- Reduce size and comparison complexity of q_exts_by_size B-tree.
Previous code used two 64-bit divisions and many other operations to
compare two B-tree elements. It created enormous overhead. This
implementation moves the math to the upper level and stores the score
in the B-tree elements themselves. Since all that we need to store in
that B-tree is the extent score and offset, those can fit into single
8 byte value instead of 24 bytes of q_exts_by_addr element and can be
compared with single operation.
- Better decouple secondary tree logic from main range_tree by moving
rt_btree_ops and related functions into dsl_scan.c as ext_size_ops.
Those functions are very small to worry about the code duplication and
range_tree does not need to know details such as rt_btree_compare.
- Instead of accounting number of pending bytes per pool, that needs
atomic on global variable per block, account the number of non-empty
per-vdev queues, that change much more rarely.
- When extent scan is interrupted by TXG end, continue it in the next
TXG instead of selecting next best extent. It allows to avoid leaving
one truncated (and so likely not the best any more) extent each TXG.
On top of some other optimizations this saves about 1.5 minutes out of
10 to scrub pool of 12 SSDs, storing 1.5TB of 4KB zvol blocks.
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tom Caputi <caputit1@tcnj.edu>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#13576
During sorted scrub multiple threads (one per vdev) are issuing many
ZIOs same time, all using the same scn->scn_zio_root ZIO as parent.
It causes huge lock contention on the single global lock on that ZIO.
Improve it by introducing per-queue null ZIOs, children to that one,
and using them instead as proxy.
For 12 SSD pool storing 1.5TB of 4KB blocks on 80-core system this
dramatically reduces lock contention and reduces scrub time from 21
minutes down to 12.5, while actual read stages (not scan) are about
3x faster, reaching 100K blocks per second per vdev.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#13553
- Make prefetch distance adaptive: up to 4MB prefetch doubles for
every, hit same as before, but after that it grows by 1/8 every time
the prefetch read does not complete in time to satisfy the demand.
My tests show that 4MB is sufficient for wide NVMe pool to saturate
single reader thread at 2.5GB/s, while new 64MB maximum allows the
same thread to reach 1.5GB/s on wide HDD pool. Further distance
increase may increase speed even more, but less dramatic and with
higher latency.
- Allow early reuse of inactive prefetch streams: streams that never
saw hits can be reused immediately if there is a demand, while others
can be reused after 1s of inactivity, starting with the oldest. After
2s of inactivity streams are deleted to free resources same as before.
This allows by several times increase strided read performance on HDD
pool in presence of simultaneous random reads, previously filling the
zfetch_max_streams limit for seconds and so blocking most of prefetch.
- Always issue intermediate indirect block reads with SYNC priority.
Each of those reads if delayed for longer may delay up to 1024 other
block prefetches, that may be not good for wide pools.
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.
Closes#13452
When calculating mg_aliquot alike to #12046 use number of unique data
disks in the vdev, not the total number of children vdev. Increase
default value of the tunable from 512KB to 1MB to compensate.
Before this change each disk in striped pool was getting 512KB of
sequential data, in 2-wide mirror -- 1MB, in 3-wide RAIDZ1 -- 768KB.
After this change in all the cases each disk should get 1MB.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#13388
Previous flushing algorithm limited only total number of log blocks to
the minimum of 256K and 4x number of metaslabs in the pool. As result,
system with 1500 disks with 1000 metaslabs each, touching several new
metaslabs each TXG could grow spacemap log to huge size without much
benefits. We've observed one of such systems importing pool for about
45 minutes.
This patch improves the situation from five sides:
- By limiting maximum period for each metaslab to be flushed to 1000
TXGs, that effectively limits maximum number of per-TXG spacemap logs
to load to the same number.
- By making flushing more smooth via accounting number of metaslabs
that were touched after the last flush and actually need another flush,
not just ms_unflushed_txg bump.
- By applying zfs_unflushed_log_block_pct to the number of metaslabs
that were touched after the last flush, not all metaslabs in the pool.
- By aggressively prefetching per-TXG spacemap logs up to 16 TXGs in
advance, making log spacemap load process for wide HDD pool CPU-bound,
accelerating it by many times.
- By reducing zfs_unflushed_log_block_max from 256K to 128K, reducing
single-threaded by nature log processing time from ~10 to ~5 minutes.
As further optimization we could skip bumping ms_unflushed_txg for
metaslabs not touched since the last flush, but that would be an
incompatible change, requiring new pool feature.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#12789
Use error thresholds from policy to control whether to scrub data
and/or metadata. If threshold is set to UINT64_MAX, then caller
probably does not care about result and we may skip that part.
By default import neither set the data error threshold nor read
the error counter, so skip the data scrub for faster import.
Metadata are still scrubbed and fail if even single error found.
While there just for symmetry return number of metadata errors in
case threshold is not set to zero and we haven't reached it.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Pavel Zakharov <pavel.zakharov@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#13022
The fnvlist versions of the functions are fatal if they fail,
saving each call from having to include checking the result.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Signed-off-by: Allan Jude <allan@klarasystems.com>
It is wrong for arc_write_ready() to use zfs_abd_scatter_enabled to
decide whether to reallocate/copy the buffer, because the answer is
OS-specific and depends on the buffer size. Instead of that use
abd_size_alloc_linear(), moved into public header.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#12425
The only reason for spa_config_*() to use refcount instead of simple
non-atomic (thanks to scl_lock) variable for scl_count is tracking,
hard disabled for the last 8 years. Switch to simple int scl_count
reduces the lock hold time by avoiding atomic, plus makes structure
fit into single cache line, reducing the locks contention.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#12287
When scrubbing a raidz/draid pool, which contains a replacing or
sparing mirror with multiple online children, only one child will
be read. This is not normally a serious concern because the DTL
records are used to determine where a good copy of the data is.
As long as the data can be read from one child the mirror vdev
will use it to repair gaps in any of its children. Furthermore,
even if the data which was read is corrupt the raidz code will
detect this and issue its own repair I/O to correct the damage
in the mirror vdev.
However, in the scenario where the DTL is wrong due to silent
data corruption (say due to overwriting one child) and the scrub
happens to read from a child with good data, then the other damaged
mirror child will not be detected nor repaired.
While this is possible for both raidz and draid vdevs, it's most
pronounced when using draid. This is because by default the zed
will sequentially rebuild a draid pool to a distributed spare,
and the distributed spare half of the mirror is always preferred
since it delivers better performance. This means the damaged
half of the mirror will go undetected even after scrubbing.
