Before r278702 prefetch was blocked for I/Os > 1MB, after -- >= 1MB.
1MB I/Os are used for bulk operations in CTL (XCOPY, VERIFY), and disabling
prefetch for them reduced the performance.
This is temporary local patch, that should be replaced when upstreamed.
Discussed with: mahrens
MFC after: 3 days
otherwise DTRACE_ANCHORED() returns false and that makes stack()
insert a bogus frame at the top.
For example:
dtrace -n 'test:dtrace_test::sdttest { stack(); }
This change is not really a solution, but just a work-around.
The real solution is to record the probe's call site and to use
that for resolving a function name.
PR: 195222
MFC after: 22 days
A change to a property on a dataset must be propagated to its descendants
in case that property is inherited. For datasets whose information is
not currently loaded into memory (e.g. a snapshot that isn't currently
mounted), there is nothing to do; the property change will take effect
the next time that dataset is loaded. To handle updates to datasets that
are in-core, ZFS registers a callback entry for each property of each
loaded dataset with the dsl directory that holds that dataset. There
is a dsl directory associated with each live dataset that references
both the live dataset and any snapshots of the live dataset. A property
change is effected by doing a traversal of the tree of dsl directories
for a pool, starting at the directory sourcing the change, and invoking
these callbacks.
The current implementation both registers and de-registers properties
individually for each loaded dataset. While registration for a property is
O(1) (insert into a list), de-registration is O(n) (search list and then
remove). The 'n' for de-registration, however, is not limited to the size
(number of snapshots + 1) of the dsl directory. The eviction portion
of the life cycle for the in core state of datasets is asynchronous,
which allows multiple copies of the dataset information to be in-core
at once. Only one of these copies is active at any time with the rest
going through tear down processing, but all copies contribute to the
cost of performing a dsl_prop_unregister().
One way to create multiple, in-flight copies of dataset information
is by performing "zfs list" operations from multiple threads
concurrently. In-core dataset information is loaded on demand and then
evicted when reference counts drops to zero. For datasets that are not
mounted, there is no persistent reference count to keep them resident.
So, a list operation will load them, compute the information required to
do the list operation, and then evict them. When performing this operation
from multiple threads it is possible that some of the in-core dataset
information will be reused, but also possible to lose the race and load
the dataset again, even while the same information is being torn down.
Compounding the performance issue further is a change made for illumos
issue 5056 which made dataset eviction single threaded. In environments
using automation to manage ZFS datasets, it is now possible to create
enough of a backlog of dataset evictions to consume excessive amounts
of kernel memory and to bog down the system.
The fix employed here is to make property de-registration O(1). With this
change in place, it is hoped that a single thread is more than sufficient
to handle eviction processing. If it isn't, the problem can be solved
by increasing the number of threads devoted to the eviction taskq.
sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dsl_dataset.c
sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dsl_dir.c:
sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dsl_prop.c:
sys/cddl/contrib/opensolaris/uts/common/fs/zfs/sys/dsl_dataset.h:
sys/cddl/contrib/opensolaris/uts/common/fs/zfs/sys/dsl_dir.h:
sys/cddl/contrib/opensolaris/uts/common/fs/zfs/sys/dsl_prop.h:
Associate dsl property callback records with both the
dsl directory and the dsl dataset that is registering the
callback. Both connections are protected by the dsl directory's
"dd_lock".
When linking callbacks into a dsl directory, group them by
the property type. This helps reduce the space penalty for the
double association (the property name pointer is stored once
per dsl_dir instead of in each record) and reduces the number of
strcmp() calls required to do callback processing when updating
a single property. Property types are stored in a linked list
since currently ZFS registers a maximum of 10 property types
for each dataset.
Note that the property buckets/records associated with a dsl
directory are created on demand, but only freed when the dsl
directory is freed. Given the static nature of property types
and their small number, there is no benefit to freeing the few
bytes of memory used to represent the property record earlier.
When a property record becomes empty, the dsl directory is either
going to become unreferenced a little later in this thread of
execution, or there is a high chance that another dataset is
going to be loaded that would recreate the bucket anyway.
Replace dsl_prop_unregister() with dsl_prop_unregister_all().
