Unlike regular receive, raw receive require destination to have the
same block structure as the source. In case of dnode reclaim this
triggers two special cases, requiring special handling:
- If dn_nlevels == 1, we can change the ibs, but dnode_set_blksz()
should not dirty the data buffer if block size does not change, or
durign receive dbuf_dirty_lightweight() will trigger assertion.
- If dn_nlevels > 1, we just can't change the ibs, dnode_set_blksz()
would fail and receive_object would trigger assertion, so we should
destroy and recreate the dnode from scratch.
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15039
If this is not done, and the pool has an ashift other than the default
(at the moment 9) then the following happens:
1) vdev_alloc() assigns the ashift of the pool to L2ARC device, but
upon export it is not stored anywhere
2) at the first import, vdev_open() sees an vdev_ashift() of 0 and
assigns the logical_ashift, which is 9
3) reading the contents of L2ARC, including the header fails
4) L2ARC buffers are not restored in ARC.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#14313Closes#14963
While commit bcd5321 adjusts the write size based on the size of the log
block, this happens after comparing the unadjusted write size to the
evicted (target) size.
In this case l2ad_hand will exceed l2ad_evict and violate an assertion
at the end of l2arc_write_buffers().
Fix this by adding the max log block size to the allocated size of the
buffer to be committed before comparing the result to the target
size.
Also reset the l2arc_trim_ahead ZFS module variable when the adjusted
write size exceeds the size of the L2ARC device.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#14936Closes#14954
l2arc_write_size() should return the write size after adjusting for trim
and overhead of the L2ARC log blocks. Also take into account the
allocated size of log blocks when deciding when to stop writing buffers
to L2ARC.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#14939
l2arc_evict() performs the adjustment of the size of buffers to be
written on L2ARC unnecessarily. l2arc_write_size() is called right
before l2arc_evict() and performs those adjustments.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#14828
482da24e2 missed arc_buf_destroy() calls on log parse errors, possibly
leaking up to 128KB of memory per dataset during ZIL replay.
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
- Clang 15 doesn't support `-fno-ipa-sra` anymore. Do a separate
check for `-fno-ipa-sra` support by $KERNEL_CC.
- Don't enable `-mgeneral-regs-only` for certain module files.
Fix#13260
- Scope `GCC diagnostic ignored` statements to GCC only. Clang
doesn't need them to compile the code.
Porting notes:
- Moved the stanzas removing -mgeneral-regs-only to Makefile.in
since they wouldn't readily work in Kbuild.in and that did.
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: szubersk <szuberskidamian@gmail.com>
Closes#13260Closes#14150Closes#14624
Ported-by: Rich Ercolani <rincebrain@gmail.com
Signed-off-by: Rich Ercolani <rincebrain@gmail.com>
As of the 4.13 kernel filemap_range_has_page() can be used to
check if there is a page mapped in a given file range. When
available this interface should be used which eliminates the
need for the zp->z_is_mapped boolean.
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#14493
Page writebacks with WB_SYNC_NONE can take several seconds to complete
since they wait for the transaction group to close before being
committed. This is usually not a problem since the caller does not
need to wait. However, if we're simultaneously doing a writeback
with WB_SYNC_ALL (e.g via msync), the latter can block for several
seconds (up to zfs_txg_timeout) due to the active WB_SYNC_NONE
writeback since it needs to wait for the transaction to complete
and the PG_writeback bit to be cleared.
This commit deals with 2 cases:
- No page writeback is active. A WB_SYNC_ALL page writeback starts
and even completes. But when it's about to check if the PG_writeback
bit has been cleared, another writeback with WB_SYNC_NONE starts.
The sync page writeback ends up waiting for the non-sync page
writeback to complete.
- A page writeback with WB_SYNC_NONE is already active when a
WB_SYNC_ALL writeback starts. The WB_SYNC_ALL writeback ends up
waiting for the WB_SYNC_NONE writeback.
The fix works by carefully keeping track of active sync/non-sync
writebacks and committing when beneficial.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Shaan Nobee <sniper111@gmail.com>
Closes#12662Closes#12790
There seems to be no reason for ZIL blocks to be limited by 128KB
other than replay code is written in such a way. This change does
not increase the limit yet, just removes the artificial limitation.
Avoided extra memcpy() may save us a second during replay.