For system administrations this behavior is non-intuitive and in
a worst case scenario could result in the only good copy of the
data being unknowingly detached from the mirror.
This change resolves the issue by reading all replacing/sparing
mirror children when scrubbing. When the BP isn't available for
verification, then compare the data buffers from each child. They
must all be identical, if not there's silent damage and an error
is returned to prompt the top-level vdev to issue a repair I/O to
rewrite the data on all of the mirror children. Since we can't
tell which child was wrong a checksum error is logged against the
replacing or sparing mirror vdev.
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#13555
Since we use two B-trees q_exts_by_size and q_exts_by_addr, we should
count 2x sizeof (range_seg_gap_t) per node. And since average B-tree
memory efficiency is about 75%, we should increase it to 3x.
Previous code under-counted up to 30% of the memory usage.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#13537
I genuinely don't know why this didn't come up before,
but adding the LZ4 early abort pointed out this flaw,
in which we're allocating a buffer of one size, and
then telling the compressor that we're handing it buffers
of a different size, which may be Very Different - say,
allocating 512b and then telling it the inputs are 128k.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Amanakis <gamanakis@gmail.com>
Signed-off-by: Rich Ercolani <rincebrain@gmail.com>
Closes#13375
It is typical, but not generally true that if log summary has more
blocks it must also have unflushed metaslabs. Normally with metaslabs
flushed in order it works, but there are known exceptions, such as
device removal or metaslab being loaded during its flush attempt.
Before 600a02b884 if spa_flush_metaslabs() hit loading metaslab it
usually stopped (unless memlimit is also exceeded), but now it may
flush more metaslabs, just skipping that particular one. This
increased chances of assertion to fire when the skipped metaslab is
flushed on next iteration if all other metaslabs in that summary
entry are already flushed out of order.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#13486Closes#13513
clang-15 emits the following error message for functions without
a prototype:
fs/zfs/os/linux/spl/spl-kmem-cache.c:1423:27: error:
a function declaration without a prototype is deprecated
in all versions of C [-Werror,-Wstrict-prototypes]
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Aidan Harris <me@aidanharr.is>
Closes#13421
Holding a dbuf is a common operation which can become highly contended
in dbuf_find() when acquiring the dbuf hash mutex. This is particularly
true on Linux when reading/writing volumes since by default up to 32
threads from the zvol_taskq may be taking a hold of the same dbuf.
This should also be observable on FreeBSD as long as there are enough
processes accessing the volume concurrently.
This is further aggregrated by the fact that only the block id will
be unique when calculating the dbuf hash for a single volume. The
objset id, object id, and level will be the same for data blocks.
This has been observed to result in a somehwat less than uniform hash
distribution and a longer than expected max hash chain depth (~20)
on a large memory system (256 GB) using volumes.
This commit improves the siutation by switching the hash mutex to
an rwlock to allow concurrent lookups, and increasing DBUF_RWLOCKS
from 2048 to 8192 to further reduce the odds of a hash collision.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#13405
The only zdb utility require to read metaslab-related data during
read-only pool import because of spacemaps validation. Add global
variable which will allow zdb read spacemaps in case of readonly
import mode.
Reviewed-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Fedor Uporov <fuporov.vstack@gmail.com>
Closes#9095Closes#12687
Lustre makes light use of the zfs_refcount interfaces which
isn't a problem when using a non-debug build of OpenZFS. However,
when debugging is enabled the required symbols are not exported.
Reviewed-by: Olaf Faaland <faaland1@llnl.gov>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#12613
Strict hole reporting was previously disabled by default as a
performance optimization. However, this has lead to confusion
over the expected behavior and a variety of workarounds being
adopted by consumers of ZFS. Change the default behavior to
always report holes and force the TXG sync.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Upstream-commit: 05b3eb6d23
Ref: #13261Closes#12746
This is a direct commit to zfs-2.1-release to fix release builds that
error out on an unused variable. The issue is avoided on master by a
huge series of commits that change how the ASSERT macros work, but that
is not feasible to backport.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Signed-off-by: Ryan Moeller <freqlabs@FreeBSD.org>
Closes#13194Closes#13196
When unlinking multiple files from a pool at 100% capacity, it was
possible for ENOSPC to be returned after the first unlink. e.g.
rm -f /mnt/fs/test1.0.0 /mnt/fs/test1.1.0 /mnt/fs/test1.2.0
rm: cannot remove '/mnt/fs/test1.1.0': No space left on device
rm: cannot remove '/mnt/fs/test1.2.0': No space left on device
After waiting for the pending deferred frees from the first unlink to
be processed the remaining files can then be unlinked. This is caused
by the quota limit in dsl_dir_tempreserve_impl() being temporarily
decreased to the allocatable pool capacity less any deferred free
space.
This is resolved using the existing mechanism of returning ERESTART
when over quota as long as we know enough space will shortly be
available after processing the pending deferred frees.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Ryan Moeller <freqlabs@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#13172
Raw sending from pool1/encrypted with ashift=9 to pool2/encrypted with
ashift=12 results to failure when mounting pool2/encrypted (Input/Output
error). Notably, the opposite, raw sending from a greater ashift to a
lower one does not fail.
This happens because zio_compress_write() falsely checks only
ZIO_FLAG_RAW_COMPRESS and not ZIO_FLAG_RAW_ENCRYPT which is also set in
encrypted raw send streams. In this case it rounds up the psize and if
not equal to the zio->io_size it modifies the block by zeroing out
the extra bytes. Because this happens in a SA attr. registration object
(type=46), the decryption fails upon mounting the filesystem, and zpool
status falsely reports an error.
