Since we are calculating a free space fragmentation, we should
weight metaslabs by the amount of their free space, not a full
size. Fragmentation of full metaslabs may not matter in presence
empty ones. The old algorithm did not differentiate metaslabs
having only one free 4KB block from metaslabs having 50% of space
free in 4KB blocks, reporting higher fragmentation.
While there, move metaslab_group_alloc_update() call after setting
mg_fragmentation, otherwise the effect may be delayed by one TXG.
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Linux 6.12 has conflicting range_tree_{find,destroy,clear} symbols.
Signed-off-by: Ivan Volosyuk <Ivan.Volosyuk@gmail.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Rob Norris <robn@despairlabs.com>
The purpose of no-op is to simulate a failure between a device cache and
its permanent store. We still want it to go through the queue and
respond in the same way to everything else.
So, inject "success" as the very last thing, and then move on to
VDEV_IO_DONE to be dequeued and so any followup work can occur.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#17029
Gang blocks have a significant impact on the long and short term
performance of a zpool, but there is not a lot of observability into
whether they're being used. This change adds gang-specific kstats to
ZFS, to better allow users to see whether ganging is happening.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Paul Dagnelie <paul.dagnelie@klarasystems.com>
Closes#17003
When you are using large recordsizes in conjunction with raidz, with
incompressible data, you can pretty reliably be making 21 MB
allocations. Unfortunately, the fragmentation metric in ZFS considers
any metaslabs with 16 MB free chunks completely unfragmented, so you can
have a metaslab report 0% fragmented and be unable to satisfy an
allocation. When using the segment-based metaslab weight, this is
inconvenient; when using the space-based one, it can seriously degrade
performance.
We expand the fragmentation table to extend up to 512MB, and redefine
the table size based on the actual table, rather than having a static
define. We also tweak the one variable that depends on fragmentation
directly.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Allan Jude <allan@klarasystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Paul Dagnelie <paul.dagnelie@klarasystems.com>
Closes#16986
As zios are reexecuted after resume from suspension, their ready and
wait states need to be propagated to wait counts on all their parents.
It's possible for those parents to have active children passing through
READY or DONE, which then end up in zio_notify_parent(), take their
parent's lock, and decrement the wait count. Without also taking a lock
here, it's possible for an increment race to occur, which leads to
either there being no references left (tripping the assert in
zio_notify_parent()), or a parent waiting forever for a nonexistent
child to complete.
To protect against this, we simply take the appropriate zio locks in
zio_reexecute() before updating the wait counts.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Allan Jude <allan@klarasystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#17016
This change will prevent prefetch to perform unnecessary ARC buffer
fill when reading from disk.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Jaydeep Kshirsagar <jkshirsagar@maxlinear.com>
Co-authored-by: Alexander Motin <mav@FreeBSD.org>
Closes#17013
Injecting a device probe failure is not possible by matching IO types,
because probe IO goes to the label regions, which is explicitly excluded
from injection. Even if it were possible, it would be awkward to do,
because a probe is sequence of reads and writes.
This commit adds a new IO "type" to match for injection, which looks for
the ZIO_FLAG_PROBE flag instead. Any probe IO will be match the
injection record and recieve the wanted error.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#16947
I'm about to add a new "type", and I need somewhere to put it!
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#16947
When building tests with zinject, it can be quite difficult to work out
if you're producing the right kind of IO to match the rules you've set
up.
So, here we extend injection records to count the number of times a
handler matched the operation, and how often an error was actually
injected (ie after frequency and other exclusions are applied).
Then, display those counts in the `zinject` output.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Closes#16938
Similar to what we saw in #16569, we need to consider that a
replacing vdev should not be considered as fully contributing
to the redundancy of a raidz vdev even though current IO has
enough redundancy.
When a failed vdev_probe() is faulting a disk, it now checks
if that disk is required, and if so it suspends the pool until
the admin can return the missing disks.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Allan Jude <allan@klarasystems.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Don Brady <don.brady@klarasystems.com>
Closes#16864
It's possible for a vdev to be flagged for async remove after the pool
has suspended. If the removed device has been returned when the pool is
resumed, the ASYNC_REMOVE task will still run at the end of txg, and
remove the device from the pool again.