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
At least for RAIDZ zio_shrink() does not reduce zio size, but reduced
wsz in that case likely results in writing uninitialized memory.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#14853
We only need to know if ZIO has any parent there. We do not care if
it has more than one, but use of zio_unique_parent() == NULL asserts
that.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#14823
- There is no reason to assert that added gang is not empty. It
may be weird to add an empty gang, but it is legal.
- When moving chain list from the added gang clear its size, or it
will trigger assertion in abd_verify() when that gang is freed.
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#14816
TX_COMMIT has no on-disk representation and does not produce any more
dirty data. It should not wait for anything, and even just skipping
the checks if not waiting gives improvement noticeable in profiler.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Prakash Surya <prakash.surya@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#14798
Gang ABDs without childred are legal, and they do have zero size.
For other ABD types zero size doesn't have much sense and likely
not working correctly now.
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#14795
In the case of a regular compilation, the compiler
raises a warning for a dsl_deadlist_merge function, that
the stack size is to large. In debug build this can
generate an error.
Move large structures to heap.
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Mariusz Zaborski <mariusz.zaborski@klarasystems.com>
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Closes#14524
A NULL pointer will occur when doing a 'zfs send -S' on a dataset that
is still being received. The problem is that the new 'send' will
rightfully fail to own the datasets (i.e. dsl_dataset_own_force() will
fail), but then dmu_send() will still do the dsl_dataset_disown().
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Luís Henriques <henrix@camandro.org>
Closes#14903Closes#14890
zpool initialize functions well for touching every free byte...once.
But if we want to do it again, we're currently out of luck.
So let's add zpool initialize -u to clear it.
Co-authored-by: Rich Ercolani <rincebrain@gmail.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rich Ercolani <rincebrain@gmail.com>
Closes#12451Closes#14873
Before allowing the ZED to mark a vdev as REMOVED due to a
hotplug event confirm that it is non-responsive with probe.
Any device which can be successfully probed should be left
ONLINE to prevent a healthy pool from being incorrectly
SUSPENDED. This may occur for at least the following two
scenarios.
1) Drive expansion (zpool online -e) in VMware environments.
If, during the partition resize operation, a partition is
removed and re-created then udev will send a removed event.
2) Re-scanning the namespaces of an NVMe device (nvme ns-rescan)
may result in a udev remove and add event being delivered.
Finally, update the ZED to only kick in a spare when the
removal was successful.
Reviewed-by: Ameer Hamza <ahamza@ixsystems.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #14859Closes#14861
For draid vdevs it was possible to initiate both the
sequential and healing resilver at same time.
This fixes the following two scenarios.
1) There's a window where a sequential rebuild can
be started via ZED even if a healing resilver has been
scheduled.
- This is fixed by adding additional check in
spa_vdev_attach() for any scheduled resilver and return
appropriate error code when a resilver is already in
progress.
2) It was possible for zpool clear to start a healing
resilver when it wasn't needed at all. This occurs because
during a vdev_open() the device is presumed to be healthy not
until the device is validated by vdev_validate() and it's set
unavailable. However, by this point an async resilver will
have already been requested if the DTL isn't empty.
- This is fixed by cancelling the SPA_ASYNC_RESILVER
request immediately at the end of vdev_reopen() when a resilver
is unneeded.
Finally, added a testcase in ZTS for verification.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Dipak Ghosh <dipak.ghosh@hpe.com>
Signed-off-by: Akash B <akash-b@hpe.com>
Closes#14881Closes#14892
Provides an interface which callers can use to declare a write when
the exact starting offset in not yet known. Since the full range
being updated is not available only the first L0 block at the
provided offset will be prefetched.
Reviewed-by: Olaf Faaland <faaland1@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#14819
If we receive a DRR_FREEOBJECTS as the first entry in an object range,
this might end up producing a hole if the freed objects were the
only existing objects in the block.
If the txg starts syncing before we've processed any following
DRR_OBJECT records, this leads to a possible race where the backing
arc_buf_t gets its psize set to 0 in the arc_write_ready() callback
while still being referenced from a dirty record in the open txg.