Fix this by checking both ZIO_FLAG_RAW_COMPRESS and ZIO_FLAG_RAW_ENCRYPT
before deciding whether to zero-pad a block.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#13067Closes#13074
There are two codepaths than can dirty final TXGs:
1) If calling spa_export_common()->spa_unload()->
spa_unload_log_sm_flush_all() after the spa_final_txg is set, then
spa_sync()->spa_flush_metaslabs() may end up dirtying the final
TXGs. Then we have the following panic:
Call Trace:
<TASK>
dump_stack_lvl+0x46/0x62
spl_panic+0xea/0x102 [spl]
dbuf_dirty+0xcd6/0x11b0 [zfs]
zap_lockdir_impl+0x321/0x590 [zfs]
zap_lockdir+0xed/0x150 [zfs]
zap_update+0x69/0x250 [zfs]
feature_sync+0x5f/0x190 [zfs]
space_map_alloc+0x83/0xc0 [zfs]
spa_generate_syncing_log_sm+0x10b/0x2f0 [zfs]
spa_flush_metaslabs+0xb2/0x350 [zfs]
spa_sync_iterate_to_convergence+0x15a/0x320 [zfs]
spa_sync+0x2e0/0x840 [zfs]
txg_sync_thread+0x2b1/0x3f0 [zfs]
thread_generic_wrapper+0x62/0xa0 [spl]
kthread+0x127/0x150
ret_from_fork+0x22/0x30
</TASK>
2) Calling vdev_*_stop_all() for a second time in spa_unload() after
spa_export_common() unnecessarily delays the final TXGs beyond what
spa_final_txg is set at.
Fix this by performing the check and call for
spa_unload_log_sm_flush_all() before the spa_final_txg is set in
spa_export_common(). Also check if the spa_final_txg has already been
set in spa_unload() and skip those calls in this case.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
External-issue: https://www.illumos.org/issues/9081Closes#13048Closes#13098
All of these externs are already #included as static inline
functions via corresponding headers.
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tomohiro Kusumi <kusumi.tomohiro@gmail.com>
Closes#13073
dmu_recv_begin_check() unconditionally sets the DS_HOLD_FLAG_DECRYPT
flag before calling dsl_dataset_hold_flags(). If the key on the
receiving side isn't loaded or the send stream contains embedded
blocks, the receive check fails for a stream which is perfectly
valid and could be received without any problem. This seems like
a remnant of the initial design, where unencrypted datasets below
encrypted ones weren't allowed.
Add a condition to set `DS_HOLD_FLAG_DECRYPT` only for encrypted
datasets, modify an existing test to detect this regression and add
a test for raw replication streams.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Amanakis <gamanakis@gmail.com>
Co-authored-by: George Amanakis <gamanakis@gmail.com>
Signed-off-by: Attila Fülöp <attila@fueloep.org>
Closes#13033Closes#13076
The following commit moved the users of `deferred` into function
dsl_pool_unreserved_space:
commit d2734cce68
Author: Serapheim Dimitropoulos <serapheim.dimitro@delphix.com>
Date: Fri Dec 16 14:11:29 2016 -0800
OpenZFS 9166 - zfs storage pool checkpoint
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <freqlabs@FreeBSD.org>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Christian Schwarz <christian.schwarz@nutanix.com>
Closes#13056
POSIX requires that set-uid and set-gid bits to be removed when an
unprivileged user writes to a file and ZFS does that during normal
operation.
The problem arrises when the write is stored in the ZIL and replayed.
During replay we have no access to original credentials of the process
doing the write, so zfs_write() will be performed with the root
credentials. When root is doing the write set-uid and set-gid bits
are not removed from the file.
To correct that, log a separate TX_SETATTR entry that removed those bits
on first write to such file.
Idea from: Christian Schwarz
Add test for ZIL replay of setuid/setgid clearing.
Improve various edge cases when clearing setid bits:
- The setid bits can be readded during a single write, so make sure to check
for them on every chunk write.
- Log TX_SETATTR record at most once per transaction group (if the setid bits
are keep coming back).
- Move zfs_log_setattr() outside of zp->z_acl_lock.
Reviewed-by: Dan McDonald <danmcd@joyent.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Co-authored-by: Christian Schwarz <me@cschwarz.com>
Signed-off-by: Pawel Jakub Dawidek <pawel@dawidek.net>
Closes#13027
In files created/modified before 4254acb there may be a corruption of
xattrs which is not reported during scrub and normal send/receive. It
manifests only as an error when raw sending/receiving. This happens
because currently only the raw receive path checks for discrepancies
between the dnode bonus length and the spill pointer flag.
In case we encounter a dnode whose bonus length is greater than the
predicted one, we should report an error. Modify in this regard
dnode_sync() with an assertion at the end, dump_dnode() to error out,
dsl_scan_recurse() to report errors during a scrub, and zstream to
report a warning when dumping. Also added a test to verify spill blocks
are sent correctly in a raw send.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#12720Closes#13014
For kernel to send snapshot mount/unmount events to zed.
For kernel to send symlink creates/removes on zvol plumbing.
(/dev/run/dsk/zvol/$pool/$zvol -> /dev/diskX)
If zed misses the ENODEV, all errors after are EINVAL. Treat any error
as kernel module failure.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Jorgen Lundman <lundman@lundman.net>
Closes#12416
Raw receiving a snapshot back to the originating dataset is currently
impossible because of user accounting being present in the originating
dataset.
One solution would be resetting user accounting when raw receiving on
the receiving dataset. However, to recalculate it we would have to dirty
all dnodes, which may not be preferable on big datasets.
Instead, we rely on the os_phys flag
OBJSET_FLAG_USERACCOUNTING_COMPLETE to indicate that user accounting is
incomplete when raw receiving. Thus, on the next mount of the receiving
dataset the local mac protecting user accounting is zeroed out.
The flag is then cleared when user accounting of the raw received
snapshot is calculated.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#12981Closes#10523Closes#11221Closes#11294Closes#12594
Issue #11300
FreeBSD's implementation of zfs_uio_fault_move() returns EFAULT when a
page fault occurs while copying data in or out of user buffers. The VFS
treats such errors specially and will retry the I/O operation (which may
have made some partial progress).
When the FreeBSD and Linux implementations of zfs_write() were merged,
the handling of errors from dmu_write_uio_dbuf() changed such that
EFAULT is not handled as a partial write. For example, when appending
to a file, the z_size field of the znode is not updated after a partial
write resulting in EFAULT.
Restore the old handling of errors from dmu_write_uio_dbuf() to fix
this. This should have no impact on Linux, which has special handling
for EFAULT already.
Reviewed-by: Andriy Gapon <avg@FreeBSD.org>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Mark Johnston <markj@FreeBSD.org>
Closes#12964
When the eviction thread goes to shrink an ARC state, it allocates a set
of marker buffers used to hold its place in the state's sublists.
This can be problematic in low memory conditions, since
1) the allocation can be substantial, as we allocate NCPU markers;
2) on at least FreeBSD, page reclamation can block in
arc_wait_for_eviction()
In particular, in stress tests it's possible to hit a deadlock on
FreeBSD when the number of free pages is very low, wherein the system is
waiting for the page daemon to reclaim memory, the page daemon is
waiting for the ARC eviction thread to finish, and the ARC eviction
thread is blocked waiting for more memory.