All callers of dsl_prop_unregister() are trying to remove
all property registrations for a given dsl dataset anyway. By
changing the API, we can avoid doing any lookups of callbacks
by property type and just traverse the list of all callbacks
for the dataset and free each one.
sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dmu_objset.c:
sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zfs_vfsops.c:
Replace use of dsl_prop_unregister() with the new
dsl_prop_unregister_all() API.
illumos/illumos-gate@03bad06fbb
Author: Justin Gibbs <gibbs@scsiguy.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Prakash Surya <prakash.surya@delphix.com>
Approved by: Dan McDonald <danmcd@omniti.com>
Illumos issue:
6171 dsl_prop_unregister() slows down dataset eviction
https://www.illumos.org/issues/6171
MFC after: 2 weeks
c546f36aa8https://www.illumos.org/issues/6220
5408 introduced a memleak in l2arc, namely the member b_thawed gets leaked when
an arc_hdr is realloced from full to l2only.
Reviewed by: Saso Kiselkov <saso.kiselkov@nexenta.com>
Reviewed by: Simon Klinkert <simon.klinkert@gmail.com>
Reviewed by: George Wilson <george@delphix.com>
Approved by: Robert Mustacchi <rm@joyent.com>
Author: Arne Jansen <sensille@gmx.net>
updated
ZFS already supports storing the vdev FRU in a vdev property. There
is code in libzfs to work with this property, and there is code in
the zfs-retire FMA module that looks for that information. But there
is no code actually setting or updating the FRU.
To address this, ZFS is changed to send a handful of new events
whenever a vdev is added, attached, cleared, or onlined, as well
as when a pool is created or imported.
Note that syseventd is not currently available on FreeBSD and thus
some work is needed to actually support the new ZFS events (e.g. in
zfsd) to actually use this capability, this changeset is mostly a
diff reduction from upstream.
illumos/illumos-gate@1437283407
Illumos issues:
5997 FRU field not set during pool creation and never updated
https://www.illumos.org/issues/5997
In r286570 (MFV of r277426) an unprotected write to b_flags to
set the compression mode was introduced. This would open a race
window where data is partially decompressed, modified, checksummed
and written to the pool, resulting in pool corruption due to the
partial decompression.
Prevent this by reintroducing b_compress
illumos/illumos-gate@d4cd038c92
Illumos issues:
6214 zpools going south
https://www.illumos.org/issues/6214
Rewrite the ZFS prefetch code to detect only forward, sequential
streams.
The following kstats have been added:
kstat.zfs.misc.arcstats.sync_wait_for_async
How many sync reads have waited for async read
to complete. (less is better)
kstat.zfs.misc.arcstats.demand_hit_predictive_prefetch
How many demand read didn't have to wait for I/O
because of predictive prefetch. (more is better)
zfetch kstats have been similified to hits, misses, and max_streams,
with max_streams representing times when we were not able to create
new stream because we already have the maximum number of sequences
for a file.
The sysctl variable/loader tunable vfs.zfs.zfetch.block_cap have been
replaced by vfs.zfs.zfetch.max_distance, which controls maximum bytes
to prefetch per stream.
illumos/illumos-gate@cf6106c8a0
Illumos ZFS issues:
5987 zfs prefetch code needs work
https://www.illumos.org/issues/5987
since on amd64 the first argument to a function is generally not on the
stack.
Revert an old DTrace bug fix to some code that assumed that
sizeof(struct amd64_frame) == 16.
Reviewed by: jhb, kib
Sponsored by: EMC / Isilon Storage Division
Differential Revision: https://reviews.freebsd.org/D3255
5930 fasttrap_pid_enable() panics when prfind() fails in forking process
Reviewed by: Adam Leventhal <ahl@delphix.com>
Reviewed by: Gordon Ross <gordon.ross@nexenta.com>
Approved by: Richard Lowe <richlowe@richlowe.net>
Author: Bryan Cantrill <bryan@joyent.com>
illumos/illumos-gate@9df7e4e12e
This makes it possible to analyze the performance of the new ZFS
write throttle with dtrace
PR: 200316
Submitted by: Lacey Powers <lacey.leanne@gmail.com>
Reviewed by: avg, smh, delphij (no objection)
Approved by: bapt (mentor)
MFC after: 1 month
Sponsored by: ScaleEngine Inc.
Differential Revision: https://reviews.freebsd.org/D3472
r286951 by reinstating changes in r274628.
In l2arc_compress_buf(), we allocate a buffer to stash away the compressed
data in 'cdata', allocated of l2hdr->b_asize bytes.
