zfs_arc_shrinker_limit (default: 10000) avoids ARC collapse
due to excessive memory reclaim. However, when the kernel is
in direct reclaim mode (ie: low on memory), limiting ARC reclaim
increases OOM risk. This is especially true on system without
(or with inadequate) swap.
This patch ignores zfs_arc_shrinker_limit when the kernel is in
direct reclaim mode, avoiding most OOM. It also restores
"echo 3 > /proc/sys/vm/drop_caches" ability to correctly drop
(almost) all ARC.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Adam Moss <c@yotes.com>
Signed-off-by: Gionatan Danti <g.danti@assyoma.it>
Closes#16313
These had minimal useful information for the admin, didn't work properly
in some places, and knew far too much about taskq internals.
With the new stats available, these should never be needed anymore.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Sponsored-by: Klara, Inc.
Sponsored-by: Syneto
Closes#16171
`l2arc_mfuonly` was added to avoid wasting L2 ARC on read-once MRU
data and metadata. However it can be useful to cache as much
metadata as possible while, at the same time, restricting data
cache to MFU buffers only.
This patch allow for such behavior by setting `l2arc_mfuonly` to 2
(or higher). The list of possible values is the following:
0: cache both MRU and MFU for both data and metadata;
1: cache only MFU for both data and metadata;
2: cache both MRU and MFU for metadata, but only MFU for data.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Gionatan Danti <g.danti@assyoma.it>
Closes#16343Closes#16402
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Allan Jude <allan@klarasystems.com>
Closes#15895
Adds a log/journal to dedup. At the end of txg, instead of writing the
entry directly to the ZAP, instead its adding to an in-memory tree and
appended to an on-disk object. The on-disk object is only read at
import, to reload the in-memory tree.
Lookups first go the the log tree before going to the ZAP, so
recently-used entries will remain close by in memory. This vastly
reduces overhead from dedup IO, as it will not have to do so many
read/update/write cycles on ZAP leaf nodes.
A flushing facility is added at end of txg, to push logged entries out
to the ZAP. There's actually two separate "logs" (in-memory tree and
on-disk object), one active (recieving updated entries) and one flushing
(writing out to disk). These are swapped (ie flushing begins) based on
memory used by the in-memory log trees and time since we last flushed
something.
The flushing facility monitors the amount of entries coming in and being
flushed out, and calibrates itself to try to flush enough each txg to
keep up with the ingest rate without competing too much with other IO.
Multiple tuneables are provided to control the flushing facility.
All the histograms and stats are update to accomodate the log as a
separate entry store. zdb gains knowledge of how to count them and dump
them. Documentation included!
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Co-authored-by: Allan Jude <allan@klarasystems.com>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Sponsored-by: Klara, Inc.
Sponsored-by: iXsystems, Inc.
Closes#15895
- When receiving memory pressure signal from OS be more strict
trying to free some memory. Otherwise kernel may come again and
request much more. Return as result how much arc_c was actually
reduced due to this request, that may be less than requested.
- On Linux when receiving direct reclaim from some file system
(that may be ZFS) instead of ignoring request completely, just
shrink the ARC, but do not wait for eviction. Waiting there may
cause deadlock. Ignoring it as before may put extra pressure on
other caches and/or swap, and cause OOM if nothing help. While
not waiting may result in more ARC evicted later, and may be too
late if OOM killer activate right now, but I hope it to be better
than doing nothing at all.
- On Linux set arc_no_grow before waiting for reclaim, not after,
or it may grow back while we are waiting.
- On Linux add new parameter zfs_arc_shrinker_seeks to balance
ARC eviction cost, relative to page cache and other subsystems.
- Slightly update Linux arc_set_sys_free() math for new kernels.
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Reviewed-by: Rob Norris <rob.norris@klarasystems.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Signed-off-by: Mateusz Piotrowski <0mp@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Optionally turn off disk's enclosure slot if an I/O is hung
triggering the deadman.
It's possible for outstanding I/O to a misbehaving SCSI disk to
neither promptly complete or return an error. This can occur due
to retry and recovery actions taken by the SCSI layer, driver, or
disk. When it occurs the pool will be unresponsive even though
there may be sufficient redundancy configured to proceeded without
this single disk.
When a hung I/O is detected by the kmods it will be posted as a
deadman event. By default an I/O is considered to be hung after
5 minutes. This value can be changed with the zfs_deadman_ziotime_ms
module parameter. If ZED_POWER_OFF_ENCLOSURE_SLOT_ON_DEADMAN is set
the disk's enclosure slot will be powered off causing the outstanding
I/O to fail. The ZED will then handle this like a normal disk failure.
