Restructure the code in zfs_log_xvattr() to use a lr_attr_end
structure when accessing lr_attr_t elements located after the
variable sized array. This makes the code more understandable
and resolves the accessing beyond the end of the field warnings.
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#13528Closes#13575
This code should be kept inline with the upstream lua version as much
as possible. Therefore, we simply want to silence the warning. This
check was enabled by default as part of -Wall in gcc 12.1.
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#13528Closes#13575
It may happen that scan bookmark points to a block that was turned
into a part of a big hole. In such case dsl_scan_visitbp() may skip
it and dsl_scan_check_resume() will not be called for it. As result
new scan suspend won't be possible until the end of the object, that
may take hours if the object is a multi-terabyte ZVOL on a slow HDD
pool, stretching TXG to all that time, creating all sorts of problems.
This patch changes the resume condition to any greater or equal block,
so even if we miss the bookmarked block, the next one we find will
delete the bookmark, allowing new suspend.
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Before this change for every valid parity column raidz_parity_verify()
allocated new buffer and copied there existing data, then recalculated
the parity and compared the result with the copy. This patch removes
the memory copy, simply swapping original buffer pointers with newly
allocated empty ones for parity recalculation and comparison. Original
buffers with potentially incorrect parity data are then just freed,
while new recalculated ones are used for repair.
On a pool of 12 4-wide raidz vdevs, storing 1.5TB of 16MB blocks, this
change reduces memory traffic during scrub by 17% and total unhalted
CPU time by 25%.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#13613
Issuing several scrub reads for a block we may use the parent ZIO
buffer for one of child ZIOs. If that read complete successfully,
then we won't need to copy the data explicitly. If block has only
one copy (typical for root vdev, which is also a mirror inside),
then we never need to copy -- succeed or fail as-is. Previous
code also copied data from buffer of every successfully completed
child ZIO, but that just does not make any sense.
On healthy N-wide mirror this saves all N+1 (or even more in case
of ditto blocks) memory copies for each scrubbed block, allowing
CPU to focus mostly on check-summing. For other vdev types it
should save one memory copy per block copy at root vdev.
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#13606
Block statistics calculation during scrub I/O issue in case of sorted
scrub accounted ditto blocks several times. Embedded blocks on other
side were not accounted at all. This change moves the accounting from
issue to scan stage, that fixes both problems and also allows to avoid
pool-wide locking and the lock contention it created.
Since this statistics is quite specific and is not even exposed now
anywhere, disable its calculation by default to not waste CPU time.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#13579
Change math to make it like the ARC, using multiplications instead.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#13591
- Introduce first element offset within a leaf. It allows to reduce
by ~50% average memmove() size when adding/removing elements. If the
added/removed element is in the first half of the leaf, we may shift
elements before it and adjust the bth_first instead of moving more
elements after it.
- Use memcpy() instead of memmove() when we know there is no overlap.
- Switch from uint64_t to uint32_t. It does not limit anything,
but 32-bit arches should appreciate it greatly in hot paths.
- Store leaf capacity in struct btree to avoid 64-bit divisions.
- Adjust zfs_btree_insert_into_leaf() to always result in balanced
leaves after splitting, no matter where the new element was inserted.
Not that we care about it much, but it should also allow B-trees with
as little as two elements per leaf instead of 4 previously.
When scrubbing pool of 12 SSDs, storing 1.5TB of 4KB zvol blocks this
reduces amount of time spent in memmove() inside the scan thread from
13.7% to 5.7% and total scrub time by ~15 seconds out of 9 minutes.
It should also reduce spacemaps load time, but I haven't measured it.
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#13582
- 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
Handle crypto_dispatch() return values same as crp->crp_etype errors.
On FreeBSD 12 many drivers returned same errors both ways, and lack
of proper handling for the first ended up in assertion panic later.
It was changed in FreeBSD 13, but there is no reason to not be safe.
While there, skip waiting for completion, including locking and
wakeup() call, for sessions on synchronous crypto drivers, such as
typical aesni and software.
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#13563
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
Modern Clang and GCC can successfully implement simple conditions
without branching with math and flag operations. Use of arrays for
translation no longer helps as much as it was 14+ years ago.
