In case of cache device removal it is possible that at the end of
l2arc_evict() we have l2ad_hand = l2ad_evict. This can lead to the
following panic in case of a debug build:
VERIFY3(dev->l2ad_hand < dev->l2ad_evict) failed (321920512 < 321920512)
Call Trace:
dump_stack+0x66/0x90
spl_panic+0xef/0x117 [spl]
l2arc_remove_vdev+0x11d/0x290 [zfs]
spa_load_l2cache+0x275/0x5b0 [zfs]
spa_vdev_remove+0x4a5/0x6e0 [zfs]
zfs_ioc_vdev_remove+0x59/0xa0 [zfs]
zfsdev_ioctl_common+0x5b3/0x630 [zfs]
zfsdev_ioctl+0x53/0xe0 [zfs]
do_vfs_ioctl+0x42e/0x6b0
ksys_ioctl+0x5e/0x90
do_syscall_64+0x5b/0x1a0
entry_SYSCALL_64_after_hwframe+0x44/0xa9
In case of cache device removal it also possible that l2ad_hand +
distance > l2ad_end since we do not iterate l2arc_evict() and l2ad_hand
is not reset. This has no functional consequence however as the cache
device is about to be removed.
Fix this by omitting the ASSERT in case of device removal.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#11205
After initial arc_c was reduced to arc_c_min it became possible that
on datasets with primarycache=metadata or none dirty data make up most
of ARC capacity and easily more than configured 50% of initial arc_c,
that causes forced txg commits by arc_tempreserve_space() and periodic
very long write delays.
This patch makes arc_tempreserve_space() to use arc_c only after ARC
warmed up once and arc_c really means something, but use arc_c_max
before that.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Matt Macy <mmacy@FreeBSD.org>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#11178
L2ARC devices of several terabytes filled with 4KB blocks may take 15
minutes to rebuild. Due to the way L2ARC log reading is implemented
it is quite likely that for all that time rebuild thread will never
sleep. At least on FreeBSD kernel threads have absolute priority and
can not be preempted by threads with lower priorities. If some thread
is also bound to that specific CPU it may not get any CPU time for all
the 15 minutes.
Reviewed-by: Cedric Berger <cedric@precidata.com>
Reviewed-by: Ryan Moeller <freqlabs@FreeBSD.org>
Reviewed-by: George Amanakis <gamanakis@gmail.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#11116
It's even documented already.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes#11094
The current l2_misses accounting behavior treats all reads to pools
without a configured l2arc as an l2arc miss, IFF there is at least
one other pool on the system which does have an l2arc configured.
This makes it extremely hard to tune for an improved l2arc hit/miss
ratio because this ratio will be modulated by reads from pools which
do not (and should not) have l2arc devices; its upper limit will
depend on the ratio of reads from l2arc'd pools and non-l2arc'd pools.
This PR prevents ARC reads affecting l2arc stats (n.b. l2_misses is
the only relevant one) where the target spa doesn't have an l2arc.
Includes new test - l2arc_l2miss_pos.ksh
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Amanakis <gamanakis@gmail.com>
Signed-off-by: Adam Moss <c@yotes.com>
Closes#10921
Instead of relying on arbitrary timers after pool export/import or cache
device off/online rely on arcstats. This makes the L2ARC tests more
robust. Also cleanup some functions related to persistent L2ARC.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Adam Moss <c@yotes.com>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#10983
In non regular use cases allocated memory might stay persistent in memory
pool. This small patch checks every minute if there are old objects which
can be released from memory pool.
Right now with regular use, the pool is checked for old objects on each
allocation attempt from this pool. so basically polling by its use. Now
consider what happens if someone writes a lot of files and stops use of
the volume or even unmounts it. So the code will no longer check if
objects can be released from the pool. Already allocated objects will
still stay in pool cache. this is no big issue for common use. But
someone discovered this issue while doing tests. personally i know this
behavior and I'm aware of it. Its no big issue. just a enhancement
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Kjeld Schouten-Lebbing <kjeld@schouten-lebbing.nl>
Signed-off-by: Sebastian Gottschall <s.gottschall@dd-wrt.com>
Closes#10938Closes#10969
Commit 45152dc removed clearing of L2CACHE flag in arc_read_done() and
moved related code in l2arc_write_eligible(). After careful code
inspection arc_read_done() is not bypassed in the case of prefetches.
Thus restore the old behavior.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: adam moss <c@yotes.com>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#10951
== Motivation and Context
The new vdev ashift optimization prevents the removal of devices when
a zfs configuration is comprised of disks which have different logical
and physical block sizes. This is caused because we set 'spa_min_ashift'
in vdev_open and then later call 'vdev_ashift_optimize'. This would
result in an inconsistency between spa's ashift calculations and that
of the top-level vdev.
In addition, the optimization logical ignores the overridden ashift
value that would be provided by '-o ashift=<val>'.
== Description
This change reworks the vdev ashift optimization so that it's only
set the first time the device is configured. It still allows the
physical and logical ahsift values to be set every time the device
is opened but those values are only consulted on first open.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Cedric Berger <cedric@precidata.com>
Signed-off-by: George Wilson <gwilson@delphix.com>
External-Issue: DLPX-71831
Closes#10932
Currently the ARC state (MFU/MRU) of cached L2ARC buffer and their
content type is unknown. Knowing this information may prove beneficial
in adjusting the L2ARC caching policy.
