- mmp_concurrent_import: added test case to verify that concurrent
import correctness. The pool may only be imported once.
- mmp_exported_import: an activity check is now required for pools
which were cleanly exported if the system and pool hostids don't
match.
- mmp_inactive_import: an activity check is now required for any
pool which wasn't cleanly exported, even if the system and pool
hostids match.
- mmp_on_uberblocks: updated expected uberblocks to take in to account
the value MMP_INTERVAL_DEFAULT is set too.
- mmp_reset_interval: reduce the number of iterations from 10 to 3.
This is sufficient to verify functionality and significantly speeds
up the test.
- mmp_on_uberblocks: adjust the thresholds and increase the runtime
to avoid false positives observed in CI.
- Update tests to use 'zhack action idle' instead of ztest to improve
the reliability of the tests.
- Add additional log_note messages to test cases which have multiple
verification steps to make it clear which portion of a test failed
when reviewing the logs.
- Replace default_setup/cleanup_noexit calls with 'zpool create' and
'zpool destroy' calls to avoid additional unnecessary dataset
creation work.
- Update activity/noactivity check helper functions to use the
ZFS_LOAD_INFO_DEBUG information now available from 'zpool import'
to determine if this activity check ran and why. This is more
reliable in the CI than measuring the runtime.
- Removed all mmp tests from the zts-report.py exceptions list.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Olaf Faaland <faaland1@llnl.gov>
Reviewed-by: Akash B <akash-b@hpe.com>
As part of SPA_LOAD_IMPORT add an additional activity check to
detect simultaneous imports from different hosts. This check is
only required when the timing is such that there's no activity
for the the read-only tryimport check to detect. This extra
safety chceck operates as follows:
1. Repeats the following MMP check 10 times:
a. Write out an MMP uberblock with the best txg and a random
sequence id to all primary pool vdevs.
b. Verify a minimum number of good writes such that even if
the pool appears degraded on the remote host it will see
at least one of the updated MMP uberblocks.
c. Wait for the MMP interval this leaves a window for other
racing hosts to make similar modifications which can be
detected.
d. Call vdev_uberblock_load() to determine the best uberblock
to use, this should be the MMP uberblock just written.
e. Verify the txg and random sequeunce number match the MMP
uberblock written in 1a.
2. Restore the original MMP uberblocks. This allows the check
to be performed again if the pool fails to import for an
unrelated reason.
This change also includes some refactoring and minor improvements.
- Never try loading earlier txgs during import when the import
fails with EREMOTEIO or EINTER. These errors don't indicate
the txg is damaged but instead that its either in use on a
remote host or the import was interactively cancelled. No
rewind is also performed for EBADD which can result from a
stale trusted config when doing a verbatim import.
- Refactor the code for consistent logging of the multihost
activity check using spa_load_note() and console messages
indicating when the activity check was trigger and the result.
- Added MMP_*_MASK and MMP_SEQ_CLEAR() macros to allow easier
modification of the sequence number in an uberblock.
- Added ZFS_LOAD_INFO_DEBUG environment variable which can be
set to log to dump to stdout the spa_load_info nvlist returned
during import. This is used by the updated mmp test cases
to determine if an activity check was run and its result.
- Standardize the mmp messages similarly to make it easier to
find all the relevent mmp lines in the debug log.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Olaf Faaland <faaland1@llnl.gov>
Reviewed-by: Akash B <akash-b@hpe.com>
Tryimport adds a unique prefix to the pool name to avoid name
collisions. This makes it awkward to log user-friendly info
during a tryimport. Add a spa_load_name() function which can
be used to report the unmodified pool name.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Olaf Faaland <faaland1@llnl.gov>
Reviewed-by: Akash B <akash-b@hpe.com>
Move the "Starting import" log message in to the import block so
it's matched with the "Fiinshed importing" debug message.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Olaf Faaland <faaland1@llnl.gov>
Reviewed-by: Akash B <akash-b@hpe.com>
For a cleanly exported pools there exists a small window where
both systems may determine it's safe to import the pool and skip
the activity check. Only allow the check to be skipped when the
last imported hostid matches the systems hostid and the pool was
cleanly exported.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Olaf Faaland <faaland1@llnl.gov>
Reviewed-by: Akash B <akash-b@hpe.com>
They aren't used outside these very small blocks, and their initial
values are never used at all.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Sponsored-by: https://despairlabs.com/sponsor/Closes#17551
Free issue threads might block waiting for synchronous DDT, BRT or
GANG header reads. So unlike other taskqs using ZTI_SCALE to scale
with number of CPUs, here we also need some amount of threads to
potentially saturate pool reads. I am not sure we always want the
96 threads we had before ZTI_SCALE introduction at #11966 on small
systems, but lets make it at least 32.
