
|
|
|
vnode_t *vp, offset_t *off)
|
2013-09-04 12:00:57 +00:00
|
|
|
{
|
|
|
|
dsl_pool_t *dp;
|
|
|
|
dsl_dataset_t *ds;
|
|
|
|
int err;
|
2013-12-11 22:33:41 +00:00
|
|
|
boolean_t owned = B_FALSE;
|
2013-09-04 12:00:57 +00:00
|
|
|
|
2013-12-11 22:33:41 +00:00
|
|
|
if (fromsnap != NULL && strpbrk(fromsnap, "@#") == NULL)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2013-09-04 12:00:57 +00:00
|
|
|
|
|
|
|
err = dsl_pool_hold(tosnap, FTAG, &dp);
|
|
|
|
if (err != 0)
|
|
|
|
return (err);
|
|
|
|
|
2013-12-11 22:33:41 +00:00
|
|
|
if (strchr(tosnap, '@') == NULL && spa_writeable(dp->dp_spa)) {
|
|
|
|
/*
|
|
|
|
* We are sending a filesystem or volume. Ensure
|
|
|
|
* that it doesn't change by owning the dataset.
|
|
|
|
*/
|
|
|
|
err = dsl_dataset_own(dp, tosnap, FTAG, &ds);
|
|
|
|
owned = B_TRUE;
|
|
|
|
} else {
|
|
|
|
err = dsl_dataset_hold(dp, tosnap, FTAG, &ds);
|
|
|
|
}
|
2013-09-04 12:00:57 +00:00
|
|
|
if (err != 0) {
|
|
|
|
dsl_pool_rele(dp, FTAG);
|
|
|
|
return (err);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (fromsnap != NULL) {
|
2013-12-11 22:33:41 +00:00
|
|
|
zfs_bookmark_phys_t zb;
|
|
|
|
boolean_t is_clone = B_FALSE;
|
|
|
|
int fsnamelen = strchr(tosnap, '@') - tosnap;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If the fromsnap is in a different filesystem, then
|
|
|
|
* mark the send stream as a clone.
|
|
|
|
*/
|
|
|
|
if (strncmp(tosnap, fromsnap, fsnamelen) != 0 ||
|
|
|
|
(fromsnap[fsnamelen] != '@' &&
|
|
|
|
fromsnap[fsnamelen] != '#')) {
|
|
|
|
is_clone = B_TRUE;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (strchr(fromsnap, '@')) {
|
|
|
|
dsl_dataset_t *fromds;
|
|
|
|
err = dsl_dataset_hold(dp, fromsnap, FTAG, &fromds);
|
|
|
|
if (err == 0) {
|
|
|
|
if (!dsl_dataset_is_before(ds, fromds, 0))
|
|
|
|
err = SET_ERROR(EXDEV);
|
|
|
|
zb.zbm_creation_time =
|
2015-04-01 15:14:34 +00:00
|
|
|
dsl_dataset_phys(fromds)->ds_creation_time;
|
2013-12-11 22:33:41 +00:00
|
|
|
zb.zbm_creation_txg =
|
2015-04-01 15:14:34 +00:00
|
|
|
dsl_dataset_phys(fromds)->ds_creation_txg;
|
|
|
|
zb.zbm_guid = dsl_dataset_phys(fromds)->ds_guid;
|
2013-12-11 22:33:41 +00:00
|
|
|
is_clone = (ds->ds_dir != fromds->ds_dir);
|
|
|
|
dsl_dataset_rele(fromds, FTAG);
|
|
|
|
}
|
|
|
|
} else {
|
|
|
|
err = dsl_bookmark_lookup(dp, fromsnap, ds, &zb);
|
|
|
|
}
|
2013-09-04 12:00:57 +00:00
|
|
|
if (err != 0) {
|
|
|
|
dsl_dataset_rele(ds, FTAG);
|
|
|
|
dsl_pool_rele(dp, FTAG);
|
|
|
|
return (err);
|
|
|
|
}
|
2014-11-03 20:15:08 +00:00
|
|
|
err = dmu_send_impl(FTAG, dp, ds, &zb, is_clone,
|
2016-07-11 17:45:52 +00:00
|
|
|
embedok, large_block_ok, compressok,
|
2016-01-06 21:22:48 +00:00
|
|
|
outfd, resumeobj, resumeoff, vp, off);
|
2013-12-11 22:33:41 +00:00
|
|
|
} else {
|
2014-11-03 20:15:08 +00:00
|
|
|
err = dmu_send_impl(FTAG, dp, ds, NULL, B_FALSE,
|
2016-07-11 17:45:52 +00:00
|
|
|
embedok, large_block_ok, compressok,
|
2016-01-06 21:22:48 +00:00
|
|
|
outfd, resumeobj, resumeoff, vp, off);
|
2013-09-04 12:00:57 +00:00
|
|
|
}
|
2013-12-11 22:33:41 +00:00
|
|
|
if (owned)
|
|
|
|
dsl_dataset_disown(ds, FTAG);
|
|
|
|
else
|
|
|
|
dsl_dataset_rele(ds, FTAG);
|
|
|
|
return (err);
|
2013-09-04 12:00:57 +00:00
|
|
|
}
|
|
|
|
|
2015-04-08 18:37:13 +00:00
|
|
|
static int
|
2016-07-11 17:45:52 +00:00
|
|
|
dmu_adjust_send_estimate_for_indirects(dsl_dataset_t *ds, uint64_t uncompressed,
|
|
|
|
uint64_t compressed, boolean_t stream_compressed, uint64_t *sizep)
|
2015-04-08 18:37:13 +00:00
|
|
|
{
|
|
|
|
int err;
|
2016-07-11 17:45:52 +00:00
|
|
|
uint64_t size;
|
2015-04-08 18:37:13 +00:00
|
|
|
/*
|
|
|
|
* Assume that space (both on-disk and in-stream) is dominated by
|
|
|
|
* data. We will adjust for indirect blocks and the copies property,
|
|
|
|
* but ignore per-object space used (eg, dnodes and DRR_OBJECT records).
|
|
|
|
*/
|
|
|
|
|
2016-07-11 17:45:52 +00:00
|
|
|
uint64_t recordsize;
|
|
|
|
uint64_t record_count;
|
2016-07-07 22:00:51 +00:00
|
|
|
objset_t *os;
|
|
|
|
VERIFY0(dmu_objset_from_ds(ds, &os));
|
2016-07-11 17:45:52 +00:00
|
|
|
|
|
|
|
/* Assume all (uncompressed) blocks are recordsize. */
|
2016-07-07 22:00:51 +00:00
|
|
|
if (os->os_phys->os_type == DMU_OST_ZVOL) {
|
|
|
|
err = dsl_prop_get_int_ds(ds,
|
|
|
|
zfs_prop_to_name(ZFS_PROP_VOLBLOCKSIZE), &recordsize);
|
|
|
|
} else {
|
|
|
|
err = dsl_prop_get_int_ds(ds,
|
|
|
|
zfs_prop_to_name(ZFS_PROP_RECORDSIZE), &recordsize);
|
|
|
|
}
|
2016-07-11 17:45:52 +00:00
|
|
|
if (err != 0)
|
|
|
|
return (err);
|
|
|
|
record_count = uncompressed / recordsize;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If we're estimating a send size for a compressed stream, use the
|
|
|
|
* compressed data size to estimate the stream size. Otherwise, use the
|
|
|
|
* uncompressed data size.
|
|
|
|
*/
|
|
|
|
size = stream_compressed ? compressed : uncompressed;
|
|
|
|
|
2015-04-08 18:37:13 +00:00
|
|
|
/*
|
|
|
|
* Subtract out approximate space used by indirect blocks.
|
|
|
|
* Assume most space is used by data blocks (non-indirect, non-dnode).
|
2016-07-11 17:45:52 +00:00
|
|
|
* Assume no ditto blocks or internal fragmentation.
|
2015-04-08 18:37:13 +00:00
|
|
|
*
|
|
|
|
* Therefore, space used by indirect blocks is sizeof(blkptr_t) per
|
2016-07-11 17:45:52 +00:00
|
|
|
* block.
|
2015-04-08 18:37:13 +00:00
|
|
|
*/
|
2016-07-11 17:45:52 +00:00
|
|
|
size -= record_count * sizeof (blkptr_t);
|
2015-04-08 18:37:13 +00:00
|
|
|
|
|
|
|
/* Add in the space for the record associated with each block. */
|
2016-07-11 17:45:52 +00:00
|
|
|
size += record_count * sizeof (dmu_replay_record_t);
|
2015-04-08 18:37:13 +00:00
|
|
|
|
|
|
|
*sizep = size;
|
|
|
|
|
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
int
|
2016-07-11 17:45:52 +00:00
|
|
|
dmu_send_estimate(dsl_dataset_t *ds, dsl_dataset_t *fromds,
|
|
|
|
boolean_t stream_compressed, uint64_t *sizep)
|
2011-11-17 18:14:36 +00:00
|
|
|
{
|
|
|
|
int err;
|
2016-07-11 17:45:52 +00:00
|
|
|
uint64_t uncomp, comp;
|
2013-09-04 12:00:57 +00:00
|
|
|
|
2015-05-13 17:50:35 +00:00
|
|
|
ASSERT(dsl_pool_config_held(ds->ds_dir->dd_pool));
|
2011-11-17 18:14:36 +00:00
|
|
|
|
|
|
|
/* tosnap must be a snapshot */
|
2015-04-02 03:44:32 +00:00
|
|
|
if (!ds->ds_is_snapshot)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2011-11-17 18:14:36 +00:00
|
|
|
|
2015-07-02 13:04:35 +00:00
|
|
|
/* fromsnap, if provided, must be a snapshot */
|
|
|
|
if (fromds != NULL && !fromds->ds_is_snapshot)
|
|
|
|
return (SET_ERROR(EINVAL));
|
|
|
|
|
2013-08-28 11:45:09 +00:00
|
|
|
/*
|
|
|
|
* fromsnap must be an earlier snapshot from the same fs as tosnap,
|
|
|
|
* or the origin's fs.
|
|
|
|
*/
|
2013-12-11 22:33:41 +00:00
|
|
|
if (fromds != NULL && !dsl_dataset_is_before(ds, fromds, 0))
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EXDEV));
|
2011-11-17 18:14:36 +00:00
|
|
|
|
2016-07-11 17:45:52 +00:00
|
|
|
/* Get compressed and uncompressed size estimates of changed data. */
|
2011-11-17 18:14:36 +00:00
|
|
|
if (fromds == NULL) {
|
2016-07-11 17:45:52 +00:00
|
|
|
uncomp = dsl_dataset_phys(ds)->ds_uncompressed_bytes;
|
|
|
|
comp = dsl_dataset_phys(ds)->ds_compressed_bytes;
|
2011-11-17 18:14:36 +00:00
|
|
|
} else {
|
2016-07-11 17:45:52 +00:00
|
|
|
uint64_t used;
|
2011-11-17 18:14:36 +00:00
|
|
|
err = dsl_dataset_space_written(fromds, ds,
|
2016-07-11 17:45:52 +00:00
|
|
|
&used, &comp, &uncomp);
|
2013-09-04 12:00:57 +00:00
|
|
|
if (err != 0)
|
2011-11-17 18:14:36 +00:00
|
|
|
return (err);
|
|
|
|
}
|
|
|
|
|
2016-07-11 17:45:52 +00:00
|
|
|
err = dmu_adjust_send_estimate_for_indirects(ds, uncomp, comp,
|
|
|
|
stream_compressed, sizep);
|
2016-07-07 22:00:51 +00:00
|
|
|
/*
|
|
|
|
* Add the size of the BEGIN and END records to the estimate.
|
|
|
|
*/
|
|
|
|
*sizep += 2 * sizeof (dmu_replay_record_t);
|
2015-04-08 18:37:13 +00:00
|
|
|
return (err);
|
|
|
|
}
|
2011-11-17 18:14:36 +00:00
|
|
|
|
2016-07-11 17:45:52 +00:00
|
|
|
struct calculate_send_arg {
|
|
|
|
uint64_t uncompressed;
|
|
|
|
uint64_t compressed;
|
|
|
|
};
|
|
|
|
|
2015-04-08 18:37:13 +00:00
|
|
|
/*
|
|
|
|
* Simple callback used to traverse the blocks of a snapshot and sum their
|
2016-07-11 17:45:52 +00:00
|
|
|
* uncompressed and compressed sizes.
|
2015-04-08 18:37:13 +00:00
|
|
|
*/
|
|
|
|
/* ARGSUSED */
|
|
|
|
static int
|
|
|
|
dmu_calculate_send_traversal(spa_t *spa, zilog_t *zilog, const blkptr_t *bp,
|
|
|
|
const zbookmark_phys_t *zb, const dnode_phys_t *dnp, void *arg)
|
|
|
|
{
|
2016-07-11 17:45:52 +00:00
|
|
|
struct calculate_send_arg *space = arg;
|
2015-04-08 18:37:13 +00:00
|
|
|
if (bp != NULL && !BP_IS_HOLE(bp)) {
|
2016-07-11 17:45:52 +00:00
|
|
|
space->uncompressed += BP_GET_UCSIZE(bp);
|
|
|
|
space->compressed += BP_GET_PSIZE(bp);
|
2015-04-08 18:37:13 +00:00
|
|
|
}
|
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Given a desination snapshot and a TXG, calculate the approximate size of a
|
|
|
|
* send stream sent from that TXG. from_txg may be zero, indicating that the
|
|
|
|
* whole snapshot will be sent.
|
|
|
|
*/
|
|
|
|
int
|
|
|
|
dmu_send_estimate_from_txg(dsl_dataset_t *ds, uint64_t from_txg,
|
2016-07-11 17:45:52 +00:00
|
|
|
boolean_t stream_compressed, uint64_t *sizep)
|
2015-04-08 18:37:13 +00:00
|
|
|
{
|
|
|
|
int err;
|
2016-07-11 17:45:52 +00:00
|
|
|
struct calculate_send_arg size = { 0 };
|
2015-04-08 18:37:13 +00:00
|
|
|
|
2015-05-13 17:50:35 +00:00
|
|
|
ASSERT(dsl_pool_config_held(ds->ds_dir->dd_pool));
|
2015-04-08 18:37:13 +00:00
|
|
|
|
|
|
|
/* tosnap must be a snapshot */
|
|
|
|
if (!dsl_dataset_is_snapshot(ds))
|
|
|
|
return (SET_ERROR(EINVAL));
|
|
|
|
|
|
|
|
/* verify that from_txg is before the provided snapshot was taken */
|
|
|
|
if (from_txg >= dsl_dataset_phys(ds)->ds_creation_txg) {
|
|
|
|
return (SET_ERROR(EXDEV));
|
|
|
|
}
|
2011-11-17 18:14:36 +00:00
|
|
|
/*
|
2015-04-08 18:37:13 +00:00
|
|
|
* traverse the blocks of the snapshot with birth times after
|
|
|
|
* from_txg, summing their uncompressed size
|
2011-11-17 18:14:36 +00:00
|
|
|
*/
|
2015-04-08 18:37:13 +00:00
|
|
|
err = traverse_dataset(ds, from_txg, TRAVERSE_POST,
|
|
|
|
dmu_calculate_send_traversal, &size);
|
2016-07-11 17:45:52 +00:00
|
|
|
|
2015-04-08 18:37:13 +00:00
|
|
|
if (err)
|
2011-11-17 18:14:36 +00:00
|
|
|
return (err);
|
|
|
|
|
2016-07-11 17:45:52 +00:00
|
|
|
err = dmu_adjust_send_estimate_for_indirects(ds, size.uncompressed,
|
|
|
|
size.compressed, stream_compressed, sizep);
|
2015-04-08 18:37:13 +00:00
|
|
|
return (err);
|
2011-11-17 18:14:36 +00:00
|
|
|
}
|
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
typedef struct dmu_recv_begin_arg {
|
|
|
|
const char *drba_origin;
|
|
|
|
dmu_recv_cookie_t *drba_cookie;
|
|
|
|
cred_t *drba_cred;
|
2013-07-29 18:55:16 +00:00
|
|
|
uint64_t drba_snapobj;
|
2013-09-04 12:00:57 +00:00
|
|
|
} dmu_recv_begin_arg_t;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
static int
|
2013-09-04 12:00:57 +00:00
|
|
|
recv_begin_check_existing_impl(dmu_recv_begin_arg_t *drba, dsl_dataset_t *ds,
|
|
|
|
uint64_t fromguid)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
|
|
|
uint64_t val;
|
2013-09-04 12:00:57 +00:00
|
|
|
int error;
|
|
|
|
dsl_pool_t *dp = ds->ds_dir->dd_pool;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
/* temporary clone name must not exist */
|
|
|
|
error = zap_lookup(dp->dp_meta_objset,
|
2015-04-01 15:14:34 +00:00
|
|
|
dsl_dir_phys(ds->ds_dir)->dd_child_dir_zapobj, recv_clone_name,
|
2013-09-04 12:00:57 +00:00
|
|
|
8, 1, &val);
|
|
|
|
if (error != ENOENT)
|
|
|
|
return (error == 0 ? EBUSY : error);
|
|
|
|
|
2010-08-26 21:24:34 +00:00
|
|
|
/* new snapshot name must not exist */
|
2013-09-04 12:00:57 +00:00
|
|
|
error = zap_lookup(dp->dp_meta_objset,
|
2015-04-01 15:14:34 +00:00
|
|
|
dsl_dataset_phys(ds)->ds_snapnames_zapobj,
|
|
|
|
drba->drba_cookie->drc_tosnap, 8, 1, &val);
|
2013-09-04 12:00:57 +00:00
|
|
|
if (error != ENOENT)
|
|
|
|
return (error == 0 ? EEXIST : error);
|
2010-08-26 21:24:34 +00:00
|
|
|
|
2015-04-01 13:07:48 +00:00
|
|
|
/*
|
|
|
|
* Check snapshot limit before receiving. We'll recheck again at the
|
|
|
|
* end, but might as well abort before receiving if we're already over
|
|
|
|
* the limit.