This commit adds L2ARC arcstats that display the aligned size
(in bytes) of L2ARC buffers according to their content type
(data/metadata) and according to their ARC state (MRU/MFU or
prefetch). It also expands the existing evict_l2_eligible arcstat to
differentiate between MFU and MRU buffers.
L2ARC caches buffers from the MRU and MFU lists of ARC. Upon caching a
buffer, its ARC state (MRU/MFU) is stored in the L2 header
(b_arcs_state). The l2_m{f,r}u_asize arcstats reflect the aligned size
(in bytes) of L2ARC buffers according to their ARC state (based on
b_arcs_state). We also account for the case where an L2ARC and ARC
cached MRU or MRU_ghost buffer transitions to MFU. The l2_prefetch_asize
reflects the alinged size (in bytes) of L2ARC buffers that were cached
while they had the prefetch flag set in ARC. This is dynamically updated
as the prefetch flag of L2ARC buffers changes.
When buffers are evicted from ARC, if they are determined to be L2ARC
eligible then their logical size is recorded in
evict_l2_eligible_m{r,f}u arcstats according to their ARC state upon
eviction.
Persistent L2ARC:
When committing an L2ARC buffer to a log block (L2ARC metadata) its
b_arcs_state and prefetch flag is also stored. If the buffer changes
its arcstate or prefetch flag this is reflected in the above arcstats.
However, the L2ARC metadata cannot currently be updated to reflect this
change.
Example: L2ARC caches an MRU buffer. L2ARC metadata and arcstats count
this as an MRU buffer. The buffer transitions to MFU. The arcstats are
updated to reflect this. Upon pool re-import or on/offlining the L2ARC
device the arcstats are cleared and the buffer will now be counted as an
MRU buffer, as the L2ARC metadata were not updated.
Bug fix:
- If l2arc_noprefetch is set, arc_read_done clears the L2CACHE flag of
an ARC buffer. However, prefetches may be issued in a way that
arc_read_done() is bypassed. Instead, move the related code in
l2arc_write_eligible() to account for those cases too.
Also add a test and update manpages for l2arc_mfuonly module parameter,
and update the manpages and code comments for l2arc_noprefetch.
Move persist_l2arc tests to l2arc.
Reviewed-by: Ryan Moeller <freqlabs@FreeBSD.org>
Reviewed-by: Richard Elling <Richard.Elling@RichardElling.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#10743
In certain workloads it may be beneficial to reduce wear of L2ARC
devices by not caching MRU metadata and data into L2ARC. This commit
introduces a new tunable l2arc_mfuonly for this purpose.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Richard Elling <Richard.Elling@RichardElling.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#10710
Duplicate io and checksum ereport events can misrepresent that
things are worse than they seem. Ideally the zpool events and the
corresponding vdev stat error counts in a zpool status should be
for unique errors -- not the same error being counted over and over.
This can be demonstrated in a simple example. With a single bad
block in a datafile and just 5 reads of the file we end up with a
degraded vdev, even though there is only one unique error in the pool.
The proposed solution to the above issue, is to eliminate duplicates
when posting events and when updating vdev error stats. We now save
recent error events of interest when posting events so that we can
easily check for duplicates when posting an error.
Reviewed by: Brad Lewis <brad.lewis@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Don Brady <don.brady@delphix.com>
Closes#10861
There are a number of places where cv_?_sig is used simply for
accounting purposes but the surrounding code has no ability to
cope with actually receiving a signal. On FreeBSD it is possible
to send signals to individual kernel threads so this could
enable undesirable behavior.
This patch adds routines on Linux that will do the same idle
accounting as _sig without making the task interruptible. On
FreeBSD cv_*_idle are all aliases for cv_*
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes#10843
use (void) to silence analyzers.
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Toomas Soome <tsoome@me.com>
Closes#10857
Since L2ARC buffers are not evicted on memory pressure, too large
amount of headers on system with irrationally large L2ARC can render
it slow or even unusable. This change limits L2ARC writes and
rebuild if unevictable L2ARC-only headers reach dangerous level.
While there, call arc_adapt() on L2ARC rebuild, so that it could
properly grow arc_c, reflecting potentially significant ARC size
increase and avoiding slow growth with hopeless eviction attempts
later when "overflow" is detected.
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reported-by: Richard Elling <Richard.Elling@RichardElling.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#10765
Many modern devices use physical allocation units that are much
larger than the minimum logical allocation size accessible by
external commands. Two prevalent examples of this are 512e disk
drives (512b logical sector, 4K physical sector) and flash devices
(512b logical sector, 4K or larger allocation block size, and 128k
or larger erase block size). Operations that modify less than the
physical sector size result in a costly read-modify-write or garbage
collection sequence on these devices.
Simply exporting the true physical sector of the device to ZFS would
yield optimal performance, but has two serious drawbacks:
1. Existing pools created with devices that have different logical
and physical block sizes, but were configured to use the logical
block size (e.g. because the OS version used for pool construction
reported the logical block size instead of the physical block
size) will suddenly find that the vdev allocation size has
increased. This can be easily tolerated for active members of
the array, but ZFS would prevent replacement of a vdev with
another identical device because it now appears that the smaller
allocation size required by the pool is not supported by the new
device.
2. The device's physical block size may be too large to be supported
by ZFS. The optimal allocation size for the vdev may be quite
large. For example, a RAID controller may export a vdev that
requires read-modify-write cycles unless accessed using 64k
aligned/sized requests. ZFS currently has an 8k minimum block
size limit.