We then ask zio_compress_data() to compress the buffer, b_l1hdr.b_tmp_cdata,
which is of l2hdr->b_asize bytes, and have the compressed size (or original
size, if compress didn't gain enough) stored in csize.
To pad the buffer to fit the optimal write size, we round up the compressed
size to L2 device's vdev_ashift.
Illumos code rounds up the size by at most SPA_MINBLOCKSIZE. Because we
know csize <= b_asize, and b_asize is integer multiple of SPA_MINBLOCKSIZE,
we are guaranteed that the rounded up csize would be <= b_asize. However,
this is not necessarily true when we round up to 1 << vdev_ashift, because
it could be larger than SPA_MINBLOCKSIZE.
So, in the worst case scenario, we are overwriting at most
(1 << vdev_ashift - SPA_MINBLOCKSIZE)
bytes of memory next to the compressed data buffer.
Andriy's original change in r274628 reorganized the code a little bit,
by moving the padding to after we determined that the compression was
beneficial. At which point, we would check rounded size against the
allocated buffer size, and the buffer overrun would not be possible.
as parent. In the case of a send or receive, the curproc would be the
userland application that issues the ioctl. This would trigger an assertion
failure introduced in Solaris compatibility shims in r196458 when kernel is
compiled with INVARIANTS.
Fix this by using p0 (proc0 or kernel) as the parent thread when creating
the kernel threads.
FreeBSD porting notes:
- only kernel-side changes are merged
- the new ioctl is not actually implemented yet
- thus, the goal is to synchronize DMU code
illumos/illumos-gate@2bcf0248e9https://www.illumos.org/issues/5692
we would like to expose the number of hole (sparse) blocks in a file.
this can be useful to for example if you want to fill in the holes with
some data; knowing the number of holes in advances allows you to report
progress on hole filling. We could use SEEK_HOLE to do that but it would
be O(n) where n is the number of holes present in the file.
Author: Max Grossman <max.grossman@delphix.com>
Reviewed by: Adam Leventhal <ahl@delphix.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Boris Protopopov <bprotopopov@hotmail.com>
Approved by: Richard Lowe <richlowe@richlowe.net>
The change that introduced the L2ARC compression support also introduced
a bug where the on-disk size of the selected buffers could end up larger
than the target size if the ashift is greater than 9. This was because
the buffer selection could did not take into account the fact that
on-disk size could be larger than the in-memory buffer size due to
the alignment requirements.
At the moment b_asize is a misnomer as it does not always represent the
allocated size: if a buffer is compressed, then the compressed size is
properly rounded (on FreeBSD), but if the compression fails or it is not
applied, then the original size is kept and it could be smaller than what
ashift requires.
For the same reasons arcstat_l2_asize and the reported used space
on the cache device could be smaller than the actual allocated size
if ashift > 9. That problem is not fixed by this change.
This change only ensures that l2ad_hand is not advanced by more
than target_sz. Otherwise we would overwrite active (unevicted)
L2ARC buffers. That problem is manifested as growing l2_cksum_bad
and l2_io_error counters.
This change also changes 'p' prefix to 'a' prefix in a few places
where variables represent allocated rather than physical size.
The resolved problem could also result in the reported allocated size
being greater than the cache device's capacity, because of the
overwritten buffers (more than one buffer claiming the same disk
space).
This change is already in ZFS-on-Linux:
zfsonlinux/zfs@ef56b0780c
PR: 198242
PR: 195746 (possibly related)
Reviewed by: mahrens (https://reviews.csiden.org/r/229/)
Tested by: gkontos@aicom.gr (most recently)
MFC after: 15 days
X-MFC note: patch does not apply as is at the moment
Relnotes: yes
Sponsored by: ClusterHQ
Differential Revision: https://reviews.freebsd.org/D2764
Reviewed by: noone (@FreeBSD.org)
The lock order reversal and a resulting deadlock were introduced
in r285021 / D2865. The problem is that zfs_register_callbacks() calls
dsl_prop_get_integer() that has to acquire spa_namespace_lock.
At the same time, spa_config_sync() is called with spa_namespace_lock
held and then it performs ZFS vnode operations that acquire
z_teardown_lock in the reader mode.
So, fix the problem by using dsl_prop_get_int_ds() instead of
dsl_prop_get_integer(). The former does not need to look up
the pool and the dataset by name.