To fix, we clear the async remove flag at reopen, just as we did for the
async fault flag in 5de3ac223.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#16921
In Linux, block devices currently lack support for `copy_file_range`
API because the kernel does not provide the necessary functionality.
However, there is an ongoing upstream effort to address this
limitation: https://patchwork.kernel.org/project/dm-devel/cover/20240520102033.9361-1-nj.shetty@samsung.com/.
We have adopted this upstream kernel patch into the TrueNAS kernel and
made some additional modifications to enable block cloning specifically
for the zvol block device. This patch implements the platform-
independent portions of these changes for inclusion in OpenZFS.
This patch does not introduce any new functionality directly into
OpenZFS. The `TX_CLONE_RANGE` replay capability is only relevant when
zvols are migrated to non-TrueNAS systems that support Clone Range
replay in the ZIL.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Ameer Hamza <ahamza@ixsystems.com>
Closes#16901
People have noted there's no way to remove a pool userprop, only zero
it. Turns vdev userprops had a method, by setting empty-string. So this
makes pool userprops follow the same behaviour.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#16887
The count of chunks in a microzap block is stored as an uint16_t
(mze_chunkid). Each chunk is 64 bytes, and the first is used to store a
header, so there are 32767 usable chunks, which is just under 2M. 1M is
the largest power-2-rounded block size under 2M, so we must set the
limit there.
If it goes higher, the loop in mzap_addent can overflow and fall into
the PANIC case.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#16888
VDEV_PROP_USERPROP is equal do VDEV_PROP_INVAL and so is not a real
property. That's why vdev_prop_readonly() does not work right for
it. In particular it may declare all vdev user properties readonly
on FreeBSD.
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#16890
We should not dereference rra after the last zio_nowait() is called.
It seems very unlikely, but ASAN in ztest managed to catch it.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16868
It seems there's no good reason for vdev_disk & vdev_geom to explicitly
detect no support for flush and set vdev_nowritecache. Instead, just
signal it by setting the error to ENOTSUP, and let zio_vdev_io_assess()
take care of it in one place.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#16855
The first time a device returns ENOTSUP in repsonse to a flush request,
we set vdev_nowritecache so we don't issue flushes in the future and
instead just pretend the succeeded. However, we still return an error
for the initial flush, even though we just decided such errors are
meaningless!
So, when setting vdev_nowritecache in response to a flush error, also
reset the error code to assume success.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#16855
In dbuf_sync_leaf, we clone the arc_buf in dr if we share it with db
except for overridden case. However, this exception causes a race where
dbuf_new_size could free the arc_buf after the last dereference of
*datap and causes use-after-free. We fix this by cloning the buf
regardless if it's overridden.
The race:
--
P0 P1
dbuf_hold_impl()
// dbuf_hold_copy passed
// because db_data_pending NULL
dbuf_sync_leaf()
// doesn't clone *datap
// *datap derefed to db_buf
dbuf_write(*datap)
dbuf_new_size()
dmu_buf_will_dirty()
dbuf_fix_old_data()
// alloc new buf for P0 dr
// but can't change *datap
arc_alloc_buf()
arc_buf_destroy()
// alloc new buf for db_buf
// and destroy old buf
dbuf_write() // continue
abd_get_from_buf(data->b_data,
arc_buf_size(data))
// use-after-free
--
Here's an example when it happens:
BUG: kernel NULL pointer dereference, address: 000000000000002e
RIP: 0010:arc_buf_size+0x1c/0x30 [zfs]
Call Trace:
dbuf_write+0x3ff/0x580 [zfs]
dbuf_sync_leaf+0x13c/0x530 [zfs]
dbuf_sync_list+0xbf/0x120 [zfs]
dnode_sync+0x3ea/0x7a0 [zfs]
sync_dnodes_task+0x71/0xa0 [zfs]
taskq_thread+0x2b8/0x4e0 [spl]
kthread+0x112/0x130
ret_from_fork+0x1f/0x30
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Chunwei Chen <david.chen@nutanix.com>
Co-authored-by: Chunwei Chen <david.chen@nutanix.com>
Closes#16854
When vdev first sees some block cloning, there is a window when
brt_maybe_exists() might already return true since something was
cloned, but bv_mos_entries is still 0 since BRT ZAP was not yet
created. In such case we should not try to look into the ZAP
and dereference NULL bv_mos_entries_dnode.
