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
During snapshot deletion ZFS may issue several reads for each deadlist
to merge them into next snapshot's or pool's bpobj. Number of the dead
lists increases with number of snapshots. On HDD pools it may take
significant time during which sync thread is blocked.
This patch introduces prescient prefetch of required blocks for up to
128 deadlists ahead. Tests show reduction of time required to delete
dataset with 720 snapshots with randomly overwritten file on wide HDD
pool from 75-85 to 22-28 seconds.
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.
Issue #14276Closes#14402
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
If we receive a DRR_FREEOBJECTS as the first entry in an object range,
this might end up producing a hole if the freed objects were the
only existing objects in the block.
If the txg starts syncing before we've processed any following
DRR_OBJECT records, this leads to a possible race where the backing
arc_buf_t gets its psize set to 0 in the arc_write_ready() callback
while still being referenced from a dirty record in the open txg.
To prevent this, we insert a txg_wait_synced call if the first
record in the range was a DRR_FREEOBJECTS that actually
resulted in one or more freed objects.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: David Hedberg <david.hedberg@findity.com>
Sponsored by: Findity AB
Closes#11893Closes#14358
In the zstream code, Coverity reported:
"The argument could be controlled by an attacker, who could invoke the
function with arbitrary values (for example, a very high or negative
buffer size)."
It did not report this in the kernel. This is likely because the
userspace code stored this in an int before passing it into the
allocator, while the kernel code stored it in a uint32_t.
However, this did reveal a potentially real problem. On 32-bit systems
and systems with only 4GB of physical memory or less in general, it is
possible to pass a large enough value that the system will hang. Even
worse, on Linux systems, the kernel memory allocator is not able to
support allocations up to the maximum 4GB allocation size that this
allows.
This had already been limited in userspace to 64MB by
`ZFS_SENDRECV_MAX_NVLIST`, but we need a hard limit in the kernel to
protect systems. After some discussion, we settle on 256MB as a hard
upper limit. Attempting to receive a stream that requires more memory
than that will result in E2BIG being returned to user space.
Reported-by: Coverity (CID-1529836)
Reported-by: Coverity (CID-1529837)
Reported-by: Coverity (CID-1529838)
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14285
Introduce four new vdev properties:
checksum_n
checksum_t
io_n
io_t
These properties can be used for configuring the thresholds of zed's
diagnosis engine and are interpeted as <N> events in T <seconds>.
When this property is set to a non-default value on a top-level vdev,
those thresholds will also apply to its leaf vdevs. This behavior can be
overridden by explicitly setting the property on the leaf vdev.
Note that, these properties do not persist across vdev replacement. For
this reason, it is advisable to set the property on the top-level vdev
instead of the leaf vdev.
The default values for zed's diagnosis engine (10 events, 600 seconds)
remains unchanged.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Allan Jude <allan@klarasystems.com>
Signed-off-by: Rob Wing <rob.wing@klarasystems.com>
Sponsored-by: Seagate Technology LLC
Closes#13805
In 2016, the authors of PVS Studio ran it on the FreeBSD kernel, which
identified a number of bugs / cleanup opportunities in the FreeBSD ZFS kernel
code. A few of them persist to the present day:
https://reviews.freebsd.org/D5245
Note that the scan was done against
freebsd/freebsd-src@46763fd4ca.
In particular, we have the following in free_blocks():
\sys\cddl\contrib\opensolaris\uts\common\fs\zfs\dnode_sync.c (174): error V547: Expression '__left >= __right' is always true. Unsigned type value is always >= 0.
\sys\cddl\contrib\opensolaris\uts\common\fs\zfs\dnode_sync.c (171): error V634: The priority of the '*' operation is higher than that of the '<<' operation. It's possible that parentheses should be used in the expression.
\sys\cddl\contrib\opensolaris\uts\common\fs\zfs\dnode_sync.c (175): error V547: Expression '__left >= __right' is always true. Unsigned type value is always >= 0.
A couple of assertions accidentally typecast the arguments they check to
unsigned in such a way that the result is always true. Also, parentheses
are missing around `1<<epbs` in `(db->db_blkid * 1<<epbs)`. This works
out to be okay due to multiplication not caring what order of operations
we use, but it is better to fix it to be `(db->db_blkid << epbs)`.