By default ZED_POWER_OFF_ENCLOSURE_SLOT_ON_DEADMAN is not set.
As part of this change `zfs_deadman_events_per_second` is added
to control the ratelimitting of deadman events independantly of
delay events. In practice, a single deadman event is sufficient
and more aren't particularly useful.
Alphabetize the zfs_deadman_* entries in zfs.4.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#16226
High priority threads are handling ZIL writes. While there is no
ZIL compression, there is encryption, checksuming and RAIDZ math.
We've found that on large systems 1 taskq with 5 threads can be
a bottleneck for throughput, IOPS or both. Instead of just bumping
number of threads with a risk of overloading CPUs and increasing
latency, switch to using TQ_FRONT mechanism to increase sync write
requests priority within standard write threads. Do not do it on
Illumos, since its TQ_FRONT implementation is inherently unfair.
FreeBSD and Linux don't have this problem, so we can do it there.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Rob Norris <robn@despairlabs.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#16146
- Reduce number of allocators on small system down to one per 4
CPU cores, keeping maximum at 4 on 16+ core systems. Small systems
should not have the lock contention multiple allocators supposed
to solve, while having several metaslabs open and modified each
TXG is not free.
- Reduce number of write issue taskqs down to one per 16 CPU
cores and an integer fraction of number of allocators. On mid-
sized systems, where multiple allocators already make sense, too
many write issue taskqs may reduce write speed on single-file
workloads, since single file is handled by only one taskq to
reduce fragmentation. On large systems, that can actually benefit
from many taskq's better IOPS, the bottleneck is less important,
since in worst case there will be at least 16 cores to handle it.
- Distribute dnodes between allocators (and taskqs) in a round-
robin fashion instead of relying on sync taskqs to be balanced.
The last is not guarantied and may depend on scheduling.
- Remove io_wr_iss_tq from struct zio. io_allocator is enough.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16130
This allows ZAPs to shrink. When there are two empty sibling leafs,
one of them is collapsed and its storage space is reused.
This improved performance on directories that at one time contained
a large number of files, but many or all of those files have since
been deleted.
This also applies to all other types of ZAPs as well.
Sponsored-by: iXsystems, Inc.
Sponsored-by: Klara, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Alexander Stetsenko <alex.stetsenko@klarasystems.com>
Closes#15888
There is no reason for these module parameters to be read-only.
Being modified they just apply on next pool import/creation, that
is useful for testing different values.
Reviewed-by: Rich Ercolani <rincebrain@gmail.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16118
Before this change speculative prefetcher was able to detect a stream
only if all of its accesses are perfectly sequential. It was easy to
implement and is perfectly fine for single-threaded applications.
Unfortunately multi-threaded network servers, such as iSCSI, SMB or
NFS usually have plenty of threads and may often reorder requests,
preventing successful speculation and prefetch.
This change allows speculative prefetcher to detect streams even if
requests are reordered by introducing a list of 9 non-contiguous
ranges up to 16MB ahead of current stream position and filling the
gaps as more requests arrive. It also allows stream to proceed
even with holes up to a certain configurable threshold (25%).
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16022
Currently, zvol uses a single taskq, resulting in throughput bottleneck
under heavy load due to lock contention on the single taskq. This patch
addresses the performance bottleneck under heavy load conditions by
utilizing multiple taskqs, thus mitigating lock contention. The number
of taskqs scale dynamically based on the available CPUs in the system,
as illustrated below:
taskq total
cpus taskqs threads threads
------- ------- ------- -------
1 1 32 32
2 1 32 32
4 1 32 32
8 2 16 32
16 3 11 33
32 5 7 35
64 8 8 64
128 11 12 132
256 16 16 256
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Signed-off-by: Ameer Hamza <ahamza@ixsystems.com>
Closes#15992
This makes the submission method selectable at module load time via the
`zfs_vdev_disk_classic` parameter, allowing this change to be backported
to 2.2 safely, and disabled in favour of the "classic" submission method
if new problems come up.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Closes#15533Closes#15588
This commit tackles a number of issues in the way BIOs (`struct bio`)
are constructed for submission to the Linux block layer.
The kernel has a hard upper limit on the number of pages/segments that
can be added to a BIO, as well as a separate limit for each device
(related to its queue depth and other scheduling characteristics).
ZFS counts the number of memory pages in the request ABD
(`abd_nr_pages_off()`, and then uses that as the number of segments to
put into the BIO, up to the hard upper limit. If it requires more than
the limit, it will create multiple BIOs.