|
|
|
|
*
|
|
|
|
* Note that we do not check the file system limit with
|
|
|
|
* dsl_dir_fscount_check because the temporary %clones don't count
|
|
|
|
* against that limit.
|
|
|
|
*/
|
|
|
|
error = dsl_fs_ss_limit_check(ds->ds_dir, 1, ZFS_PROP_SNAPSHOT_LIMIT,
|
|
|
|
NULL, drba->drba_cred);
|
|
|
|
if (error != 0)
|
|
|
|
return (error);
|
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
if (fromguid != 0) {
|
2013-07-29 18:55:16 +00:00
|
|
|
dsl_dataset_t *snap;
|
2015-04-01 15:14:34 +00:00
|
|
|
uint64_t obj = dsl_dataset_phys(ds)->ds_prev_snap_obj;
|
2013-07-29 18:55:16 +00:00
|
|
|
|
|
|
|
/* Find snapshot in this dir that matches fromguid. */
|
|
|
|
while (obj != 0) {
|
|
|
|
error = dsl_dataset_hold_obj(dp, obj, FTAG,
|
|
|
|
&snap);
|
|
|
|
if (error != 0)
|
|
|
|
return (SET_ERROR(ENODEV));
|
|
|
|
if (snap->ds_dir != ds->ds_dir) {
|
|
|
|
dsl_dataset_rele(snap, FTAG);
|
|
|
|
return (SET_ERROR(ENODEV));
|
|
|
|
}
|
2015-04-01 15:14:34 +00:00
|
|
|
if (dsl_dataset_phys(snap)->ds_guid == fromguid)
|
2013-07-29 18:55:16 +00:00
|
|
|
break;
|
2015-04-01 15:14:34 +00:00
|
|
|
obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
|
2013-07-29 18:55:16 +00:00
|
|
|
dsl_dataset_rele(snap, FTAG);
|
|
|
|
}
|
|
|
|
if (obj == 0)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(ENODEV));
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2013-07-29 18:55:16 +00:00
|
|
|
if (drba->drba_cookie->drc_force) {
|
|
|
|
drba->drba_snapobj = obj;
|
|
|
|
} else {
|
|
|
|
/*
|
|
|
|
* If we are not forcing, there must be no
|
|
|
|
* changes since fromsnap.
|
|
|
|
*/
|
|
|
|
if (dsl_dataset_modified_since_snap(ds, snap)) {
|
2010-05-28 20:45:14 +00:00
|
|
|
dsl_dataset_rele(snap, FTAG);
|
2013-07-29 18:55:16 +00:00
|
|
|
return (SET_ERROR(ETXTBSY));
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
2013-07-29 18:55:16 +00:00
|
|
|
drba->drba_snapobj = ds->ds_prev->ds_object;
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
2013-07-29 18:55:16 +00:00
|
|
|
|
|
|
|
dsl_dataset_rele(snap, FTAG);
|
2010-05-28 20:45:14 +00:00
|
|
|
} else {
|
2015-07-02 13:03:31 +00:00
|
|
|
/* if full, then must be forced */
|
|
|
|
if (!drba->drba_cookie->drc_force)
|
|
|
|
return (SET_ERROR(EEXIST));
|
|
|
|
/* start from $ORIGIN@$ORIGIN, if supported */
|
|
|
|
drba->drba_snapobj = dp->dp_origin_snap != NULL ?
|
|
|
|
dp->dp_origin_snap->ds_object : 0;
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
return (0);
|
2013-09-04 12:00:57 +00:00
|
|
|
|
|
|
|
}
|
|
|
|
|
|
|
|
static int
|
|
|
|
dmu_recv_begin_check(void *arg, dmu_tx_t *tx)
|
|
|
|
{
|
|
|
|
dmu_recv_begin_arg_t *drba = arg;
|
|
|
|
dsl_pool_t *dp = dmu_tx_pool(tx);
|
|
|
|
struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
|
|
|
|
uint64_t fromguid = drrb->drr_fromguid;
|
|
|
|
int flags = drrb->drr_flags;
|
|
|
|
int error;
|
2014-06-05 21:19:08 +00:00
|
|
|
uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
|
2013-09-04 12:00:57 +00:00
|
|
|
dsl_dataset_t *ds;
|
|
|
|
const char *tofs = drba->drba_cookie->drc_tofs;
|
|
|
|
|
|
|
|
/* already checked */
|
|
|
|
ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
|
2016-01-06 21:22:48 +00:00
|
|
|
ASSERT(!(featureflags & DMU_BACKUP_FEATURE_RESUMING));
|
2013-09-04 12:00:57 +00:00
|
|
|
|
|
|
|
if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
|
|
|
|
DMU_COMPOUNDSTREAM ||
|
|
|
|
drrb->drr_type >= DMU_OST_NUMTYPES ||
|
|
|
|
((flags & DRR_FLAG_CLONE) && drba->drba_origin == NULL))
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2013-09-04 12:00:57 +00:00
|
|
|
|
|
|
|
/* Verify pool version supports SA if SA_SPILL feature set */
|
2014-06-05 21:19:08 +00:00
|
|
|
if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
|
|
|
|
spa_version(dp->dp_spa) < SPA_VERSION_SA)
|
|
|
|
return (SET_ERROR(ENOTSUP));
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
if (drba->drba_cookie->drc_resumable &&
|
|
|
|
!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EXTENSIBLE_DATASET))
|
|
|
|
return (SET_ERROR(ENOTSUP));
|
|
|
|
|
2014-06-05 21:19:08 +00:00
|
|
|
/*
|
|
|
|
* The receiving code doesn't know how to translate a WRITE_EMBEDDED
|
2016-07-11 17:45:52 +00:00
|
|
|
* record to a plain WRITE record, so the pool must have the
|
2014-06-05 21:19:08 +00:00
|
|
|
* EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
|
|
|
|
* records. Same with WRITE_EMBEDDED records that use LZ4 compression.
|
|
|
|
*/
|
|
|
|
if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
|
|
|
|
!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
|
|
|
|
return (SET_ERROR(ENOTSUP));
|
2016-07-11 17:45:52 +00:00
|
|
|
if ((featureflags & DMU_BACKUP_FEATURE_LZ4) &&
|
2014-06-05 21:19:08 +00:00
|
|
|
!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(ENOTSUP));
|
2013-09-04 12:00:57 +00:00
|
|
|
|
2014-11-03 20:15:08 +00:00
|
|
|
/*
|
|
|
|
* The receiving code doesn't know how to translate large blocks
|
|
|
|
* to smaller ones, so the pool must have the LARGE_BLOCKS
|
2017-07-26 01:52:40 +00:00
|
|
|
* feature enabled if the stream has LARGE_BLOCKS. Same with
|
|
|
|
* large dnodes.
|
2014-11-03 20:15:08 +00:00
|
|
|
*/
|
|
|
|
if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
|
|
|
|
!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS))
|
|
|
|
return (SET_ERROR(ENOTSUP));
|
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 ((featureflags & DMU_BACKUP_FEATURE_LARGE_DNODE) &&
|
|
|
|
!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_DNODE))
|
|
|
|
return (SET_ERROR(ENOTSUP));
|
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
|
|
|
|
if (error == 0) {
|
|
|
|
/* target fs already exists; recv into temp clone */
|
|
|
|
|
|
|
|
/* Can't recv a clone into an existing fs */
|
2016-06-09 18:18:16 +00:00
|
|
|
if (flags & DRR_FLAG_CLONE || drba->drba_origin) {
|
2013-09-04 12:00:57 +00:00
|
|
|
dsl_dataset_rele(ds, FTAG);
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2013-09-04 12:00:57 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
error = recv_begin_check_existing_impl(drba, ds, fromguid);
|
|
|
|
dsl_dataset_rele(ds, FTAG);
|
|
|
|
} else if (error == ENOENT) {
|
|
|
|
/* target fs does not exist; must be a full backup or clone */
|
2016-06-15 21:28:36 +00:00
|
|
|
char buf[ZFS_MAX_DATASET_NAME_LEN];
|
2013-09-04 12:00:57 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* If it's a non-clone incremental, we are missing the
|
|
|
|
* target fs, so fail the recv.
|
|
|
|
*/
|
2015-12-22 01:31:57 +00:00
|
|
|
if (fromguid != 0 && !(flags & DRR_FLAG_CLONE ||
|
|
|
|
drba->drba_origin))
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(ENOENT));
|
2013-09-04 12:00:57 +00:00
|
|
|
|
2016-06-09 18:18:16 +00:00
|
|
|
/*
|
|
|
|
* If we're receiving a full send as a clone, and it doesn't
|
|
|
|
* contain all the necessary free records and freeobject
|
|
|
|
* records, reject it.
|
|
|
|
*/
|
|
|
|
if (fromguid == 0 && drba->drba_origin &&
|
|
|
|
!(flags & DRR_FLAG_FREERECORDS))
|
|
|
|
return (SET_ERROR(EINVAL));
|
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
/* Open the parent of tofs */
|
2016-06-15 21:28:36 +00:00
|
|
|
ASSERT3U(strlen(tofs), <, sizeof (buf));
|
2013-09-04 12:00:57 +00:00
|
|
|
(void) strlcpy(buf, tofs, strrchr(tofs, '/') - tofs + 1);
|
|
|
|
error = dsl_dataset_hold(dp, buf, FTAG, &ds);
|
|
|
|
if (error != 0)
|
|
|
|
return (error);
|
|
|
|
|
2015-04-01 13:07:48 +00:00
|
|
|
/*
|
|
|
|
* Check filesystem and snapshot limits before receiving. We'll
|
|
|
|
* recheck snapshot limits again at the end (we create the
|
|
|
|
* filesystems and increment those counts during begin_sync).
|
|
|
|
*/
|
|
|
|
error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
|
|
|
|
ZFS_PROP_FILESYSTEM_LIMIT, NULL, drba->drba_cred);
|
|
|
|
if (error != 0) {
|
|
|
|
dsl_dataset_rele(ds, FTAG);
|
|
|
|
return (error);
|
|
|
|
}
|
|
|
|
|
|
|
|
error = dsl_fs_ss_limit_check(ds->ds_dir, 1,
|
|
|
|
ZFS_PROP_SNAPSHOT_LIMIT, NULL, drba->drba_cred);
|
|
|
|
if (error != 0) {
|
|
|
|
dsl_dataset_rele(ds, FTAG);
|
|
|
|
return (error);
|
|
|
|
}
|
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
if (drba->drba_origin != NULL) {
|
|
|
|
dsl_dataset_t *origin;
|
|
|
|
error = dsl_dataset_hold(dp, drba->drba_origin,
|
|
|
|
FTAG, &origin);
|
|
|
|
if (error != 0) {
|
|
|
|
dsl_dataset_rele(ds, FTAG);
|
|
|
|
return (error);
|
|
|
|
}
|
2015-04-02 03:44:32 +00:00
|
|
|
if (!origin->ds_is_snapshot) {
|
2013-09-04 12:00:57 +00:00
|
|
|
dsl_dataset_rele(origin, FTAG);
|
|
|
|
dsl_dataset_rele(ds, FTAG);
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2013-09-04 12:00:57 +00:00
|
|
|
}
|
2016-06-09 18:18:16 +00:00
|
|
|
if (dsl_dataset_phys(origin)->ds_guid != fromguid &&
|
|
|
|
fromguid != 0) {
|
2013-09-04 12:00:57 +00:00
|
|
|
dsl_dataset_rele(origin, FTAG);
|
|
|
|
dsl_dataset_rele(ds, FTAG);
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(ENODEV));
|
2013-09-04 12:00:57 +00:00
|
|
|
}
|
|
|
|
dsl_dataset_rele(origin, FTAG);
|
|
|
|
}
|
|
|
|
dsl_dataset_rele(ds, FTAG);
|
|
|
|
error = 0;
|
|
|
|
}
|
|
|
|
return (error);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void
|
2013-09-04 12:00:57 +00:00
|
|
|
dmu_recv_begin_sync(void *arg, dmu_tx_t *tx)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
2013-09-04 12:00:57 +00:00
|
|
|
dmu_recv_begin_arg_t *drba = arg;
|
|
|
|
dsl_pool_t *dp = dmu_tx_pool(tx);
|
2016-01-06 21:22:48 +00:00
|
|
|
objset_t *mos = dp->dp_meta_objset;
|
2013-09-04 12:00:57 +00:00
|
|
|
struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
|
|
|
|
const char *tofs = drba->drba_cookie->drc_tofs;
|
|
|
|
dsl_dataset_t *ds, *newds;
|
2008-11-20 20:01:55 +00:00
|
|
|
uint64_t dsobj;
|
2013-09-04 12:00:57 +00:00
|
|
|
int error;
|
2016-01-06 21:22:48 +00:00
|
|
|
uint64_t crflags = 0;
|
2013-09-04 12:00:57 +00:00
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
if (drrb->drr_flags & DRR_FLAG_CI_DATA)
|
|
|
|
crflags |= DS_FLAG_CI_DATASET;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
|
|
|
|
if (error == 0) {
|
|
|
|
/* create temporary clone */
|
2013-07-29 18:55:16 +00:00
|
|
|
dsl_dataset_t *snap = NULL;
|
|
|
|
if (drba->drba_snapobj != 0) {
|
|
|
|
VERIFY0(dsl_dataset_hold_obj(dp,
|
|
|
|
drba->drba_snapobj, FTAG, &snap));
|
|
|
|
}
|
2013-09-04 12:00:57 +00:00
|
|
|
dsobj = dsl_dataset_create_sync(ds->ds_dir, recv_clone_name,
|
2013-07-29 18:55:16 +00:00
|
|
|
snap, crflags, drba->drba_cred, tx);
|
2016-02-05 21:47:48 +00:00
|
|
|
if (drba->drba_snapobj != 0)
|
|
|
|
dsl_dataset_rele(snap, FTAG);
|
2013-09-04 12:00:57 +00:00
|
|
|
dsl_dataset_rele(ds, FTAG);
|
|
|
|
} else {
|
|
|
|
dsl_dir_t *dd;
|
|
|
|
const char *tail;
|
|
|
|
dsl_dataset_t *origin = NULL;
|
|
|
|
|
|
|
|
VERIFY0(dsl_dir_hold(dp, tofs, FTAG, &dd, &tail));
|
|
|
|
|
|
|
|
if (drba->drba_origin != NULL) {
|
|
|
|
VERIFY0(dsl_dataset_hold(dp, drba->drba_origin,
|
|
|
|
FTAG, &origin));
|
|
|
|
}
|
|
|
|
|
|
|
|
/* Create new dataset. */
|
|
|
|
dsobj = dsl_dataset_create_sync(dd,
|
|
|
|
strrchr(tofs, '/') + 1,
|
|
|
|
origin, crflags, drba->drba_cred, tx);
|
|
|
|
if (origin != NULL)
|
|
|
|
dsl_dataset_rele(origin, FTAG);
|
|
|
|
dsl_dir_rele(dd, FTAG);
|
|
|
|
drba->drba_cookie->drc_newfs = B_TRUE;
|
|
|
|
}
|
|
|
|
VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &newds));
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
if (drba->drba_cookie->drc_resumable) {
|
|
|
|
uint64_t one = 1;
|
|
|
|
uint64_t zero = 0;
|
|
|
|
|
|
|
|
dsl_dataset_zapify(newds, tx);
|
|
|
|
if (drrb->drr_fromguid != 0) {
|
|
|
|
VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_FROMGUID,
|
|
|
|
8, 1, &drrb->drr_fromguid, tx));
|
|
|
|
}
|
|
|
|
VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TOGUID,
|
|
|
|
8, 1, &drrb->drr_toguid, tx));
|
|
|
|
VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_TONAME,
|
|
|
|
1, strlen(drrb->drr_toname) + 1, drrb->drr_toname, tx));
|
|
|
|
VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OBJECT,
|
|
|
|
8, 1, &one, tx));
|
|
|
|
VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_OFFSET,
|
|
|
|
8, 1, &zero, tx));
|
|
|
|
VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_BYTES,
|
|
|
|
8, 1, &zero, tx));
|
2016-07-11 17:45:52 +00:00
|
|
|
if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
|
|
|
|
DMU_BACKUP_FEATURE_LARGE_BLOCKS) {
|
|
|
|
VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_LARGEBLOCK,
|
2016-12-12 18:46:26 +00:00
|
|
|
8, 1, &one, tx));
|
2016-07-11 17:45:52 +00:00
|
|
|
}
|
2016-01-06 21:22:48 +00:00
|
|
|
if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
|
|
|
|
DMU_BACKUP_FEATURE_EMBED_DATA) {
|
|
|
|
VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_EMBEDOK,
|
|
|
|
8, 1, &one, tx));
|
|
|
|
}
|
2016-07-11 17:45:52 +00:00
|
|
|
if (DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo) &
|
|
|
|
DMU_BACKUP_FEATURE_COMPRESSED) {
|
|
|
|
VERIFY0(zap_add(mos, dsobj, DS_FIELD_RESUME_COMPRESSOK,
|
|
|
|
8, 1, &one, tx));
|
|
|
|
}
|
2016-01-06 21:22:48 +00:00
|
|
|
}
|
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
dmu_buf_will_dirty(newds->ds_dbuf, tx);
|
2015-04-01 15:14:34 +00:00
|
|
|
dsl_dataset_phys(newds)->ds_flags |= DS_FLAG_INCONSISTENT;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2010-05-28 20:45:14 +00:00
|
|
|
/*
|
|
|
|
* If we actually created a non-clone, we need to create the
|
|
|
|
* objset in our new dataset.