Try to reduce the likelihood of such deadlocks by pre-allocating markers
for the eviction thread at ARC initialization time. When evicting
buffers from an ARC state, check to see if the current thread is the ARC
eviction thread, and use the pre-allocated markers for that purpose
rather than dynamically allocating them.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: George Amanakis <gamanakis@gmail.com>
Signed-off-by: Mark Johnston <markj@FreeBSD.org>
Closes#12985
Verify that all empty sectors are zero filled before using them to
calculate parity. Failure to do so can result in incorrect parity
columns being generated and written to disk if the contents of an
empty sector are non-zero. This was possible because the checksum
only protects the data portions of the buffer, not the empty sector
padding.
This issue has been addressed by updating raidz_parity_verify() to
check that all dRAID empty sectors are zero filled. Any sectors
which are non-zero will be fixed, repair IO issued, and a checksum
error logged. They can then be safely used to verify the parity.
This specific type of damage is unlikely to occur since it requires
a disk to have silently returned bad data, for an empty sector, while
performing a scrub. However, if a pool were to have been damaged
in this way, scrubbing the pool with this change applied will repair
both the empty sector and parity columns as long as the data checksum
is valid. Checksum errors will be reported in the `zpool status`
output for any repairs which are made.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#12857
On FreeBSD vnode reclamation is single-threaded, protected by single
global lock. Linux seems to be able to use a thread per mount point,
but at this time it creates more harm than good.
Reduce number of threads to 1, adding tunable in case somebody wants
to try more.
Reviewed-by: Ryan Moeller <ryan@ixsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Chris Dunlop <chris@onthe.net.au>
Reviewed-by: Ahelenia Ziemiańska <nabijaczleweli@nabijaczleweli.xyz>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#12896
Issue #9966
Improve the ability of zfs send to determine if a block is compressed
or not by using information contained in the blkptr.
Reviewed-by: Rich Ercolani <rincebrain@gmail.com>
Reviewed-by: Matthew Ahrens <matthew.ahrens@delphix.com>
Signed-off-by: Paul Dagnelie <pcd@delphix.com>
Closes#12770
Due to a possible lock inversion the zvol open call path on Linux
needs to be able to retry in the case where the spa_namespace_lock
cannot be acquired.
For Linux 5.12 an older kernel this was accomplished by returning
-ERESTARTSYS from zvol_open() to request that blkdev_get() drop
the bdev->bd_mutex lock, reaquire it, then call the open callback
again. However, as of the 5.13 kernel this behavior was removed.
Therefore, for 5.12 and older kernels we preserved the existing
retry logic, but for 5.13 and newer kernels we retry internally in
zvol_open(). This should always succeed except in the case where
a pool's vdev are layed on zvols, in which case it may fail. To
handle this case vdev_disk_open() has been updated to retry when
opening a device when -ERESTARTSYS is returned.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #12301Closes#12759
Userland figures out which encryption-root keys are required to load,
and issues ZFS_IOC_LOAD_KEY.
The tail section of spa_keystore_load_wkey() will call
zvol_create_minors() on the encryption-root object.
Any clones of the encrypted zvol will not be plumbed. This commits
adds additional logic to detect if zvol has clones, and is encrypted,
then adds these to the list of zvols to call zvol_create_minors() on.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Jorgen Lundman <lundman@lundman.net>
Closes#12471
In addition to flushing memory mapped regions when checking holes,
commit de198f2d95 modified the dirty dnode detection logic to check
the dn->dn_dirty_records instead of the dn->dn_dirty_link. Relying
on the dirty record has not be reliable, switch back to the previous
method.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #11900Closes#12745
When using lseek(2) to report data/holes memory mapped regions of
the file were ignored. This could result in incorrect results.
To handle this zfs_holey_common() was updated to asynchronously
writeback any dirty mmap(2) regions prior to reporting holes.
Additionally, while not strictly required, the dn_struct_rwlock is
now held over the dirty check to prevent the dnode structure from
changing. This ensures that a clean dnode can't be dirtied before
the data/hole is located. The range lock is now also taken to
ensure the call cannot race with zfs_write().
Furthermore, the code was refactored to provide a dnode_is_dirty()
helper function which checks the dnode for any dirty records to
determine its dirtiness.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Rich Ercolani <rincebrain@gmail.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #11900Closes#12724
When you create a pool, zfs writes vd->vdev_enc_sysfs_path with the
enclosure sysfs path to the fault LEDs, like:
vdev_enc_sysfs_path = /sys/class/enclosure/0:0:1:0/SLOT8
However, this enclosure path doesn't get updated on successive imports
even if enclosure path to the disk changes. This patch fixes the issue.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tony Hutter <hutter2@llnl.gov>
Closes#11950Closes#12095
As of the Linux 5.9 kernel a fallthrough macro has been added which
should be used to anotate all intentional fallthrough paths. Once
all of the kernel code paths have been updated to use fallthrough
the -Wimplicit-fallthrough option will because the default. To
avoid warnings in the OpenZFS code base when this happens apply
the fallthrough macro.
Additional reading: https://lwn.net/Articles/794944/
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#12441
Errors in zil_lwb_write_done() are not propagated to
zil_lwb_flush_vdevs_done() which can result in zil_commit_impl()
not returning an error to applications even when zfs was not able
to write data to the disk.
Remove the ZIO_FLAG_DONT_PROPAGATE flag from zio_rewrite() to
allow errors to propagate and consolidate the error handling for
flush and write errors to a single location (rather than having
error handling split between the "write done" and "flush done"
handlers).
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Prakash Surya <prakash.surya@delphix.com>
Signed-off-by: Arun KV <arun.kv@datacore.com>
Closes#12391Closes#12443
The block pointer verification check in arc_read() should also
cover embedded block pointers. While highly unlikely, accessing
a damaged block pointer can result in panic. To further harden
the code extend the existing check to include embedded block
pointers and add a comment explaining the rational for this
sanity check. Lastly, correct a flaw in zfs_blkptr_verify()
so the error count is checked even when checking a untrusted
config to verify the non-pool-specific portions of a block
pointer.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#12535
When zfs_send_corrupt_data is set, use the TRAVERSE_HARD flag,
so traverse_visitbp() will not fail with ECKSUM if a blockpointer
cannot be read, but rather will continue and send the objects it can.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Signed-off-by: Allan Jude <allan@klarasystems.com>
Sponsored-By: Klara Inc.