Reported by: many
Reviewed by: delphij
Tested by: delphij, Jens Schweikhardt <schweikh@schweikhardt.net>
MFC after: 5 days
X-MFC with: r285021
- Add
nvlist_{add,get,take,move,exists,free}_{number,bool,string,nvlist,
descriptor} functions.
- Add support for (un)packing arrays.
- Add the nvl_array_next field to the nvlist structure.
If an array is added by the nvlist_{move,add}_nvlist_array function
this field will contains next element in the array.
- Add the nitems field to the nvpair and nvpair_header structure.
This field contains number of elements in the array.
- Add special flag (NV_FLAG_IN_ARRAY) which is set if nvlist is a part of
an array.
- Add special type (NV_TYPE_NVLIST_ARRAY_NEXT).This type is used only
on packing/unpacking.
- Add new API for traversing arrays (nvlist_get_array_next).
- Add the nvlist_get_pararr function which combines the
nvlist_get_array_next and nvlist_get_parent functions. If nvlist is in
the array it will return next element from array. If nvlist is last
element in array or it isn't in array it will return his
container (parent). This function should simplify traveling over nvlist.
- Add tests for new features.
- Add documentation for new functions.
- Add my copyright.
- Regenerate the sys/cddl/compat/opensolaris/sys/nvpair.h file.
PR: 191083
Reviewed by: allanjude (doc)
Approved by: pjd (mentor)
Reviewed by: Jason King <jason.brian.king@gmail.com>
Reviewed by: Josef Sipek <jeffpc@josefsipek.net>
Approved by: Richard Lowe <richlowe@richlowe.net>
Author: Bart Coddens <bart.coddens@gmail.com>
illumos/illumos-gate@fc98fea58e
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Paul Dagnelie <paul.dagnelie@delphix.com>
Reviewed by: Adam Leventhal <ahl@delphix.com>
Reviewed by: Alex Reece <alex@delphix.com>
Reviewed by: Richard Elling <richard.elling@richardelling.com>
Approved by: Garrett D'Amore <garrett@damore.org>
Author: Prakash Surya <prakash.surya@delphix.com>
illumos/illumos-gate@2fd872a734
As a way to make it more difficult to introduce bugs into the ARC, and to
make it easier to diagnose issues when bugs do creep in, it would be
beneficial to change the type of the arc_state_t's arcs_size field to be
a refcount_t instead of a uint64_t. This would allow us to make stricter
checks when incrementing and decrementing the value with debugging enabled,
but still fallback to simple, fast atomic operations when debugging is
disabled.
when adjusting MFU size.
illumos/illumos-gate@31c46cf23chttps://www.illumos.org/issues/6033
When we're looking for the list containing oldest buffer we never
actually look at the MFU lists even when we try to evict from MFU.
looks like a copy paste error, the fix is here:
Reviewed by: Saso Kiselkov <saso.kiselkov@nexenta.com>
Reviewed by: Xin Li <delphij@delphij.net>
Reviewed by: Prakash Surya <me@prakashsurya.com>
Approved by: Matthew Ahrens <mahrens@delphix.com>
Author: Alek Pinchuk <alek@nexenta.com>
Obtained from: illumos
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Richard Elling <richard.elling@richardelling.com>
Approved by: Dan McDonald <danmcd@omniti.com>
Author: Prakash Surya <prakash.surya@delphix.com>
illumos/illumos-gate@244781f10d
This patch attempts to reduce lock contention on the current arc_state_t
mutexes. These mutexes are used liberally to protect the number of LRU
lists within the ARC (e.g. ARC_mru, ARC_mfu, etc). The granularity at
which these locks are acquired has been shown to greatly affect the
performance of highly concurrent, cached workloads.