|
|
|
|
*/
|
2017-01-27 19:43:42 +00:00
|
|
|
rrw_enter(&newds->ds_bp_rwlock, RW_READER, FTAG);
|
2013-09-04 12:00:57 +00:00
|
|
|
if (BP_IS_HOLE(dsl_dataset_get_blkptr(newds))) {
|
2010-05-28 20:45:14 +00:00
|
|
|
(void) dmu_objset_create_impl(dp->dp_spa,
|
2013-09-04 12:00:57 +00:00
|
|
|
newds, dsl_dataset_get_blkptr(newds), drrb->drr_type, tx);
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
2017-01-27 19:43:42 +00:00
|
|
|
rrw_exit(&newds->ds_bp_rwlock, FTAG);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
drba->drba_cookie->drc_ds = newds;
|
2010-05-28 20:45:14 +00:00
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
spa_history_log_internal_ds(newds, "receive", tx, "");
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
static int
|
|
|
|
dmu_recv_resume_begin_check(void *arg, dmu_tx_t *tx)
|
|
|
|
{
|
|
|
|
dmu_recv_begin_arg_t *drba = arg;
|
|
|
|
dsl_pool_t *dp = dmu_tx_pool(tx);
|
|
|
|
struct drr_begin *drrb = drba->drba_cookie->drc_drrb;
|
|
|
|
int error;
|
|
|
|
uint64_t featureflags = DMU_GET_FEATUREFLAGS(drrb->drr_versioninfo);
|
|
|
|
dsl_dataset_t *ds;
|
|
|
|
const char *tofs = drba->drba_cookie->drc_tofs;
|
|
|
|
uint64_t val;
|
|
|
|
|
2016-06-15 21:28:36 +00:00
|
|
|
/* 6 extra bytes for /%recv */
|
|
|
|
char recvname[ZFS_MAX_DATASET_NAME_LEN + 6];
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
/* already checked */
|
|
|
|
ASSERT3U(drrb->drr_magic, ==, DMU_BACKUP_MAGIC);
|
|
|
|
ASSERT(featureflags & DMU_BACKUP_FEATURE_RESUMING);
|
|
|
|
|
|
|
|
if (DMU_GET_STREAM_HDRTYPE(drrb->drr_versioninfo) ==
|
|
|
|
DMU_COMPOUNDSTREAM ||
|
|
|
|
drrb->drr_type >= DMU_OST_NUMTYPES)
|
|
|
|
return (SET_ERROR(EINVAL));
|
|
|
|
|
|
|
|
/* Verify pool version supports SA if SA_SPILL feature set */
|
|
|
|
if ((featureflags & DMU_BACKUP_FEATURE_SA_SPILL) &&
|
|
|
|
spa_version(dp->dp_spa) < SPA_VERSION_SA)
|
|
|
|
return (SET_ERROR(ENOTSUP));
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The receiving code doesn't know how to translate a WRITE_EMBEDDED
|
|
|
|
* record to a plain WRITE record, so the pool must have the
|
|
|
|
* EMBEDDED_DATA feature enabled if the stream has WRITE_EMBEDDED
|
|
|
|
* records. Same with WRITE_EMBEDDED records that use LZ4 compression.
|
|
|
|
*/
|
|
|
|
if ((featureflags & DMU_BACKUP_FEATURE_EMBED_DATA) &&
|
|
|
|
!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_EMBEDDED_DATA))
|
|
|
|
return (SET_ERROR(ENOTSUP));
|
2016-07-11 17:45:52 +00:00
|
|
|
if ((featureflags & DMU_BACKUP_FEATURE_LZ4) &&
|
2016-01-06 21:22:48 +00:00
|
|
|
!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LZ4_COMPRESS))
|
|
|
|
return (SET_ERROR(ENOTSUP));
|
|
|
|
|
2017-07-26 01:52:40 +00:00
|
|
|
/*
|
|
|
|
* The receiving code doesn't know how to translate large blocks
|
|
|
|
* to smaller ones, so the pool must have the LARGE_BLOCKS
|
|
|
|
* feature enabled if the stream has LARGE_BLOCKS. Same with
|
|
|
|
* large dnodes.
|
|
|
|
*/
|
|
|
|
if ((featureflags & DMU_BACKUP_FEATURE_LARGE_BLOCKS) &&
|
|
|
|
!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_BLOCKS))
|
|
|
|
return (SET_ERROR(ENOTSUP));
|
|
|
|
if ((featureflags & DMU_BACKUP_FEATURE_LARGE_DNODE) &&
|
|
|
|
!spa_feature_is_enabled(dp->dp_spa, SPA_FEATURE_LARGE_DNODE))
|
|
|
|
return (SET_ERROR(ENOTSUP));
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
(void) snprintf(recvname, sizeof (recvname), "%s/%s",
|
|
|
|
tofs, recv_clone_name);
|
|
|
|
|
|
|
|
if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) {
|
|
|
|
/* %recv does not exist; continue in tofs */
|
|
|
|
error = dsl_dataset_hold(dp, tofs, FTAG, &ds);
|
|
|
|
if (error != 0)
|
|
|
|
return (error);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* check that ds is marked inconsistent */
|
|
|
|
if (!DS_IS_INCONSISTENT(ds)) {
|
|
|
|
dsl_dataset_rele(ds, FTAG);
|
|
|
|
return (SET_ERROR(EINVAL));
|
|
|
|
}
|
|
|
|
|
|
|
|
/* check that there is resuming data, and that the toguid matches */
|
|
|
|
if (!dsl_dataset_is_zapified(ds)) {
|
|
|
|
dsl_dataset_rele(ds, FTAG);
|
|
|
|
return (SET_ERROR(EINVAL));
|
|
|
|
}
|
|
|
|
error = zap_lookup(dp->dp_meta_objset, ds->ds_object,
|
|
|
|
DS_FIELD_RESUME_TOGUID, sizeof (val), 1, &val);
|
|
|
|
if (error != 0 || drrb->drr_toguid != val) {
|
|
|
|
dsl_dataset_rele(ds, FTAG);
|
|
|
|
return (SET_ERROR(EINVAL));
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Check if the receive is still running. If so, it will be owned.
|
|
|
|
* Note that nothing else can own the dataset (e.g. after the receive
|
|
|
|
* fails) because it will be marked inconsistent.
|
|
|
|
*/
|
|
|
|
if (dsl_dataset_has_owner(ds)) {
|
|
|
|
dsl_dataset_rele(ds, FTAG);
|
|
|
|
return (SET_ERROR(EBUSY));
|
|
|
|
}
|
|
|
|
|
|
|
|
/* There should not be any snapshots of this fs yet. */
|
|
|
|
if (ds->ds_prev != NULL && ds->ds_prev->ds_dir == ds->ds_dir) {
|
|
|
|
dsl_dataset_rele(ds, FTAG);
|
|
|
|
return (SET_ERROR(EINVAL));
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Note: resume point will be checked when we process the first WRITE
|
|
|
|
* record.
|
|
|
|
*/
|
|
|
|
|
|
|
|
/* check that the origin matches */
|
|
|
|
val = 0;
|
|
|
|
(void) zap_lookup(dp->dp_meta_objset, ds->ds_object,
|
|
|
|
DS_FIELD_RESUME_FROMGUID, sizeof (val), 1, &val);
|
|
|
|
if (drrb->drr_fromguid != val) {
|
|
|
|
dsl_dataset_rele(ds, FTAG);
|
|
|
|
return (SET_ERROR(EINVAL));
|
|
|
|
}
|
|
|
|
|
|
|
|
dsl_dataset_rele(ds, FTAG);
|
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
|
|
|
static void
|
|
|
|
dmu_recv_resume_begin_sync(void *arg, dmu_tx_t *tx)
|
|
|
|
{
|
|
|
|
dmu_recv_begin_arg_t *drba = arg;
|
|
|
|
dsl_pool_t *dp = dmu_tx_pool(tx);
|
|
|
|
const char *tofs = drba->drba_cookie->drc_tofs;
|
|
|
|
dsl_dataset_t *ds;
|
|
|
|
uint64_t dsobj;
|
2016-06-15 21:28:36 +00:00
|
|
|
/* 6 extra bytes for /%recv */
|
|
|
|
char recvname[ZFS_MAX_DATASET_NAME_LEN + 6];
|
2016-01-06 21:22:48 +00:00
|
|
|
|
|
|
|
(void) snprintf(recvname, sizeof (recvname), "%s/%s",
|
|
|
|
tofs, recv_clone_name);
|
|
|
|
|
|
|
|
if (dsl_dataset_hold(dp, recvname, FTAG, &ds) != 0) {
|
|
|
|
/* %recv does not exist; continue in tofs */
|
|
|
|
VERIFY0(dsl_dataset_hold(dp, tofs, FTAG, &ds));
|
|
|
|
drba->drba_cookie->drc_newfs = B_TRUE;
|
|
|
|
}
|
|
|
|
|
|
|
|
/* clear the inconsistent flag so that we can own it */
|
|
|
|
ASSERT(DS_IS_INCONSISTENT(ds));
|
|
|
|
dmu_buf_will_dirty(ds->ds_dbuf, tx);
|
|
|
|
dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
|
|
|
|
dsobj = ds->ds_object;
|
|
|
|
dsl_dataset_rele(ds, FTAG);
|
|
|
|
|
|
|
|
VERIFY0(dsl_dataset_own_obj(dp, dsobj, dmu_recv_tag, &ds));
|
|
|
|
|
|
|
|
dmu_buf_will_dirty(ds->ds_dbuf, tx);
|
|
|
|
dsl_dataset_phys(ds)->ds_flags |= DS_FLAG_INCONSISTENT;
|
|
|
|
|
2017-01-27 19:43:42 +00:00
|
|
|
rrw_enter(&ds->ds_bp_rwlock, RW_READER, FTAG);
|
2016-01-06 21:22:48 +00:00
|
|
|
ASSERT(!BP_IS_HOLE(dsl_dataset_get_blkptr(ds)));
|
2017-01-27 19:43:42 +00:00
|
|
|
rrw_exit(&ds->ds_bp_rwlock, FTAG);
|
2016-01-06 21:22:48 +00:00
|
|
|
|
|
|
|
drba->drba_cookie->drc_ds = ds;
|
|
|
|
|
|
|
|
spa_history_log_internal_ds(ds, "resume receive", tx, "");
|
|
|
|
}
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
/*
|
|
|
|
* NB: callers *MUST* call dmu_recv_stream() if dmu_recv_begin()
|
|
|
|
* succeeds; otherwise we will leak the holds on the datasets.
|
|
|
|
*/
|
|
|
|
int
|
2016-01-06 21:22:48 +00:00
|
|
|
dmu_recv_begin(char *tofs, char *tosnap, dmu_replay_record_t *drr_begin,
|
|
|
|
boolean_t force, boolean_t resumable, char *origin, dmu_recv_cookie_t *drc)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
2013-09-04 12:00:57 +00:00
|
|
|
dmu_recv_begin_arg_t drba = { 0 };
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
bzero(drc, sizeof (dmu_recv_cookie_t));
|
2016-01-06 21:22:48 +00:00
|
|
|
drc->drc_drr_begin = drr_begin;
|
|
|
|
drc->drc_drrb = &drr_begin->drr_u.drr_begin;
|
2008-11-20 20:01:55 +00:00
|
|
|
drc->drc_tosnap = tosnap;
|
2013-09-04 12:00:57 +00:00
|
|
|
drc->drc_tofs = tofs;
|
2008-11-20 20:01:55 +00:00
|
|
|
drc->drc_force = force;
|
2016-01-06 21:22:48 +00:00
|
|
|
drc->drc_resumable = resumable;
|
2015-04-01 13:07:48 +00:00
|
|
|
drc->drc_cred = CRED();
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
if (drc->drc_drrb->drr_magic == BSWAP_64(DMU_BACKUP_MAGIC)) {
|
2013-09-04 12:00:57 +00:00
|
|
|
drc->drc_byteswap = B_TRUE;
|
2016-07-22 15:52:49 +00:00
|
|
|
(void) fletcher_4_incremental_byteswap(drr_begin,
|
2013-09-04 12:00:57 +00:00
|
|
|
sizeof (dmu_replay_record_t), &drc->drc_cksum);
|
2016-01-06 21:22:48 +00:00
|
|
|
byteswap_record(drr_begin);
|
|
|
|
} else if (drc->drc_drrb->drr_magic == DMU_BACKUP_MAGIC) {
|
2016-07-22 15:52:49 +00:00
|
|
|
(void) fletcher_4_incremental_native(drr_begin,
|
2013-09-04 12:00:57 +00:00
|
|
|
sizeof (dmu_replay_record_t), &drc->drc_cksum);
|
2016-01-06 21:22:48 +00:00
|
|
|
} else {
|
|
|
|
return (SET_ERROR(EINVAL));
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
drba.drba_origin = origin;
|
|
|
|
drba.drba_cookie = drc;
|
|
|
|
drba.drba_cred = CRED();
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
if (DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo) &
|
|
|
|
DMU_BACKUP_FEATURE_RESUMING) {
|
|
|
|
return (dsl_sync_task(tofs,
|
|
|
|
dmu_recv_resume_begin_check, dmu_recv_resume_begin_sync,
|
|
|
|
&drba, 5, ZFS_SPACE_CHECK_NORMAL));
|
|
|
|
} else {
|
|
|
|
return (dsl_sync_task(tofs,
|
|
|
|
dmu_recv_begin_check, dmu_recv_begin_sync,
|
|
|
|
&drba, 5, ZFS_SPACE_CHECK_NORMAL));
|
|
|
|
}
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
struct receive_record_arg {
|
|
|
|
dmu_replay_record_t header;
|
|
|
|
void *payload; /* Pointer to a buffer containing the payload */
|
|
|
|
/*
|
|
|
|
* If the record is a write, pointer to the arc_buf_t containing the
|
|
|
|
* payload.
|
|
|
|
*/
|
|
|
|
arc_buf_t *write_buf;
|
|
|
|
int payload_size;
|
2016-01-06 21:22:48 +00:00
|
|
|
uint64_t bytes_read; /* bytes read from stream when record created */
|
2015-12-22 01:31:57 +00:00
|
|
|
boolean_t eos_marker; /* Marks the end of the stream */
|
|
|
|
bqueue_node_t node;
|
|
|
|
};
|
|
|
|
|
|
|
|
struct receive_writer_arg {
|
2015-07-06 03:20:31 +00:00
|
|
|
objset_t *os;
|
2013-09-04 12:00:57 +00:00
|
|
|
boolean_t byteswap;
|
2015-12-22 01:31:57 +00:00
|
|
|
bqueue_t q;
|
2016-01-06 21:22:48 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
/*
|
|
|
|
* These three args are used to signal to the main thread that we're
|
|
|
|
* done.
|
|
|
|
*/
|
|
|
|
kmutex_t mutex;
|
|
|
|
kcondvar_t cv;
|
|
|
|
boolean_t done;
|
2016-01-06 21:22:48 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
int err;
|
|
|
|
/* A map from guid to dataset to help handle dedup'd streams. */
|
|
|
|
avl_tree_t *guid_to_ds_map;
|
2016-01-06 21:22:48 +00:00
|
|
|
boolean_t resumable;
|
|
|
|
uint64_t last_object, last_offset;
|
|
|
|
uint64_t bytes_read; /* bytes read when current record created */
|
2015-12-22 01:31:57 +00:00
|
|
|
};
|
2015-07-06 03:20:31 +00:00
|
|
|
|
2016-06-09 18:18:16 +00:00
|
|
|
struct objlist {
|
|
|
|
list_t list; /* List of struct receive_objnode. */
|
|
|
|
/*
|
|
|
|
* Last object looked up. Used to assert that objects are being looked
|
|
|
|
* up in ascending order.
|
|
|
|
*/
|
|
|
|
uint64_t last_lookup;
|
|
|
|
};
|
|
|
|
|
|
|
|
struct receive_objnode {
|
|
|
|
list_node_t node;
|
|
|
|
uint64_t object;
|
|
|
|
};
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
struct receive_arg {
|
|
|
|
objset_t *os;
|
|
|
|
vnode_t *vp; /* The vnode to read the stream from */
|
|
|
|
uint64_t voff; /* The current offset in the stream */
|
2016-01-06 21:22:48 +00:00
|
|
|
uint64_t bytes_read;
|
2015-12-22 01:31:57 +00:00
|
|
|
/*
|
|
|
|
* A record that has had its payload read in, but hasn't yet been handed
|
|
|
|
* off to the worker thread.
|
|
|
|
*/
|
|
|
|
struct receive_record_arg *rrd;
|
|
|
|
/* A record that has had its header read in, but not its payload. */
|
|
|
|
struct receive_record_arg *next_rrd;
|
2008-11-20 20:01:55 +00:00
|
|
|
zio_cksum_t cksum;
|
2015-07-06 03:20:31 +00:00
|
|
|
zio_cksum_t prev_cksum;
|
2015-12-22 01:31:57 +00:00
|
|
|
int err;
|
|
|
|
boolean_t byteswap;
|
|
|
|
/* Sorted list of objects not to issue prefetches for. */
|
2016-06-09 18:18:16 +00:00
|
|
|
struct objlist ignore_objlist;
|
2008-11-20 20:01:55 +00:00
|
|
|
};
|
|
|
|
|
2010-05-28 20:45:14 +00:00
|
|
|
typedef struct guid_map_entry {
|
|
|
|
uint64_t guid;
|
|
|
|
dsl_dataset_t *gme_ds;
|
|
|
|
avl_node_t avlnode;
|
|
|
|
} guid_map_entry_t;
|
|
|
|
|
|
|
|
static int
|
|
|
|
guid_compare(const void *arg1, const void *arg2)
|
|
|
|
{
|
2016-08-27 18:12:53 +00:00
|
|
|
const guid_map_entry_t *gmep1 = (const guid_map_entry_t *)arg1;
|
|
|
|
const guid_map_entry_t *gmep2 = (const guid_map_entry_t *)arg2;
|
2010-05-28 20:45:14 +00:00
|
|
|
|
2016-08-27 18:12:53 +00:00
|
|
|
return (AVL_CMP(gmep1->guid, gmep2->guid));
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
|
|
|
|
2010-08-26 21:24:34 +00:00
|
|
|
static void
|
|
|
|
free_guid_map_onexit(void *arg)
|
|
|
|
{
|
|
|
|
avl_tree_t *ca = arg;
|
|
|
|
void *cookie = NULL;
|
|
|
|
guid_map_entry_t *gmep;
|
|
|
|
|
|
|
|
while ((gmep = avl_destroy_nodes(ca, &cookie)) != NULL) {
|
2013-09-04 12:00:57 +00:00
|
|
|
dsl_dataset_long_rele(gmep->gme_ds, gmep);
|
2013-04-10 21:54:56 +00:00
|
|
|
dsl_dataset_rele(gmep->gme_ds, gmep);
|
2010-08-26 21:24:34 +00:00
|
|
|
kmem_free(gmep, sizeof (guid_map_entry_t));
|
|
|
|
}
|
|
|
|
avl_destroy(ca);
|
|
|
|
kmem_free(ca, sizeof (avl_tree_t));
|
|
|
|
}
|
|
|
|
|
2015-07-06 03:20:31 +00:00
|
|
|
static int
|
2015-12-22 01:31:57 +00:00
|
|
|
receive_read(struct receive_arg *ra, int len, void *buf)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
|
|
|
int done = 0;
|
|
|
|
|
2016-06-09 19:07:01 +00:00
|
|
|
/*
|
|
|
|
* The code doesn't rely on this (lengths being multiples of 8). See
|
|
|
|
* comment in dump_bytes.