Sponsored-By: WHC Online Solutions Inc.
Closes#12541
Unfortunately, there was an overzealous assertion that was (in pretty
specific circumstances) false, causing failure. This assertion was
added in error, so we're removing it.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rich Ercolani <rincebrain@gmail.com>
Closes#9897Closes#12020Closes#12246
We round up the psize to the nearest multiple of the asize or to the
lsize, whichever is smaller. Once that's done, we allocate a new
buffer of the appropriate size, zero the tail, and copy the data
into it. This adds a small performance cost to these kinds of writes,
but fixes the bookkeeping problems.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Co-authored-by: Matthew Ahrens <matthew.ahrens@delphix.com>
Signed-off-by: Paul Dagnelie <pcd@delphix.com>
Closes#12522Closes#8462
benchmark_raidz() allocates a row to benchmark parity calculation and
reconstruction. In the latter case, the parity blocks are left
uninitialized, leading to reports from KMSAN.
Initialize parity blocks to 0xAA as we do for the data earlier in the
function. This does not affect the selected RAID-Z implementation on
any of several systems tested.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Mark Johnston <markj@FreeBSD.org>
Closes#12473
When a header is allocated for full overwrite it is a waste of time
to allocate b_pabd/b_rabd for it, since arc_write() will free them
without ever being touched. If it is a read or a partial overwrite
then arc_read() and arc_hdr_decrypt() allocate them explicitly.
Reduced memory allocation in user threads also reduces ARC eviction
throttling there, proportionally increasing it in ZIO threads, that
is not good. To minimize or even avoid it introduce ARC allocation
reserve, allowing certain arc_get_data_abd() callers to allocate a
bit longer in situations where user threads will already throttle.
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#12398
It is very expensive and not informative to call multilist_is_empty()
for each arc_change_state() on debug builds to check for impossible.
Instead implement special index function for arc_l2c_only->arcs_list,
multilists, panicking on any attempt to use it.
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#12421
Instead of clearing stats inside arc_buf_alloc_impl() do it inside
arc_hdr_alloc() and arc_release(). It fixes statistics being wiped
every time a new dbuf is filled from the ARC.
Remove b_l1hdr.b_l2_hits. L2ARC hits are accounted at b_l2hdr.b_hits.
Since the hits are accounted under hash lock, replace atomics with
simple increments.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Wilson <george.wilson@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#12422
vq_lock is already too congested for two more operations per I/O.
Instead of dropping and reacquiring it inside vdev_queue_aggregate()
delegate the zio_vdev_io_bypass() and zio_execute() calls for parent
I/Os to callers, that drop the lock any way to execute the new I/O.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#12297
Use atomic_load_64() for zfs_refcount_count() to prevent torn reads
on 32-bit platforms. On 64-bit ones it should not change anything.
When built with ZFS_DEBUG but running without tracking enabled use
atomics instead of mutexes same as for builds without ZFS_DEBUG.
Since rc_tracked can't change live we can check it without lock.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#12420
Before OpenZFS 2.0, trying to set the FreeBSD sysctl vfs.zfs.arc_max
to a disallowed value would return an error.
Since the switch, it instead only generates WARN_IF_TUNING_IGNORED
Keep the ability to set the sysctl's specifically to 0, even though
that is less than the minimum, because some tests depend on this.
Also lost, was the ability to set vfs.zfs.arc_max to a value less
than the default vfs.zfs.arc_min at boot time. Restore this as well.
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Reviewed-by: Ryan Moeller <ryan@ixsystems.com>
Signed-off-by: Allan Jude <allan@klarasystems.com>
Closes#12161
Run arc_evict thread at higher priority, nice=0, to give it more CPU
time which can improve performance for workload with high ARC evict
activities.
On mixed read/write and sequential read workloads, I've seen between
10-40% better performance.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Tony Nguyen <tony.nguyen@delphix.com>
Closes#12397
In l2arc_add_vdev() first decide whether the device is eligible for
L2ARC rebuild or whole device trim and then add it to the list of cache
devices. Otherwise l2arc_feed_thread() might already start writing on
the device invalidating previous content as l2ad_hand = l2ad_start.
However l2arc_rebuild_vdev() needs the device present in the cache
device list to figure out its l2arc_dev_t. Fix this by moving most of
l2arc_rebuild_vdev() in a new function l2arc_rebuild_dev() which does
not need to search in the cache device list.
In contrast to l2arc_add_vdev() we do not have to worry about
l2arc_feed_thread() invalidating previous content when onlining a
cache device. The device parameters (l2ad*) are not cleared when
offlining the device and writing new buffers will not invalidate
all previous content. In worst case only buffers that have not had
their log block written to the device will be lost.
Retire persist_l2arc_00{4,5,8} tests since they cover code already
covered by the remaining ones. Test persist_l2arc_006 is renamed to
persist_l2arc_004 and persist_l2arc_007 is renamed to persist_l2arc_005.
Fix a typo in persist_l2arc_004, and remove an assertion that is not
always true from l2arc_arcstats_pos. Also update an assertion in
persist_l2arc_005 and explain why in a comment.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#12365
It seems nothing ensures that this array is zeroed when a dnode is
freshly allocated, so in principle it retains the values from the
previous allocation. In practice it seems to be the case that the
fields should end up zeroed, but we can zero the field anyway for
consistency.
This was found using KMSAN.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Mark Johnston <markj@FreeBSD.org>
Closes#12383
When logging a TX_WRITE record in the case where file data has to be
copied from the DMU, we pad the log record size to a multiple of 8
bytes. In this case, any padding bytes should be zeroed, otherwise the
contents of uninitialized memory are written to the ZIL.
This was found using KMSAN.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Mark Johnston <markj@FreeBSD.org>
Closes#12383
When allocating a record, we round up the allocation size to a multiple
of 8. In this case, any padding bytes should be zeroed, otherwise the
contents of uninitialized memory are written to the ZIL.
This was found using KMSAN.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Mark Johnston <markj@FreeBSD.org>
Closes#12383
When logging TX_SETATTR, we could otherwise fail to initialize part of
the corresponding ZIL record depending on which fields are present in
the xvattr. Initialize the creation time and the AV scan timestamp to
zero so that uninitialized bytes are not written to the ZIL.