Leaving aside the fact that page count method is wrong (see below), not
limiting to the device segment max means that the device driver will
need to split the BIO in half. This is alone is not necessarily a
problem, but it interacts with another issue to cause a much larger
problem.
The kernel function to add a segment to a BIO (`bio_add_page()`) takes a
`struct page` pointer, and offset+len within it. `struct page` can
represent a run of contiguous memory pages (known as a "compound page").
In can be of arbitrary length.
The ZFS functions that count ABD pages and load them into the BIO
(`abd_nr_pages_off()`, `bio_map()` and `abd_bio_map_off()`) will never
consider a page to be more than `PAGE_SIZE` (4K), even if the `struct
page` is for multiple pages. In this case, it will load the same `struct
page` into the BIO multiple times, with the offset adjusted each time.
With a sufficiently large ABD, this can easily lead to the BIO being
entirely filled much earlier than it could have been. This is also
further contributes to the problem caused by the incorrect segment limit
calculation, as its much easier to go past the device limit, and so
require a split.
Again, this is not a problem on its own.
The logic for "never submit more than `PAGE_SIZE`" is actually a little
more subtle. It will actually never submit a buffer that crosses a 4K
page boundary.
In practice, this is fine, as most ABDs are scattered, that is a list of
complete 4K pages, and so are loaded in as such.
Linear ABDs are typically allocated from slabs, and for small sizes they
are frequently not aligned to page boundaries. For example, a 12K
allocation can span four pages, eg:
-- 4K -- -- 4K -- -- 4K -- -- 4K --
| | | | |
:## ######## ######## ######: [1K, 4K, 4K, 3K]
Such an allocation would be loaded into a BIO as you see:
[1K, 4K, 4K, 3K]
This tends not to be a problem in practice, because even if the BIO were
filled and needed to be split, each half would still have either a start
or end aligned to the logical block size of the device (assuming 4K at
least).
---
In ideal circumstances, these shortcomings don't cause any particular
problems. Its when they start to interact with other ZFS features that
things get interesting.
Aggregation will create a "gang" ABD, which is simply a list of other
ABDs. Iterating over a gang ABD is just iterating over each ABD within
it in turn.
Because the segments are simply loaded in order, we can end up with
uneven segments either side of the "gap" between the two ABDs. For
example, two 12K ABDs might be aggregated and then loaded as:
[1K, 4K, 4K, 3K, 2K, 4K, 4K, 2K]
Should a split occur, each individual BIO can end up either having an
start or end offset that is not aligned to the logical block size, which
some drivers (eg SCSI) will reject. However, this tends not to happen
because the default aggregation limit usually keeps the BIO small enough
to not require more than one split, and most pages are actually full 4K
pages, so hitting an uneven gap is very rare anyway.
If the pool is under particular memory pressure, then an IO can be
broken down into a "gang block", a 512-byte block composed of a header
and up to three block pointers. Each points to a fragment of the
original write, or in turn, another gang block, breaking the original
data up over and over until space can be found in the pool for each of
them.
Each gang header is a separate 512-byte memory allocation from a slab,
that needs to be written down to disk. When the gang header is added to
the BIO, its a single 512-byte segment.
Pulling all this together, consider a large aggregated write of gang
blocks. This results a BIO containing lots of 512-byte segments. Given
our tendency to overfill the BIO, a split is likely, and most possible
split points will yield a pair of BIOs that are misaligned. Drivers that
care, like the SCSI driver, will reject them.
---
This commit is a substantial refactor and rewrite of much of `vdev_disk`
to sort all this out.
`vdev_bio_max_segs()` now returns the ideal maximum size for the device,
if available. There's also a tuneable `zfs_vdev_disk_max_segs` to
override this, to assist with testing.
We scan the ABD up front to count the number of pages within it, and to
confirm that if we submitted all those pages to one or more BIOs, it
could be split at any point with creating a misaligned BIO. If the
pages in the BIO are not usable (as in any of the above situations), the
ABD is linearised, and then checked again. This is the same technique
used in `vdev_geom` on FreeBSD, adjusted for Linux's variable page size
and allocator quirks.
`vbio_t` is a cleanup and enhancement of the old `dio_request_t`. The
idea is simply that it can hold all the state needed to create, submit
and return multiple BIOs, including all the refcounts, the ABD copy if
it was needed, and so on. Apart from what I hope is a clearer interface,
the major difference is that because we know how many BIOs we'll need up
front, we don't need the old overflow logic that would grow the BIO
array, throw away all the old work and restart. We can get it right from
the start.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Closes#15533Closes#15588
Similar to DDT make BRT data and indirect block sizes configurable
via module parameters. I am not sure what would be the best yet,
but similar to DDT 4KB blocks kill all chances of compression on
vdev with ashift=12 or more, that on my tests reaches 3x.