|
|
|
|
*/
|
2013-05-10 21:17:03 +00:00
|
|
|
ASSERT0(len % 8);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
while (done < len) {
|
|
|
|
ssize_t resid;
|
|
|
|
|
|
|
|
ra->err = vn_rdwr(UIO_READ, ra->vp,
|
2015-07-06 03:20:31 +00:00
|
|
|
(char *)buf + done, len - done,
|
2008-11-20 20:01:55 +00:00
|
|
|
ra->voff, UIO_SYSSPACE, FAPPEND,
|
|
|
|
RLIM64_INFINITY, CRED(), &resid);
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
if (resid == len - done) {
|
|
|
|
/*
|
|
|
|
* Note: ECKSUM indicates that the receive
|
|
|
|
* was interrupted and can potentially be resumed.
|
|
|
|
*/
|
|
|
|
ra->err = SET_ERROR(ECKSUM);
|
|
|
|
}
|
2008-11-20 20:01:55 +00:00
|
|
|
ra->voff += len - done - resid;
|
|
|
|
done = len - resid;
|
2013-09-04 12:00:57 +00:00
|
|
|
if (ra->err != 0)
|
2015-07-06 03:20:31 +00:00
|
|
|
return (ra->err);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
ra->bytes_read += len;
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
ASSERT3U(done, ==, len);
|
2015-07-06 03:20:31 +00:00
|
|
|
return (0);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
2010-08-26 17:58:36 +00:00
|
|
|
noinline static void
|
2015-07-06 03:20:31 +00:00
|
|
|
byteswap_record(dmu_replay_record_t *drr)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
|
|
|
#define DO64(X) (drr->drr_u.X = BSWAP_64(drr->drr_u.X))
|
|
|
|
#define DO32(X) (drr->drr_u.X = BSWAP_32(drr->drr_u.X))
|
|
|
|
drr->drr_type = BSWAP_32(drr->drr_type);
|
|
|
|
drr->drr_payloadlen = BSWAP_32(drr->drr_payloadlen);
|
2015-07-06 03:20:31 +00:00
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
switch (drr->drr_type) {
|
|
|
|
case DRR_BEGIN:
|
|
|
|
DO64(drr_begin.drr_magic);
|
2010-05-28 20:45:14 +00:00
|
|
|
DO64(drr_begin.drr_versioninfo);
|
2008-11-20 20:01:55 +00:00
|
|
|
DO64(drr_begin.drr_creation_time);
|
|
|
|
DO32(drr_begin.drr_type);
|
|
|
|
DO32(drr_begin.drr_flags);
|
|
|
|
DO64(drr_begin.drr_toguid);
|
|
|
|
DO64(drr_begin.drr_fromguid);
|
|
|
|
break;
|
|
|
|
case DRR_OBJECT:
|
|
|
|
DO64(drr_object.drr_object);
|
|
|
|
DO32(drr_object.drr_type);
|
|
|
|
DO32(drr_object.drr_bonustype);
|
|
|
|
DO32(drr_object.drr_blksz);
|
|
|
|
DO32(drr_object.drr_bonuslen);
|
2010-05-28 20:45:14 +00:00
|
|
|
DO64(drr_object.drr_toguid);
|
2008-11-20 20:01:55 +00:00
|
|
|
break;
|
|
|
|
case DRR_FREEOBJECTS:
|
|
|
|
DO64(drr_freeobjects.drr_firstobj);
|
|
|
|
DO64(drr_freeobjects.drr_numobjs);
|
2010-05-28 20:45:14 +00:00
|
|
|
DO64(drr_freeobjects.drr_toguid);
|
2008-11-20 20:01:55 +00:00
|
|
|
break;
|
|
|
|
case DRR_WRITE:
|
|
|
|
DO64(drr_write.drr_object);
|
|
|
|
DO32(drr_write.drr_type);
|
|
|
|
DO64(drr_write.drr_offset);
|
2016-07-11 17:45:52 +00:00
|
|
|
DO64(drr_write.drr_logical_size);
|
2010-05-28 20:45:14 +00:00
|
|
|
DO64(drr_write.drr_toguid);
|
2015-07-06 03:20:31 +00:00
|
|
|
ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write.drr_key.ddk_cksum);
|
2010-05-28 20:45:14 +00:00
|
|
|
DO64(drr_write.drr_key.ddk_prop);
|
2016-07-11 17:45:52 +00:00
|
|
|
DO64(drr_write.drr_compressed_size);
|
2010-05-28 20:45:14 +00:00
|
|
|
break;
|
|
|
|
case DRR_WRITE_BYREF:
|
|
|
|
DO64(drr_write_byref.drr_object);
|
|
|
|
DO64(drr_write_byref.drr_offset);
|
|
|
|
DO64(drr_write_byref.drr_length);
|
|
|
|
DO64(drr_write_byref.drr_toguid);
|
|
|
|
DO64(drr_write_byref.drr_refguid);
|
|
|
|
DO64(drr_write_byref.drr_refobject);
|
|
|
|
DO64(drr_write_byref.drr_refoffset);
|
2015-07-06 03:20:31 +00:00
|
|
|
ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_write_byref.
|
|
|
|
drr_key.ddk_cksum);
|
2010-05-28 20:45:14 +00:00
|
|
|
DO64(drr_write_byref.drr_key.ddk_prop);
|
2008-11-20 20:01:55 +00:00
|
|
|
break;
|
2014-06-05 21:19:08 +00:00
|
|
|
case DRR_WRITE_EMBEDDED:
|
|
|
|
DO64(drr_write_embedded.drr_object);
|
|
|
|
DO64(drr_write_embedded.drr_offset);
|
|
|
|
DO64(drr_write_embedded.drr_length);
|
|
|
|
DO64(drr_write_embedded.drr_toguid);
|
|
|
|
DO32(drr_write_embedded.drr_lsize);
|
|
|
|
DO32(drr_write_embedded.drr_psize);
|
|
|
|
break;
|
2008-11-20 20:01:55 +00:00
|
|
|
case DRR_FREE:
|
|
|
|
DO64(drr_free.drr_object);
|
|
|
|
DO64(drr_free.drr_offset);
|
|
|
|
DO64(drr_free.drr_length);
|
2010-05-28 20:45:14 +00:00
|
|
|
DO64(drr_free.drr_toguid);
|
|
|
|
break;
|
|
|
|
case DRR_SPILL:
|
|
|
|
DO64(drr_spill.drr_object);
|
|
|
|
DO64(drr_spill.drr_length);
|
|
|
|
DO64(drr_spill.drr_toguid);
|
2008-11-20 20:01:55 +00:00
|
|
|
break;
|
|
|
|
case DRR_END:
|
2010-05-28 20:45:14 +00:00
|
|
|
DO64(drr_end.drr_toguid);
|
2015-07-06 03:20:31 +00:00
|
|
|
ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_end.drr_checksum);
|
2008-11-20 20:01:55 +00:00
|
|
|
break;
|
2010-08-26 16:52:41 +00:00
|
|
|
default:
|
|
|
|
break;
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
2015-07-06 03:20:31 +00:00
|
|
|
|
|
|
|
if (drr->drr_type != DRR_BEGIN) {
|
|
|
|
ZIO_CHECKSUM_BSWAP(&drr->drr_u.drr_checksum.drr_checksum);
|
|
|
|
}
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
#undef DO64
|
|
|
|
#undef DO32
|
|
|
|
}
|
|
|
|
|
2014-09-12 03:28:35 +00:00
|
|
|
static inline uint8_t
|
|
|
|
deduce_nblkptr(dmu_object_type_t bonus_type, uint64_t bonus_size)
|
|
|
|
{
|
|
|
|
if (bonus_type == DMU_OT_SA) {
|
|
|
|
return (1);
|
|
|
|
} else {
|
|
|
|
return (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_OLD_MAX_BONUSLEN -
|
|
|
|
MIN(DN_OLD_MAX_BONUSLEN, bonus_size)) >> SPA_BLKPTRSHIFT));
|
2014-09-12 03:28:35 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
static void
|
|
|
|
save_resume_state(struct receive_writer_arg *rwa,
|
|
|
|
uint64_t object, uint64_t offset, dmu_tx_t *tx)
|
|
|
|
{
|
|
|
|
int txgoff = dmu_tx_get_txg(tx) & TXG_MASK;
|
|
|
|
|
|
|
|
if (!rwa->resumable)
|
|
|
|
return;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* We use ds_resume_bytes[] != 0 to indicate that we need to
|
|
|
|
* update this on disk, so it must not be 0.
|
|
|
|
*/
|
|
|
|
ASSERT(rwa->bytes_read != 0);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* We only resume from write records, which have a valid
|
|
|
|
* (non-meta-dnode) object number.
|
|
|
|
*/
|
|
|
|
ASSERT(object != 0);
|
|
|
|
|
|
|
|
/*
|
|
|
|
* For resuming to work correctly, we must receive records in order,
|
|
|
|
* sorted by object,offset. This is checked by the callers, but
|
|
|
|
* assert it here for good measure.
|
|
|
|
*/
|
|
|
|
ASSERT3U(object, >=, rwa->os->os_dsl_dataset->ds_resume_object[txgoff]);
|
|
|
|
ASSERT(object != rwa->os->os_dsl_dataset->ds_resume_object[txgoff] ||
|
|
|
|
offset >= rwa->os->os_dsl_dataset->ds_resume_offset[txgoff]);
|
|
|
|
ASSERT3U(rwa->bytes_read, >=,
|
|
|
|
rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff]);
|
|
|
|
|
|
|
|
rwa->os->os_dsl_dataset->ds_resume_object[txgoff] = object;
|
|
|
|
rwa->os->os_dsl_dataset->ds_resume_offset[txgoff] = offset;
|
|
|
|
rwa->os->os_dsl_dataset->ds_resume_bytes[txgoff] = rwa->bytes_read;
|
|
|
|
}
|
|
|
|
|
2010-08-26 17:58:36 +00:00
|
|
|
noinline static int
|
2015-12-22 01:31:57 +00:00
|
|
|
receive_object(struct receive_writer_arg *rwa, struct drr_object *drro,
|
2017-01-20 21:17:55 +00:00
|
|
|
void *data)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
2014-09-12 03:28:35 +00:00
|
|
|
dmu_object_info_t doi;
|
2008-11-20 20:01:55 +00:00
|
|
|
dmu_tx_t *tx;
|
2014-09-12 03:28:35 +00:00
|
|
|
uint64_t object;
|
|
|
|
int err;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
if (drro->drr_type == DMU_OT_NONE ||
|
2012-12-13 23:24:15 +00:00
|
|
|
!DMU_OT_IS_VALID(drro->drr_type) ||
|
|
|
|
!DMU_OT_IS_VALID(drro->drr_bonustype) ||
|
2010-05-28 20:45:14 +00:00
|
|
|
drro->drr_checksumtype >= ZIO_CHECKSUM_FUNCTIONS ||
|
2008-11-20 20:01:55 +00:00
|
|
|
drro->drr_compress >= ZIO_COMPRESS_FUNCTIONS ||
|
|
|
|
P2PHASE(drro->drr_blksz, SPA_MINBLOCKSIZE) ||
|
|
|
|
drro->drr_blksz < SPA_MINBLOCKSIZE ||
|
2015-12-22 01:31:57 +00:00
|
|
|
drro->drr_blksz > spa_maxblocksize(dmu_objset_spa(rwa->os)) ||
|
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
|
|
|
drro->drr_bonuslen >
|
2017-07-26 01:52:40 +00:00
|
|
|
DN_BONUS_SIZE(spa_maxdnodesize(dmu_objset_spa(rwa->os))) ||
|
|
|
|
drro->drr_dn_slots >
|
|
|
|
(spa_maxdnodesize(dmu_objset_spa(rwa->os)) >> DNODE_SHIFT)) {
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
err = dmu_object_info(rwa->os, drro->drr_object, &doi);
|
2009-07-02 22:44:48 +00:00
|
|
|
|
|
|
|
if (err != 0 && err != ENOENT)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2014-09-12 03:28:35 +00:00
|
|
|
object = err == 0 ? drro->drr_object : DMU_NEW_OBJECT;
|
2009-07-02 22:44:48 +00:00
|
|
|
|
2014-09-12 03:28:35 +00:00
|
|
|
/*
|
|
|
|
* If we are losing blkptrs or changing the block size this must
|
|
|
|
* be a new file instance. We must clear out the previous file
|
|
|
|
* contents before we can change this type of metadata in the dnode.
|
|
|
|
*/
|
|
|
|
if (err == 0) {
|
|
|
|
int nblkptr;
|
|
|
|
|
|
|
|
nblkptr = deduce_nblkptr(drro->drr_bonustype,
|
|
|
|
drro->drr_bonuslen);
|
|
|
|
|
|
|
|
if (drro->drr_blksz != doi.doi_data_block_size ||
|
|
|
|
nblkptr < doi.doi_nblkptr) {
|
2015-12-22 01:31:57 +00:00
|
|
|
err = dmu_free_long_range(rwa->os, drro->drr_object,
|
2014-09-12 03:28:35 +00:00
|
|
|
0, DMU_OBJECT_END);
|
|
|
|
if (err != 0)
|
|
|
|
return (SET_ERROR(EINVAL));
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
2014-09-12 03:28:35 +00:00
|
|
|
}
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
tx = dmu_tx_create(rwa->os);
|
2014-09-12 03:28:35 +00:00
|
|
|
dmu_tx_hold_bonus(tx, object);
|
|
|
|
err = dmu_tx_assign(tx, TXG_WAIT);
|
|
|
|
if (err != 0) {
|
|
|
|
dmu_tx_abort(tx);
|
|
|
|
return (err);
|
|
|
|
}
|
|
|
|
|
|
|
|
if (object == DMU_NEW_OBJECT) {
|
|
|
|
/* currently free, want to be allocated */
|
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
|
|
|
err = dmu_object_claim_dnsize(rwa->os, drro->drr_object,
|
2008-11-20 20:01:55 +00:00
|
|
|
drro->drr_type, drro->drr_blksz,
|
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
|
|
|
drro->drr_bonustype, drro->drr_bonuslen,
|
|
|
|
drro->drr_dn_slots << DNODE_SHIFT, tx);
|
2014-09-12 03:28:35 +00:00
|
|
|
} else if (drro->drr_type != doi.doi_type ||
|
|
|
|
drro->drr_blksz != doi.doi_data_block_size ||
|
|
|
|
drro->drr_bonustype != doi.doi_bonus_type ||
|
|
|
|
drro->drr_bonuslen != doi.doi_bonus_size) {
|
|
|
|
/* currently allocated, but with different properties */
|
2015-12-22 01:31:57 +00:00
|
|
|
err = dmu_object_reclaim(rwa->os, drro->drr_object,
|
2008-11-20 20:01:55 +00:00
|
|
|
drro->drr_type, drro->drr_blksz,
|
2014-09-12 03:28:35 +00:00
|
|
|
drro->drr_bonustype, drro->drr_bonuslen, tx);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
2013-09-04 12:00:57 +00:00
|
|
|
if (err != 0) {
|
2014-09-12 03:28:35 +00:00
|
|
|
dmu_tx_commit(tx);
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
2009-07-02 22:44:48 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
dmu_object_set_checksum(rwa->os, drro->drr_object,
|
2015-07-06 03:20:31 +00:00
|
|
|
drro->drr_checksumtype, tx);
|
2015-12-22 01:31:57 +00:00
|
|
|
dmu_object_set_compress(rwa->os, drro->drr_object,
|
2015-07-06 03:20:31 +00:00
|
|
|
drro->drr_compress, tx);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2008-12-03 20:09:06 +00:00
|
|
|
if (data != NULL) {
|
2008-11-20 20:01:55 +00:00
|
|
|
dmu_buf_t *db;
|
2008-12-03 20:09:06 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
VERIFY0(dmu_bonus_hold(rwa->os, drro->drr_object, FTAG, &db));
|
2008-11-20 20:01:55 +00:00
|
|
|
dmu_buf_will_dirty(db, tx);
|
|
|
|
|
|
|
|
ASSERT3U(db->db_size, >=, drro->drr_bonuslen);
|
|
|
|
bcopy(data, db->db_data, drro->drr_bonuslen);
|
2015-12-22 01:31:57 +00:00
|
|
|
if (rwa->byteswap) {
|
2012-12-13 23:24:15 +00:00
|
|
|
dmu_object_byteswap_t byteswap =
|
|
|
|
DMU_OT_BYTESWAP(drro->drr_bonustype);
|
|
|
|
dmu_ot_byteswap[byteswap].ob_func(db->db_data,
|
2008-11-20 20:01:55 +00:00
|
|
|
drro->drr_bonuslen);
|
|
|
|
}
|
|
|
|
dmu_buf_rele(db, FTAG);
|
|
|
|
}
|
|
|
|
dmu_tx_commit(tx);
|
2016-01-06 21:22:48 +00:00
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
|
|
|
/* ARGSUSED */
|
2010-08-26 17:58:36 +00:00
|
|
|
noinline static int
|
2015-12-22 01:31:57 +00:00
|
|
|
receive_freeobjects(struct receive_writer_arg *rwa,
|
2008-11-20 20:01:55 +00:00
|
|
|
struct drr_freeobjects *drrfo)
|
|
|
|
{
|
|
|
|
uint64_t obj;
|
2016-06-09 18:18:16 +00:00
|
|
|
int next_err = 0;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
if (drrfo->drr_firstobj + drrfo->drr_numobjs < drrfo->drr_firstobj)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
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 (obj = drrfo->drr_firstobj == 0 ? 1 : drrfo->drr_firstobj;
|
2016-06-09 18:18:16 +00:00
|
|
|
obj < drrfo->drr_firstobj + drrfo->drr_numobjs && next_err == 0;
|
|
|
|
next_err = dmu_object_next(rwa->os, &obj, FALSE, 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
|
|
|
dmu_object_info_t doi;
|
2008-11-20 20:01:55 +00:00
|
|
|
int err;
|
|
|
|
|
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
|
|
|
err = dmu_object_info(rwa->os, obj, &doi);
|
|
|
|
if (err == ENOENT) {
|
|
|
|
obj++;
|
2008-11-20 20:01:55 +00:00
|
|
|
continue;
|
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 if (err != 0) {
|
|
|
|
return (err);
|
|
|
|
}
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
err = dmu_free_long_object(rwa->os, obj);
|
2013-09-04 12:00:57 +00:00
|
|
|
if (err != 0)
|
2008-11-20 20:01:55 +00:00
|
|
|
return (err);
|
|
|
|
}
|
2016-06-09 18:18:16 +00:00
|
|
|
if (next_err != ESRCH)
|
|
|
|
return (next_err);
|
2008-11-20 20:01:55 +00:00
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
2010-08-26 17:58:36 +00:00
|
|
|
noinline static int
|
2015-12-22 01:31:57 +00:00
|
|
|
receive_write(struct receive_writer_arg *rwa, struct drr_write *drrw,
|
2017-01-20 21:17:55 +00:00
|
|
|
arc_buf_t *abuf)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
|
|
|
dmu_tx_t *tx;
|
2014-09-13 14:02:18 +00:00
|
|
|
dmu_buf_t *bonus;
|
2008-11-20 20:01:55 +00:00
|
|
|
int err;
|
|
|
|
|
2016-07-11 17:45:52 +00:00
|
|
|
if (drrw->drr_offset + drrw->drr_logical_size < drrw->drr_offset ||
|
2012-12-13 23:24:15 +00:00
|
|
|
!DMU_OT_IS_VALID(drrw->drr_type))
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
/*
|
|
|
|
* For resuming to work, records must be in increasing order
|
|
|
|
* by (object, offset).