Reporting both the logical and physical allocation sizes for vdevs
solves these problems. A device may be used so long as the logical
block size is compatible with the configuration. By comparing the
logical and physical block sizes, new configurations can be optimized
and administrators can be notified of any existing pools that are
sub-optimal.
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Co-authored-by: Matthew Macy <mmacy@freebsd.org>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes#10619
Removing other_size from arc_stats breaks top in 11.x jails
running on HEAD.
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes#10745
This PR adds two new compression types, based on ZStandard:
- zstd: A basic ZStandard compression algorithm Available compression.
Levels for zstd are zstd-1 through zstd-19, where the compression
increases with every level, but speed decreases.
- zstd-fast: A faster version of the ZStandard compression algorithm
zstd-fast is basically a "negative" level of zstd. The compression
decreases with every level, but speed increases.
Available compression levels for zstd-fast:
- zstd-fast-1 through zstd-fast-10
- zstd-fast-20 through zstd-fast-100 (in increments of 10)
- zstd-fast-500 and zstd-fast-1000
For more information check the man page.
Implementation details:
Rather than treat each level of zstd as a different algorithm (as was
done historically with gzip), the block pointer `enum zio_compress`
value is simply zstd for all levels, including zstd-fast, since they all
use the same decompression function.
The compress= property (a 64bit unsigned integer) uses the lower 7 bits
to store the compression algorithm (matching the number of bits used in
a block pointer, as the 8th bit was borrowed for embedded block
pointers). The upper bits are used to store the compression level.
It is necessary to be able to determine what compression level was used
when later reading a block back, so the concept used in LZ4, where the
first 32bits of the on-disk value are the size of the compressed data
(since the allocation is rounded up to the nearest ashift), was
extended, and we store the version of ZSTD and the level as well as the
compressed size. This value is returned when decompressing a block, so
that if the block needs to be recompressed (L2ARC, nop-write, etc), that
the same parameters will be used to result in the matching checksum.
All of the internal ZFS code ( `arc_buf_hdr_t`, `objset_t`,
`zio_prop_t`, etc.) uses the separated _compress and _complevel
variables. Only the properties ZAP contains the combined/bit-shifted
value. The combined value is split when the compression_changed_cb()
callback is called, and sets both objset members (os_compress and
os_complevel).
The userspace tools all use the combined/bit-shifted value.
Additional notes:
zdb can now also decode the ZSTD compression header (flag -Z) and
inspect the size, version and compression level saved in that header.
For each record, if it is ZSTD compressed, the parameters of the decoded
compression header get printed.
ZSTD is included with all current tests and new tests are added
as-needed.
Per-dataset feature flags now get activated when the property is set.
If a compression algorithm requires a feature flag, zfs activates the
feature when the property is set, rather than waiting for the first
block to be born. This is currently only used by zstd but can be
extended as needed.
Portions-Sponsored-By: The FreeBSD Foundation
Co-authored-by: Allan Jude <allanjude@freebsd.org>
Co-authored-by: Brian Behlendorf <behlendorf1@llnl.gov>
Co-authored-by: Sebastian Gottschall <s.gottschall@dd-wrt.com>
Co-authored-by: Kjeld Schouten-Lebbing <kjeld@schouten-lebbing.nl>
Co-authored-by: Michael Niewöhner <foss@mniewoehner.de>
Signed-off-by: Allan Jude <allan@klarasystems.com>
Signed-off-by: Allan Jude <allanjude@freebsd.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Sebastian Gottschall <s.gottschall@dd-wrt.com>
Signed-off-by: Kjeld Schouten-Lebbing <kjeld@schouten-lebbing.nl>
Signed-off-by: Michael Niewöhner <foss@mniewoehner.de>
Closes#6247Closes#9024Closes#10277Closes#10278
Commit 85ec5cbae updated abd_update_scatter_stats() such that it
calls arc_space_consume() and arc_space_return() when updating the
scatter stats. This requires that the global aggsum value for the
ARC be initialized. Normally this is not an issue, however during
module unload the l2arc_do_free_on_write() function was called in
l2arc_cleanup() after arc_state_fini() destroyed the aggsum values.
We can resolve this issue by performing l2arc_do_free_on_write()
slightly earlier in arc_fini().
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#10739
The ARC caches data in scatter ABD's, which are collections of pages,
which are typically 4K. Therefore, the space used to cache each block
is rounded up to a multiple of 4K. The ABD subsystem tracks this wasted
memory in the `scatter_chunk_waste` kstat. However, the ARC's `size` is
not aware of the memory used by this round-up, it only accounts for the
size that it requested from the ABD subsystem.
Therefore, the ARC is effectively using more memory than it is aware of,
due to the `scatter_chunk_waste`. This impacts observability, e.g.
`arcstat` will show that the ARC is using less memory than it
effectively is. It also impacts how the ARC responds to memory
pressure. As the amount of `scatter_chunk_waste` changes, it appears to
the ARC as memory pressure, so it needs to resize `arc_c`.
If the sector size (`1<<ashift`) is the same as the page size (or
larger), there won't be any waste. If the (compressed) block size is
relatively large compared to the page size, the amount of
`scatter_chunk_waste` will be small, so the problematic effects are
minimal.