While here, fix documentation for respective DDT parameters.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15967
This changes taskq_thread_should_stop() to limit maximum exit rate
for idle threads to one per 5 seconds. I believe the previous one
was broken, not allowing any thread exits for tasks arriving more
than one at a time and so completing while others are running.
Also while there:
- Remove taskq_thread_spawn() calls on task allocation errors.
- Remove extra taskq_thread_should_stop() call.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Rich Ercolani <rincebrain@gmail.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15873
On Linux the ioctl_ficlonerange() and ioctl_ficlone() system calls
are expected to either fully clone the specified range or return an
error. The range may be for an entire file. While internally ZFS
supports cloning partial ranges there's no way to return the length
cloned to the caller so we need to make this all or nothing.
As part of this change support for the REMAP_FILE_CAN_SHORTEN flag
has been added. When REMAP_FILE_CAN_SHORTEN is set zfs_clone_range()
will return a shortened range when encountering pending dirty records.
When it's clear zfs_clone_range() will block and wait for the records
to be written out allowing the blocks to be cloned.
Furthermore, the file range lock is held over the region being cloned
to prevent it from being modified while cloning. This doesn't quite
provide an atomic semantics since if an error is encountered only a
portion of the range may be cloned. This will be converted to an
error if REMAP_FILE_CAN_SHORTEN was not provided and returned to the
caller. However, the destination file range is left in an undefined
state.
A test case has been added which exercises this functionality by
verifying that `cp --reflink=never|auto|always` works correctly.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#15728Closes#15842
We are finding that as customers get larger and faster machines
(hundreds of cores, large NVMe-backed pools) they keep hitting
relatively low performance ceilings. Our profiling work almost always
finds that they're running into bottlenecks on the SPA IO taskqs.
Unfortunately there's often little we can advise at that point, because
there's very few ways to change behaviour without patching.
This commit adds two load-time parameters `zio_taskq_read` and
`zio_taskq_write` that can configure the READ and WRITE IO taskqs
directly.
This achieves two goals: it gives operators (and those that support
them) a way to tune things without requiring a custom build of OpenZFS,
which is often not possible, and it lets us easily try different config
variations in a variety of environments to inform the development of
better defaults for these kind of systems.
Because tuning the IO taskqs really requires a fairly deep understanding
of how IO in ZFS works, and generally isn't needed without a pretty
serious workload and an ability to identify bottlenecks, only minimal
documentation is provided. Its expected that anyone using this is going
to have the source code there as well.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#15675
Copy the disable parameter that FreeBSD implemented, and extend it to
work on Linux as well, until we're sure this is stable.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rich Ercolani <rincebrain@gmail.com>
Closes#15529
It is unused for 3 years since #10576.
Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15507
Current L2ARC write rate and headroom parameters are very conservative:
l2arc_write_max=8M and l2arc_headroom=2 (ie: a full L2ARC writes at
8 MB/s, scanning 16/32 MB of ARC tail each time; a warming L2ARC runs
at 2x these rates).
These values were selected 15+ years ago based on then-current SSDs
size, performance and endurance. Today we have multi-TB, fast and
cheap SSDs which can sustain much higher read/write rates.
For this reason, this patch increases l2arc_write_max to 32M and
l2arc_headroom to 8 (4x increase for both).
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Gionatan Danti <g.danti@assyoma.it>
Closes#15457
This feature allows disks to be added one at a time to a RAID-Z group,
expanding its capacity incrementally. This feature is especially useful
for small pools (typically with only one RAID-Z group), where there
isn't sufficient hardware to add capacity by adding a whole new RAID-Z
group (typically doubling the number of disks).
== Initiating expansion ==
A new device (disk) can be attached to an existing RAIDZ vdev, by
running `zpool attach POOL raidzP-N NEW_DEVICE`, e.g. `zpool attach tank
raidz2-0 sda`. The new device will become part of the RAIDZ group. A
"raidz expansion" will be initiated, and the new device will contribute
additional space to the RAIDZ group once the expansion completes.
The `feature@raidz_expansion` on-disk feature flag must be `enabled` to
initiate an expansion, and it remains `active` for the life of the pool.
In other words, pools with expanded RAIDZ vdevs can not be imported by
older releases of the ZFS software.
== During expansion ==
The expansion entails reading all allocated space from existing disks in
the RAIDZ group, and rewriting it to the new disks in the RAIDZ group
(including the newly added device).
The expansion progress can be monitored with `zpool status`.