While here, make free taskqs configurable, similar to read and
write ones.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Rob Norris <robn@despairlabs.com>
Signed-off-by: Alexander Motin <alexander.motin@TrueNAS.com>
Closes#17903
ZFS typically preserves proper LIFO ordering regarding map/unmap
operations that wrap the Linux kernel's kmap interfaces that
require such ordering, but one instance in abd_raidz_gen_iterate()
did not.
Similar issues have been fixed in the Linux kernel in the past,
see for instance CVE-2025-39899 for userfaultfd.
Reviewed-by: RageLtMan <rageltman@sempervictus>
Reviewed-by: Rob Norris <robn@despairlabs.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: bspengler-oss <94915855+bspengler-oss@users.noreply.github.com>
Closes#15668Closes#18030
The concurrent execution of feature_sync() can lead to a panic due
to an unprotected update of the feature refcount. Resolve this by
using the spa->spa_feat_stats_lock to synchronize the update of the
refcount.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Youzhong Yang <yyang@mathworks.com>
Closes#17184Closes#17632
This changes the basic search algorithm from a single search up and down
the tree to a full depth-first traversal to handle conditions where the
tree matches at a higher level but not a lower level.
Normally higher level blocks always point to matching blocks, but there
are cases where this does not happen:
1. Racing block pointer updates from dbuf_write_ready.
Before f664f1ee7f (#8946), both dbuf_write_ready and
dnode_next_offset held dn_struct_rwlock which protected against
pointer writes from concurrent syncs.
This no longer applies, so sync context can f.e. clear or fill all
L1->L0 BPs before the L2->L1 BP and higher BP's are updated.
dnode_free_range in particular can reach this case and skip over L1
blocks that need to be dirtied. Later, sync will panic in
free_children when trying to clear a non-dirty indirect block.
This case was found with ztest.
2. txg > 0, non-hole case. This is #11196.
Freeing blocks/dnodes breaks the assumption that a match at a higher
level implies a match at a lower level when filtering txg > 0.
Whenever some but not all L0 blocks are freed, the parent L1 block is
rewritten. Its updated L2->L1 BP reflects a newer birth txg.
Later when searching by txg, if the L1 block matches since the txg is
newer, it is possible that none of the remaining L1->L0 BPs match if
none have been updated.
The same behavior is possible with dnode search at L0.
This is reachable from dsl_destroy_head for synchronous freeing.
When this happens open context fails to free objects leaving sync
context stuck freeing potentially many objects.
This is also reachable from traverse_pool for extreme rewind where it
is theoretically possible that datasets not dirtied after txg are
skipped if the MOS has high enough indirection to trigger this case.
In both of these cases, without backtracking the search ends prematurely
as ESRCH result implies no more matches in the entire object.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Akash B <akash-b@hpe.com>
Signed-off-by: Robert Evans <evansr@google.com>
Closes#16025Closes#11196
Provide an interface to retrieve the lowest and highest minimum
allocation size for the normal allocation class. This can be used
by external consumers of the DMU to estimate potential wasted
capacity when setting the recordsize for an object.
The new "min_alloc" and "max_alloc" keys are added to the pool
configuration and used by default_volblocksize() to warn when
an ineffecient block size is requested. For older kmods which
don't yet include the new keys fallback to the previous logic.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Alexander Motin <alexander.motin@TrueNAS.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#17758
Systems with a large number of CPU cores (192+) may trigger the large
allocation warning in multilist_create() on Linux. Silence the warning
by converting the allocation to vmem_alloc().
On Linux this results in a call to kvalloc() which will alloc vmem
for large allocations and kmem for small allocations.