However, if using 512B sectors (`ashift=9`), and the (compressed) block
size is small (e.g. `compression=on` with the default `volblocksize=8k`
or a decreased `recordsize`), the amount of `scatter_chunk_waste` can be
very large. On a production system, with `arc_size` at a constant 50%
of memory, `scatter_chunk_waste` has been been observed to be 10-30% of
memory.
This commit adds `scatter_chunk_waste` to `arc_size`, and adds a new
`waste` field to `arcstat`. As a result, the ARC's memory usage is more
observable, and `arc_c` does not need to be adjusted as frequently.
Reviewed-by: Pavel Zakharov <pavel.zakharov@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#10701
`thread_create` on FreeBSD stringifies the argument passed as the
thread function to create a name for the thread. The thread name for
`l2arc_dev_rebuild_start` ended up with `(void (*)(void *))` in it.
Change the type signature so the function does not need to be cast
when creating the thread. Rename the function to
`l2arc_dev_rebuild_thread` for clarity and consistency, as well.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Amanakis <gamanakis@gmail.com>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes#10716
When reading compressed blocks from the L2ARC, with
compressed ARC disabled, arc_hdr_size() returns
LSIZE rather than PSIZE, but the actual read is PSIZE.
This causes l2arc_read_done() to compare the checksum
against the wrong size, resulting in checksum failure.
This manifests as an increase in the kstat l2_cksum_bad
and the read being retried from the main pool, making the
L2ARC ineffective.
Add new L2ARC tests with Compressed ARC enabled/disabled
Blocks are handled differently depending on the state of the
zfs_compressed_arc_enabled tunable.
If a block is compressed on-disk, and compressed_arc is enabled:
- the block is read from disk
- It is NOT decompressed
- It is added to the ARC in its compressed form
- l2arc_write_buffers() may write it to the L2ARC (as is)
- l2arc_read_done() compares the checksum to the BP (compressed)
However, if compressed_arc is disabled:
- the block is read from disk
- It is decompressed
- It is added to the ARC (uncompressed)
- l2arc_write_buffers() will use l2arc_apply_transforms() to
recompress the block, before writing it to the L2ARC
- l2arc_read_done() compares the checksum to the BP (compressed)
- l2arc_read_done() will use l2arc_untransform() to uncompress it
This test writes out a test file to a pool consisting of one disk
and one cache device, then randomly reads from it. Since the arc_max
in the tests is low, this will feed the L2ARC, and result in reads
from the L2ARC.
We compare the value of the kstat l2_cksum_bad before and after
to determine if any blocks failed to survive the trip through the
L2ARC.
Sponsored-by: The FreeBSD Foundation
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Allan Jude <allanjude@freebsd.org>
Closes#10693
The arc_adapt() function tunes LRU/MLU balance according to 4 types of
cache hits (which is passed as state agrument): ghost LRU, LRU, MRU,
ghost MRU. If this function is called with wrong cache hit (state),
adaptation will be sub-optimal and performance will suffer.
Some time ago upstream received this commit:
6950 ARC should cache compressed data) in arc_read() do next
sequence (access to ghost buffer)
Before this commit, hit to any ghost list was passed arc_adapt() before
call to arc_access() which revive element in cache and change state from
ghost to real hit.
After this commit, the order of calls was reverted and arc_adapt() is
now called only with «real» hits even if hit was in one of two ghost
lists, which renders ghost lists useless and breaks the ARC algorithm.
FreeBSD fixed this problem locally in Change D19094 / Commit r348772.
This change is an adaptation of the above commit to the current arc
code.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes#10548Closes#10618
Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Allan Jude <allanjude@freebsd.org>
Closes#10694
In case the L2ARC rebuild was canceled, do not log to spa history
log as the pool may be in the process of being removed and a panic
may occur:
BUG: kernel NULL pointer dereference, address: 0000000000000018
RIP: 0010:spa_history_log_internal+0xb1/0x120 [zfs]
Call Trace:
l2arc_rebuild+0x464/0x7c0 [zfs]
l2arc_dev_rebuild_start+0x2d/0x130 [zfs]
? l2arc_rebuild+0x7c0/0x7c0 [zfs]
thread_generic_wrapper+0x78/0xb0 [spl]
kthread+0xfb/0x130
? IS_ERR+0x10/0x10 [spl]
? kthread_park+0x90/0x90
ret_from_fork+0x35/0x40
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#10659
A collection of header changes to enable FreeBSD to build
with vendored OpenZFS.
Reviewed-by: Ryan Moeller <ryan@ixsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes#10635
The ARC shrinker callback `arc_shrinker_count/_scan()` is invoked by the
kernel's shrinker mechanism when the system is running low on free
pages. This happens via 2 code paths:
1. "direct reclaim": The system is attempting to allocate a page, but we
are low on memory. The ARC shrinker callback is invoked from the
page-allocation code path.
2. "indirect reclaim": kswapd notices that there aren't many free pages,
so it invokes the ARC shrinker callback.
In both cases, the kernel's shrinker code requests that the ARC shrinker
callback release some of its cache, and then it measures how many pages
were released. However, it's measurement of released pages does not
include pages that are freed via `__free_pages()`, which is how the ARC
releases memory (via `abd_free_chunks()`). Rather, the kernel shrinker
code is looking for pages to be placed on the lists of reclaimable pages
(which is separate from actually-free pages).
Because the kernel shrinker code doesn't detect that the ARC has
released pages, it may call the ARC shrinker callback many times,
resulting in the ARC "collapsing" down to `arc_c_min`. This has several
negative impacts:
1. ZFS doesn't use RAM to cache data effectively.
2. In the direct reclaim case, a single page allocation may wait a long
time (e.g. more than a minute) while we evict the entire ARC.