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Gordon Ross <gordon.w.ross@gmail.com>
Reviewed by: George Wilson <gwilson@zfsmail.com>
Approved by: Robert Mustacchi <rm@joyent.com>
illumos/illumos-gate@8f5190a540
6093 zfsctl_shares_lookup should only VN_RELE() on zfs_zget() success
Reviewed by: Gordon Ross <gwr@nexenta.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: George Wilson <george.wilson@delphix.com>
Approved by: Robert Mustacchi <rm@joyent.com>
Author: Dan McDonald <danmcd@omniti.com>
illumos/illumos-gate@0f92170f1e
Reviewed by: Toomas Soome <tsoome@me.com>
Reviewed by: George Wilson <george@delphix.com>
Reviewed by: Alex Reece <alex@delphix.com>
Approved by: Richard Lowe <richlowe@richlowe.net>
Author: Matthew Ahrens <mahrens@delphix.com>
illumos/illumos-gate@ca0cc3918a
A ZFS feature flags (large blocks) tracks its refcounts as the number of
datasets that have ever used the feature. Several features of this type
are planned to be added (new checksum functions). This code should be made
common infrastructure rather than duplicating the code for each feature.
5925 zfs receive -o origin=
Reviewed by: Prakash Surya <prakash.surya@delphix.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Author: Paul Dagnelie <pcd@delphix.com>
While running 'zfs recv' we noticed that every 128th 8K block required a
read. We were seeing that restore_write() was calling dmu_tx_hold_write()
and the indirect block was not cached. We should prefetch upcoming indirect
blocks to avoid having to go to disk and blocking the restore_write().
Allow an incremental send stream to be received as a clone, even if the
stream does not mark it as a clone.
illumos/illumos-gate@1d3f896f54https://www.illumos.org/issues/5981
When dmu_objset_find_dp gets called with a read lock held, it fans out
the work to the task queue. Each task in turn acquires its own read
lock before calling the callback. If during this process anyone tries
to a acquire a write lock, it will stall all read lock requests.Thus
the tasks will never finish, the read lock of the caller will never
get freed and the write lock never acquired. deadlock.
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Dan McDonald <danmcd@omniti.com>
Approved by: Robert Mustacchi <rm@joyent.com>
Author: Arne Jansen <jansen@webgods.de>
illumos/illumos-gate@12380e1e70https://www.illumos.org/issues/5269
When importing a pool (at boot or with zpool import) with many
filesystem, the process can take minutes. It doesn't matter whether
the pool has been exported cleanly or uncleanly. The problem is that
each dataset has its own log chain. On import, all datasets have to be
checked if there are logs to replay. The idea is to speed up this
process by paralellizing it.
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: George Wilson <george@delphix.com>
Reviewed by: Dan McDonald <danmcd@omniti.com>
Approved by: Dan McDonald <danmcd@omniti.com>
Author: Arne Jansen <jansen@webgods.de>
lzc_send_space when source is a bookmark
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Christopher Siden <christopher.siden@delphix.com>
Reviewed by: Steven Hartland <killing@multiplay.co.uk>
Reviewed by: Bayard Bell <buffer.g.overflow@gmail.com>
Approved by: Albert Lee <trisk@nexenta.com>
Author: Max Grossman <max.grossman@delphix.com>
illumos/illumos-gate@643da460c8
illumos/illumos-gate@70163ac57ehttps://www.illumos.org/issues/5695
In dmu_sync_ready(), a hole block pointer will have it's logical size
explicitly set as it's necessary for replay purposes. To "undo" this,
dmu_sync_done() will zero out any hole that it finds. This becomes a
problem when using the "hole_birth" feature, as this will also wipe out
any birth time that might have happened to be set on the hole.
...
As a fix, the logic to zero out a hole is only applied to old style
holes with a birth time of zero. Holes created with the "hole_birth"
feature enabled will have a non-zero birth time, and will be skipped
(thus preserving the ltime, type, and level information as well).
In addition, zdb was updated to also print the ltime, type, and level
information for these new style holes. Previously, only the logical
birth time would be printed.
Author: Prakash Surya <prakash.surya@delphix.com>
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: George Wilson <george@delphix.com>
Reviewed by: Christopher Siden <christopher.siden@delphix.com>
Reviewed by: Bayard Bell <buffer.g.overflow@gmail.com>
Approved by: Dan McDonald <danmcd@omniti.com>
5376 arc_kmem_reap_now() should not result in clearing arc_no_grow
Reviewed by: Christopher Siden <christopher.siden@delphix.com>
Reviewed by: George Wilson <george.wilson@delphix.com>
Reviewed by: Steven Hartland <killing@multiplay.co.uk>
Reviewed by: Richard Elling <richard.elling@richardelling.com>
Approved by: Dan McDonald <danmcd@omniti.com>
Author: Matthew Ahrens <mahrens@delphix.com>
illumos/illumos-gate@2ec99e3e98