Data redundancy is maintained during (and after) the expansion. If a
disk fails while the expansion is in progress, the expansion pauses
until the health of the RAIDZ vdev is restored (e.g. by replacing the
failed disk and waiting for reconstruction to complete).
The pool remains accessible during expansion. Following a reboot or
export/import, the expansion resumes where it left off.
== After expansion ==
When the expansion completes, the additional space is available for use,
and is reflected in the `available` zfs property (as seen in `zfs list`,
`df`, etc).
Expansion does not change the number of failures that can be tolerated
without data loss (e.g. a RAIDZ2 is still a RAIDZ2 even after
expansion).
A RAIDZ vdev can be expanded multiple times.
After the expansion completes, old blocks remain with their old
data-to-parity ratio (e.g. 5-wide RAIDZ2, has 3 data to 2 parity), but
distributed among the larger set of disks. New blocks will be written
with the new data-to-parity ratio (e.g. a 5-wide RAIDZ2 which has been
expanded once to 6-wide, has 4 data to 2 parity). However, the RAIDZ
vdev's "assumed parity ratio" does not change, so slightly less space
than is expected may be reported for newly-written blocks, according to
`zfs list`, `df`, `ls -s`, and similar tools.
Sponsored-by: The FreeBSD Foundation
Sponsored-by: iXsystems, Inc.
Sponsored-by: vStack
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Authored-by: Matthew Ahrens <mahrens@delphix.com>
Contributions-by: Fedor Uporov <fuporov.vstack@gmail.com>
Contributions-by: Stuart Maybee <stuart.maybee@comcast.net>
Contributions-by: Thorsten Behrens <tbehrens@outlook.com>
Contributions-by: Fmstrat <nospam@nowsci.com>
Contributions-by: Don Brady <dev.fs.zfs@gmail.com>
Signed-off-by: Don Brady <dev.fs.zfs@gmail.com>
Closes#15022
As part of transaction group commit, dsl_pool_sync() sequentially calls
dsl_dataset_sync() for each dirty dataset, which subsequently calls
dmu_objset_sync(). dmu_objset_sync() in turn uses up to 75% of CPU
cores to run sync_dnodes_task() in taskq threads to sync the dirty
dnodes (files).
There are two problems:
1. Each ZVOL in a pool is a separate dataset/objset having a single
dnode. This means the objsets are synchronized serially, which
leads to a bottleneck of ~330K blocks written per second per pool.
2. In the case of multiple dirty dnodes/files on a dataset/objset on a
big system they will be sync'd in parallel taskq threads. However,
it is inefficient to to use 75% of CPU cores of a big system to do
that, because of (a) bottlenecks on a single write issue taskq, and
(b) allocation throttling. In addition, if not for the allocation
throttling sorting write requests by bookmarks (logical address),
writes for different files may reach space allocators interleaved,
leading to unwanted fragmentation.
The solution to both problems is to always sync no more and (if
possible) no fewer dnodes at the same time than there are allocators
the pool.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Edmund Nadolski <edmund.nadolski@ixsystems.com>
Closes#15197
Commits 518b487 and 23bdb07 changed the default ARC size limit on
Linux systems to 1/2 of physical memory, which has become too
strict for modern systems with large amounts of RAM. This patch
changes the default limit to match that of FreeBSD, so ZFS may
have a unified value on both platforms.
Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Edmund Nadolski <edmund.nadolski@ixsystems.com>
Closes#15437
This reverts commit aefb6a2bd6.
aefb6a2bd temporally disabled blk-mq until we could fix a fix for
Signed-off-by: Tony Hutter <hutter2@llnl.gov>
Closes#15439
... by checking that previous block is fully written and flushed.
It allows to skip commit delays since we can give up on aggregation
in that case. This removes zil_min_commit_timeout parameter, since
for single-threaded workloads it is not needed at all, while on very
fast devices even some multi-threaded workloads may get detected as
single-threaded and still bypass the wait. To give multi-threaded
workloads more aggregation chances increase zfs_commit_timeout_pct
from 5 to 10%, as they should suffer less from additional latency.