Disassemble of the code generated by Clang 13.0.0 on FreeBSD 13.1,
Clang 14.0.4 on FreeBSD 14 and GCC 10.2.1 on Debian 11 with this
change still shows no branching instructions.
Profiling of CPU-bound scan stage of sorted scrub shows reproducible
reduction of time spent inside avl_find() from 6.52% to 4.58%.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#13540
- 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
As of the Linux 5.19 kernel the readpage() address space operation
has been replaced by read_folio().
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#13515
Linux 5.19 commit torvalds/linux@44abff2c0 splits the secure
erase functionality from the blkdev_issue_discard() function.
The blkdev_issue_secure_erase() must now be issued to issue
a secure erase.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#13515
Linux 5.19 commit torvalds/linux@44abff2c0 removed the
blk_queue_secure_erase() helper function. The preferred
interface is to now use the bdev_max_secure_erase_sectors()
function to check for discard support.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#13515
Linux 5.19 commit torvalds/linux@70200574cc removed the
blk_queue_discard() helper function. The preferred interface
is to now use the bdev_max_discard_sectors() function to check
for discard support.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#13515
As for the Linux 5.18 kernel bio_alloc() expects a block_device struct
as an argument. This removes the need for the bio_set_dev() compatibility
code for 5.18 and newer kernels.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#13515
Refcount creation for abd_zero_scatter->abd_children is redundant in
abd_alloc_zero_scatter, as it has been done in abd_init_struct.
In addition, abd_children is undefined when ZFS_DEBUG is disabled, the
reference of abd_children in abd_alloc_zero_scatter breaks build of
libzpool when ZFS_DEBUG is disabled.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Ping Huang <huangping@smartx.com>
Closes#13429
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
Linux 5.12 PPC 5.12 get_user() and __copy_from_user_inatomic()
inline helpers very indirectly include a reference to the GPL'd
array mmu_feature_keys[] and fails to build. Workaround this by
using copy_from_user() and throwing EFAULT for any calls to
__copy_from_user_inatomic(). This is a workaround until a fix
for Linux commit 7613f5a66becfd0e43a0f34de8518695888f5458
"powerpc/64s/kuap: Use mmu_has_feature()" is fully addressed.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Authored-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: szubersk <szuberskidamian@gmail.com>
Closes#11958Closes#12590Closes#13367
On some architectures ZERO_PAGE is unavailable because it references
a GPL exported symbol of empty_zero_page. Originally e08b993 removed
the call to PAGE_ZERO(0) for assignment to the abd_zero_page. However,
a simple check can be done to avoid a kernel allocation and free for
the abd_zero_page if ZERO_PAGE is available.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Brian Atkinson <batkinson@lanl.gov>
Closes#13199
This adds supports for hole-punching facilities in the FreeBSD kernel
starting from __FreeBSD_version 1400032.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Ka Ho Ng <khng@FreeBSD.org>
Sponsored-by: The FreeBSD Foundation
Closes#12458
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
Clang 13.0.0 added support for `Wunused-but-set-parameter` and
`-Wunused-but-set-variable` which correctly detects two unused
variables in zstd resulting in a build failure. This commit
annotates these instances accordingly.
https://releases.llvm.org/13.0.1/tools/clang/docs/ReleaseNotes.html#id6
In FSE_createCTable(), malloc() is intentionally defined as NULL when
compiled in the kernel so the variable is unused.
zstd/lib/compress/fse_compress.c:307:12: error: variable 'size'
set but not used [-Werror,-Wunused-but-set-variable]
Additionally, in ZSTD_seqDecompressedSize() the assert is compiled
out similarly resulting in an unused variable.
zstd/lib/compress/zstd_compress_superblock.c:412:12: error: variable
'litLengthSum' set but not used [-Werror,-Wunused-but-set-variable]
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Brian Behlendorf
наб
Tino Reichardt
Alexander Motin
Allan Jude
Ryan Moeller
Rich Ercolani
hping
Aidan Harris
Mateusz Guzik
szubersk
Brian Atkinson
Ka Ho Ng