|
|
|
|
*/
|
|
|
|
if (drrw->drr_object < rwa->last_object ||
|
|
|
|
(drrw->drr_object == rwa->last_object &&
|
|
|
|
drrw->drr_offset < rwa->last_offset)) {
|
|
|
|
return (SET_ERROR(EINVAL));
|
|
|
|
}
|
|
|
|
rwa->last_object = drrw->drr_object;
|
|
|
|
rwa->last_offset = drrw->drr_offset;
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
if (dmu_object_info(rwa->os, drrw->drr_object, NULL) != 0)
|
2014-09-13 14:02:18 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
tx = dmu_tx_create(rwa->os);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
dmu_tx_hold_write(tx, drrw->drr_object,
|
2016-07-11 17:45:52 +00:00
|
|
|
drrw->drr_offset, drrw->drr_logical_size);
|
2008-11-20 20:01:55 +00:00
|
|
|
err = dmu_tx_assign(tx, TXG_WAIT);
|
2013-09-04 12:00:57 +00:00
|
|
|
if (err != 0) {
|
2008-11-20 20:01:55 +00:00
|
|
|
dmu_tx_abort(tx);
|
|
|
|
return (err);
|
|
|
|
}
|
2015-12-22 01:31:57 +00:00
|
|
|
if (rwa->byteswap) {
|
2012-12-13 23:24:15 +00:00
|
|
|
dmu_object_byteswap_t byteswap =
|
|
|
|
DMU_OT_BYTESWAP(drrw->drr_type);
|
2015-07-06 03:20:31 +00:00
|
|
|
dmu_ot_byteswap[byteswap].ob_func(abuf->b_data,
|
2016-07-11 17:45:52 +00:00
|
|
|
DRR_WRITE_PAYLOAD_SIZE(drrw));
|
2012-12-13 23:24:15 +00:00
|
|
|
}
|
2015-07-06 03:20:31 +00:00
|
|
|
|
2016-07-11 17:45:52 +00:00
|
|
|
/* use the bonus buf to look up the dnode in dmu_assign_arcbuf */
|
2015-12-22 01:31:57 +00:00
|
|
|
if (dmu_bonus_hold(rwa->os, drrw->drr_object, FTAG, &bonus) != 0)
|
2015-07-06 03:20:31 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2014-09-13 14:02:18 +00:00
|
|
|
dmu_assign_arcbuf(bonus, drrw->drr_offset, abuf, tx);
|
2016-01-06 21:22:48 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Note: If the receive fails, we want the resume stream to start
|
|
|
|
* with the same record that we last successfully received (as opposed
|
|
|
|
* to the next record), so that we can verify that we are
|
|
|
|
* resuming from the correct location.
|
|
|
|
*/
|
|
|
|
save_resume_state(rwa, drrw->drr_object, drrw->drr_offset, tx);
|
2008-11-20 20:01:55 +00:00
|
|
|
dmu_tx_commit(tx);
|
2014-09-13 14:02:18 +00:00
|
|
|
dmu_buf_rele(bonus, FTAG);
|
2016-01-06 21:22:48 +00:00
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
2010-05-28 20:45:14 +00:00
|
|
|
/*
|
|
|
|
* Handle a DRR_WRITE_BYREF record. This record is used in dedup'ed
|
|
|
|
* streams to refer to a copy of the data that is already on the
|
|
|
|
* system because it came in earlier in the stream. This function
|
|
|
|
* finds the earlier copy of the data, and uses that copy instead of
|
|
|
|
* data from the stream to fulfill this write.
|
|
|
|
*/
|
|
|
|
static int
|
2015-12-22 01:31:57 +00:00
|
|
|
receive_write_byref(struct receive_writer_arg *rwa,
|
|
|
|
struct drr_write_byref *drrwbr)
|
2010-05-28 20:45:14 +00:00
|
|
|
{
|
|
|
|
dmu_tx_t *tx;
|
|
|
|
int err;
|
|
|
|
guid_map_entry_t gmesrch;
|
|
|
|
guid_map_entry_t *gmep;
|
2014-06-05 21:19:08 +00:00
|
|
|
avl_index_t where;
|
2010-05-28 20:45:14 +00:00
|
|
|
objset_t *ref_os = NULL;
|
|
|
|
dmu_buf_t *dbp;
|
|
|
|
|
|
|
|
if (drrwbr->drr_offset + drrwbr->drr_length < drrwbr->drr_offset)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2010-05-28 20:45:14 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* If the GUID of the referenced dataset is different from the
|
|
|
|
* GUID of the target dataset, find the referenced dataset.
|
|
|
|
*/
|
|
|
|
if (drrwbr->drr_toguid != drrwbr->drr_refguid) {
|
|
|
|
gmesrch.guid = drrwbr->drr_refguid;
|
2015-12-22 01:31:57 +00:00
|
|
|
if ((gmep = avl_find(rwa->guid_to_ds_map, &gmesrch,
|
2010-05-28 20:45:14 +00:00
|
|
|
&where)) == NULL) {
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
|
|
|
if (dmu_objset_from_ds(gmep->gme_ds, &ref_os))
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2010-05-28 20:45:14 +00:00
|
|
|
} else {
|
2015-12-22 01:31:57 +00:00
|
|
|
ref_os = rwa->os;
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
|
|
|
|
2010-08-26 16:52:42 +00:00
|
|
|
err = dmu_buf_hold(ref_os, drrwbr->drr_refobject,
|
|
|
|
drrwbr->drr_refoffset, FTAG, &dbp, DMU_READ_PREFETCH);
|
2014-06-05 21:19:08 +00:00
|
|
|
if (err != 0)
|
2010-05-28 20:45:14 +00:00
|
|
|
return (err);
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
tx = dmu_tx_create(rwa->os);
|
2010-05-28 20:45:14 +00:00
|
|
|
|
|
|
|
dmu_tx_hold_write(tx, drrwbr->drr_object,
|
|
|
|
drrwbr->drr_offset, drrwbr->drr_length);
|
|
|
|
err = dmu_tx_assign(tx, TXG_WAIT);
|
2013-09-04 12:00:57 +00:00
|
|
|
if (err != 0) {
|
2010-05-28 20:45:14 +00:00
|
|
|
dmu_tx_abort(tx);
|
|
|
|
return (err);
|
|
|
|
}
|
2015-12-22 01:31:57 +00:00
|
|
|
dmu_write(rwa->os, drrwbr->drr_object,
|
2010-05-28 20:45:14 +00:00
|
|
|
drrwbr->drr_offset, drrwbr->drr_length, dbp->db_data, tx);
|
|
|
|
dmu_buf_rele(dbp, FTAG);
|
2016-01-06 21:22:48 +00:00
|
|
|
|
|
|
|
/* See comment in restore_write. */
|
|
|
|
save_resume_state(rwa, drrwbr->drr_object, drrwbr->drr_offset, tx);
|
2010-05-28 20:45:14 +00:00
|
|
|
dmu_tx_commit(tx);
|
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
2014-06-05 21:19:08 +00:00
|
|
|
static int
|
2015-12-22 01:31:57 +00:00
|
|
|
receive_write_embedded(struct receive_writer_arg *rwa,
|
2016-01-06 21:22:48 +00:00
|
|
|
struct drr_write_embedded *drrwe, void *data)
|
2014-06-05 21:19:08 +00:00
|
|
|
{
|
|
|
|
dmu_tx_t *tx;
|
|
|
|
int err;
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
if (drrwe->drr_offset + drrwe->drr_length < drrwe->drr_offset)
|
2014-06-05 21:19:08 +00:00
|
|
|
return (EINVAL);
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
if (drrwe->drr_psize > BPE_PAYLOAD_SIZE)
|
2014-06-05 21:19:08 +00:00
|
|
|
return (EINVAL);
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
if (drrwe->drr_etype >= NUM_BP_EMBEDDED_TYPES)
|
2014-06-05 21:19:08 +00:00
|
|
|
return (EINVAL);
|
2016-01-06 21:22:48 +00:00
|
|
|
if (drrwe->drr_compression >= ZIO_COMPRESS_FUNCTIONS)
|
2014-06-05 21:19:08 +00:00
|
|
|
return (EINVAL);
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
tx = dmu_tx_create(rwa->os);
|
2014-06-05 21:19:08 +00:00
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
dmu_tx_hold_write(tx, drrwe->drr_object,
|
|
|
|
drrwe->drr_offset, drrwe->drr_length);
|
2014-06-05 21:19:08 +00:00
|
|
|
err = dmu_tx_assign(tx, TXG_WAIT);
|
|
|
|
if (err != 0) {
|
|
|
|
dmu_tx_abort(tx);
|
|
|
|
return (err);
|
|
|
|
}
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
dmu_write_embedded(rwa->os, drrwe->drr_object,
|
|
|
|
drrwe->drr_offset, data, drrwe->drr_etype,
|
|
|
|
drrwe->drr_compression, drrwe->drr_lsize, drrwe->drr_psize,
|
2015-12-22 01:31:57 +00:00
|
|
|
rwa->byteswap ^ ZFS_HOST_BYTEORDER, tx);
|
2014-06-05 21:19:08 +00:00
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
/* See comment in restore_write. */
|
|
|
|
save_resume_state(rwa, drrwe->drr_object, drrwe->drr_offset, tx);
|
2014-06-05 21:19:08 +00:00
|
|
|
dmu_tx_commit(tx);
|
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
2010-05-28 20:45:14 +00:00
|
|
|
static int
|
2015-12-22 01:31:57 +00:00
|
|
|
receive_spill(struct receive_writer_arg *rwa, struct drr_spill *drrs,
|
|
|
|
void *data)
|
2010-05-28 20:45:14 +00:00
|
|
|
{
|
|
|
|
dmu_tx_t *tx;
|
|
|
|
dmu_buf_t *db, *db_spill;
|
|
|
|
int err;
|
|
|
|
|
|
|
|
if (drrs->drr_length < SPA_MINBLOCKSIZE ||
|
2015-12-22 01:31:57 +00:00
|
|
|
drrs->drr_length > spa_maxblocksize(dmu_objset_spa(rwa->os)))
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2010-05-28 20:45:14 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
if (dmu_object_info(rwa->os, drrs->drr_object, NULL) != 0)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2010-05-28 20:45:14 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
VERIFY0(dmu_bonus_hold(rwa->os, drrs->drr_object, FTAG, &db));
|
2010-05-28 20:45:14 +00:00
|
|
|
if ((err = dmu_spill_hold_by_bonus(db, FTAG, &db_spill)) != 0) {
|
|
|
|
dmu_buf_rele(db, FTAG);
|
|
|
|
return (err);
|
|
|
|
}
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
tx = dmu_tx_create(rwa->os);
|
2010-05-28 20:45:14 +00:00
|
|
|
|
|
|
|
dmu_tx_hold_spill(tx, db->db_object);
|
|
|
|
|
|
|
|
err = dmu_tx_assign(tx, TXG_WAIT);
|
2013-09-04 12:00:57 +00:00
|
|
|
if (err != 0) {
|
2010-05-28 20:45:14 +00:00
|
|
|
dmu_buf_rele(db, FTAG);
|
|
|
|
dmu_buf_rele(db_spill, FTAG);
|
|
|
|
dmu_tx_abort(tx);
|
|
|
|
return (err);
|
|
|
|
}
|
|
|
|
dmu_buf_will_dirty(db_spill, tx);
|
|
|
|
|
|
|
|
if (db_spill->db_size < drrs->drr_length)
|
|
|
|
VERIFY(0 == dbuf_spill_set_blksz(db_spill,
|
|
|
|
drrs->drr_length, tx));
|
|
|
|
bcopy(data, db_spill->db_data, drrs->drr_length);
|
|
|
|
|
|
|
|
dmu_buf_rele(db, FTAG);
|
|
|
|
dmu_buf_rele(db_spill, FTAG);
|
|
|
|
|
|
|
|
dmu_tx_commit(tx);
|
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
/* ARGSUSED */
|
2010-08-26 17:58:36 +00:00
|
|
|
noinline static int
|
2015-12-22 01:31:57 +00:00
|
|
|
receive_free(struct receive_writer_arg *rwa, struct drr_free *drrf)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
|
|
|
int err;
|
|
|
|
|
|
|
|
if (drrf->drr_length != -1ULL &&
|
|
|
|
drrf->drr_offset + drrf->drr_length < drrf->drr_offset)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
if (dmu_object_info(rwa->os, drrf->drr_object, NULL) != 0)
|
2013-03-08 18:41:28 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
err = dmu_free_long_range(rwa->os, drrf->drr_object,
|
2008-11-20 20:01:55 +00:00
|
|
|
drrf->drr_offset, drrf->drr_length);
|
2015-12-22 01:31:57 +00:00
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
return (err);
|
|
|
|
}
|
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
/* used to destroy the drc_ds on error */
|
|
|
|
static void
|
|
|
|
dmu_recv_cleanup_ds(dmu_recv_cookie_t *drc)
|
|
|
|
{
|
2016-01-06 21:22:48 +00:00
|
|
|
if (drc->drc_resumable) {
|
|
|
|
/* wait for our resume state to be written to disk */
|
|
|
|
txg_wait_synced(drc->drc_ds->ds_dir->dd_pool, 0);
|
|
|
|
dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
|
|
|
|
} else {
|
2016-06-15 21:28:36 +00:00
|
|
|
char name[ZFS_MAX_DATASET_NAME_LEN];
|
2016-01-06 21:22:48 +00:00
|
|
|
dsl_dataset_name(drc->drc_ds, name);
|
|
|
|
dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
|
|
|
|
(void) dsl_destroy_head(name);
|
|
|
|
}
|
2013-09-04 12:00:57 +00:00
|
|
|
}
|
|
|
|
|
2015-07-06 03:20:31 +00:00
|
|
|
static void
|
2015-12-22 01:31:57 +00:00
|
|
|
receive_cksum(struct receive_arg *ra, int len, void *buf)
|
2015-07-06 03:20:31 +00:00
|
|
|
{
|
|
|
|
if (ra->byteswap) {
|
2016-07-22 15:52:49 +00:00
|
|
|
(void) fletcher_4_incremental_byteswap(buf, len, &ra->cksum);
|
2015-07-06 03:20:31 +00:00
|
|
|
} else {
|
2016-07-22 15:52:49 +00:00
|
|
|
(void) fletcher_4_incremental_native(buf, len, &ra->cksum);
|
2015-07-06 03:20:31 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
2015-12-22 01:31:57 +00:00
|
|
|
* Read the payload into a buffer of size len, and update the current record's
|
|
|
|
* payload field.
|
|
|
|
* Allocate ra->next_rrd and read the next record's header into
|
|
|
|
* ra->next_rrd->header.
|
2015-07-06 03:20:31 +00:00
|
|
|
* Verify checksum of payload and next record.