On FreeBSD both vmem_alloc and kmem_alloc internally use the same
allocator so there is no functional change.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Alexander Motin <alexander.motin@TrueNAS.com>
Reviewed-by: Rob Norris <robn@despairlabs.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#17616
All this machinery is there to try to understand when there an async
writeback waiting to complete because the intent log callbacks are still
outstanding, and force them with a timely zil_commit(). The next commit
fixes this properly, so there's no need for all this extra housekeeping.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <alexander.motin@TrueNAS.com>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#17584
This allows to change the meaning of priority differences in FreeBSD
without requiring code changes in ZFS.
This upstreams commit fd141584cf89d7d2 from FreeBSD src.
Sponsored-by: The FreeBSD Foundation
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Olivier Certner <olce@FreeBSD.org>
Closes#17489
It is not right, but there are few examples when TX is aborted
after being assigned in case of error. To handle it better on
production systems add extra cleanup steps.
While here, replace couple dmu_tx_abort() in simple cases.
Reviewed-by: Rob Norris <robn@despairlabs.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#17438
This allows to rewrite content of specified file(s) as-is without
modifications, but at a different location, compression, checksum,
dedup, copies and other parameter values. It is faster than read
plus write, since it does not require data copying to user-space.
It is also faster for sync=always datasets, since without data
modification it does not require ZIL writing. Also since it is
protected by normal range range locks, it can be done under any
other load. Also it does not affect file's modification time or
other properties.
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Rob Norris <robn@despairlabs.com>
Avoid calling dbuf_evict_one() from memory reclaim contexts (e.g. Linux
kswapd, FreeBSD pagedaemon). This prevents deadlock caused by reclaim
threads waiting for the dbuf hash lock in the call sequence:
dbuf_evict_one -> dbuf_destroy -> arc_buf_destroy
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <alexander.motin@TrueNAS.com>
Signed-off-by: Kaitlin Hoang <kthoang@amazon.com>
Closes#17561
Linux kernel shrinker in the context of null/root memcg does not scan
dentry and inode caches added by a task running in non-root memcg. For
ZFS this means that dnode cache routinely overflows, evicting valuable
meta/data and putting additional memory pressure on the system.
This patch restores zfs_prune_aliases as fallback when the kernel
shrinker does nothing, enabling zfs to actually free dnodes. Moreover,
it (indirectly) calls arc_evict when dnode_size > dnode_limit.
Reviewed-by: Rob Norris <robn@despairlabs.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Gionatan Danti <g.danti@assyoma.it>
Closes#17487Closes#17542
Loss of one indirect block of the meta dnode likely means loss of
the whole dataset. It is worse than one file that the man page
promises, and in my opinion is not much better than "none" mode.
This change restores redundancy of the meta-dnode indirect blocks,
while same time still corrects expectations in the man page.
Reviewed-by: Akash B <akash-b@hpe.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Rob Norris <robn@despairlabs.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#17339
As discussed in the comments of PR #17004, you can theoretically run
into a case where a gang child has more copies than the gang header,
which can lead to some odd accounting behavior (and even trip a
VERIFY). While the accounting code could be changed to handle this, it
fundamentally doesn't seem to make a lot of sense to allow this to
happen. If the data is supposed to have a certain level of reliability,
that isn't actually achieved unless the gang_copies property is set to
match it.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Paul Dagnelie <paul.dagnelie@klarasystems.com>
Closes#17484
The redundant_metadata setting in ZFS allows users to trade resilience
for performance and space savings. This applies to all data and metadata
blocks in zfs, with one exception: gang blocks. Gang blocks currently
just take the copies property of the IO being ganged and, if it's 1,
sets it to 2. This means that we always make at least two copies of a
gang header, which is good for resilience. However, if the users care
more about performance than resilience, their gang blocks will be even
more of a penalty than usual.
We add logic to calculate the number of gang headers copies directly,
and store it as a separate IO property. This is stored in the IO
properties and not calculated when we decide to gang because by that
point we may not have easy access to the relevant information about what
kind of block is being stored. We also check the redundant_metadata
property when doing so, and use that to decide whether to store an extra
copy of the gang headers, compared to the underlying blocks.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Signed-off-by: Paul Dagnelie <paul.dagnelie@klarasystems.com>
Co-authored-by: Paul Dagnelie <paul.dagnelie@klarasystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Sponsored-by: Klara, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <alexander.motin@TrueNAS.com>
Signed-off-by: Igor Ostapenko <igor.ostapenko@klarasystems.com>
Closes#17581
While booting, only the needed 256KiB benchmarks are done now.