A few of the function local variables probably never should have been
32-bit in the first place, so we make them 64-bit. We also replace the
existing assertions with additional assertions to ensure that 64-bit
unsigned arithmetic is safe.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14407
When zfs_file_read returns error, resid may be uninitialized.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Signed-off-by: Chunwei Chen <david.chen@nutanix.com>
Closes#14404
This is only ever used with unsigned data, so the type itself should be
unsigned. Also, PVS Studio's 2016 FreeBSD kernel report correctly
identified the following assertion as always being true, so we can drop
it:
ASSERT3U(dd->dd_space_towrite[i & TXG_MASK], >=, 0);
The reason it was always true is because it would do casts to give us
unsigned comparisons. This could have been fixed by switching to
`ASSERT3S()`, but upon inspection, it turned out that this variable
never should have been allowed to be signed in the first place.
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14408
Use the saved property index instead of looking it up once per DSL
directory when traversing up towards the root.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Allan Jude <allan@klarasystems.com>
Reviewed-by: Akash B <akash-b@hpe.com>
Signed-off-by: Mark Johnston <markj@FreeBSD.org>
Sponsored-by: The FreeBSD Foundation
Closes#14397
Reported-by: KMSAN
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Allan Jude <allan@klarasystems.com>
Reviewed-by: Akash B <akash-b@hpe.com>
Signed-off-by: Mark Johnston <markj@FreeBSD.org>
Sponsored-by: The FreeBSD Foundation
Closes#14397
I recently gained the ability to run Clang's static analyzer on the
linux kernel modules via a few hacks. This extended coverage to code
that was previously missed since Clang's static analyzer only looked at
code that we built in userspace. Running it against the Linux kernel
modules built from my local branch produced a total of 72 reports
against my local branch. Of those, 50 were reports of logic errors and
22 were reports of dead code. Since we already had cleaned up all of
the previous dead code reports, I felt it would be a good next step to
clean up these dead code reports. Clang did a further breakdown of the
dead code reports into:
Dead assignment 15
Dead increment 2
Dead nested assignment 5
The benefit of cleaning these up, especially in the case of dead nested
assignment, is that they can expose places where our error handling is
incorrect. A number of them were fairly straight forward. However
several were not:
In vdev_disk_physio_completion(), not only were we not using the return
value from the static function vdev_disk_dio_put(), but nothing used it,
so I changed it to return void and removed the existing (void) cast in
the other area where we call it in addition to no longer storing it to a
stack value.
In FSE_createDTable(), the function is dead code. Its helper function
FSE_freeDTable() is also dead code, as are the CPP definitions in
`module/zstd/include/zstd_compat_wrapper.h`. We just delete it all.
In zfs_zevent_wait(), we have an optimization opportunity. cv_wait_sig()
returns 0 if there are waiting signals and 1 if there are none. The
Linux SPL version literally returns `signal_pending(current) ? 0 : 1)`
and FreeBSD implements the same semantics, we can just do
`!cv_wait_sig()` in place of `signal_pending(current)` to avoid
unnecessarily calling it again.
zfs_setattr() on FreeBSD version did not have error handling issue
because the code was removed entirely from FreeBSD version. The error is
from updating the attribute directory's files. After some thought, I
decided to propapage errors on it to userspace.
In zfs_secpolicy_tmp_snapshot(), we ignore a lack of permission from the
first check in favor of checking three other permissions. I assume this
is intentional.
In zfs_create_fs(), the return value of zap_update() was not checked
despite setting an important version number. I see no backward
compatibility reason to permit failures, so we add an assertion to catch
failures. Interestingly, Linux is still using ASSERT(error == 0) from
OpenSolaris while FreeBSD has switched to the improved ASSERT0(error)
from illumos, although illumos has yet to adopt it here. ASSERT(error ==
0) was used on Linux while ASSERT0(error) was used on FreeBSD since the
entire file needs conversion and that should be the subject of
another patch.
dnode_move()'s issue was caused by us not having implemented
POINTER_IS_VALID() on Linux. We have a stub in
`include/os/linux/spl/sys/kmem_cache.h` for it, when it really should be
in `include/os/linux/spl/sys/kmem.h` to be consistent with
Illumos/OpenSolaris. FreeBSD put both `POINTER_IS_VALID()` and
`POINTER_INVALIDATE()` in `include/os/freebsd/spl/sys/kmem.h`, so we
copy what it did.