|
|
|
|
*/
|
|
|
|
static int
|
2015-12-22 01:31:57 +00:00
|
|
|
receive_read_payload_and_next_header(struct receive_arg *ra, int len, void *buf)
|
2015-07-06 03:20:31 +00:00
|
|
|
{
|
|
|
|
int err;
|
|
|
|
zio_cksum_t cksum_orig;
|
|
|
|
zio_cksum_t *cksump;
|
|
|
|
|
|
|
|
if (len != 0) {
|
2015-12-22 01:31:57 +00:00
|
|
|
ASSERT3U(len, <=, SPA_MAXBLOCKSIZE);
|
2016-01-06 21:22:48 +00:00
|
|
|
err = receive_read(ra, len, buf);
|
2015-07-06 03:20:31 +00:00
|
|
|
if (err != 0)
|
|
|
|
return (err);
|
2016-01-06 21:22:48 +00:00
|
|
|
receive_cksum(ra, len, buf);
|
|
|
|
|
|
|
|
/* note: rrd is NULL when reading the begin record's payload */
|
|
|
|
if (ra->rrd != NULL) {
|
|
|
|
ra->rrd->payload = buf;
|
|
|
|
ra->rrd->payload_size = len;
|
|
|
|
ra->rrd->bytes_read = ra->bytes_read;
|
|
|
|
}
|
2015-07-06 03:20:31 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
ra->prev_cksum = ra->cksum;
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
ra->next_rrd = kmem_zalloc(sizeof (*ra->next_rrd), KM_SLEEP);
|
|
|
|
err = receive_read(ra, sizeof (ra->next_rrd->header),
|
|
|
|
&ra->next_rrd->header);
|
2016-01-06 21:22:48 +00:00
|
|
|
ra->next_rrd->bytes_read = ra->bytes_read;
|
2015-12-22 01:31:57 +00:00
|
|
|
if (err != 0) {
|
|
|
|
kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
|
|
|
|
ra->next_rrd = NULL;
|
2015-07-06 03:20:31 +00:00
|
|
|
return (err);
|
2015-12-22 01:31:57 +00:00
|
|
|
}
|
|
|
|
if (ra->next_rrd->header.drr_type == DRR_BEGIN) {
|
|
|
|
kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
|
|
|
|
ra->next_rrd = NULL;
|
2015-07-06 03:20:31 +00:00
|
|
|
return (SET_ERROR(EINVAL));
|
2015-12-22 01:31:57 +00:00
|
|
|
}
|
2015-07-06 03:20:31 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Note: checksum is of everything up to but not including the
|
|
|
|
* checksum itself.
|
|
|
|
*/
|
|
|
|
ASSERT3U(offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
|
|
|
|
==, sizeof (dmu_replay_record_t) - sizeof (zio_cksum_t));
|
2015-12-22 01:31:57 +00:00
|
|
|
receive_cksum(ra,
|
2015-07-06 03:20:31 +00:00
|
|
|
offsetof(dmu_replay_record_t, drr_u.drr_checksum.drr_checksum),
|
2015-12-22 01:31:57 +00:00
|
|
|
&ra->next_rrd->header);
|
2015-07-06 03:20:31 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
cksum_orig = ra->next_rrd->header.drr_u.drr_checksum.drr_checksum;
|
|
|
|
cksump = &ra->next_rrd->header.drr_u.drr_checksum.drr_checksum;
|
2015-07-06 03:20:31 +00:00
|
|
|
|
|
|
|
if (ra->byteswap)
|
2015-12-22 01:31:57 +00:00
|
|
|
byteswap_record(&ra->next_rrd->header);
|
2015-07-06 03:20:31 +00:00
|
|
|
|
|
|
|
if ((!ZIO_CHECKSUM_IS_ZERO(cksump)) &&
|
2015-12-22 01:31:57 +00:00
|
|
|
!ZIO_CHECKSUM_EQUAL(ra->cksum, *cksump)) {
|
|
|
|
kmem_free(ra->next_rrd, sizeof (*ra->next_rrd));
|
|
|
|
ra->next_rrd = NULL;
|
2015-07-06 03:20:31 +00:00
|
|
|
return (SET_ERROR(ECKSUM));
|
2015-12-22 01:31:57 +00:00
|
|
|
}
|
2015-07-06 03:20:31 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
receive_cksum(ra, sizeof (cksum_orig), &cksum_orig);
|
2015-07-06 03:20:31 +00:00
|
|
|
|
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
2016-06-09 18:18:16 +00:00
|
|
|
static void
|
|
|
|
objlist_create(struct objlist *list)
|
|
|
|
{
|
|
|
|
list_create(&list->list, sizeof (struct receive_objnode),
|
|
|
|
offsetof(struct receive_objnode, node));
|
|
|
|
list->last_lookup = 0;
|
|
|
|
}
|
|
|
|
|
|
|
|
static void
|
|
|
|
objlist_destroy(struct objlist *list)
|
|
|
|
{
|
|
|
|
struct receive_objnode *n;
|
|
|
|
|
|
|
|
for (n = list_remove_head(&list->list);
|
|
|
|
n != NULL; n = list_remove_head(&list->list)) {
|
|
|
|
kmem_free(n, sizeof (*n));
|
|
|
|
}
|
|
|
|
list_destroy(&list->list);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* This function looks through the objlist to see if the specified object number
|
|
|
|
* is contained in the objlist. In the process, it will remove all object
|
|
|
|
* numbers in the list that are smaller than the specified object number. Thus,
|
|
|
|
* any lookup of an object number smaller than a previously looked up object
|
|
|
|
* number will always return false; therefore, all lookups should be done in
|
|
|
|
* ascending order.
|
|
|
|
*/
|
|
|
|
static boolean_t
|
|
|
|
objlist_exists(struct objlist *list, uint64_t object)
|
|
|
|
{
|
|
|
|
struct receive_objnode *node = list_head(&list->list);
|
|
|
|
ASSERT3U(object, >=, list->last_lookup);
|
|
|
|
list->last_lookup = object;
|
|
|
|
while (node != NULL && node->object < object) {
|
|
|
|
VERIFY3P(node, ==, list_remove_head(&list->list));
|
|
|
|
kmem_free(node, sizeof (*node));
|
|
|
|
node = list_head(&list->list);
|
|
|
|
}
|
|
|
|
return (node != NULL && node->object == object);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The objlist is a list of object numbers stored in ascending order. However,
|
|
|
|
* the insertion of new object numbers does not seek out the correct location to
|
|
|
|
* store a new object number; instead, it appends it to the list for simplicity.
|
|
|
|
* Thus, any users must take care to only insert new object numbers in ascending
|
|
|
|
* order.
|
|
|
|
*/
|
|
|
|
static void
|
|
|
|
objlist_insert(struct objlist *list, uint64_t object)
|
|
|
|
{
|
|
|
|
struct receive_objnode *node = kmem_zalloc(sizeof (*node), KM_SLEEP);
|
|
|
|
node->object = object;
|
|
|
|
#ifdef ZFS_DEBUG
|
|
|
|
{
|
|
|
|
struct receive_objnode *last_object = list_tail(&list->list);
|
|
|
|
uint64_t last_objnum = (last_object != NULL ? last_object->object : 0);
|
|
|
|
ASSERT3U(node->object, >, last_objnum);
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
list_insert_tail(&list->list, node);
|
|
|
|
}
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
/*
|
|
|
|
* Issue the prefetch reads for any necessary indirect blocks.
|
|
|
|
*
|
|
|
|
* We use the object ignore list to tell us whether or not to issue prefetches
|
|
|
|
* for a given object. We do this for both correctness (in case the blocksize
|
|
|
|
* of an object has changed) and performance (if the object doesn't exist, don't
|
|
|
|
* needlessly try to issue prefetches). We also trim the list as we go through
|
|
|
|
* the stream to prevent it from growing to an unbounded size.
|
|
|
|
*
|
|
|
|
* The object numbers within will always be in sorted order, and any write
|
|
|
|
* records we see will also be in sorted order, but they're not sorted with
|
|
|
|
* respect to each other (i.e. we can get several object records before
|
|
|
|
* receiving each object's write records). As a result, once we've reached a
|
|
|
|
* given object number, we can safely remove any reference to lower object
|
|
|
|
* numbers in the ignore list. In practice, we receive up to 32 object records
|
|
|
|
* before receiving write records, so the list can have up to 32 nodes in it.
|
|
|
|
*/
|
|
|
|
/* ARGSUSED */
|
|
|
|
static void
|
|
|
|
receive_read_prefetch(struct receive_arg *ra,
|
|
|
|
uint64_t object, uint64_t offset, uint64_t length)
|
|
|
|
{
|
2016-06-09 18:18:16 +00:00
|
|
|
if (!objlist_exists(&ra->ignore_objlist, object)) {
|
2015-12-22 01:31:57 +00:00
|
|
|
dmu_prefetch(ra->os, object, 1, offset, length,
|
|
|
|
ZIO_PRIORITY_SYNC_READ);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Read records off the stream, issuing any necessary prefetches.
|
|
|
|
*/
|
2015-07-06 03:20:31 +00:00
|
|
|
static int
|
2015-12-22 01:31:57 +00:00
|
|
|
receive_read_record(struct receive_arg *ra)
|
2015-07-06 03:20:31 +00:00
|
|
|
{
|
|
|
|
int err;
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
switch (ra->rrd->header.drr_type) {
|
2015-07-06 03:20:31 +00:00
|
|
|
case DRR_OBJECT:
|
|
|
|
{
|
2015-12-22 01:31:57 +00:00
|
|
|
struct drr_object *drro = &ra->rrd->header.drr_u.drr_object;
|
|
|
|
uint32_t size = P2ROUNDUP(drro->drr_bonuslen, 8);
|
|
|
|
void *buf = kmem_zalloc(size, KM_SLEEP);
|
|
|
|
dmu_object_info_t doi;
|
|
|
|
err = receive_read_payload_and_next_header(ra, size, buf);
|
|
|
|
if (err != 0) {
|
|
|
|
kmem_free(buf, size);
|
2015-07-06 03:20:31 +00:00
|
|
|
return (err);
|
2015-12-22 01:31:57 +00:00
|
|
|
}
|
|
|
|
err = dmu_object_info(ra->os, drro->drr_object, &doi);
|
|
|
|
/*
|
|
|
|
* See receive_read_prefetch for an explanation why we're
|
|
|
|
* storing this object in the ignore_obj_list.
|
|
|
|
*/
|
|
|
|
if (err == ENOENT ||
|
|
|
|
(err == 0 && doi.doi_data_block_size != drro->drr_blksz)) {
|
2016-06-09 18:18:16 +00:00
|
|
|
objlist_insert(&ra->ignore_objlist, drro->drr_object);
|
2015-12-22 01:31:57 +00:00
|
|
|
err = 0;
|
|
|
|
}
|
|
|
|
return (err);
|
2015-07-06 03:20:31 +00:00
|
|
|
}
|
|
|
|
case DRR_FREEOBJECTS:
|
|
|
|
{
|
2015-12-22 01:31:57 +00:00
|
|
|
err = receive_read_payload_and_next_header(ra, 0, NULL);
|
|
|
|
return (err);
|
2015-07-06 03:20:31 +00:00
|
|
|
}
|
|
|
|
case DRR_WRITE:
|
|
|
|
{
|
2015-12-22 01:31:57 +00:00
|
|
|
struct drr_write *drrw = &ra->rrd->header.drr_u.drr_write;
|
2016-07-11 17:45:52 +00:00
|
|
|
arc_buf_t *abuf;
|
|
|
|
boolean_t is_meta = DMU_OT_IS_METADATA(drrw->drr_type);
|
|
|
|
if (DRR_WRITE_COMPRESSED(drrw)) {
|
|
|
|
ASSERT3U(drrw->drr_compressed_size, >, 0);
|
|
|
|
ASSERT3U(drrw->drr_logical_size, >=,
|
|
|
|
drrw->drr_compressed_size);
|
|
|
|
ASSERT(!is_meta);
|
|
|
|
abuf = arc_loan_compressed_buf(
|
|
|
|
dmu_objset_spa(ra->os),
|
|
|
|
drrw->drr_compressed_size, drrw->drr_logical_size,
|
|
|
|
drrw->drr_compressiontype);
|
|
|
|
} else {
|
|
|
|
abuf = arc_loan_buf(dmu_objset_spa(ra->os),
|
|
|
|
is_meta, drrw->drr_logical_size);
|
|
|
|
}
|
2015-07-06 03:20:31 +00:00
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
err = receive_read_payload_and_next_header(ra,
|
2016-07-11 17:45:52 +00:00
|
|
|
DRR_WRITE_PAYLOAD_SIZE(drrw), abuf->b_data);
|
2015-12-22 01:31:57 +00:00
|
|
|
if (err != 0) {
|
2015-07-06 03:20:31 +00:00
|
|
|
dmu_return_arcbuf(abuf);
|
2015-12-22 01:31:57 +00:00
|
|
|
return (err);
|
|
|
|
}
|
|
|
|
ra->rrd->write_buf = abuf;
|
|
|
|
receive_read_prefetch(ra, drrw->drr_object, drrw->drr_offset,
|
2016-07-11 17:45:52 +00:00
|
|
|
drrw->drr_logical_size);
|
2015-07-06 03:20:31 +00:00
|
|
|
return (err);
|
|
|
|
}
|
|
|
|
case DRR_WRITE_BYREF:
|
|
|
|
{
|
2015-12-22 01:31:57 +00:00
|
|
|
struct drr_write_byref *drrwb =
|
|
|
|
&ra->rrd->header.drr_u.drr_write_byref;
|
|
|
|
err = receive_read_payload_and_next_header(ra, 0, NULL);
|
|
|
|
receive_read_prefetch(ra, drrwb->drr_object, drrwb->drr_offset,
|
|
|
|
drrwb->drr_length);
|
|
|
|
return (err);
|
2015-07-06 03:20:31 +00:00
|
|
|
}
|
|
|
|
case DRR_WRITE_EMBEDDED:
|
|
|
|
{
|
|
|
|
struct drr_write_embedded *drrwe =
|
2015-12-22 01:31:57 +00:00
|
|
|
&ra->rrd->header.drr_u.drr_write_embedded;
|
|
|
|
uint32_t size = P2ROUNDUP(drrwe->drr_psize, 8);
|
|
|
|
void *buf = kmem_zalloc(size, KM_SLEEP);
|
|
|
|
|
|
|
|
err = receive_read_payload_and_next_header(ra, size, buf);
|
|
|
|
if (err != 0) {
|
|
|
|
kmem_free(buf, size);
|
2015-07-06 03:20:31 +00:00
|
|
|
return (err);
|
2015-12-22 01:31:57 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
receive_read_prefetch(ra, drrwe->drr_object, drrwe->drr_offset,
|
|
|
|
drrwe->drr_length);
|
|
|
|
return (err);
|
2015-07-06 03:20:31 +00:00
|
|
|
}
|
|
|
|
case DRR_FREE:
|
|
|
|
{
|
2015-12-22 01:31:57 +00:00
|
|
|
/*
|
|
|
|
* It might be beneficial to prefetch indirect blocks here, but
|
|
|
|
* we don't really have the data to decide for sure.
|
|
|
|
*/
|
|
|
|
err = receive_read_payload_and_next_header(ra, 0, NULL);
|
|
|
|
return (err);
|
2015-07-06 03:20:31 +00:00
|
|
|
}
|
|
|
|
case DRR_END:
|
|
|
|
{
|
2015-12-22 01:31:57 +00:00
|
|
|
struct drr_end *drre = &ra->rrd->header.drr_u.drr_end;
|
2015-07-06 03:20:31 +00:00
|
|
|
if (!ZIO_CHECKSUM_EQUAL(ra->prev_cksum, drre->drr_checksum))
|
2016-01-06 21:22:48 +00:00
|
|
|
return (SET_ERROR(ECKSUM));
|
2015-07-06 03:20:31 +00:00
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
case DRR_SPILL:
|
|
|
|
{
|
2015-12-22 01:31:57 +00:00
|
|
|
struct drr_spill *drrs = &ra->rrd->header.drr_u.drr_spill;
|
|
|
|
void *buf = kmem_zalloc(drrs->drr_length, KM_SLEEP);
|
|
|
|
err = receive_read_payload_and_next_header(ra, drrs->drr_length,
|
|
|
|
buf);
|
2015-07-06 03:20:31 +00:00
|
|
|
if (err != 0)
|
2015-12-22 01:31:57 +00:00
|
|
|
kmem_free(buf, drrs->drr_length);
|
|
|
|
return (err);
|
|
|
|
}
|
|
|
|
default:
|
|
|
|
return (SET_ERROR(EINVAL));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Commit the records to the pool.
|
|
|
|
*/
|
|
|
|
static int
|
|
|
|
receive_process_record(struct receive_writer_arg *rwa,
|
|
|
|
struct receive_record_arg *rrd)
|
|
|
|
{
|
|
|
|
int err;
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
/* Processing in order, therefore bytes_read should be increasing. */
|
|
|
|
ASSERT3U(rrd->bytes_read, >=, rwa->bytes_read);
|
|
|
|
rwa->bytes_read = rrd->bytes_read;
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
switch (rrd->header.drr_type) {
|
|
|
|
case DRR_OBJECT:
|
|
|
|
{
|
|
|
|
struct drr_object *drro = &rrd->header.drr_u.drr_object;
|
|
|
|
err = receive_object(rwa, drro, rrd->payload);
|
|
|
|
kmem_free(rrd->payload, rrd->payload_size);
|
|
|
|
rrd->payload = NULL;
|
|
|
|
return (err);
|
|
|
|
}
|
|
|
|
case DRR_FREEOBJECTS:
|
|
|
|
{
|
|
|
|
struct drr_freeobjects *drrfo =
|
|
|
|
&rrd->header.drr_u.drr_freeobjects;
|
|
|
|
return (receive_freeobjects(rwa, drrfo));
|
|
|
|
}
|
|
|
|
case DRR_WRITE:
|
|
|
|
{
|
|
|
|
struct drr_write *drrw = &rrd->header.drr_u.drr_write;
|
|
|
|
err = receive_write(rwa, drrw, rrd->write_buf);
|
|
|
|
/* if receive_write() is successful, it consumes the arc_buf */
|
|
|
|
if (err != 0)
|
|
|
|
dmu_return_arcbuf(rrd->write_buf);
|
|
|
|
rrd->write_buf = NULL;
|
|
|
|
rrd->payload = NULL;
|
|
|
|
return (err);
|
|
|
|
}
|
|
|
|
case DRR_WRITE_BYREF:
|
|
|
|
{
|
|
|
|
struct drr_write_byref *drrwbr =
|
|
|
|
&rrd->header.drr_u.drr_write_byref;
|
|
|
|
return (receive_write_byref(rwa, drrwbr));
|
|
|
|
}
|
|
|
|
case DRR_WRITE_EMBEDDED:
|
|
|
|
{
|
|
|
|
struct drr_write_embedded *drrwe =
|
|
|
|
&rrd->header.drr_u.drr_write_embedded;
|
|
|
|
err = receive_write_embedded(rwa, drrwe, rrd->payload);
|
|
|
|
kmem_free(rrd->payload, rrd->payload_size);
|
|
|
|
rrd->payload = NULL;
|
|
|
|
return (err);
|
|
|
|
}
|
|
|
|
case DRR_FREE:
|
|
|
|
{
|
|
|
|
struct drr_free *drrf = &rrd->header.drr_u.drr_free;
|
|
|
|
return (receive_free(rwa, drrf));
|
|
|
|
}
|
|
|
|
case DRR_SPILL:
|
|
|
|
{
|
|
|
|
struct drr_spill *drrs = &rrd->header.drr_u.drr_spill;
|
|
|
|
err = receive_spill(rwa, drrs, rrd->payload);
|
|
|
|
kmem_free(rrd->payload, rrd->payload_size);
|
|
|
|
rrd->payload = NULL;
|
|
|
|
return (err);
|
2015-07-06 03:20:31 +00:00
|
|
|
}
|
|
|
|
default:
|
|
|
|
return (SET_ERROR(EINVAL));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2008-11-20 20:01:55 +00:00
|
|
|
/*
|
2015-12-22 01:31:57 +00:00
|
|
|
* dmu_recv_stream's worker thread; pull records off the queue, and then call
|
|
|
|
* receive_process_record When we're done, signal the main thread and exit.