This was found using KMSAN.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Mark Johnston <markj@FreeBSD.org>
Closes#12383
spa_prop_find() may fail to find the specified property, in which case
it suppresses ENOENT from zap_lookup(). In this case, the return value
is left uninitialized, so spa_autoreplace was being initialized using an
uninitialized stack variable.
This was found using KMSAN. It appears to be a regression from commit
9eb7b46ed0, which removed the initialization of "autoreplace" from the
definition.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Mark Johnston <markj@FreeBSD.org>
Closes#12383
After 1325434b, we can in certain circumstances end up calling
spa_update_dspace with vd->vdev_mg NULL, which ends poorly during
vdev removal.
So let's not do that further space adjustment when we can't.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rich Ercolani <rincebrain@gmail.com>
Closes#12380Closes#12428
Remove mc_lock use from metaslab_class_throttle_*(). The math there
is based on refcounts and so atomic, so the only race possible there
is between zfs_refcount_count() and zfs_refcount_add(). But in most
cases metaslab_class_throttle_reserve() is called with the allocator
lock held, which covers the race. In cases where the lock is not
held, GANG_ALLOCATION() or METASLAB_MUST_RESERVE are set, and so we
do not use zfs_refcount_count(). And even if we assume some other
non-existing scenario, the worst that may happen from this race is
few more I/Os get to allocation earlier, that is not a problem.
Move locks and data of different allocators into different cache
lines to avoid false sharing. Group spa_alloc_* arrays together
into single array of aligned struct spa_alloc spa_allocs. Align
struct metaslab_class_allocator.
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Don Brady <don.brady@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#12314
Remove unneeded global, practically constant, state pointer variables
(arc_anon, arc_mru, etc.), replacing them with macros of real state
variables addresses (&ARC_anon, &ARC_mru, etc.).
Change ARC_EVICT_ALL from -1ULL to UINT64_MAX, not requiring special
handling in inner loop of ARC reclamation. Respectively change bytes
argument of arc_evict_state() from int64_t to uint64_t.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#12348
Ensure all calls to bqueue_init() has a corresponding call to bqueue_destroy()
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Co-authored-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Jorgen Lundman <lundman@lundman.net>
Closes#12118
* zio: avoid callback typecasting
* zil: avoid zil_itxg_clean() callback typecasting
* zpl: decouple zpl_readpage() into two separate callbacks
* nvpair: explicitly declare callbacks for xdr_array()
* linux/zfs_nvops: don't use external iput() as a callback
* zcp_synctask: don't use fnvlist_free() as a callback
* zvol: don't use ops->zv_free() as a callback for taskq_dispatch()
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Signed-off-by: Alexander Lobakin <alobakin@pm.me>
Closes#12260
We don't use or need the pool name or value source in the zvol tasks.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes#12361
Most of dsl_dir_diduse_space() and dsl_dir_transfer_space() CPU time
is a dd_lock overhead and time spent in dmu_buf_will_dirty(). Calling
them one after another is a waste of time and even more contention.
Doing that twice for each rewritten block within dbuf_write_done()
via dsl_dataset_block_kill() and dsl_dataset_block_born() created one
of the biggest CPU overheads in case of small blocks rewrite.
dsl_dir_diduse_transfer_space() combines functionality of these two
functions for cases where it is needed, but without double overhead,
practically for the cost of dsl_dir_diduse_space() or even cheaper.
While there, optimize dsl_dir_phys() calls in dsl_dir_diduse_space()
and dsl_dir_transfer_space(). It seems Clang detects some aliasing
there, repeating dd->dd_dbuf->db_data dereference multiple times,
increasing dd_lock scope and contention.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Author: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#12300
* Tinker with slop space accounting with dedup
Do not include the deduplicated space usage in the slop space
reservation, it leads to surprising outcomes.
* Update spa_dedup_dspace sometimes
Sometimes, we get into spa_get_slop_space() with
spa_dedup_dspace=~0ULL, AKA "unset", while spa_dspace is correctly set.
So call the code to update it before we use it if we hit that case.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Signed-off-by: Rich Ercolani <rincebrain@gmail.com>
Closes#12271
arc_evict_hdr() returns number of evicted bytes in scope of specific
state. For ghost states it does not mean the amount of really freed
memory, but the logical buffer size. It is correct for the eviction
process, but not for waking up threads waiting for ARC size reduction,
as added in "Revise ARC shrinker algorithm" commit, causing premature
wakeups while ARC is still overflowed, allowing even bigger overflow,
plus processing overhead when next allocation will also get blocked,
probably also for too short time.
To fix that make arc_evict_hdr() also return the amount of really
freed memory, which for the ghost states is only the header, and use
it to update arc_evict_count instead. Originally I was thinking to
not return it at all, since arc_get_data_impl() does not account for
the headers, but decided that some slow allocation progress is better
than long waits, reaching on my tests up to 100ms.
To reduce negative latency effects of long time periods when reclaim
thread can free little real memory, start reclamation process earlier,
before we actually reached the overflow threshold, when we have to
throttle new allocations. We can also do it without taking global
arc_evict_lock, reducing the contention.
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Allan Jude <allan@klarasystems.com>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#12279
Callers of zfs_file_get and zfs_file_put can corrupt the reference
counts for the file structure resulting in a panic or a soft lockup.
When zfs send/recv runs, it will add a reference count to the
open file, and begin to send or recv the stream. If the file descriptor
is closed, then when dmu_recv_stream() or dmu_send() return we will
call zfs_file_put to remove the reference we placed on the file
structure. Unfortunately, because zfs_file_put() uses the file
descriptor to lookup the file structure, it may end up finding that
the file descriptor table no longer contains the file struct, thus
leaking the file structure. Or it might end up finding a file
descriptor for a different file and blindly updating its reference
counts. Other failure modes probably exists.
This change reworks the zfs_file_[get|put] interface to not rely
on the file descriptor but instead pass the zfs_file_t pointer around.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Co-authored-by: Allan Jude <allan@klarasystems.com>
Signed-off-by: George Wilson <gwilson@delphix.com>
External-issue: DLPX-76119
Closes#12299
This dramatically reduces the lock contention on systems with slower
(non-TSC) timecounters. With TSC the difference is minimal, but since
this lock is pretty congested, any improvement counts. Plus I don't
see any reason to do it under the lock other than the latency of the
lock itself, which this change actually reduces.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#12281
With default dbuf cache size of 1/32 of ARC, it makes no sense to have
hash table of the same size (or even bigger on Linux). Reduce it to
1/8 of ARC's one, still leaving some slack, assuming higher I/O rate
via dbuf cache than via ARC.