Whenever a report was in platform-specific code, I checked the FreeBSD
version to see if it also applied to FreeBSD, but it was only relevant a
few times.
Lastly, the patch that enabled Clang's static analyzer to be run on the
Linux kernel modules needs more work before it can be put into a PR. I
plan to do that in the future as part of the on-going static analysis
work that I am doing.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14380
There is an external assembly declaration extension in GNU C that glibc
uses when building with ieee128 floating point support on ppc64le.
Marking that as volatile makes no sense, so the build breaks.
It does not make sense to only mark this as volatile on Linux, since if
do not want the compiler reordering things on Linux, we do not want the
compiler reordering things on any other platform, so we stop treating
Linux specially and just manually inline the CPP macro so that we can
eliminate it. This should fix the build on ppc64le.
Tested-by: @gyakovlev
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14308Closes#14384
The Linux 5.16.14 kernel's coccicheck caught this. The semantic
patch that caught it was:
./scripts/coccinelle/null/badzero.cocci
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14372
The Linux 5.16.14 kernel's coccicheck caught this. The semantic
patch that caught it was:
./scripts/coccinelle/misc/semicolon.cocci
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14372
The Linux 5.16.14 kernel's coccicheck caught this. The semantic
patch that caught it was:
./scripts/coccinelle/misc/minmax.cocci
There was a third opportunity to use `MIN()`, but that was in
`FSE_minTableLog()` in `module/zstd/lib/compress/fse_compress.c`.
Upstream zstd has yet to make this change and I did not want to change
header includes just for MIN, or do a one off, so I left it alone.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14372
The Linux 5.16.14 kernel's coccicheck caught this. The semantic
patch that caught it was:
./scripts/coccinelle/misc/flexible_array.cocci
However, unlike the cases where the GNU zero length array extension had
been used, coccicheck would not suggest patches for the older style
single member arrays. That was good because blindly changing them would
break size calculations in most cases.
Therefore, this required care to make sure that we did not break size
calculations. In the case of `indirect_split_t`, we use
`offsetof(indirect_split_t, is_child[is->is_children])` to calculate
size. This might be subtly wrong according to an old mailing list
thread:
https://inbox.sourceware.org/gcc-prs/20021226123454.27019.qmail@sources.redhat.com/T/
That is because the C99 specification should consider the flexible array
members to start at the end of a structure, but compilers prefer to put
padding at the end. A suggestion was made to allow compilers to allocate
padding after the VLA like compilers already did:
http://std.dkuug.dk/JTC1/SC22/WG14/www/docs/n983.htm
However, upon thinking about it, whether or not we allocate end of
structure padding does not matter, so using offsetof() to calculate the
size of the structure is fine, so long as we do not mix it with sizeof()
on structures with no array members.
In the case that we mix them and padding causes offsetof(struct_t,
vla_member[0]) to differ from sizeof(struct_t), we would be doing unsafe
operations if we underallocate via `offsetof()` and then overcopy via
sizeof().
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14372
The Linux 5.16.14 kernel's coccicheck caught this. The semantic
patch that caught it was:
./scripts/coccinelle/misc/flexible_array.cocci
The Linux kernel's documentation makes a good case for why we should not
use these:
https://www.kernel.org/doc/html/latest/process/deprecated.html#zero-length-and-one-element-arrays
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14372
The Linux 5.16.14 kernel's coccicheck caught these. The semantic patch
that caught them was:
./scripts/coccinelle/api/alloc/alloc_cast.cocci
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14372
When activating filesystem features after receiving a snapshot, do
so only in syncing context.
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#14304Closes#14252
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
The Blocking Queue (bqueue) code is used by zfs send/receive to send
messages between the various threads. It uses a shared linked list,
which is locked whenever we enqueue or dequeue. For workloads which
process many blocks per second, the locking on the shared list can be
quite expensive.
This commit changes the bqueue logic to have 3 linked lists:
1. An enquing list, which is used only by the (single) enquing thread,
and thus needs no locks.
2. A shared list, with an associated lock.
3. A dequing list, which is used only by the (single) dequing thread,
and thus needs no locks.
The entire enquing list can be moved to the shared list in constant
time, and the entire shared list can be moved to the dequing list in
constant time. These operations only happen when the `fill_fraction` is
reached, or on an explicit flush request. Therefore, the lock only
needs to be acquired infrequently.