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#16851
cstyle can handle these cases now, so we don't need to disable it.
Sponsored-by: https://despairlabs.com/sponsor/
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Closes#16840
- Instead of copying one ashift-sized block per ZIO, copy as much
as we have contiguous data up to 16MB per old vdev. To avoid data
moves use gang ABDs, so that read ZIOs can directly fill buffers
for write ZIOs. ABDs have much smaller overhead than ZIOs in both
memory usage and processing time, plus big I/Os do not depend on
I/O aggregation and scheduling to reach decent performance on HDDs.
- Reduce raidz_expand_max_copy_bytes to 16MB on 32bit platforms.
- Use 32bit range tree when possible (practically always now) to
slightly reduce memory usage.
- Use ZIO_PRIORITY_REMOVAL for early stages of expansion, same as
for main ones.
- Fix rate overflows in `zpool status` reporting.
With these changes expanding RAIDZ1 from 4 to 5 children I am able
to reach 6-12GB/s rate on SSDs and ~500MB/s on HDDs, both are
limited by devices instead of CPU.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15680Closes#16819
Same as writes block cloning can increase block size and number of
indirection levels. That means it can dirty block 0 at level 0 or
at new top indirection level without explicitly holding them.
A block cloning test case for large offsets has been added.
Reviewed-by: Rob Norris <robn@despairlabs.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Co-authored-by: Ameer Hamza <ahamza@ixsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16825
This also includes removing L2 vdevs asynchronously.
This commit also guarantees that spa_load_guid is unique.
The zpool reguid feature introduced the spa_load_guid, which is a
transient value used for runtime identification purposes in the ARC.
This value is not the same as the spa's persistent pool guid.
However, the value is seeded from spa_generate_load_guid() which
does not check for uniqueness against the spa_load_guid from other
pools. Although extremely rare, you can end up with two different
pools sharing the same spa_load_guid value! So we guarantee that
the value is always unique and additionally not still in use by an
async arc flush task.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Allan Jude <allan@klarasystems.com>
Signed-off-by: Don Brady <don.brady@klarasystems.com>
Closes#16215
- Issue prescient prefetches for demand indirect blocks after the
first one. It should be quite rare for reads/writes, but much more
useful for cloning due to much bigger (up to 1022 blocks) accesses.
It covers the gap during the first couple accesses when we can not
speculate yet, but we know what is needed right now. It reduces
dbuf_hold() sync read delays in dmu_buf_hold_array_by_dnode().
- Increase maximum prefetch distance for indirect blocks from 64
to 128MB. It should cover the maximum 1022 blocks of block cloning
access size in case of default 128KB recordsize used. In case of
bigger recordsize the above prescient prefetch should also help.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16814
In some cases like dsl_dataset_hold_obj() it is possible to handle
those errors, so failure to hold dataset should be better than
kernel panic. Some other places where these errors are still not
handled but asserted should be less dangerous just as unreachable.
We have a user report about pool corruption leading to assertions
on these errors. Hopefully this will make behavior a bit nicer.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16836
Some users might want to scrub only new data because they would like
to know if the new write wasn't corrupted. This PR adds possibility
scrub only newly written data.
This introduces new `last_scrubbed_txg` property, indicating the
transaction group (TXG) up to which the most recent scrub operation
has checked and repaired the dataset, so users can run scrub only
from the last saved point. We use a scn_max_txg and scn_min_txg
which are already built into scrub, to accomplish that.
Reviewed-by: Allan Jude <allan@klarasystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Mariusz Zaborski <mariusz.zaborski@klarasystems.com>
Sponsored-By: Wasabi Technology, Inc.
Sponsored-By: Klara Inc.
Closes#16301
Direct I/O implementation added condition to call dbuf_undirty()
only in case of block cloning. But the condition is not right if
the block is no longer dirty in this TXG, but still in DB_NOFILL
state. It resulted in block not reverting to DB_UNCACHED and
following NULL de-reference on attempt to access absent db_data.
While there, add assertions for db_data to make debugging easier.