To prevent this, we insert a txg_wait_synced call if the first
record in the range was a DRR_FREEOBJECTS that actually
resulted in one or more freed objects.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: David Hedberg <david.hedberg@findity.com>
Sponsored by: Findity AB
Closes#11893Closes#14358
spa_import() relies on a pool config fetched by spa_try_import() for
spare/cache devices. Import flags are not passed to spa_tryimport(),
which makes it return early due to a missing log device and missing
retrieving the cache device and spare eventually. Passing
ZFS_IMPORT_MISSING_LOG to spa_tryimport() makes it fetch the correct
configuration regardless of the missing log device.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ameer Hamza <ahamza@ixsystems.com>
Closes#14794
At our site we have seen cases when multi-modifier protection is enabled
(multihost=on) on our pool and the pool gets suspended due to a single
disk that is failing and responding very slowly. Our pools have 90 disks
in them and we expect disks to fail. The current version of MMP requires
that we wait for other writers before moving on. When a disk is
responding very slowly, we observed that waiting here was bad enough to
cause the pool to suspend. This change allows the MMP thread to bypass
waiting for other threads and reduces the chances the pool gets
suspended.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Herb Wartens <hawartens@gmail.com>
Closes#14659
When testing distributed rebuild performance with more capable
hardware it was observed than increasing the zfs_rebuild_vdev_limit
to 64M reduced the rebuild time by 17%. Beyond 64MB there was
some improvement (~2%) but it was not significant when weighed
against the increased memory usage. Memory usage is capped at 1/4
of arc_c_max.
Additionally, vr_bytes_inflight_max has been moved so it's updated
per-metaslab to allow the size to be adjust while a rebuild is
running.
Reviewed-by: Akash B <akash-b@hpe.com>
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#14428
For HDD based pools the default zfs_scan_vdev_limit of 4M
per-vdev can significantly limit the maximum scrub performance.
Increasing the default to 16M can double the scrub speed from
80 MB/s per disk to 160 MB/s per disk.
This does increase the memory footprint during scrub/resilver
but given the performance win this is a reasonable trade off.
Memory usage is capped at 1/4 of arc_c_max. Note that number
of outstanding I/Os has not changed and is still limited by
zfs_vdev_scrub_max_active.
Reviewed-by: Akash B <akash-b@hpe.com>
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#14428
When resilvering the estimated time remaining is calculated using
the average issue rate over the current pass. Where the current
pass starts when a scan was started, or restarted, if the pool
was exported/imported.
For dRAID pools in particular this can result in wildly optimistic
estimates since the issue rate will be very high while scanning
when non-degraded regions of the pool are scanned. Once repair
I/O starts being issued performance drops to a realistic number
but the estimated performance is still significantly skewed.
To address this we redefine a pass such that it starts after a
scanning phase completes so the issue rate is more reflective of
recent performance. Additionally, the zfs_scan_report_txgs
module option can be set to reset the pass statistics more often.
Reviewed-by: Akash B <akash-b@hpe.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#14410
Spare vdev should detach from the pool when a disk is reinserted.
However, spare detachment depends on the completion of resilvering,
and if resilver does not schedule, the spare vdev keeps attached to
the pool until the next resilvering. When a zfs pool contains
several disks (25+ mirror), resilvering does not always happen when
a disk is reinserted. In this patch, spare vdev is manually detached
from the pool when resilvering does not occur and it has been tested
on both Linux and FreeBSD.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Ameer Hamza <ahamza@ixsystems.com>
Closes#14722
This reverts commit 4b3133e671.
Users identified this commit as a possible source of data
corruption:
https://github.com/openzfs/zfs/issues/14753
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tony Hutter <hutter2@llnl.gov>
Issue #14753Closes#14761
Holding the zp->z_rangelock as a RL_READER over the range
0-UINT64_MAX is sufficient to prevent the dnode from being
re-dirtied by concurrent writers. To avoid potentially
looping multiple times for external caller which do not
take the rangelock holes are not reported after the first
sync. While not optimal this is always functionally correct.
This change adds the missing rangelock calls on FreeBSD to
zvol_cdev_ioctl().