The API already allows for dequing to block until an explicit flush, so
callers don't need to be changed.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Wilson <george.wilson@delphix.com>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#14121
ECHRNG is returned when the channel program encounters a runtime
error. For example, this can happen when a snapshot doesn't exist.
We handle this error the same way as the existing EEXIST and ENOENT
error checks.
Additionally, improve the internal debug message to include the
error describing why a pool couldn't be opened.
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#14351
Encrypted blocks can have up to 2 DVA's, as the third DVA is reserved
for the salt+IV. However, dmu_write_policy() allows non-encrypted
blocks (e.g. DMU_OT_OBJSET) inside encrypted datasets to request and
allocate 3 DVA's, since they don't need a salt+IV (they are merely
authenicated).
However, if such a block becomes a gang block, the gang code incorrectly
limits the gang block header to 2 DVA's. This leads to a "NDVAs
inversion", where a parent block (the gang block header) has less DVA's
than its children (the gang members), causing an assertion failure in
zio_write_gang_member_ready().
This commit addresses the problem by only restricting the gang block
header to 2 DVA's if the block is actually encrypted (and thus its gang
block members can have at most 2 DVA's).
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#14250Closes#14356
This commit supports for spare vdev hotplug. The
spare vdev associated with all the pools will be
marked as "Removed" when the drive is physically
detached and will become "Available" when the
drive is reattached. Currently, the spare vdev
status does not change on the drive removal and
the same is the case with reattachment.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ameer Hamza <ahamza@ixsystems.com>
Closes#14295
Every ARC buffer holds a reference on the header. It means headers with
buffers are never evictable. When we are evicting a header, there can
be no more buffers to free. Just assert that.
b_evict_lock seems not protecting anything now. Remove it.
Buffers checksum should also be freed with the last uncompressed buffer,
so it should not be there also when we are evicting the header.
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
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
Previously the primarycache property was handled only in the dbuf
layer. Since the speculative prefetcher is implemented in the ARC,
it had to be disabled for uncacheable buffers.
This change gives the ARC knowledge about uncacheable buffers
via arc_read() and arc_write(). So when remove_reference() drops
the last reference on the ARC header, it can either immediately destroy
it, or if it is marked as prefetch, put it into a new arc_uncached state.
That state is scanned every second, evicting stale buffers that were
not demand read.
This change also tracks dbufs that were read from the beginning,
but not to the end. It is assumed that such buffers may receive further
reads, and so they are stored in dbuf cache. If a following
reads reaches the end of the buffer, it is immediately evicted.
Otherwise it will follow regular dbuf cache eviction. Since the dbuf
layer does not know actual file sizes, this logic is not applied to
the final buffer of a dnode.
Since uncacheable buffers should no longer stay in the ARC for long,
this patch also tries to optimize I/O by allocating ARC physical
buffers as linear to allow buffer sharing.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Wilson <george.wilson@delphix.com>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#14243
ARC code was many times significantly modified over the years, that
created significant amount of tangled and potentially broken code.
This should make arc_access()/arc_read() code some more readable.
- Decouple prefetch status tracking from b_refcnt. It made sense
originally, but became highly cryptic over the years. Move all the
logic into arc_access(). While there, clean up and comment state
transitions in arc_access(). Some transitions were weird IMO.
- Unify arc_access() calls to arc_read() instead of sometimes calling
it from arc_read_done(). To avoid extra state changes and checks add
one more b_refcnt for ARC_FLAG_IO_IN_PROGRESS.
- Reimplement ARC_FLAG_WAIT in case of ARC_FLAG_IO_IN_PROGRESS with
the same callback mechanism to not falsely account them as hits. Count
those as "iohits", an intermediate between "hits" and "misses". While
there, call read callbacks in original request order, that should be
good for fairness and random speculations/allocations/aggregations.
- Introduce additional statistic counters for prefetch, accounting
predictive vs prescient and hits vs iohits vs misses.
- Remove hash_lock argument from functions not needing it.
- Remove ARC_FLAG_PREDICTIVE_PREFETCH, since it should be opposite
to ARC_FLAG_PRESCIENT_PREFETCH if ARC_FLAG_PREFETCH is set. We may
wish to add ARC_FLAG_PRESCIENT_PREFETCH to few more places.
- Fix few false positive tests found in the process.
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#14123
There is a lock order inversion deadlock between `spa_errlog_lock` and
`dp_config_rwlock`:
A thread in `spa_delete_dataset_errlog()` is running from a sync task.