The delay for checking all checksums occurs when requested via:
- Linux: cat /proc/spl/kstat/zfs/chksum_bench
- FreeBSD: sysctl kstat.zfs.misc.chksum_bench
Reported by: Lahiru Gunathilake <gunathilakebllg@gmail.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Tino Reichardt <milky-zfs@mcmilk.de>
Co-authored-by: Colin Percival <cperciva@tarsnap.com>
Closes#17563Closes#17560
When we're passivating a metaslab group we start by passivating the
metaslabs that have been activated for each of the allocators. To do
that, we need to provide a weight. However, currently this erroneously
always uses a segment-based weight, even if segment-based weighting is
disabled.
Use the normal weight function, which will decide which type of weight
to use.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <alexander.motin@TrueNAS.com>
Signed-off-by: Paul Dagnelie <paul.dagnelie@klarasystems.com>
Closes#17566
Update the default FICLONE and FICLONERANGE ioctl behavior to wait
on dirty blocks. While this does remove some control from the
application, in practice ZFS is better positioned to the optimial
thing and immediately force a TXG sync.
Reviewed-by: Rob Norris <robn@despairlabs.com>
Reviewed-by: Alexander Motin <alexander.motin@TrueNAS.com>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#17455
Under parallel workloads ZIL may delay writes of open LWBs that
are not full enough. On suspend we do not expect anything new to
appear since zil_get_commit_list() will not let it pass, only
returning TXG number to wait for. But I suspect that waiting for
the TXG commit without having the last LWB issued may not wait for
its completion, resulting in panic described in #17509.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Rob Norris <rob.norris@klarasystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#17521
Currently, after a failed allocation, the metaslab code recalculates the
weight for a metaslab. However, for space-based metaslabs, it uses the
maximum free segment size instead of the normal weighting
algorithm. This is presumably because the normal metaslab weight is
(roughly) intended to estimate the size of the largest free segment, but
it doesn't do that reliably at most fragmentation levels. This means
that recalculated metaslabs are forced to a weight that isn't really
using the same units as the rest of them, resulting in undesirable
behaviors. We switch this to use the normal space-weighting function.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Paul Dagnelie <paul.dagnelie@klarasystems.com>
Sponsored-by: Wasabi Technology, Inc.
Sponsored-by: Klara, Inc.
Closes#17531
There are still a variety of bugs involving the vdev_nonrot property
that will cause problems if you try to run the test suite with
segment-based weighting disabled, and with other things in the weighting
code. Parents' nonrot property need to be updated when children are
added. When vdevs are expanded and more metaslabs are added, the weights
have to be recalculated (since the number of metaslabs is an input to
the lba bias function). When opening, faulted or unopenable children
should not be considered for whether a vdev is nonrot or not (since the
nonrot property is determined during a successful open, this can cause
false negatives). And draid spares need to have the nonrot property set
correctly.
Sponsored-by: Eshtek, creators of HexOS
Sponsored-by: Klara, Inc.
Reviewed-by: Allan Jude <allan@klarasystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Rob Norris <robn@despairlabs.com>
Signed-off-by: Paul Dagnelie <paul.dagnelie@klarasystems.com>
Closes#17469
dbuf_verify(): Don't need the lock, since we only compare pointers.
dbuf_findbp(): Don't need the lock, since aside of unneeded assert
we only produce the pointer, but don't de-reference it.
dnode_next_offset_level(): When working on top level indirection
should lock dnode buffer's db_rwlock, since it is our parent. If
dnode has no buffer, then it is meta-dnode or one of quotas and we
should lock the dataset's ds_bp_rwlock instead.
Reviewed-by: Alan Somers <asomers@gmail.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#17441
The `scn_min_txg` can now be used not only with resilver. Instead
of checking `scn_min_txg` to determine whether it’s a resilver or
a scrub, simply check which function is defined. Thanks to this
change, a scrub_finish event is generated when performing a scrub
from the saved txg.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Mariusz Zaborski <mariusz.zaborski@klarasystems.com>
Closes#17432
- Don't drop L2ARC header if we have more buffers in this header.
Since we leave them the header, leave them the L2ARC header also.
Honestly we are not required to drop it even if there are no other
buffers, but then we'd need to allocate it a separate header, which
we might drop soon if the old block is really deleted. Multiple
buffers in a header likely mean active snapshots or dedup, so we
know that the block in L2ARC will remain valid. It might be rare,
but why not?