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#14512Closes#14641
zfsd fetches new pool configuration through ZFS_IOC_POOL_STATS but
it does not get updated nvlist configuration for spare vdev since
the configuration is read by spa_spares->sav_config. In this commit,
updating the vdev state for spare vdev that is consumed by zfsd on
spare disk hotplug.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ameer Hamza <ahamza@ixsystems.com>
Closes#14653
This commit supports for spare vdev hotplug. The
spare vdev associated with all the pools will be
marked as "Removed" when the drive is physically
detached and will become "Available" when the
drive is reattached. Currently, the spare vdev
status does not change on the drive removal and
the same is the case with reattachment.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ameer Hamza <ahamza@ixsystems.com>
Closes#14295
In order for zed to process the removal event correctly,
udev change event needs to be posted to sync the blkid
information. spa_create() and spa_config_update() posts
the event already through spa_write_cachefile(). Doing
the same for spa_vdev_attach() that handles the case
for vdev attachment and replacement.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Ameer Hamza <ahamza@ixsystems.com>
Closes#14172
ZED does not take any action for disk removal events if there is no
spare VDEV available. Added zpool_vdev_remove_wanted() in libzfs
and vdev_remove_wanted() in vdev.c to remove the VDEV through ZED
on removal event. This means that if you are running zed and
remove a disk, it will be propertly marked as REMOVED.
Signed-off-by: Ameer Hamza <ahamza@ixsystems.com>
Debugging reported NULL de-reference panic in dnode_hold_impl() I found
that for certain types of errors arc_read() may only return error code,
but not properly report it via done and pio arguments. Lack of done
calls may result in reference and/or memory leaks in higher level code.
Lack of error reporting via pio may result in unnoticed errors there.
For example, dbuf_read(), where dbuf_read_impl() ignores arc_read()
return, relies completely on the pio mechanism and missed the errors.
This patch makes arc_read() to always call done callback and always
propagate errors to parent zio, if either is provided.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
`lseek(SEEK_DATA | SEEK_HOLE)` are only accurate when the on-disk blocks
reflect all writes, i.e. when there are no dirty data blocks. To ensure
this, if the target dnode is dirty, they wait for the open txg to be
synced, so we can call them "stabilizing operations". If they cause
txg_wait_synced often, it can be detrimental to performance.
Typically, a group of files are all modified, and then SEEK_DATA/HOLE
are performed on them. In this case, the first SEEK does a
txg_wait_synced(), and subsequent SEEKs don't need to wait, so
performance is good.
However, if a workload involves an interleaved metadata modification,
the subsequent SEEK may do a txg_wait_synced() unnecessarily. For
example, if we do a `read()` syscall to each file before we do its SEEK.
This applies even with `relatime=on`, when the `read()` is the first
read after the last write. The txg_wait_synced() is unnecessary because
the SEEK operations only care that the structure of the tree of indirect
and data blocks is up to date on disk. They don't care about metadata
like the contents of the bonus or spill blocks. (They also don't care
if an existing data block is modified, but this would be more involved
to filter out.)
This commit changes the behavior of SEEK_DATA/HOLE operations such that
they do not call txg_wait_synced() if there is only a pending change to
the bonus or spill block.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#13368
Issue #14594
Issue #14512
Issue #14009
The default_bs and default_ibs tunables control the default block size
and indirect block size.
So far, default_bs and default_ibs were tunable only on FreeBSD, e.g.,
sysctl vfs.zfs.default_ibs
Remove the FreeBSD-specific sysctl code and expose default_bs and
default_ibs as tunables on both Linux and FreeBSD using
ZFS_MODULE_PARAM.
One of the use cases for changing the values of those tunables is to
lower the indirect block size, which may improve performance of large
directories (as discussed during the OpenZFS Leadership Meeting
on 2022-08-16).
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Signed-off-by: Mateusz Piotrowski <mateusz.piotrowski@klarasystems.com>
Sponsored-by: Wasabi Technology, Inc.
Closes#14293
dc5c8006f6 was recently merged to prefetch
up to 128 deadlists. Unfortunately, a loop was missing an increment,
such that it will prefetch all deadlists. The performance properties of
that patch probably should be re-evaluated.
This was caught by CodeQL's cpp/constant-comparison check in an
experimental branch where I am testing the security-and-extended
queries. It complained about the `i < 128` part of the loop condition
always evaluating to the same thing. The standard CodeQL configuration
we use missed this because it does not include that check.
Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14573
by placing the most common use case (no special vdevs) first and avoid
allocating new variables.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#14494Closes#14563
With some pathological access patterns it is possible to make ZFS
accumulate almost unlimited amount of speculative prefetch ZIOs.
Combined with linear ABD allocations in RAIDZ code, it appears to
be possible to exhaust system KVA, triggering kernel panic.
Address this by introducing a system-wide counter of active prefetch
requests and blocking prefetch distance doubling per stream hits if
the number of active requests is higher that ~6% of ARC size.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#14516
During snapshot deletion ZFS may issue several reads for each deadlist
to merge them into next snapshot's or pool's bpobj. Number of the dead
lists increases with number of snapshots. On HDD pools it may take
significant time during which sync thread is blocked.