Also single-threaded workloads detection allows in perspective better
prediction of the next block size.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Prakash Surya <prakash.surya@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15381
For synchronous write workloads with large IO sizes, a pool configured
with a slog performs worse than one with an embedded zil:
sequential_writes 1m sync ios, 16 threads
Write IOPS: 1292 438 -66.10%
Write Bandwidth: 1323570 448910 -66.08%
Write Latency: 12128400 36330970 3.0x
sequential_writes 1m sync ios, 32 threads
Write IOPS: 1293 430 -66.74%
Write Bandwidth: 1324184 441188 -66.68%
Write Latency: 24486278 74028536 3.0x
The reason is the `zil_slog_bulk` variable. In `zil_lwb_write_open`,
if a zil block is greater than 768K, the priority of the write is
downgraded from sync to async. Increasing the value allows greater
throughput. To select a value for this PR, I ran an fio workload with
the following values for `zil_slog_bulk`:
zil_slog_bulk KiB/s
1048576 422132
2097152 478935
4194304 533645
8388608 623031
12582912 827158
16777216 1038359
25165824 1142210
33554432 1211472
50331648 1292847
67108864 1308506
100663296 1306821
134217728 1304998
At 64M, the results with a slog are now improved to parity with an
embedded zil:
sequential_writes 1m sync ios, 16 threads
Write IOPS: 438 1288 2.9x
Write Bandwidth: 448910 1319062 2.9x
Write Latency: 36330970 12163408 -66.52%
sequential_writes 1m sync ios, 32 threads
Write IOPS: 430 1290 3.0x
Write Bandwidth: 441188 1321693 3.0x
Write Latency: 74028536 24519698 -66.88%
None of the other tests in the performance suite (run with a zil or
slog) had a significant change, including the random_write_zil tests,
which use multiple datasets.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Signed-off-by: John Wren Kennedy <john.kennedy@delphix.com>
Closes#14378
There was a report of zvol data loss (#15351) after enabling blk-mq on a
zvol backed with 16k physical block sized disks. Out of an abundance of
caution, do not allow the user to enable blk-mq until we can look into
the issue.
Note that blk-mq was not enabled by default on zvols. It was always
opt-in via the zvol_use_blk_mq module parameter.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Signed-off-by: Tony Hutter <hutter2@llnl.gov>
Addresses: #15351Closes#15378
Benchmarks show that at certain write sizes range lock/unlock take
not so much time as extra memory copy. The exact threshold is not
obvious due to other overheads, but it is definitely lower than
~63KB used before. Make it configurable, defaulting at 7.5KB,
that is 8KB of nearest malloc() size minus itx and lr structs.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15353
Before this change ZFS created threads for 50% of CPUs for each top-
level vdev. Plus it created the same number of threads for embedded
log groups (that have only one metaslab and don't need any preload).
As result, on system with 80 CPUs and pool of 60 vdevs this resulted
in 4800 metaslab preload threads, that is absolutely insane.
This patch changes the preload threads to 50% of CPUs in one taskq
per pool, so on the mentioned system it will be only 40 threads.
Among other things this fixes zdb on the mentioned system and pool
on FreeBSD, that failed to create so many threads in one process.
Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15319
metaslab_force_ganging isn't enough to actually force ganging, because
it still only forces 3% of the time. This adds
metaslab_force_ganging_pct so we can configure how often to force
ganging.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Closes#15088
- Reduce maximum prefetch distance for 32bit platforms to 8MB as it
was previously. Those systems didn't grow much probably, so better
stay conservative there.
- Retire array_rd_sz tunable, blocking prefetch for large requests.
We should not penalize applications trying to be more efficient. The
speculative prefetcher by itself has reasonable distance limits, and
1MB is not much at all these days.
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#15072
The DDT is really inefficient on 4k and up vdevs, because it always
allocates 4k blocks, and while compression could save us somewhat
at ashift 9, that stops being true.
So let's change the default to 32 KiB, which seems like a reasonable
compromise between improved space savings and inflated write sizes
for DDT updates.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rich Ercolani <rincebrain@gmail.com>
Closes#14654
Switch FIFO queues (SYNC/TRIM) and active queue of vdev queue from
time-sorted AVL-trees to simple lists. AVL-trees are too expensive
for such a simple task. To change I/O priority without searching
through the trees, add io_queue_state field to struct zio.
To not check number of queued I/Os for each priority add vq_cqueued
bitmap to struct vdev_queue. Update it when adding/removing I/Os.
Make vq_cactive a separate array instead of struct vdev_queue_class
member. Together those allow to avoid lots of cache misses when
looking for work in vdev_queue_class_to_issue().
Introduce deadline of ~0.5s for LBA-sorted queues. Before this I
saw some I/Os waiting in a queue for up to 8 seconds and possibly
more due to starvation. With this change I no longer see it. I
had to slightly more complicate the comparison function, but since
it uses all the same cache lines the difference is minimal. For a
sequential I/Os the new code in vdev_queue_io_to_issue() actually
often uses more simple avl_first(), falling back to avl_find() and
avl_nearest() only when needed.