- Remove some impossible assertions and conditions.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#17126
This PR condenses the FDT dedup log syncing into a single sync
pass. This reduces the overhead of modifying indirect blocks for the
dedup table multiple times per txg. In addition, changes were made to
the formula for how much to sync per txg. We now also consider the
backlog we have to clear, to prevent it from growing too large, or
remaining large on an idle system.
Sponsored-by: Klara, Inc.
Sponsored-by: iXsystems, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Authored-by: Don Brady <don.brady@klarasystems.com>
Authored-by: Paul Dagnelie <paul.dagnelie@klarasystems.com>
Signed-off-by: Paul Dagnelie <paul.dagnelie@klarasystems.com>
Closes#17038
ZIL introduced dependencies between its write ZIOs to permit flush
defer, when we flush vdev caches only once all the write ZIOs has
completed. But it was recently spotted that it serializes not only
ZIO completions handling, but also their ready stage. It means ZIO
pipeline can't calculate checksums for the following ZIOs until all
the previous are checksumed, even though it is not required. On a
systems where memory throughput of a single CPU core is limited,
it creates single-core CPU bottleneck, which is difficult to see
due to ZIO pipeline design with many taskqueue threads.
While it would be great to bypass the ready stage waits, it would
require changes to ZIO code, and I haven't found a clean way to do
it. But I've noticed that we don't need any dependency between
the write ZIOs if the previous one has some waiters, which means
it won't defer any flushes and work as a barrier for the earlier
ones.
Bypassing it won't help large single-thread writes, since all the
write ZIOs except the last in that case won't have waiters, and
so will be dependent. But in that case the ZIO processing might
not be a bottleneck, since there will be only one thread populating
the write buffers, that will likely be the bottleneck.
But bypassing the ZIO dependency on multi-threaded write workloads
really allows them to scale beyond the checksuming throughput of
one CPU core.
My tests with writing 12 files on a same dataset on a pool with
4 striped NVMes as SLOGs from 12 threads with 1MB blocks on a
system with Xeon Silver 4114 CPU show total throughput increase
from 4.3GB/s to 8.5GB/s, increasing the SLOGs busy from ~30% to
~70%.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Rob Norris <robn@despairlabs.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#17458
It makes no sense to limit read size below the block size, since
DMU will any way consume resources for the whole block, while the
current zfs_vnops_read_chunk_size is only 1MB, which is smaller
that maximum block size of 16MB. Plus in case of misaligned
Uncached I/O the buffer may get evicted between the chunks,
requiring repeating I/Os.
On 64-bit platforms increase zfs_vnops_read_chunk_size to 32MB.
It allows to less depend on speculative prefetcher if application
requests specific size, first not waiting for prefetcher to start
and later not prefetching more than needed.
Also while there, we don't need to align reads to the chunk size,
but only to a block size, which is smaller and so more forgiving.
My profiles show ~4% of CPU time saving when reading 16MB blocks.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed by: Igor Kozhukhov <ikozhukhov@gmail.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#17415
It's been many years, we can probably do without.
Sponsored-by: https://despairlabs.com/sponsor/
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Pavel Snajdr <snajpa@snajpa.net>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Closes#17376
On systems with enormous amounts of memory, the single arc_evict thread
can become a bottleneck if reads and writes are stuck behind it, waiting
for old data to be evicted before new data can take its place.
This commit adds support for evicting from multiple ARC lists in
parallel, by farming the evict work out to some number of threads and
then accumulating their results.
A new tuneable, zfs_arc_evict_threads, sets the number of threads. By
default, it will scale based on the number of CPUs.
Sponsored-by: Expensify, Inc.
Sponsored-by: Klara, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Youzhong Yang <youzhong@gmail.com>
Signed-off-by: Allan Jude <allan@klarasystems.com>
Signed-off-by: Mateusz Piotrowski <mateusz.piotrowski@klarasystems.com>
Signed-off-by: Alexander Stetsenko <alex.stetsenko@klarasystems.com>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Co-authored-by: Rob Norris <rob.norris@klarasystems.com>
Co-authored-by: Mateusz Piotrowski <mateusz.piotrowski@klarasystems.com>
Co-authored-by: Alexander Stetsenko <alex.stetsenko@klarasystems.com>
Closes#16486
This also includes removing L2 vdevs asynchronously.