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16829
This allowed to debug #16714, fixed in #16782. Without assertions
added here it is difficult to figure out what logs cause the problem,
since the assertion happens in sync thread context.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Pavel Snajdr <snajpa@snajpa.net>
Co-authored-by: Alexander Motin <mav@FreeBSD.org>
Closes#16795
This fixes assertion in brt_sync_table() on debug builds when last
cloned block on the vdev is freed and bv_meta_dirty is cleared,
while bv_entcount_dirty is not. Should not matter in production.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16791
Previous implementation of zap_leaf_array_free() put chunks on the
free list in reverse order. Also zap_leaf_transfer_entry() and
zap_entry_remove() were freeing name and value arrays in reverse
order. Together this created a mess in the free list, making
following allocations much more fragmented than necessary.
This patch re-implements zap_leaf_array_free() to keep existing
chunks order, and implements non-destructive zap_leaf_array_copy()
to be used in zap_leaf_transfer_entry() to allow properly ordered
freeing name and value arrays there and in zap_entry_remove().
With this change test of some writes and deletes shows percent of
non-contiguous chunks in DDT reducing from 61% and 47% to 0% and
17% for arrays and frees respectively. Sure some explicit sorting
could do even better, especially for ZAPs with variable-size arrays,
but it would also cost much more, while this should be very cheap.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16766
zio_delay_interrupt(), apparently used for fault injection, is executed
in the I/O pipeline. It can cause the calling thread to go to sleep,
which is not allowed on FreeBSD. This happens only for small delays,
though, and there's no apparent reason to avoid deferring to a taskqueue
in that case, as it already does otherwise.
Simply go to sleep unconditionally. This fixes an occasional panic I
see when running the ZTS on FreeBSD. Also remove an unhelpful comment
referencing the non-existent timeout_generic().
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Mark Johnston <markj@FreeBSD.org>
Closes#16785
- With both pending and current AVL-trees being per-vdev and having
effectively identical comparison functions (pending tree compared
also birth time, but I don't believe it is possible for them to be
different for the same offset within one transaction group), it
makes no sense to move entries from one to another. Instead inline
dramatically simplified brt_entry_addref() into brt_pending_apply().
It no longer requires bv_lock, since there is nothing concurrent
to it at the time. And it does not need to search the tree for the
previous entries, since it is the same tree, we already have the
entry and we know it is unique.
- Put brt_vdev_lookup() and brt_vdev_addref() into different tree
traversals to avoid false positives in the first due to the second
entcount modifications. It saves dramatic amount of time when a
file cloned first time by not looking for non-existent ZAP entries.
- Remove avl_is_empty(bv_tree) check from brt_maybe_exists(). I
don't think it is needed, since by the time all added entries are
already accounted in bv_entcount. The extra check must be producing
too many false positives for no reason. Also we don't need bv_lock
there, since bv_entcount pointer must be table at this point, and
we don't care about false positive races here, while false negative
should be impossible, since all brt_vdev_addref() have already
completed by this point. This dramatically reduces lock contention
on massive deletes of cloned blocks. The only remaining one is
between multiple parallel free threads calling brt_entry_decref().
- Do not update ZAP if net change for a block over the TXG was 0.
In combination with above it makes file move between datasets as
cheap operation as originally intended if it fits into one TXG.
- Do not allocate vdevs on pool creation or import if it did not
have active block cloning. This allows to save a bit in few cases.
- While here, add proper error handling in brt_load() on pool
import instead of assertions.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16773
Without doing that there is a race window on export when history
log write by completed rebuild dirties transaction beyond final,
triggering assertion.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Amanakis <gamanakis@gmail.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16714Closes#16782
Those values require global atomics to get current hash_elements
values in few of the hottest code paths, while in all the years I
never cared about it. If somebody wants, it should be easy to
get it by periodic sampling, since neither ARC header nor DBUF
counts change so fast that it would be difficult to catch.
For now I've left hash_elements_max kstat for ARC, since it was
used/reported by arc_summary and it would break older versions,
but now it just reports the current value.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16759
Compression names actually aren't used in dedup table names, but
checksum names are.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#16776
While block cloning operation from the beginning was made per-vdev,
before this change most of its data were protected by two pool-
wide locks. It created lots of lock contention in many workload.