3. Even with the improvements made in 67c0f0dedc ("ARC shrinking blocks
reads/writes"), occasionally `arc_size` may stay above `arc_c` for the
entire time of the ARC collapse, thus blocking ZFS read/write operations
in `arc_get_data_impl()`.
To address these issues, this commit limits the ways that the ARC
shrinker callback can be used by the kernel shrinker code, and mitigates
the impact of arc_is_overflowing() on ZFS read/write operations.
With this commit:
1. We limit the amount of data that can be reclaimed from the ARC via
the "direct reclaim" shrinker. This limits the amount of time it takes
to allocate a single page.
2. We do not allow the ARC to shrink via kswapd (indirect reclaim).
Instead we rely on `arc_evict_zthr` to monitor free memory and reduce
the ARC target size to keep sufficient free memory in the system. Note
that we can't simply rely on limiting the amount that we reclaim at once
(as for the direct reclaim case), because kswapd's "boosted" logic can
invoke the callback an unlimited number of times (see
`balance_pgdat()`).
3. When `arc_is_overflowing()` and we want to allocate memory,
`arc_get_data_impl()` will wait only for a multiple of the requested
amount of data to be evicted, rather than waiting for the ARC to no
longer be overflowing. This allows ZFS reads/writes to make progress
even while the ARC is overflowing, while also ensuring that the eviction
thread makes progress towards reducing the total amount of memory used
by the ARC.
4. The amount of memory that the ARC always tries to keep free for the
rest of the system, `arc_sys_free` is increased.
5. Now that the shrinker callback is able to provide feedback to the
kernel's shrinker code about our progress, we can safely enable
the kswapd hook. This will allow the arc to receive notifications
when memory pressure is first detected by the kernel. We also
re-enable the appropriate kstats to track these callbacks.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Co-authored-by: George Wilson <george.wilson@delphix.com>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#10600
Renamed to avoid conflicting with refcount.h when a different
implementation is already provided by the platform.
Reviewed-by: Ryan Moeller <ryan@ixsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes#10620
When debugging issues or generally analyzing the runtime of
a system it would be nice to be able to tell the different
ZTHRs running by name rather than having to analyze their
stack.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Co-authored-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Closes#10630
The process of evicting data from the ARC is referred to as
`arc_adjust`.
This commit changes the term to `arc_evict`, which is more specific.
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#10592
The SPL kmem_cache implementation provides a mechanism, `skc_reclaim`,
whereby individual caches can register a callback to be invoked when
there is memory pressure. This mechanism is used in only one place: the
ARC registers the `hdr_recl()` reclaim function. This function wakes up
the `arc_reap_zthr`, whose job is to call `kmem_cache_reap()` and
`arc_reduce_target_size()`.
The `skc_reclaim` callbacks are invoked only by shrinker callbacks and
`arc_reap_zthr`, and only callback only wakes up `arc_reap_zthr`. When
called from `arc_reap_zthr`, waking `arc_reap_zthr` is a no-op. When
called from shrinker callbacks, we are already aware of memory pressure
and responding to it. Therefore there is little benefit to ever calling
the `hdr_recl()` `skc_reclaim` callback.
The `arc_reap_zthr` also wakes once a second, and if memory is low when
allocating an ARC buffer. Therefore, additionally waking it from the
shrinker calbacks has little benefit.
The shrinker callbacks can be invoked very frequently, e.g. 10,000 times
per second. Additionally, for invocation of the shrinker callback,
skc_reclaim is invoked many times. Therefore, this mechanism consumes
significant amounts of CPU time.
The kmem_cache shrinker calls `spl_kmem_cache_reap_now()`, which,
in addition to invoking `skc_reclaim()`, does two things to attempt to
free pages for use by the system:
1. Return free objects from the magazine layer to the slab layer
2. Return entirely-free slabs to the page layer (i.e. free pages)
These actions apply only to caches implemented by the SPL, not those
that use the underlying kernel SLAB/SLUB caches. The SPL caches are
used for objects >=32KB, which are primarily linear ABD's cached in the
DBUF cache.
These actions (freeing objects from the magazine layer and returning
entirely-free slabs) are also taken whenever a `kmem_cache_free()` call
finds a full magazine. So there would typically be zero entirely-free
slabs, and the number of objects in magazines is limited (typically no
more than 64 objects per magazine, and there's one magazine per CPU).
Therefore the benefit of `spl_kmem_cache_reap_now()`, while nonzero, is
modest.
We also call `spl_kmem_cache_reap_now()` from the `arc_reap_zthr`, when
memory pressure is detected. Therefore, calling
`spl_kmem_cache_reap_now()` from the kmem_cache shrinker is not needed.
This commit removes the `skc_reclaim` mechanism, its only callback
`hdr_recl()`, and the kmem_cache shrinker callback.
Reviewed-By: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Pavel Zakharov <pavel.zakharov@delphix.com>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#10576
In case l2arc_write_done() handles a zio that was not successful check
that the list of log block pointers is not empty when restoring them
in the device header. Otherwise zero them out. In any case perform the
actual write updating the device header after the zio of
l2arc_write_buffers() completes as l2arc_write_done() may have touched
the memory holding the log block pointers in the device header.