This commit also guarantees that spa_load_guid is unique.
The zpool reguid feature introduced the spa_load_guid, which is a
transient value used for runtime identification purposes in the ARC.
This value is not the same as the spa's persistent pool guid.
However, the value is seeded from spa_generate_load_guid() which
does not check for uniqueness against the spa_load_guid from other
pools. Although extremely rare, you can end up with two different
pools sharing the same spa_load_guid value! So we guarantee that
the value is always unique and additionally not still in use by an
async arc flush task.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Allan Jude <allan@klarasystems.com>
Signed-off-by: Don Brady <don.brady@klarasystems.com>
Closes#16215
If a write is split across mutliple itxs, we only want the callback on
the last one, otherwise it will be called for every itx associated with
this single write, which makes it very hard to know what to clean up.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Mark Johnston <markj@FreeBSD.org>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#17445
Previous dmu_tx_count_clone() was broken, stating that cloning is
similar to free. While they might be from some points, cloning
is not net-free. It will likely consume space and memory, and
unlike free it will do it no matter whether the destination has
the blocks or not (usually not, so previous code did nothing).
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#17431
Looking on txg_wait_synced(, 0) I've noticed that it always syncs
5 TXGs: 3 TXG_CONCURRENT_STATES + 2 TXG_DEFER_SIZE. But in case
of dmu_offset_next() we do not care about deferred frees. And even
concurrent TXGs we might not need sync all 3 if the dnode was not
dirtied in last few TXGs.
This patch makes dmu_offset_next() to sync one TXG at a time until
the dnode is clean, but no more than 3 TXG_CONCURRENT_STATES times.
My tests with random simultaneous writes and seeks over many files
on HDD pool show 7-14% performance increase.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Rob Norris <robn@despairlabs.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#17434
It is not right, but there are few examples when TX is aborted
after being assigned in case of error. To handle it better on
production systems add extra cleanup steps.
While here, replace couple dmu_tx_abort() in simple cases.
Reviewed-by: Rob Norris <robn@despairlabs.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#17438
Having high-refcount dedup entries for zero blocks is inefficient
when they could be recorded as a holes instead. Normally, zero
compression is not done if compression is disabled to not confuse
naive benchmarks. But with dedup enabled, it is expected that the
write will be skipped anyway, so we are just optimizing the way it
is skipped.
Reviewed-by: Rob Norris <robn@despairlabs.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#17435
I've noticed that after some dedup tests system reboot ends up in
assertion about ms_defer tree not free. It seems to be caused by
DDT flushing still freeing some blocks while ZFS trying to reach
a final steady state due to spa_final_txg, while being set by
spa_export_common() on pool export, is not set when spa_unload()
is called by spa_evict_all() on system reboot/shutdown. Setting
spa_final_txg in spa_unload() fixes this issue.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Rob Norris <rob.norris@klarasystems.com>
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#17395
This patch fixes a race where vdev_remove_wanted may be set after probe
initiation, which could otherwise trigger redundant fault and removal.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Ameer Hamza <ahamza@ixsystems.com>
Closes#17400
Three occurences with an 'e', and all of them mine. Maybe it's an
British thing?
Sponsored-by: https://despairlabs.com/sponsor/
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Pavel Snajdr <snajpa@snajpa.net>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Closes#17377
If a variable is only available in the kernel, then the tunable should
also only be available there.
This matters very little so long as we don't have userspace tunables,
but its still good hygeine.
Sponsored-by: https://despairlabs.com/sponsor/
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Pavel Snajdr <snajpa@snajpa.net>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Closes#17377
It actually doesn't matter if it's not initialised when we first query
the current value; it just returns empty-string. A crash is quite
obnoxious even if it is a rare case.
Sponsored-by: https://despairlabs.com/sponsor/
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Pavel Snajdr <snajpa@snajpa.net>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Closes#17377
Likely it's only int64 for comparison with ssize_t, which is signed.
However, it would make no sense for it to be less than 0 or greater than
4G, so making it a regular uint will make it safe for comparison and
remove the only S64 tunable in core.
Sponsored-by: https://despairlabs.com/sponsor/
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Pavel Snajdr <snajpa@snajpa.net>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Closes#17377