Arrange members in struct zio to access only one cache line when
searching through vdev queues. While there, remove io_alloc_node,
reusing the io_queue_node instead. Those two are never used same
time.
Remove zfs_vdev_aggregate_trim parameter. It was disabled for 4
years since implemented, while still wasted time maintaining the
offset-sorted tree of TRIM requests. Just remove the tree.
Remove locking from txg_all_lists_empty(). It is racy by design,
while 2 pair of locks/unlocks take noticeable time under the vdev
queue lock.
With these changes in my tests with volblocksize=4KB I measure vdev
queue lock spin time reduction by 50% on read and 75% on write.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#14925
It's been observed that in certain workloads (zvol-related being a
big one), ZFS will end up spending a large amount of time spinning
up taskqs only to tear them down again almost immediately, then
spin them up again...
I noticed this when I looked at what my mostly-idle system was doing
and wondered how on earth taskq creation/destroy was a bunch of time...
So I added a configurable delay to avoid it tearing down tasks the
first time it notices them idle, and the total number of threads at
steady state went up, but the amount of time being burned just
tearing down/turning up new ones almost vanished.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rich Ercolani <rincebrain@gmail.com>
Closes#14938
It was a vdev level read cache, designed to aggregate many small
reads by speculatively issuing bigger reads instead and caching
the result. But since it has almost no idea about what is going
on with exception of ZIO_FLAG_DONT_CACHE flag set by higher layers,
it was found to make more harm than good, for which reason it was
disabled for the past 12 years. These days we have much better
instruments to enlarge the I/Os, such as speculative and prescient
prefetches, I/O scheduler, I/O aggregation etc.
Besides just the dead code removal this removes one extra mutex
lock/unlock per write inside vdev_cache_write(), not otherwise
disabled and trying to do some work.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#14953
Added a flag '-e' in zpool scrub to scrub only blocks in error log. A
user can pause, resume and cancel the error scrub by passing additional
command line arguments -p -s just like a regular scrub. This involves
adding a new flag, creating new libzfs interfaces, a new ioctl, and the
actual iteration and read-issuing logic. Error scrubbing is executed in
multiple txg to make sure pool performance is not affected.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Co-authored-by: TulsiJain tulsi.jain@delphix.com
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#8995Closes#12355
It was previously available only to FreeBSD.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Sponsored-by: Klara, Inc.
Sponsored-by: Seagate Technology LLC
Closes#14718
Traditionally ARC adaptation was limited to MRU/MFU distribution. But
for years people with metadata-centric workload demanded mechanisms to
also manage data/metadata distribution, that in original ZFS was just
a FIFO. As result ZFS effectively got separate states for data and
metadata, minimum and maximum metadata limits etc, but it all required
manual tuning, was not adaptive and in its heart remained a bad FIFO.
This change removes most of existing eviction logic, rewriting it from
scratch. This makes MRU/MFU adaptation individual for data and meta-
data, same as the distribution between data and metadata themselves.
Since most of required states separation was already done, it only
required to make arcs_size state field specific per data/metadata.
The adaptation logic is still based on previous concept of ghost hits,
just now it balances ARC capacity between 4 states: MRU data, MRU
metadata, MFU data and MFU metadata. To simplify arc_c changes instead
of arc_p measured in bytes, this code uses 3 variable arc_meta, arc_pd
and arc_pm, representing ARC balance between metadata and data, MRU and
MFU for data, and MRU and MFU for metadata respectively as 32-bit fixed
point fractions. Since we care about the math result only when need to
evict, this moves all the logic from arc_adapt() to arc_evict(), that
reduces per-block overhead, since per-block operations are limited to
stats collection, now moved from arc_adapt() to arc_access() and using
cheaper wmsums. This also allows to remove ugly ARC_HDR_DO_ADAPT flag
from many places.
This change also removes number of metadata specific tunables, part of
which were actually not functioning correctly, since not all metadata
are equal and some (like L2ARC headers) are not really evictable.
Instead it introduced single opaque knob zfs_arc_meta_balance, tuning
ARC's reaction on ghost hits, allowing administrator give more or less
preference to metadata without setting strict limits.
Some of old code parts like arc_evict_meta() are just removed, because
since introduction of ABD ARC they really make no sense: only headers
referenced by small number of buffers are not evictable, and they are
really not evictable no matter what this code do. Instead just call
arc_prune_async() if too much metadata appear not evictable.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Allan Jude <allan@klarasystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#14359
When testing distributed rebuild performance with more capable
hardware it was observed than increasing the zfs_rebuild_vdev_limit
to 64M reduced the rebuild time by 17%. Beyond 64MB there was
some improvement (~2%) but it was not significant when weighed
against the increased memory usage. Memory usage is capped at 1/4
of arc_c_max.