Reviewed-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#10540Closes#10543
OS-specific code (e.g. under `module/os/linux`) does not need to share
its code structure with any other operating systems. In particular, the
ARC and kmem code need not be similar to the code in illumos, because we
won't be syncing this OS-specific code between operating systems. For
example, if/when illumos support is added to the common repo, we would
add a file `module/os/illumos/zfs/arc_os.c` for the illumos versions of
this code.
Therefore, we can simplify the code in the OS-specific ARC and kmem
routines.
These changes do not impact system behavior, they are purely code
cleanup. The changes are:
Arenas are not used on Linux or FreeBSD (they are always `NULL`), so
`heap_arena`, `zio_arena`, and `zio_alloc_arena` can be removed, along
with code that uses them.
In `arc_available_memory()`:
* `desfree` is unused, remove it
* rename `freemem` to avoid conflict with pre-existing `#define`
* remove checks related to arenas
* use units of bytes, rather than converting from bytes to pages and
then back to bytes
`SPL_KMEM_CACHE_REAP` is unused, remove it.
`skc_reap` is unused, remove it.
The `count` argument to `spl_kmem_cache_reap_now()` is unused, remove
it.
`vmem_size()` and associated type and macros are unused, remove them.
In `arc_memory_throttle()`, use a less confusing variable name to store
the result of `arc_free_memory()`.
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Pavel Zakharov <pavel.zakharov@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@ixsystems.com>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#10499
ZFS registers a memory hook, `__arc_shrinker_func`, which is supposed to
allow the ARC to shrink when the kernel experiences memory pressure.
The ARC shrinker changes `arc_c` via a call to
`arc_reduce_target_size()`. Before commit 3ec34e5527, the ARC
shrinker would also evict data from the ARC to bring `arc_size` down to
the new `arc_c`. However, that commit (seemingly inadvertently) made it
so that the ARC shrinker no longer evicts any data or waits for eviction
to complete.
Repeated calls to the ARC shrinker can reduce `arc_c` drastically, often
all the way to `arc_c_min`. Since it doesn't wait for the actual
eviction of data from the ARC, this creates a situation where `arc_size`
is more than `arc_c` for the several seconds/minutes it takes for
`arc_adjust_zthr` to evict data from the ARC. During this time,
arc_get_data_impl() will block, so ZFS can't process read/write requests
(e.g. from iSCSI, NFS, or read/write syscalls).
To ensure that `arc_c` doesn't shrink faster than the adjust thread can
keep up, this commit makes the ARC shrinker wait for the eviction to
complete, resulting in similar behavior to what we had before commit
3ec34e5527.
Note: commit 3ec34e5527 is `OpenZFS 9284 - arc_reclaim_thread
has 2 jobs` and was integrated in December 2018, and is part of ZoL
0.8.x but not 0.7.x.
Additionally, when the ARC size is reduced drastically, the
`arc_adjust_zthr` can be on-CPU for many seconds without blocking. Any
threads that are bound to the same CPU that arc_adjust_zthr is running
on will not able to run for a long time.
To ensure that CPU-bound threads can make progress, this commit changes
`arc_evict_state_impl()` make a voluntary preemption call,
`cond_resched()`.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Prakash Surya <prakash.surya@delphix.com>
Reviewed-by: Pavel Zakharov <pavel.zakharov@delphix.com>
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
External-issue: DLPX-70703
Closes#10496
Mark functions used only in the same translation unit as static. This
only includes functions that do not have a prototype in a header file
either.
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Arvind Sankar <nivedita@alum.mit.edu>
Closes#10470
Apparently missed in the initial port integration was
the need to reap the abd_chunk_cache on FreeBSD. This
change addresses that oversight.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes#10474
For at least 15 years since OpenSolaris arc_c was set by default to
arc_c_max, later decreased under memory pressure. I've noticed that
if arc_c was set high enough to cause memory pressure as considered
by ZFS, setting of arc_no_grow to TRUE in arc_reap_cb_check() makes
no effect until both arc_kmem_reap_soon() and delay(reap_retry_ms)
return. All that time ZFS can continue increasing its effective ARC
size, causing more memory pressure, potentially up to the point when
OS low memory handler activates and reduces arc_c, requesting fast
reclamation of just allocated memory.
The problem seems to be more serious on FreeBSD and I guess Linux,
since neither of them implement/use asynchronous kmem reclamation,
so arc_kmem_reap_soon() can take more time. On older FreeBSD 11 not
supporting multiple memory domains system with lots of RAM can get
completely unresponsive for minutes due to heavy lock congestion
between ARC reclamation and page daemon kmem reclamation threads.
With this change to more conservative arc_c value ARC stops growing
just it time and does not need later reclamation.
Also while there, since now growing arc_c is a more often situation,
use aggsum_upper_bound() instead of aggsum_compare() in arc_adapt()
to reduce lock congestion. It is also getting in sync with code in
arc_get_data_impl().
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Allan Jude <allanjude@freebsd.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#10437
Correct various typos in the comments and tests.
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Andrea Gelmini <andrea.gelmini@gelma.net>
Closes#10423
The l2arc_evict() function is responsible for evicting buffers which
reference the next bytes of the L2ARC device to be overwritten. Teach
this function to additionally TRIM that vdev space before it is
overwritten if the device has been filled with data. This is done by
vdev_trim_simple() which trims by issuing a new type of TRIM,
TRIM_TYPE_SIMPLE.