It is holding the `dp_config_rwlock` for writer (see
`dsl_sync_task_sync()`), and waiting for the `spa_errlog_lock`.
A thread in `dsl_pool_config_enter()` is holding the `spa_errlog_lock`
(see `spa_get_errlog_size()`) and waiting for the `dp_config_rwlock` (as
reader).
Note that this was introduced by #12812.
This commit address this by defining the lock ordering to be
dp_config_rwlock first, then spa_errlog_lock / spa_errlist_lock.
spa_get_errlog() and spa_get_errlog_size() can acquire the locks in this
order, and then process_error_block() and get_head_and_birth_txg() can
verify that the dp_config_rwlock is already held.
Additionally, a buffer overrun in `spa_get_errlog()` is corrected. Many
code paths didn't check if `*count` got to zero, instead continuing to
overwrite past the beginning of the userspace buffer at `uaddr`.
Tested by having some errors in the pool (via `zinject -t data
/path/to/file`), one thread running `zpool iostat 0.001`, and another
thread runs `zfs destroy` (in a loop, although it hits the first time).
This reproduces the problem easily without the fix, and works with the
fix.
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Reviewed-by: George Amanakis <gamanakis@gmail.com>
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#14239Closes#14289
This fixes a kernel stack leak.
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Tested-by: Nicholas Sherlock <n.sherlock@gmail.com>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#13778Closes#14255
We currently compute a 64-bit hash three times, which consumes 0.8% CPU
time on ARC eviction heavy workloads. Caching the 64-bit value in the
dbuf allows us to avoid that overhead.
Sponsored-By: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: Richard Yao <richard.yao@klarasystems.com>
Closes#14251
If the fields to be listed and sorted by are constrained to those
populated by dsl_dataset_fast_stat(), then zfs list is much faster,
as it does not need to open each objset and reads its properties.
A previous optimization by Pawel Dawidek
(0cee24064a) took advantage
of this to make listing snapshot names sorted only by name much faster.
However, it was limited to `-o name -s name`, this work extends this
optimization to work with:
- name
- guid
- createtxg
- numclones
- inconsistent
- redacted
- origin
and could be further extended to any other properties supported by
dsl_dataset_fast_stat() or similar, that do not require extra locking
or reading from disk.
This was committed before (9a9e2e343dfa2af28bf7910de77ae73aa006de62),
but was reverted due to a regression when used with an older kernel.
If the kernel does not populate zc->zc_objset_stats, we now fallback
to getting the properties via the slower interface, to avoid problems
with newer userland and older kernels.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Allan Jude <allan@klarasystems.com>
Closes#14110
Context:
We recently had a scenario where a customer with 2x10TB disks at 95+%
fragmentation and capacity, wanted to migrate their disks to a 2x20TB
setup. So they added the 2 new disks and submitted the removal of the
first 10TB disk. The removal took a lot more than expected (order of
more than a week to 2 weeks vs a couple of days) and once it was done it
generated a huge indirect mappign table in RAM (~16GB vs expected ~1GB).
Root-Cause:
The removal code calls `metaslab_alloc_dva()` to allocate a new block
for each evacuating block in the removing device and it tries to batch
them into 16MB segments. If it can't find such a segment it tries for
8MBs, 4MBs, all the way down to 512 bytes.
In our scenario what would happen is that `metaslab_alloc_dva()` from
the removal thread pick the new devices initially but wouldn't allocate
from them because of throttling in their metaslab allocation queue's
depth (see `metaslab_group_allocatable()`) as these devices are new and
favored for most types of allocations because of their free space. So
then the removal thread would look at the old fragmented disk for
allocations and wouldn't find any contiguous space and finally retry
with a smaller allocation size until it would to the low KB range. This
caused a lot of small mappings to be generated blowing up the size of
the indirect table. It also wasted a lot of CPU while the removal was
active making everything slow.
This patch:
Make all allocations coming from the device removal thread bypass the
throttle checks. These allocations are not even counted in the metaslab
allocation queues anyway so why check them?