Remove padding from ARC hash locks array. The idea behind padding
is to avoid false sharing between locks. It would have sense if
there would be a limited number of very busy locks. But since we
have no limit on the number, using the same memory for more locks we
can achieve even lower lock contention with the same false sharing,
or we can use less memory for the same contention level.
Reduce number of hash locks from 8192 to 2048. The number is still
big enough to not cause contention, but reduced memory size improves
cache hit rate for mutex_tryenter() in ARC eviction thread, saving
about 1% of the thread time.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#12289
Fix a leak of abd_t that manifested mostly when using
raidzN with at least as many columns as N (e.g. a
four-disk raidz2 but not a three-disk raidz2).
Sufficiently heavy raidz use would eventually run a system
out of memory.
Additionally:
* Switch abd_cache arena to FIRSTFIT, which empirically
improves perofrmance.
* Make abd_chunk_cache more performant and debuggable.
* Allocate the abd_zero_buf from abd_chunk_cache rather
than the heap.
* Don't try to reap non-existent qcaches in abd_cache arena.
* KM_PUSHPAGE->KM_SLEEP when allocating chunks from their
own arena
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Jorgen Lundman <lundman@lundman.net>
Co-authored-by: Sean Doran <smd@use.net>
Closes#12295
dmu_zfetch_stream_fini() is missing calls to destroy the refcounts,
leaking them and the mutex inside.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Jorgen Lundman <lundman@lundman.net>
Closes#12294
In all places except two spa_get_random() is used for small values,
and the consumers do not require well seeded high quality values.
Switch those two exceptions directly to random_get_pseudo_bytes()
and optimize spa_get_random(), renaming it to random_in_range(),
since it is not related to SPA or ZFS in general.
On FreeBSD directly map random_in_range() to new prng32_bounded() KPI
added in FreeBSD 13. On Linux and in user-space just reduce the type
used to uint32_t to avoid more expensive 64bit division.
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#12183
The number of sublists in a multilist is relatively small. We dont need
64 bits to calculate an index. 32 bits is sufficient and makes the
code more efficient.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#12288
While abd_verify() does nothing when built without debug, compiler
can't optimize it out by itself due to calls to external list_*()
and abd_verify_scatter(). This commit makes it explicit.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Adam Moss <c@yotes.com>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#12280
Unlike most other properties the 'compatibility' property is stored
in the pool config object and not the DMU_OT_POOL_PROPS object.
This had the advantage that the compatibility information is available
without needing to fully import the pool (it can be read with zdb).
However, this means we need to make sure to update both the copy of
the config in the MOS and the cache file. This wasn't being done.
This commit adds a call to spa_async_request() to ensure the copy of
the config in the cache file gets updated as well as the one stored
in the pool. This same change is made for the 'comment' property
which suffers from the same inconsistency.
Reviewed-by: Sean Eric Fagan <sef@ixsystems.com>
Reviewed-by: Colm Buckley <colm@tuatha.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#12261Closes#12276
According to current zfs man page zfs_metaslab_mem_limit should be
25 instead of 75.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Signed-off-by: jumbi77@users.noreply.github.comCloses#12273
ZFS loves using %llu for uint64_t, but that requires a cast to not
be noisy - which is even done in many, though not all, places.
Also a couple places used %u for uint64_t, which were promoted
to %llu.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rich Ercolani <rincebrain@gmail.com>
Closes#12233
This reverts commit 13fac09868.
Per the discussion in #11531, the reverted commit---which intended only
to be a cleanup commit---introduced a subtle, unintended change in
behavior.
Care was taken to partially revert and then reapply 10b3c7f5e4
which would otherwise have caused a conflict. These changes were
squashed in to this commit.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Suggested-by: @chrisrd
Suggested-by: robn@despairlabs.com
Signed-off-by: Antonio Russo <aerusso@aerusso.net>
Closes#11531Closes#12227
wmsum was designed exactly for cases like these with many updates
and rare reads. It allows to completely avoid atomic operations on
congested global variables.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#12172
In case we have I/O and try to remove an L2ARC device a deadlock might
occur. arc_read()->zio_read()->zfs_blkptr_verify() waits for SCL_VDEV
to be dropped while holding the hash_lock. However, spa_l2cache_load()
holds SCL_ALL and waits for the hash_lock in l2arc_evict().
Fix this by moving zfs_blkptr_verify() to the top top arc_read() before
the hash_lock is taken. Verify the block pointer and return a checksum
error if damaged rather than halting the system, by using
BLK_VERIFY_LOG instead of BLK_VERIFY_HALT.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#12054
vdev_draid_min_asize() returns the minimum size of a child vdev. This
is used when determining if a disk is big enough to replace a child.
It's also used by zdb to determine how big of a child to make to test
replacement.
vdev_draid_min_asize() says that the child’s asize has to be at least
1/Nth of the entire draid’s asize, which is the same logic as raidz.
However, this contradicts the code in vdev_draid_open(), which
calculates the draid’s asize based on a reduced child size:
An additional 32MB of scratch space is reserved at the end of each
child for use by the dRAID expansion feature
So the problem is that you can replace a draid disk with one that’s
vdev_draid_min_asize(), but it actually needs to be larger to accommodate
the additional 32MB. The replacement is allowed and everything works at
first (since the reserved space is at the end, and we don’t try to use
it yet), but when you try to close and reopen the pool,
vdev_draid_open() calculates a smaller asize for the draid, because of
the smaller leaf, which is not allowed.
I think the confusion is that vdev_draid_min_asize() is correctly
returning the amount of required *allocatable* space in a leaf, but the
actual *size* of the leaf needs to be at least 32MB more than that.
ztest_vdev_attach_detach() assumes that it can attach that size of
device, and it actually can (the kernel/libzpool accepts it), but it
then later causes zdb to not be able to open the pool.
This commit changes vdev_draid_min_asize() to return the required size
of the leaf, not the size that draid will make available to the metaslab
allocator.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#11459Closes#12221
This commit partially reverts changes to multilists in PR 7968
(multi-threaded spa-sync()) and adds some cache line alignments to
separate read-only multilists and heavily modified refcount's to
different cache lines.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-by: iXsystems, Inc.