This patch introduces prescient prefetch of required blocks for up to
128 deadlists ahead. Tests show reduction of time required to delete
dataset with 720 snapshots with randomly overwritten file on wide HDD
pool from 75-85 to 22-28 seconds.
Reviewed-by: Allan Jude <allan@klarasystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Issue #14276Closes#14402
Despite all optimizations, tests on actual hardware show that FreeBSD
kernel can't sleep for less then ~2us. Similar tests on Linux show
~50us delay at least from nanosleep() (haven't tested inside kernel).
It means that on very fast log device ZIL may not be able to satisfy
zfs_commit_timeout_pct block commit timeout, increasing log latency
more than desired.
Handle that by introduction of zil_min_commit_timeout parameter,
specifying minimal timeout value where additional delays to aggregate
writes may be skipped. Also skip delays if the LWB is more than 7/8
full, that often happens if I/O sizes are constant and match one of
LWB sizes. Both things are applied only if there were no already
outstanding log blocks, that may indicate single-threaded workload,
that by definition can not benefit from the commit delays.
While there, add short time moving average to zl_last_lwb_latency to
make it more stable.
Tests of single-threaded 4KB writes to NVDIMM SLOG on FreeBSD show IOPS
increase by 9% instead of expected 5%. For zfs_commit_timeout_pct of
1 there IOPS increase by 5.5% instead of expected 1%.
Reviewed-by: Allan Jude <allan@klarasystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Prakash Surya <prakash.surya@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#14418
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#14199
I've noticed that some of those counters are used in hot paths like
dnode_hold_impl(), and results of this change is visible in profiler.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#14198
atomic_dec_32() should be a bit lighter than atomic_dec_32_nv().
Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#14200
Microzap on-disk format does not include a hash tree, expecting one to
be built in RAM during mzap_open(). The built tree is linked to DMU
user buffer, freed when original DMU buffer is dropped from cache. I've
found that workloads accessing many large directories and having active
eviction from DMU cache spend significant amount of time building and
then destroying the trees. I've also found that for each 64 byte mzap
element additional 64 byte tree element is allocated, that is a waste
of memory and CPU caches.
Improve memory efficiency of the hash tree by switching from AVL-tree
to B-tree. It allows to save 24 bytes per element just on pointers.
Save 32 bits on mze_hash by storing only upper 32 bits since lower 32
bits are always zero for microzaps. Save 16 bits on mze_chunkid, since
microzap can never have so many elements. Respectively with the 16 bits
there can be no more than 16 bits of collision differentiators. As
result, struct mzap_ent now drops from 48 (rounded to 64) to 8 bytes.
Tune B-trees for small data. Reduce BTREE_CORE_ELEMS from 128 to 126
to allow struct zfs_btree_core in case of 8 byte elements to pack into
2KB instead of 4KB. Aside of the microzaps it should also help 32bit
range trees. Allow custom B-tree leaf size to reduce memmove() time.
Split zap_name_alloc() into zap_name_alloc() and zap_name_init_str().
It allows to not waste time allocating/freeing memory when processing
multiple names in a loop during mzap_open().
Together on a pool with 10K directories of 1800 files each and DMU
cache limited to 128MB this reduces time of `find . -name zzz` by 41%
from 7.63s to 4.47s, and saves additional ~30% of CPU time on the DMU
cache reclamation.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#14039
(cherry picked from commit 9dcdee7889)
Otherwise the dataset may be freed after the last dmu_buf_rele() leading
to a panic.
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#14522Closes#14523
With commit 34ce4c42f applied, there is no need for eee9362a7.
Revert that aside from the test. All tests introduced in those commits
pass.
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#14502
The zio returned from arc_write() in dmu_objset_sync() uses
zio_nowait(). However we may reach the end of dsl_dataset_sync()
which checks if we need to activate features in the filesystem
without knowing if that zio has even run through the ZIO pipeline yet.
In that case we will flag features to be activated in
dsl_dataset_block_born() but dsl_dataset_sync() has already
completed its run and those features will not actually be activated.
Mitigate this by moving the feature activation code in
dsl_dataset_sync_done(). Also add new ASSERTs in
dsl_scan_visitbp() checking if a block contradicts any filesystem
flags.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#13816