Additionally, vr_bytes_inflight_max has been moved so it's updated
per-metaslab to allow the size to be adjust while a rebuild is
running.
Reviewed-by: Akash B <akash-b@hpe.com>
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#14428
For HDD based pools the default zfs_scan_vdev_limit of 4M
per-vdev can significantly limit the maximum scrub performance.
Increasing the default to 16M can double the scrub speed from
80 MB/s per disk to 160 MB/s per disk.
This does increase the memory footprint during scrub/resilver
but given the performance win this is a reasonable trade off.
Memory usage is capped at 1/4 of arc_c_max. Note that number
of outstanding I/Os has not changed and is still limited by
zfs_vdev_scrub_max_active.
Reviewed-by: Akash B <akash-b@hpe.com>
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#14428
When resilvering the estimated time remaining is calculated using
the average issue rate over the current pass. Where the current
pass starts when a scan was started, or restarted, if the pool
was exported/imported.
For dRAID pools in particular this can result in wildly optimistic
estimates since the issue rate will be very high while scanning
when non-degraded regions of the pool are scanned. Once repair
I/O starts being issued performance drops to a realistic number
but the estimated performance is still significantly skewed.
To address this we redefine a pass such that it starts after a
scanning phase completes so the issue rate is more reflective of
recent performance. Additionally, the zfs_scan_report_txgs
module option can be set to reset the pass statistics more often.
Reviewed-by: Akash B <akash-b@hpe.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#14410
Despite all optimizations, tests on actual hardware show that FreeBSD
kernel can't sleep for less then ~2us. Similar tests on Linux show
~50us delay at least from nanosleep() (haven't tested inside kernel).
It means that on very fast log device ZIL may not be able to satisfy
zfs_commit_timeout_pct block commit timeout, increasing log latency
more than desired.
Handle that by introduction of zil_min_commit_timeout parameter,
specifying minimal timeout value where additional delays to aggregate
writes may be skipped. Also skip delays if the LWB is more than 7/8
full, that often happens if I/O sizes are constant and match one of
LWB sizes. Both things are applied only if there were no already
outstanding log blocks, that may indicate single-threaded workload,
that by definition can not benefit from the commit delays.
While there, add short time moving average to zl_last_lwb_latency to
make it more stable.
Tests of single-threaded 4KB writes to NVDIMM SLOG on FreeBSD show IOPS
increase by 9% instead of expected 5%. For zfs_commit_timeout_pct of
1 there IOPS increase by 5.5% instead of expected 1%.
Reviewed-by: Allan Jude <allan@klarasystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Prakash Surya <prakash.surya@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#14418
It was changed in e99932f7de,
but without docs update.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Signed-off-by: George Melikov <mail@gmelikov.ru>
Closes#14400
The default_bs and default_ibs tunables control the default block size
and indirect block size.
So far, default_bs and default_ibs were tunable only on FreeBSD, e.g.,
sysctl vfs.zfs.default_ibs
Remove the FreeBSD-specific sysctl code and expose default_bs and
default_ibs as tunables on both Linux and FreeBSD using
ZFS_MODULE_PARAM.
One of the use cases for changing the values of those tunables is to
lower the indirect block size, which may improve performance of large
directories (as discussed during the OpenZFS Leadership Meeting
on 2022-08-16).
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Signed-off-by: Mateusz Piotrowski <mateusz.piotrowski@klarasystems.com>
Sponsored-by: Wasabi Technology, Inc.
Closes#14293
This change turns `MZAP_MAX_BLKSZ` into a `ZFS_MODULE_PARAM()` called
`zap_micro_max_size`. As a result, we can experiment with different
micro ZAP sizes to improve directory size scaling.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Co-authored-by: Mateusz Piotrowski <mateuszpiotrowski@klarasystems.com>
Co-authored-by: Toomas Soome <toomas.soome@klarasystems.com>
Signed-off-by: Mateusz Piotrowski <mateuszpiotrowski@klarasystems.com>
Sponsored-by: Wasabi Technology, Inc.
Closes#14292
This saves 40 bytes per full ARC header, reducing it on FreeBSD from
240 to 200 bytes on production bits.
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#14315
shodanshok
Rob Norris
Allan Jude
Alexander Motin
Mateusz Piotrowski
Brian Behlendorf
Ameer Hamza
Rob N
Rich Ercolani
Don Brady
ednadolski-ix
Tony Hutter
John Wren Kennedy
George Amanakis
George Melikov
Mateusz Piotrowski