Closes#12158
This mostly reverts "3537 want pool io kstats" commit of 8 years ago.
From one side this code using pool-wide locks became pretty bad for
performance, creating significant lock contention in I/O pipeline.
From another, there are more efficient ways now to obtain detailed
statistics, while this statistics is illumos-specific and much less
usable on Linux and FreeBSD, reported only via procfs/sysctls.
This commit does not remove KSTAT_TYPE_IO implementation, that may
be removed later together with already unused KSTAT_TYPE_INTR and
KSTAT_TYPE_TIMER.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#12212
`getfsstat(2)` is used to retrieve the list of mounted file systems,
which libzfs uses when fetching properties like mountpoint, atime,
setuid, etc. The `mode` parameter may be `MNT_NOWAIT`, which uses
information in the VFS's cache, or `MNT_WAIT`, which effectively does a
`statfs` on every single mounted file system in order to fetch the most
up-to-date information. As far as I can tell, the only fields that
libzfs cares about are the filesystem's name, mountpoint, fstypename,
and mount flags. Those things are always updated on mount and unmount,
so they will always be accurate in the VFS's mount cache except in two
circumstances:
1) When a file system is busy unmounting
2) When a ZFS file system changes the value of a mount-overridable
property like atime or setuid, but doesn't remount the file system.
Right now that only happens when the property is changed by an
unprivileged user who has delegated authority to change the property
but not to mount the dataset. But perhaps libzfs could choose to do
it for other reasons in the future.
Switching to `MNT_NOWAIT` will greatly improve speed with no downside,
as long as we explicitly update the mount cache whenever we change a
mount-overridable property.
For comparison, Illumos gets this information using the native
`getmntany` and `getmntent` functions, which also use cached
information. The illumos function that would refresh the cache,
`resetmnttab`, is never called by libzfs.
And on GNU/Linux, `getmntany` and `getmntent` don't even communicate
with the kernel directly. They simply parse the file they are given,
which is usually /etc/mtab or /proc/mounts. Perhaps the implementation
of /proc/mounts is synchronous, ala MNT_WAIT; I don't know.
Sponsored-by: Axcient
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alan Somers <asomers@gmail.com>
Closes: #12091
Update the logic to handle the dedup-case of consecutive
FREEs in the livelist code. The logic still ensures that
all the FREE entries are matched up with a respective
ALLOC by keeping a refcount for each FREE blkptr that we
encounter and ensuring that this refcount gets to zero
by the time we are done processing the livelist.
zdb -y no longer panics when encountering double frees
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: John Kennedy <john.kennedy@delphix.com>
Reviewed-by: Don Brady <don.brady@delphix.com>
Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Closes#11480Closes#12177
- Avoid atomic_add() when updating as_lower_bound/as_upper_bound.
Previous code was excessively strong on 64bit systems while not
strong enough on 32bit ones. Instead introduce and use real
atomic_load() and atomic_store() operations, just an assignments
on 64bit machines, but using proper atomics on 32bit ones to avoid
torn reads/writes.
- Reduce number of buckets on large systems. Extra buckets not as
much improve add speed, as hurt reads. Unlike wmsum for aggsum
reads are still important.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#12145
wmsum counters are a reduced version of aggsum counters, optimized for
write-mostly scenarios. They do not provide optimized read functions,
but instead allow much cheaper add function. The primary usage is
infrequently read statistic counters, not requiring exact precision.
The Linux implementation is directly mapped into percpu_counter KPI.
The FreeBSD implementation is directly mapped into counter(9) KPI.
In user-space due to lack of better implementation mapped to aggsum.
Unfortunately neither Linux percpu_counter nor FreeBSD counter(9)
provide sufficient functionality to completelly replace aggsum, so
it still remains to be used for several hot counters.
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#12114
Previously, ZFS scaled maxinflight_bytes based on total number of
disks in the pool. A 3-wide mirror was receiving a queue depth of 3
disks, which it should not, since it reads from all the disks inside.
For wide raidz the situation was slightly better, but still a 3-wide
raidz1 received a depth of 3 disks instead of 2.
The new code counts only unique data disks, i.e. 1 disk for mirrors
and non-parity disks for raidz/draid. For draid the math is still
imperfect, since vdev_get_nparity() returns number of parity disks
per group, not per vdev, but still some better than it was.
This should slightly reduce scrub influence on payload for some pool
topologies by avoiding excessive queuing.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closing #12046
Propagate vdev child state to parents on invalid label
Add VDEV_AUX_BAD_LABEL to print_import_config()
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Co-authored-by: Srikanth N S <srikanth.nagasubbaraoseetharaman@hpe.com>
Signed-off-by: Vipin Kumar Verma <vipin.verma@hpe.com>
Closes#12088
This change addresses two distinct scenarios which are possible
when performing a sequential resilver to a dRAID pool with vdevs
that contain silent unknown damage. Which in this circumstance
took the form of the devices being intentionally overwritten with
zeros. However, it could also result from a device returning incorrect
data while a sequential resilver was in progress.
Scenario 1) A sequential resilver is performed while all of the
dRAID vdevs are ONLINE and there is silent damage present on the
vdev being resilvered. In this case, nothing will be repaired
by vdev_raidz_io_done_reconstruct_known_missing() because
rc->rc_error isn't set on any of the raid columns. To address
this vdev_draid_io_start_read() has been updated to always mark
the resilvering column as ESTALE for sequential resilver IO.
Scenario 2) Multiple columns contain silent damage for the same
block and a sequential resilver is performed. In this case it's
impossible to generate the correct data from parity unless all of
the damaged columns are being sequentially resilvered (and thus
only good data is used to generate parity). This is as expected
and there's nothing which can be done about it. However, we need
to be careful not to make to situation worse. Since we can't
verify the data is actually good without a checksum, we must
only repair the devices which are being sequentially resilvered.
Otherwise, an incorrect repair to a device which previously
contained good data could effectively lock in the damage and
make reconstruction impossible. A check for this was added to
vdev_raidz_io_done_verified() along with a new test case.
Lastly, this change updates the redundancy_draid_spare1 and
redundancy_draid_spare3 test cases to be more representative
of normal dRAID replacement operation. Specifically, what we
care about is that the scrub run after a sequential resilver
does not find additional blocks which need repair. This would
indicate the sequential resilver failed to rebuild a section of
one of the devices. Note also the tests were switched to using
the verify_pool() function which still checks for checksum errors.
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#12061