We also implement a "Trim Ahead" feature. It is a zfs module parameter,
expressed in % of the current write size. This trims ahead of the
current write size. A minimum of 64MB will be trimmed. The default is 0
which disables TRIM on L2ARC as it can put significant stress to
underlying storage devices. To enable TRIM on L2ARC we set
l2arc_trim_ahead > 0.
We also implement TRIM of the whole cache device upon addition to a
pool, pool creation or when the header of the device is invalid upon
importing a pool or onlining a cache device. This is dependent on
l2arc_trim_ahead > 0. TRIM of the whole device is done with
TRIM_TYPE_MANUAL so that its status can be monitored by zpool status -t.
We save the TRIM state for the whole device and the time of completion
on-disk in the header, and restore these upon L2ARC rebuild so that
zpool status -t can correctly report them. Whole device TRIM is done
asynchronously so that the user can export of the pool or remove the
cache device while it is trimming (ie if it is too slow).
We do not TRIM the whole device if persistent L2ARC has been disabled by
l2arc_rebuild_enabled = 0 because we may not want to lose all cached
buffers (eg we may want to import the pool with
l2arc_rebuild_enabled = 0 only once because of memory pressure). If
persistent L2ARC has been disabled by setting the module parameter
l2arc_rebuild_blocks_min_l2size to a value greater than the size of the
cache device then the whole device is trimmed upon creation or import of
a pool if l2arc_trim_ahead > 0.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Adam D. Moss <c@yotes.com>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#9713Closes#9789Closes#10224
FreeBSD needs arc_adjust_zthr to run periodically for kstats to be
updated. A comment in the code suggests this may have been the
original intent in illumos as well:
c946d5a913/module/zfs/arc.c (L4697-L4700)
Create the thread with a 1 second timer.
Reviewed-by: Matt Macy <mmacy@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes#10371
Due to hotplug support or BIOS bugs sometimes max_ncpus can be
an absurdly high value. I have a system with 32 cores/threads
but reports max_ncpus == 440. This many threads potentially
cripples the system during arc_prune floods for example.
boot_ncpus is the number of working CPUs when called so use
that instead.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: DHE <git@dehacked.net>
Closes#10282
Functional changes:
We implement refcounts of log blocks and their aligned size on the
cache device along with two corresponding arcstats. The refcounts are
reflected in the header of the device and provide valuable information
as to whether log blocks are accounted for correctly. These are
dynamically adjusted as log blocks are committed/evicted. zdb also uses
this information in the device header and compares it to the
corresponding values as reported by dump_l2arc_log_blocks() which
emulates l2arc_rebuild(). If the refcounts saved in the device header
report higher values, zdb exits with an error. For this feature to work
correctly there should be no active writes on the device. This is also
employed in the tests of persistent L2ARC. We extend the structure of
the cache device header by adding the two new variables mirroring the
refcounts after the existing variables to preserve backward
compatibility in terms of persistent L2ARC.
1) a new arcstat "l2_log_blk_asize" and refcount "l2ad_lb_asize" which
reflect the total aligned size of log blocks on the device. This is
also reflected in the header of the cache device as "dh_lb_asize".
2) a new arcstat "l2arc_log_blk_count" and refcount "l2ad_lb_count"
which reflect the total number of L2ARC log blocks present on cache
devices. It is also reflected in the header of the cache device as
"dh_lb_count".
In l2arc_rebuild_vdev() if the amount of committed log entries in a log
block is 0 and the device header is valid we update the device header.
This will facilitate trimming of the whole device in this case when
TRIM for L2ARC is implemented.
Improve loop protection in l2arc_rebuild() by using the starting offset
of the payload of each log block instead of the starting offset of the
log block.
If the zio in l2arc_write_buffers() fails, restore the lbps array in the
header of the device to its previous state in l2arc_write_done().
If l2arc_rebuild() ends the rebuild process without restoring any L2ARC
log blocks in ARC and without any other error, this means that the lbps
array in the header is pointing to non-existent or invalid log blocks.
Reset the device header in this case.
In l2arc_rebuild() change the zfs_dbgmsg messages to
spa_history_log_internal() making them user visible with zpool history
command.
Non-functional changes:
Make the first test in persistent L2ARC use `zdb -lll` to increase
coverage in `zdb.c`.
Rename psize with asize when referring to log blocks, since
L2ARC_SET_PSIZE stores the vdev aligned size for log blocks. Also
rename dh_log_blk_entries to dh_log_entries to make it clear that
it is a mirror of l2ad_log_entries. Added comments for both changes.
Fix inaccurate comments for example in l2arc_log_blk_restore().
Add asserts at the end in l2arc_evict() and l2arc_write_buffers().
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#10228
Minor fixes on persistent L2ARC improving code readability and fixing
a typo in zdb.c when byte-swapping a log block. It also improves the
pesist_l2arc_007_pos.ksh test by giving it more time to retrieve log
blocks on the cache device.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Adam D. Moss <c@yotes.com>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#10210
This commit makes the L2ARC persistent across reboots. We implement
a light-weight persistent L2ARC metadata structure that allows L2ARC
contents to be recovered after a reboot. This significantly eases the
impact a reboot has on read performance on systems with large caches.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Co-authored-by: Saso Kiselkov <skiselkov@gmail.com>
Co-authored-by: Jorgen Lundman <lundman@lundman.net>
Co-authored-by: George Amanakis <gamanakis@gmail.com>
Ported-by: Yuxuan Shui <yshuiv7@gmail.com>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#925Closes#1823Closes#2672Closes#3744Closes#9582
Set arc_c_min before arc_c_max so that when zfs_arc_min is set lower
than the default allmem/32 zfs_arc_max can also be set lower.