|
|
|
|
*/
|
|
|
|
static void
|
|
|
|
receive_writer_thread(void *arg)
|
|
|
|
{
|
|
|
|
struct receive_writer_arg *rwa = arg;
|
|
|
|
struct receive_record_arg *rrd;
|
2016-08-21 13:22:32 +00:00
|
|
|
fstrans_cookie_t cookie = spl_fstrans_mark();
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
for (rrd = bqueue_dequeue(&rwa->q); !rrd->eos_marker;
|
|
|
|
rrd = bqueue_dequeue(&rwa->q)) {
|
|
|
|
/*
|
|
|
|
* If there's an error, the main thread will stop putting things
|
|
|
|
* on the queue, but we need to clear everything in it before we
|
|
|
|
* can exit.
|
|
|
|
*/
|
|
|
|
if (rwa->err == 0) {
|
|
|
|
rwa->err = receive_process_record(rwa, rrd);
|
|
|
|
} else if (rrd->write_buf != NULL) {
|
|
|
|
dmu_return_arcbuf(rrd->write_buf);
|
|
|
|
rrd->write_buf = NULL;
|
|
|
|
rrd->payload = NULL;
|
|
|
|
} else if (rrd->payload != NULL) {
|
|
|
|
kmem_free(rrd->payload, rrd->payload_size);
|
|
|
|
rrd->payload = NULL;
|
|
|
|
}
|
|
|
|
kmem_free(rrd, sizeof (*rrd));
|
|
|
|
}
|
|
|
|
kmem_free(rrd, sizeof (*rrd));
|
|
|
|
mutex_enter(&rwa->mutex);
|
|
|
|
rwa->done = B_TRUE;
|
|
|
|
cv_signal(&rwa->cv);
|
|
|
|
mutex_exit(&rwa->mutex);
|
2016-08-21 13:22:32 +00:00
|
|
|
spl_fstrans_unmark(cookie);
|
2017-01-18 23:10:35 +00:00
|
|
|
thread_exit();
|
2015-12-22 01:31:57 +00:00
|
|
|
}
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
static int
|
|
|
|
resume_check(struct receive_arg *ra, nvlist_t *begin_nvl)
|
|
|
|
{
|
|
|
|
uint64_t val;
|
|
|
|
objset_t *mos = dmu_objset_pool(ra->os)->dp_meta_objset;
|
|
|
|
uint64_t dsobj = dmu_objset_id(ra->os);
|
|
|
|
uint64_t resume_obj, resume_off;
|
|
|
|
|
|
|
|
if (nvlist_lookup_uint64(begin_nvl,
|
|
|
|
"resume_object", &resume_obj) != 0 ||
|
|
|
|
nvlist_lookup_uint64(begin_nvl,
|
|
|
|
"resume_offset", &resume_off) != 0) {
|
|
|
|
return (SET_ERROR(EINVAL));
|
|
|
|
}
|
|
|
|
VERIFY0(zap_lookup(mos, dsobj,
|
|
|
|
DS_FIELD_RESUME_OBJECT, sizeof (val), 1, &val));
|
|
|
|
if (resume_obj != val)
|
|
|
|
return (SET_ERROR(EINVAL));
|
|
|
|
VERIFY0(zap_lookup(mos, dsobj,
|
|
|
|
DS_FIELD_RESUME_OFFSET, sizeof (val), 1, &val));
|
|
|
|
if (resume_off != val)
|
|
|
|
return (SET_ERROR(EINVAL));
|
|
|
|
|
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
2015-12-22 01:31:57 +00:00
|
|
|
/*
|
|
|
|
* Read in the stream's records, one by one, and apply them to the pool. There
|
|
|
|
* are two threads involved; the thread that calls this function will spin up a
|
|
|
|
* worker thread, read the records off the stream one by one, and issue
|
|
|
|
* prefetches for any necessary indirect blocks. It will then push the records
|
|
|
|
* onto an internal blocking queue. The worker thread will pull the records off
|
|
|
|
* the queue, and actually write the data into the DMU. This way, the worker
|
|
|
|
* thread doesn't have to wait for reads to complete, since everything it needs
|
|
|
|
* (the indirect blocks) will be prefetched.
|
|
|
|
*
|
2008-11-20 20:01:55 +00:00
|
|
|
* NB: callers *must* call dmu_recv_end() if this succeeds.
|
|
|
|
*/
|
|
|
|
int
|
2010-08-26 21:24:34 +00:00
|
|
|
dmu_recv_stream(dmu_recv_cookie_t *drc, vnode_t *vp, offset_t *voffp,
|
|
|
|
int cleanup_fd, uint64_t *action_handlep)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
2015-07-06 03:20:31 +00:00
|
|
|
int err = 0;
|
2016-05-09 19:15:30 +00:00
|
|
|
struct receive_arg *ra;
|
|
|
|
struct receive_writer_arg *rwa;
|
2010-05-28 20:45:14 +00:00
|
|
|
int featureflags;
|
2016-01-06 21:22:48 +00:00
|
|
|
uint32_t payloadlen;
|
|
|
|
void *payload;
|
|
|
|
nvlist_t *begin_nvl = NULL;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2016-05-09 19:15:30 +00:00
|
|
|
ra = kmem_zalloc(sizeof (*ra), KM_SLEEP);
|
|
|
|
rwa = kmem_zalloc(sizeof (*rwa), KM_SLEEP);
|
|
|
|
|
|
|
|
ra->byteswap = drc->drc_byteswap;
|
|
|
|
ra->cksum = drc->drc_cksum;
|
|
|
|
ra->vp = vp;
|
|
|
|
ra->voff = *voffp;
|
2016-01-06 21:22:48 +00:00
|
|
|
|
|
|
|
if (dsl_dataset_is_zapified(drc->drc_ds)) {
|
|
|
|
(void) zap_lookup(drc->drc_ds->ds_dir->dd_pool->dp_meta_objset,
|
|
|
|
drc->drc_ds->ds_object, DS_FIELD_RESUME_BYTES,
|
|
|
|
sizeof (ra->bytes_read), 1, &ra->bytes_read);
|
|
|
|
}
|
|
|
|
|
2016-06-09 18:18:16 +00:00
|
|
|
objlist_create(&ra->ignore_objlist);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
/* these were verified in dmu_recv_begin */
|
2013-09-04 12:00:57 +00:00
|
|
|
ASSERT3U(DMU_GET_STREAM_HDRTYPE(drc->drc_drrb->drr_versioninfo), ==,
|
2010-05-28 20:45:14 +00:00
|
|
|
DMU_SUBSTREAM);
|
2013-09-04 12:00:57 +00:00
|
|
|
ASSERT3U(drc->drc_drrb->drr_type, <, DMU_OST_NUMTYPES);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
/*
|
|
|
|
* Open the objset we are modifying.
|
|
|
|
*/
|
2016-05-09 19:15:30 +00:00
|
|
|
VERIFY0(dmu_objset_from_ds(drc->drc_ds, &ra->os));
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2015-04-01 15:14:34 +00:00
|
|
|
ASSERT(dsl_dataset_phys(drc->drc_ds)->ds_flags & DS_FLAG_INCONSISTENT);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2010-05-28 20:45:14 +00:00
|
|
|
featureflags = DMU_GET_FEATUREFLAGS(drc->drc_drrb->drr_versioninfo);
|
|
|
|
|
|
|
|
/* if this stream is dedup'ed, set up the avl tree for guid mapping */
|
|
|
|
if (featureflags & DMU_BACKUP_FEATURE_DEDUP) {
|
2010-08-26 21:24:34 +00:00
|
|
|
minor_t minor;
|
|
|
|
|
|
|
|
if (cleanup_fd == -1) {
|
2016-05-09 19:15:30 +00:00
|
|
|
ra->err = SET_ERROR(EBADF);
|
2010-08-26 21:24:34 +00:00
|
|
|
goto out;
|
|
|
|
}
|
2016-05-09 19:15:30 +00:00
|
|
|
ra->err = zfs_onexit_fd_hold(cleanup_fd, &minor);
|
|
|
|
if (ra->err != 0) {
|
2010-08-26 21:24:34 +00:00
|
|
|
cleanup_fd = -1;
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (*action_handlep == 0) {
|
2016-05-09 19:15:30 +00:00
|
|
|
rwa->guid_to_ds_map =
|
2010-08-26 21:24:34 +00:00
|
|
|
kmem_alloc(sizeof (avl_tree_t), KM_SLEEP);
|
2016-05-09 19:15:30 +00:00
|
|
|
avl_create(rwa->guid_to_ds_map, guid_compare,
|
2010-08-26 21:24:34 +00:00
|
|
|
sizeof (guid_map_entry_t),
|
|
|
|
offsetof(guid_map_entry_t, avlnode));
|
2015-07-06 03:20:31 +00:00
|
|
|
err = zfs_onexit_add_cb(minor,
|
2016-05-09 19:15:30 +00:00
|
|
|
free_guid_map_onexit, rwa->guid_to_ds_map,
|
2010-08-26 21:24:34 +00:00
|
|
|
action_handlep);
|
2016-05-09 19:15:30 +00:00
|
|
|
if (ra->err != 0)
|
2010-08-26 21:24:34 +00:00
|
|
|
goto out;
|
|
|
|
} else {
|
2015-07-06 03:20:31 +00:00
|
|
|
err = zfs_onexit_cb_data(minor, *action_handlep,
|
2016-05-09 19:15:30 +00:00
|
|
|
(void **)&rwa->guid_to_ds_map);
|
|
|
|
if (ra->err != 0)
|
2010-08-26 21:24:34 +00:00
|
|
|
goto out;
|
|
|
|
}
|
2011-10-16 06:41:05 +00:00
|
|
|
|
2016-05-09 19:15:30 +00:00
|
|
|
drc->drc_guid_to_ds_map = rwa->guid_to_ds_map;
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
|
|
|
|
2016-01-06 21:22:48 +00:00
|
|
|
payloadlen = drc->drc_drr_begin->drr_payloadlen;
|
|
|
|
payload = NULL;
|
|
|
|
if (payloadlen != 0)
|
|
|
|
payload = kmem_alloc(payloadlen, KM_SLEEP);
|
|
|
|
|
|
|
|
err = receive_read_payload_and_next_header(ra, payloadlen, payload);
|
|
|
|
if (err != 0) {
|
|
|
|
if (payloadlen != 0)
|
|
|
|
kmem_free(payload, payloadlen);
|
2015-07-06 03:20:31 +00:00
|
|
|
goto out;
|
2016-01-06 21:22:48 +00:00
|
|
|
}
|
|
|
|
if (payloadlen != 0) {
|
|
|
|
err = nvlist_unpack(payload, payloadlen, &begin_nvl, KM_SLEEP);
|
|
|
|
kmem_free(payload, payloadlen);
|
|
|
|
if (err != 0)
|
|
|
|
goto out;
|
|
|
|
}
|
|
|
|
|
|
|
|
if (featureflags & DMU_BACKUP_FEATURE_RESUMING) {
|
|
|
|
err = resume_check(ra, begin_nvl);
|
|
|
|
if (err != 0)
|
|
|
|
goto out;
|
|
|
|
}
|
2015-07-06 03:20:31 +00:00
|
|
|
|
2016-05-09 19:15:30 +00:00
|
|
|
(void) bqueue_init(&rwa->q, zfs_recv_queue_length,
|
2015-12-22 01:31:57 +00:00
|
|
|
offsetof(struct receive_record_arg, node));
|
2016-05-09 19:15:30 +00:00
|
|
|
cv_init(&rwa->cv, NULL, CV_DEFAULT, NULL);
|
|
|
|
mutex_init(&rwa->mutex, NULL, MUTEX_DEFAULT, NULL);
|
|
|
|
rwa->os = ra->os;
|
|
|
|
rwa->byteswap = drc->drc_byteswap;
|
2016-01-06 21:22:48 +00:00
|
|
|
rwa->resumable = drc->drc_resumable;
|
2015-12-22 01:31:57 +00:00
|
|
|
|
2016-05-09 19:15:30 +00:00
|
|
|
(void) thread_create(NULL, 0, receive_writer_thread, rwa, 0, curproc,
|
2015-12-22 01:31:57 +00:00
|
|
|
TS_RUN, minclsyspri);
|
|
|
|
/*
|
2016-05-09 19:15:30 +00:00
|
|
|
* We're reading rwa->err without locks, which is safe since we are the
|
2015-12-22 01:31:57 +00:00
|
|
|
* only reader, and the worker thread is the only writer. It's ok if we
|
|
|
|
* miss a write for an iteration or two of the loop, since the writer
|
|
|
|
* thread will keep freeing records we send it until we send it an eos
|
|
|
|
* marker.
|
|
|
|
*
|
2016-05-09 19:15:30 +00:00
|
|
|
* We can leave this loop in 3 ways: First, if rwa->err is
|
2015-12-22 01:31:57 +00:00
|
|
|
* non-zero. In that case, the writer thread will free the rrd we just
|
|
|
|
* pushed. Second, if we're interrupted; in that case, either it's the
|
2016-01-06 21:22:48 +00:00
|
|
|
* first loop and ra->rrd was never allocated, or it's later and ra->rrd
|
2015-12-22 01:31:57 +00:00
|
|
|
* has been handed off to the writer thread who will free it. Finally,
|
|
|
|
* if receive_read_record fails or we're at the end of the stream, then
|
2016-05-09 19:15:30 +00:00
|
|
|
* we free ra->rrd and exit.
|
2015-12-22 01:31:57 +00:00
|
|
|
*/
|
2016-05-09 19:15:30 +00:00
|
|
|
while (rwa->err == 0) {
|
2008-11-20 20:01:55 +00:00
|
|
|
if (issig(JUSTLOOKING) && issig(FORREAL)) {
|
2015-07-06 03:20:31 +00:00
|
|
|
err = SET_ERROR(EINTR);
|
|
|
|
break;
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
2016-05-09 19:15:30 +00:00
|
|
|
ASSERT3P(ra->rrd, ==, NULL);
|
|
|
|
ra->rrd = ra->next_rrd;
|
|
|
|
ra->next_rrd = NULL;
|
|
|
|
/* Allocates and loads header into ra->next_rrd */
|
|
|
|
err = receive_read_record(ra);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2016-05-09 19:15:30 +00:00
|
|
|
if (ra->rrd->header.drr_type == DRR_END || err != 0) {
|
|
|
|
kmem_free(ra->rrd, sizeof (*ra->rrd));
|
|
|
|
ra->rrd = NULL;
|
2010-05-28 20:45:14 +00:00
|
|
|
break;
|
2015-12-22 01:31:57 +00:00
|
|
|
}
|
|
|
|
|
2016-05-09 19:15:30 +00:00
|
|
|
bqueue_enqueue(&rwa->q, ra->rrd,
|
|
|
|
sizeof (struct receive_record_arg) + ra->rrd->payload_size);
|
|
|
|
ra->rrd = NULL;
|
2015-12-22 01:31:57 +00:00
|
|
|
}
|
2016-05-09 19:15:30 +00:00
|
|
|
if (ra->next_rrd == NULL)
|
|
|
|
ra->next_rrd = kmem_zalloc(sizeof (*ra->next_rrd), KM_SLEEP);
|
|
|
|
ra->next_rrd->eos_marker = B_TRUE;
|
|
|
|
bqueue_enqueue(&rwa->q, ra->next_rrd, 1);
|
2015-12-22 01:31:57 +00:00
|
|
|
|
2016-05-09 19:15:30 +00:00
|
|
|
mutex_enter(&rwa->mutex);
|
|
|
|
while (!rwa->done) {
|
|
|
|
cv_wait(&rwa->cv, &rwa->mutex);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
2016-05-09 19:15:30 +00:00
|
|
|
mutex_exit(&rwa->mutex);
|
2015-12-22 01:31:57 +00:00
|
|
|
|
2016-05-09 19:15:30 +00:00
|
|
|
cv_destroy(&rwa->cv);
|
|
|
|
mutex_destroy(&rwa->mutex);
|
|
|
|
bqueue_destroy(&rwa->q);
|
2015-12-22 01:31:57 +00:00
|
|
|
if (err == 0)
|
2016-05-09 19:15:30 +00:00
|
|
|
err = rwa->err;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
|
|
|
out:
|
2016-01-06 21:22:48 +00:00
|
|
|
nvlist_free(begin_nvl);
|
2010-08-26 21:24:34 +00:00
|
|
|
if ((featureflags & DMU_BACKUP_FEATURE_DEDUP) && (cleanup_fd != -1))
|
|
|
|
zfs_onexit_fd_rele(cleanup_fd);
|
|
|
|
|
2015-07-06 03:20:31 +00:00
|
|
|
if (err != 0) {
|
2008-11-20 20:01:55 +00:00
|
|
|
/*
|
2016-01-06 21:22:48 +00:00
|
|
|
* Clean up references. If receive is not resumable,
|
|
|
|
* destroy what we created, so we don't leave it in
|
|
|
|
* the inconsistent state.