This change makes most of block cloning data structures per-vdev,
which allows to lock them separately. The only pool-wide lock now
it spa_brt_lock, protecting array of per-vdev pointers and in most
cases taken as reader. Also this splits per-vdev locks into three
different ones: bv_pending_lock protects the AVL-tree of pending
operations in open context, bv_mos_entries_lock protects BRT ZAP
object from while being prefetched, and bv_lock protects the rest
of per-vdev context during TXG commit process. There should be
no functional difference aside of some optimizations.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Pawel Jakub Dawidek <pjd@FreeBSD.org>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16740
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Pawel Jakub Dawidek <pjd@FreeBSD.org>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16740
We are doing exactly the same checks around all brt_pending_add().
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Pawel Jakub Dawidek <pjd@FreeBSD.org>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16740
dsl_free() calls zio_free() to free the block. For most blocks, this
simply calls metaslab_free() without doing any IO or putting anything on
the IO pipeline.
Some blocks however require additional IO to free. This at least
includes gang, dedup and cloned blocks. For those, zio_free() will issue
a ZIO_TYPE_FREE IO and return.
If a huge number of blocks are being freed all at once, it's possible
for dsl_dataset_block_kill() to be called millions of time on a single
transaction (eg a 2T object of 128K blocks is 16M blocks). If those are
all IO-inducing frees, that then becomes 16M FREE IOs placed on the
pipeline. At time of writing, a zio_t is 1280 bytes, so for just one 2T
object that requires a 20G allocation of resident memory from the
zio_cache. If that can't be satisfied by the kernel, an out-of-memory
condition is raised.
This would be better handled by improving the cases that the
dmu_tx_assign() throttle will handle, or by reducing the overheads
required by the IO pipeline, or with a better central facility for
freeing blocks.
For now, we simply check for the cases that would cause zio_free() to
create a FREE IO, and instead put the block on the pool's freelist. This
is the same place that blocks from destroyed datasets go, and the async
destroy machinery will automatically see them and trickle them out as
normal.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#6783Closes#16708Closes#16722Closes#16697
..., before we make the header or the log block visible to others.
It should fix assertion on allocated space going negative if the
header is freed once the lock is dropped, while the write is still
going.
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#16040Closes#16743
Currently, even though send_reader_thread prefetches spill block,
do_dump() will not use it and issues its own blocking arc_read. This
causes significant performance degradation when sending datasets with
lots of spill blocks.
For unmodified spill blocks, we also create send_range struct for them
in send_reader_thread and issue prefetches for them. We piggyback them
on the dnode send_range instead of enqueueing them so we don't break
send_range_after check.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Chunwei Chen <david.chen@nutanix.com>
Co-authored-by: david.chen <david.chen@nutanix.com>
Closes#16701
Small block workloads may use a very large number of dirty records.
During simple block cloning test due to BRT still using 4KB blocks
I can easily see up to 2.5M of those used. Before this change
dbuf_dirty_record_t structures representing them were allocated via
kmem_zalloc(), that rounded their size up to 512 bytes.
Introduction of specialized kmem cache allows to reduce the size
from 512 to 408 bytes. Additionally, since override and raw params
in dirty records are mutually exclusive, puting them into a union
allows to reduce structure size down to 368 bytes, increasing the
saving to 28%, that can be a 0.5GB or more of RAM.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16694
I think we've done enough experiments.
Sponsored-by: https://despairlabs.com/sponsor/
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Closes#16189Closes#16712
Now that we can handle these different alignments, we don't this
workaround.
This reverts commit aefc2da8a5.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#16687
While reading some code @grwilson came across the above function that
seemingly had no consumers besides a ztest callback that ensures that
the tx_callback infrastructure works correctly. It turns out that Lustre
is the main (and potentially the only) consumer of this. Refer to
`osd_trans_commit_cb` of `lustre/osd-zfs/osd_handler.c` in the Lustre
repo for more info. Let's add a comment highlighting this before someone
removes it by mistake.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Serapheim Dimitropoulos <serapheimd@gmail.com>
Closes#16698
If on the first open device's logical ashift is bigger than set
by pool's ashift property, ignore the last as unusable instead of
creating vdev that will fail most of I/Os due to misalignment.
Reviewed-by: Rob Norris <robn@despairlabs.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ameer Hamza <ahamza@ixsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16690