Add warning messages when tunables are being ignored.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes#10157Closes#10158
Increasing l2arc_write_size or l2arc_write_boost can result in
l2arc_write_buffers() not having enough space to perform its writes and
panic zio_write_phys().
Instead of resetting l2ad_hand to l2ad_start at the end of
l2arc_write_buffers() and not taking into account a possible
user-mediated increase of l2arc_write_max, we do this in l2arc_evict(),
right after l2arc_write_size() has run. If there is not enough space to
evict (ie we will exceed l2ad_end) we evict to the end of the device,
reset l2ad_hand to l2ad_start, set l2ad_first to 0 and iterate
l2arc_evict(). We avoid infinite iteration of l2arc_evict() by making
sure in l2arc_write_size() that l2ad_start + size does not exceed
l2ad_end.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#10154
Linux changed the default max ARC size to 1/2 of physical memory to
deal with shortcomings of the Linux SLUB allocator. Other platforms
do not require the same logic.
Implement an arc_default_max() function to determine a default max ARC
size in platform code.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Closes#10155
Make the cityhash code compile into libzfs, in preparation for the new
"zstream" command.
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#10152
Using zfs with Lustre, an arc_read can trigger kernel memory allocation
that in turn leads to a memory reclaim callback and a deadlock within a
single zfs process. This change uses spl_fstrans_mark and
spl_trans_unmark to prevent the reclaim attempt and the deadlock
(https://zfsonlinux.topicbox.com/groups/zfs-devel/T4db2c705ec1804ba).
The stack trace observed is:
__schedule at ffffffff81610f2e
schedule at ffffffff81611558
schedule_preempt_disabled at ffffffff8161184a
__mutex_lock at ffffffff816131e8
arc_buf_destroy at ffffffffa0bf37d7 [zfs]
dbuf_destroy at ffffffffa0bfa6fe [zfs]
dbuf_evict_one at ffffffffa0bfaa96 [zfs]
dbuf_rele_and_unlock at ffffffffa0bfa561 [zfs]
dbuf_rele_and_unlock at ffffffffa0bfa32b [zfs]
osd_object_delete at ffffffffa0b64ecc [osd_zfs]
lu_object_free at ffffffffa06d6a74 [obdclass]
lu_site_purge_objects at ffffffffa06d7fc1 [obdclass]
lu_cache_shrink_scan at ffffffffa06d81b8 [obdclass]
shrink_slab at ffffffff811ca9d8
shrink_node at ffffffff811cfd94
do_try_to_free_pages at ffffffff811cfe63
try_to_free_pages at ffffffff811d01c4
__alloc_pages_slowpath at ffffffff811be7f2
__alloc_pages_nodemask at ffffffff811bf3ed
new_slab at ffffffff81226304
___slab_alloc at ffffffff812272ab
__slab_alloc at ffffffff8122740c
kmem_cache_alloc at ffffffff81227578
spl_kmem_cache_alloc at ffffffffa048a1fd [spl]
arc_buf_alloc_impl at ffffffffa0befba2 [zfs]
arc_read at ffffffffa0bf0924 [zfs]
dbuf_read at ffffffffa0bf9083 [zfs]
dmu_buf_hold_by_dnode at ffffffffa0c04869 [zfs]
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Mark Roper <markroper@gmail.com>
Closes#9987
When doing a zfs send on a dataset with small recordsize (e.g. 8K),
performance is dominated by the per-block overheads. This is especially
true with `zfs send --compressed`, which further reduces the amount of
data sent, for the same number of blocks. Several threads are involved,
but the limiting factor is the `send_prefetch` thread, which is 100% on
CPU.
The main job of the `send_prefetch` thread is to issue zio's for the
data that will be needed by the main thread. It does this by calling
`arc_read(ARC_FLAG_PREFETCH)`. This has an immediate cost of creating
an arc_hdr, which takes around 14% of one CPU. It also induces later
costs by other threads:
* Since the data was only prefetched, dmu_send()->dmu_dump_write() will
need to call arc_read() again to get the data. This will have to
look up the arc_hdr in the hash table and copy the data from the
scatter ABD in the arc_hdr to a linear ABD in arc_buf. This takes
27% of one CPU.
* dmu_dump_write() needs to arc_buf_destroy() This takes 11% of one
CPU.
* arc_adjust() will need to evict this arc_hdr, taking about 50% of one
CPU.
All of these costs can be avoided by bypassing the ARC if the data is
not already cached. This commit changes `zfs send` to check for the
data in the ARC, and if it is not found then we directly call
`zio_read()`, reading the data into a linear ABD which is used by
dmu_dump_write() directly.
The performance improvement is best expressed in terms of how many
blocks can be processed by `zfs send` in one second. This change
increases the metric by 50%, from ~100,000 to ~150,000. When the amount
of data per block is small (e.g. 2KB), there is a corresponding
reduction in the elapsed time of `zfs send >/dev/null` (from 86 minutes
to 58 minutes in this test case).
In addition to improving the performance of `zfs send`, this change
makes `zfs send` not pollute the ARC cache. In most cases the data will
not be reused, so this allows us to keep caching useful data in the MRU
(hit-once) part of the ARC.
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#10067