|
2008-11-20 20:01:55 +00:00
|
|
|
*/
|
2013-09-04 12:00:57 +00:00
|
|
|
dmu_recv_cleanup_ds(drc);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
2016-05-09 19:15:30 +00:00
|
|
|
*voffp = ra->voff;
|
2016-06-09 18:18:16 +00:00
|
|
|
objlist_destroy(&ra->ignore_objlist);
|
2016-05-09 19:15:30 +00:00
|
|
|
kmem_free(ra, sizeof (*ra));
|
|
|
|
kmem_free(rwa, sizeof (*rwa));
|
2015-07-06 03:20:31 +00:00
|
|
|
return (err);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static int
|
2013-09-04 12:00:57 +00:00
|
|
|
dmu_recv_end_check(void *arg, dmu_tx_t *tx)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
2013-09-04 12:00:57 +00:00
|
|
|
dmu_recv_cookie_t *drc = arg;
|
|
|
|
dsl_pool_t *dp = dmu_tx_pool(tx);
|
|
|
|
int error;
|
|
|
|
|
|
|
|
ASSERT3P(drc->drc_ds->ds_owner, ==, dmu_recv_tag);
|
|
|
|
|
|
|
|
if (!drc->drc_newfs) {
|
|
|
|
dsl_dataset_t *origin_head;
|
|
|
|
|
|
|
|
error = dsl_dataset_hold(dp, drc->drc_tofs, FTAG, &origin_head);
|
|
|
|
if (error != 0)
|
|
|
|
return (error);
|
2013-07-29 18:55:16 +00:00
|
|
|
if (drc->drc_force) {
|
|
|
|
/*
|
|
|
|
* We will destroy any snapshots in tofs (i.e. before
|
|
|
|
* origin_head) that are after the origin (which is
|
|
|
|
* the snap before drc_ds, because drc_ds can not
|
|
|
|
* have any snaps of its own).
|
|
|
|
*/
|
2015-04-01 15:14:34 +00:00
|
|
|
uint64_t obj;
|
|
|
|
|
|
|
|
obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
|
|
|
|
while (obj !=
|
|
|
|
dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
|
2013-07-29 18:55:16 +00:00
|
|
|
dsl_dataset_t *snap;
|
|
|
|
error = dsl_dataset_hold_obj(dp, obj, FTAG,
|
|
|
|
&snap);
|
|
|
|
if (error != 0)
|
2015-07-02 13:03:58 +00:00
|
|
|
break;
|
2013-07-29 18:55:16 +00:00
|
|
|
if (snap->ds_dir != origin_head->ds_dir)
|
|
|
|
error = SET_ERROR(EINVAL);
|
|
|
|
if (error == 0) {
|
|
|
|
error = dsl_destroy_snapshot_check_impl(
|
|
|
|
snap, B_FALSE);
|
|
|
|
}
|
2015-04-01 15:14:34 +00:00
|
|
|
obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
|
2013-07-29 18:55:16 +00:00
|
|
|
dsl_dataset_rele(snap, FTAG);
|
|
|
|
if (error != 0)
|
2015-07-02 13:03:58 +00:00
|
|
|
break;
|
|
|
|
}
|
|
|
|
if (error != 0) {
|
|
|
|
dsl_dataset_rele(origin_head, FTAG);
|
|
|
|
return (error);
|
2013-07-29 18:55:16 +00:00
|
|
|
}
|
|
|
|
}
|
2013-09-04 12:00:57 +00:00
|
|
|
error = dsl_dataset_clone_swap_check_impl(drc->drc_ds,
|
2013-07-27 17:50:07 +00:00
|
|
|
origin_head, drc->drc_force, drc->drc_owner, tx);
|
2013-09-04 12:00:57 +00:00
|
|
|
if (error != 0) {
|
|
|
|
dsl_dataset_rele(origin_head, FTAG);
|
|
|
|
return (error);
|
|
|
|
}
|
|
|
|
error = dsl_dataset_snapshot_check_impl(origin_head,
|
2015-04-01 13:07:48 +00:00
|
|
|
drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
|
2013-09-04 12:00:57 +00:00
|
|
|
dsl_dataset_rele(origin_head, FTAG);
|
|
|
|
if (error != 0)
|
|
|
|
return (error);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
error = dsl_destroy_head_check_impl(drc->drc_ds, 1);
|
|
|
|
} else {
|
|
|
|
error = dsl_dataset_snapshot_check_impl(drc->drc_ds,
|
2015-04-01 13:07:48 +00:00
|
|
|
drc->drc_tosnap, tx, B_TRUE, 1, drc->drc_cred);
|
2013-09-04 12:00:57 +00:00
|
|
|
}
|
|
|
|
return (error);
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
static void
|
2013-09-04 12:00:57 +00:00
|
|
|
dmu_recv_end_sync(void *arg, dmu_tx_t *tx)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
2013-09-04 12:00:57 +00:00
|
|
|
dmu_recv_cookie_t *drc = arg;
|
|
|
|
dsl_pool_t *dp = dmu_tx_pool(tx);
|
|
|
|
|
|
|
|
spa_history_log_internal_ds(drc->drc_ds, "finish receiving",
|
|
|
|
tx, "snap=%s", drc->drc_tosnap);
|
|
|
|
|
|
|
|
if (!drc->drc_newfs) {
|
|
|
|
dsl_dataset_t *origin_head;
|
|
|
|
|
|
|
|
VERIFY0(dsl_dataset_hold(dp, drc->drc_tofs, FTAG,
|
|
|
|
&origin_head));
|
2013-07-29 18:55:16 +00:00
|
|
|
|
|
|
|
if (drc->drc_force) {
|
|
|
|
/*
|
|
|
|
* Destroy any snapshots of drc_tofs (origin_head)
|
|
|
|
* after the origin (the snap before drc_ds).
|
|
|
|
*/
|
2015-04-01 15:14:34 +00:00
|
|
|
uint64_t obj;
|
|
|
|
|
|
|
|
obj = dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
|
|
|
|
while (obj !=
|
|
|
|
dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj) {
|
2013-07-29 18:55:16 +00:00
|
|
|
dsl_dataset_t *snap;
|
|
|
|
VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG,
|
|
|
|
&snap));
|
|
|
|
ASSERT3P(snap->ds_dir, ==, origin_head->ds_dir);
|
2015-04-01 15:14:34 +00:00
|
|
|
obj = dsl_dataset_phys(snap)->ds_prev_snap_obj;
|
2013-07-29 18:55:16 +00:00
|
|
|
dsl_destroy_snapshot_sync_impl(snap,
|
|
|
|
B_FALSE, tx);
|
|
|
|
dsl_dataset_rele(snap, FTAG);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
VERIFY3P(drc->drc_ds->ds_prev, ==,
|
|
|
|
origin_head->ds_prev);
|
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
dsl_dataset_clone_swap_sync_impl(drc->drc_ds,
|
|
|
|
origin_head, tx);
|
|
|
|
dsl_dataset_snapshot_sync_impl(origin_head,
|
|
|
|
drc->drc_tosnap, tx);
|
|
|
|
|
|
|
|
/* set snapshot's creation time and guid */
|
|
|
|
dmu_buf_will_dirty(origin_head->ds_prev->ds_dbuf, tx);
|
2015-04-01 15:14:34 +00:00
|
|
|
dsl_dataset_phys(origin_head->ds_prev)->ds_creation_time =
|
2013-09-04 12:00:57 +00:00
|
|
|
drc->drc_drrb->drr_creation_time;
|
2015-04-01 15:14:34 +00:00
|
|
|
dsl_dataset_phys(origin_head->ds_prev)->ds_guid =
|
2013-09-04 12:00:57 +00:00
|
|
|
drc->drc_drrb->drr_toguid;
|
2015-04-01 15:14:34 +00:00
|
|
|
dsl_dataset_phys(origin_head->ds_prev)->ds_flags &=
|
2013-09-04 12:00:57 +00:00
|
|
|
~DS_FLAG_INCONSISTENT;
|
|
|
|
|
|
|
|
dmu_buf_will_dirty(origin_head->ds_dbuf, tx);
|
2015-04-01 15:14:34 +00:00
|
|
|
dsl_dataset_phys(origin_head)->ds_flags &=
|
|
|
|
~DS_FLAG_INCONSISTENT;
|
2013-09-04 12:00:57 +00:00
|
|
|
|
2017-02-04 17:10:24 +00:00
|
|
|
drc->drc_newsnapobj =
|
|
|
|
dsl_dataset_phys(origin_head)->ds_prev_snap_obj;
|
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
dsl_dataset_rele(origin_head, FTAG);
|
|
|
|
dsl_destroy_head_sync_impl(drc->drc_ds, tx);
|
2013-07-27 17:50:07 +00:00
|
|
|
|
|
|
|
if (drc->drc_owner != NULL)
|
|
|
|
VERIFY3P(origin_head->ds_owner, ==, drc->drc_owner);
|
2013-09-04 12:00:57 +00:00
|
|
|
} else {
|
|
|
|
dsl_dataset_t *ds = drc->drc_ds;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
dsl_dataset_snapshot_sync_impl(ds, drc->drc_tosnap, tx);
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
/* set snapshot's creation time and guid */
|
|
|
|
dmu_buf_will_dirty(ds->ds_prev->ds_dbuf, tx);
|
2015-04-01 15:14:34 +00:00
|
|
|
dsl_dataset_phys(ds->ds_prev)->ds_creation_time =
|
2013-09-04 12:00:57 +00:00
|
|
|
drc->drc_drrb->drr_creation_time;
|
2015-04-01 15:14:34 +00:00
|
|
|
dsl_dataset_phys(ds->ds_prev)->ds_guid =
|
|
|
|
drc->drc_drrb->drr_toguid;
|
|
|
|
dsl_dataset_phys(ds->ds_prev)->ds_flags &=
|
|
|
|
~DS_FLAG_INCONSISTENT;
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
dmu_buf_will_dirty(ds->ds_dbuf, tx);
|
2015-04-01 15:14:34 +00:00
|
|
|
dsl_dataset_phys(ds)->ds_flags &= ~DS_FLAG_INCONSISTENT;
|
2016-01-06 21:22:48 +00:00
|
|
|
if (dsl_dataset_has_resume_receive_state(ds)) {
|
|
|
|
(void) zap_remove(dp->dp_meta_objset, ds->ds_object,
|
|
|
|
DS_FIELD_RESUME_FROMGUID, tx);
|
|
|
|
(void) zap_remove(dp->dp_meta_objset, ds->ds_object,
|
|
|
|
DS_FIELD_RESUME_OBJECT, tx);
|
|
|
|
(void) zap_remove(dp->dp_meta_objset, ds->ds_object,
|
|
|
|
DS_FIELD_RESUME_OFFSET, tx);
|
|
|
|
(void) zap_remove(dp->dp_meta_objset, ds->ds_object,
|
|
|
|
DS_FIELD_RESUME_BYTES, tx);
|
|
|
|
(void) zap_remove(dp->dp_meta_objset, ds->ds_object,
|
|
|
|
DS_FIELD_RESUME_TOGUID, tx);
|
|
|
|
(void) zap_remove(dp->dp_meta_objset, ds->ds_object,
|
|
|
|
DS_FIELD_RESUME_TONAME, tx);
|
|
|
|
}
|
2017-02-04 17:10:24 +00:00
|
|
|
drc->drc_newsnapobj =
|
|
|
|
dsl_dataset_phys(drc->drc_ds)->ds_prev_snap_obj;
|
2013-09-04 12:00:57 +00:00
|
|
|
}
|
2014-03-22 09:07:14 +00:00
|
|
|
zvol_create_minors(dp->dp_spa, drc->drc_tofs, B_TRUE);
|
2013-09-04 12:00:57 +00:00
|
|
|
/*
|
|
|
|
* Release the hold from dmu_recv_begin. This must be done before
|
|
|
|
* we return to open context, so that when we free the dataset's dnode,
|
|
|
|
* we can evict its bonus buffer.
|
|
|
|
*/
|
|
|
|
dsl_dataset_disown(drc->drc_ds, dmu_recv_tag);
|
|
|
|
drc->drc_ds = NULL;
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
|
|
|
|
2011-10-16 06:41:05 +00:00
|
|
|
static int
|
2013-09-04 12:00:57 +00:00
|
|
|
add_ds_to_guidmap(const char *name, avl_tree_t *guid_map, uint64_t snapobj)
|
2011-10-16 06:41:05 +00:00
|
|
|
{
|
2013-09-04 12:00:57 +00:00
|
|
|
dsl_pool_t *dp;
|
2011-10-16 06:41:05 +00:00
|
|
|
dsl_dataset_t *snapds;
|
|
|
|
guid_map_entry_t *gmep;
|
|
|
|
int err;
|
|
|
|
|
|
|
|
ASSERT(guid_map != NULL);
|
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
err = dsl_pool_hold(name, FTAG, &dp);
|
|
|
|
if (err != 0)
|
|
|
|
return (err);
|
2013-04-10 21:54:56 +00:00
|
|
|
gmep = kmem_alloc(sizeof (*gmep), KM_SLEEP);
|
|
|
|
err = dsl_dataset_hold_obj(dp, snapobj, gmep, &snapds);
|
2011-10-16 06:41:05 +00:00
|
|
|
if (err == 0) {
|
2015-04-01 15:14:34 +00:00
|
|
|
gmep->guid = dsl_dataset_phys(snapds)->ds_guid;
|
2011-10-16 06:41:05 +00:00
|
|
|
gmep->gme_ds = snapds;
|
|
|
|
avl_add(guid_map, gmep);
|
2013-09-04 12:00:57 +00:00
|
|
|
dsl_dataset_long_hold(snapds, gmep);
|
2013-04-10 21:54:56 +00:00
|
|
|
} else {
|
|
|
|
kmem_free(gmep, sizeof (*gmep));
|
2011-10-16 06:41:05 +00:00
|
|
|
}
|
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
dsl_pool_rele(dp, FTAG);
|
2011-10-16 06:41:05 +00:00
|
|
|
return (err);
|
|
|
|
}
|
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
static int dmu_recv_end_modified_blocks = 3;
|
|
|
|
|
2010-05-28 20:45:14 +00:00
|
|
|
static int
|
|
|
|
dmu_recv_existing_end(dmu_recv_cookie_t *drc)
|
2008-11-20 20:01:55 +00:00
|
|
|
{
|
2013-09-04 12:00:57 +00:00
|
|
|
#ifdef _KERNEL
|
|
|
|
/*
|
|
|
|
* We will be destroying the ds; make sure its origin is unmounted if
|
|
|
|
* necessary.
|
|
|
|
*/
|
2016-06-15 21:28:36 +00:00
|
|
|
char name[ZFS_MAX_DATASET_NAME_LEN];
|
2013-09-04 12:00:57 +00:00
|
|
|
dsl_dataset_name(drc->drc_ds, name);
|
|
|
|
zfs_destroy_unmount_origin(name);
|
|
|
|
#endif
|
2008-11-20 20:01:55 +00:00
|
|
|
|
2017-02-04 17:10:24 +00:00
|
|
|
return (dsl_sync_task(drc->drc_tofs,
|
2013-09-04 12:00:57 +00:00
|
|
|
dmu_recv_end_check, dmu_recv_end_sync, drc,
|
2017-02-04 17:10:24 +00:00
|
|
|
dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL));
|
2008-11-20 20:01:55 +00:00
|
|
|
}
|
2010-05-28 20:45:14 +00:00
|
|
|
|
|
|
|
static int
|
|
|
|
dmu_recv_new_end(dmu_recv_cookie_t *drc)
|
2017-02-04 17:10:24 +00:00
|
|
|
{
|
|
|
|
return (dsl_sync_task(drc->drc_tofs,
|
|
|
|
dmu_recv_end_check, dmu_recv_end_sync, drc,
|
|
|
|
dmu_recv_end_modified_blocks, ZFS_SPACE_CHECK_NORMAL));
|
|
|
|
}
|
|
|
|
|
|
|
|
int
|
|
|
|
dmu_recv_end(dmu_recv_cookie_t *drc, void *owner)
|
2010-05-28 20:45:14 +00:00
|
|
|
{
|
2013-09-04 12:00:57 +00:00
|
|
|
int error;
|
2010-05-28 20:45:14 +00:00
|
|
|
|
2017-02-04 17:10:24 +00:00
|
|
|
drc->drc_owner = owner;
|
|
|
|
|
|
|
|
if (drc->drc_newfs)
|
|
|
|
error = dmu_recv_new_end(drc);
|
|
|
|
else
|
|
|
|
error = dmu_recv_existing_end(drc);
|
2010-05-28 20:45:14 +00:00
|
|
|
|
2013-09-04 12:00:57 +00:00
|
|
|
if (error != 0) {
|
|
|
|
dmu_recv_cleanup_ds(drc);
|
|
|
|
} else if (drc->drc_guid_to_ds_map != NULL) {
|
|
|
|
(void) add_ds_to_guidmap(drc->drc_tofs,
|
|
|
|
drc->drc_guid_to_ds_map,
|
|
|
|
drc->drc_newsnapobj);
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
2013-09-04 12:00:57 +00:00
|
|
|
return (error);
|
2010-05-28 20:45:14 +00:00
|
|
|
}
|
|
|
|
|
2013-07-29 18:58:53 +00:00
|
|
|
/*
|
|
|
|
* Return TRUE if this objset is currently being received into.
|
|
|
|
*/
|
|
|
|
boolean_t
|
|
|
|
dmu_objset_is_receiving(objset_t *os)
|
|
|
|
{
|
|
|
|
return (os->os_dsl_dataset != NULL &&
|
|
|
|
os->os_dsl_dataset->ds_owner == dmu_recv_tag);
|
|
|
|
}
|
2013-12-17 21:53:52 +00:00
|
|
|
|
|
|
|
#if defined(_KERNEL)
|
|
|
|
module_param(zfs_send_corrupt_data, int, 0644);
|
|
|
|
MODULE_PARM_DESC(zfs_send_corrupt_data, "Allow sending corrupt data");
|
|
|
|
#endif
|