Side-Fix:
Allocations with METASLAB_DONT_THROTTLE in their flags would not be
accounted at the throttle queues but they'd still abide by the
throttling rules which seems wrong. This patch fixes this by checking
for that flag in `metaslab_group_allocatable()`. I did a quick check to
see where else this flag is used and it doesn't seem like this change
would cause issues.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Closes#14159
If the bp is NULL, we have a hole. However, when we build with
assertions, we will dereference bp when `blkid == DMU_SPILL_BLKID`. When
this happens on a hole, we will have a NULL pointer dereference.
Reported-by: Coverity (CID-1524670)
Reviewed-by: Damian Szuberski <szuberskidamian@gmail.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14264
dsl_dataset_snapshot_sync_impl() declares `static zil_header_t zero_zil
__maybe_unused;`, but this is also declared globally. This wastes
memory.
CodeQL's cpp/local-variable-hides-global-variable check caught this.
Reviewed-by: Damian Szuberski <szuberskidamian@gmail.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14263
When doing a device removal on a pool with gang blocks, the zio pipeline
can deadlock when trying to free blocks from a device which is being
removed with a stack similar to this:
0xffff8ab9a13a1740 UNINTERRUPTIBLE 4
__schedule+0x2e5
__schedule+0x2e5
schedule+0x33
schedule_preempt_disabled+0xe
__mutex_lock.isra.12+0x2a7
__mutex_lock.isra.12+0x2a7
__mutex_lock_slowpath+0x13
mutex_lock+0x2c
free_from_removing_vdev+0x61
metaslab_free_impl+0xd6
metaslab_free_dva+0x5e
metaslab_free+0x196
zio_free_sync+0xe4
zio_free_gang+0x38
zio_gang_tree_issue+0x42
zio_gang_tree_issue+0xa2
zio_gang_issue+0x6d
zio_execute+0x94
zio_execute+0x94
taskq_thread+0x23b
kthread+0x120
ret_from_fork+0x1f
Since there are gang blocks we have to read the gang members as part of
the free. This can be seen with a zio dependency tree that looks like
this:
sdb> echo 0xffff900c24f8a700 | zio -rc | zio
ADDRESS TYPE STAGE WAITER
0xffff900c24f8a700 NULL CHECKSUM_VERIFY 0xffff900ddfd31740
0xffff900c24f8c920 FREE GANG_ASSEMBLE -
0xffff900d93d435a0 READ DONE
In the illustration above we are processing frees but because of gang
block we have to read the constituents blocks. Once we finish the READ
in the zio pipeline we will execute the parent. In this case the parent
is a FREE but the zio taskq is a READ and we continue to process the
pipeline leading to the stack above. In the stack above, we are blocked
waiting for the svr_lock so as a result a READ interrupt taskq thread
is now consumed. Eventually, all of the READ taskq threads end up
blocked and we're unable to complete any read requests.
In zio_notify_parent there is an optimization to continue to use
the taskq thread to exectue the parent's pipeline. To resolve the
deadlock above, we only allow this optimization if the parent's
zio type matches the child which just completed.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: George Wilson <gwilson@delphix.com>
External-issue: DLPX-80130
Closes#14236
After a device has been removed, any nopwrites for blocks on that
indirect vdev should be ignored and a new block should be allocated. The
original code attempted to handle this but used the wrong block pointer
when checking for indirect vdevs and failed to check all DVAs.
This change corrects both of these issues and modifies the test case
to ensure that it properly tests nopwrites with device removal.
Reviewed-by: Prakash Surya <prakash.surya@delphix.com>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Wilson <gwilson@delphix.com>
Closes#14235
The checksum error counter is incremented after reporting to ZED. This
leads ZED to receiving a checksum error report with 0 checksum errors.
To avoid this, bump the checksum error counter before reporting to ZED.
Sponsored-by: Seagate Technology LLC
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Wing <rob.wing@klarasystems.com>
Closes#14190
- Clang 15 doesn't support `-fno-ipa-sra` anymore. Do a separate
check for `-fno-ipa-sra` support by $KERNEL_CC.
- Don't enable `-mgeneral-regs-only` for certain module files.
Fix#13260
- Scope `GCC diagnostic ignored` statements to GCC only. Clang
doesn't need them to compile the code.
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: szubersk <szuberskidamian@gmail.com>
Closes#13260Closes#14150
When ZFS is built with assertions, a prefetch is done on a redacted
blkptr and `dpa->dpa_dnode` is NULL, we will have a NULL pointer
dereference in `dbuf_prefetch_indirect_done()`.
Both Coverity and Clang's Static Analyzer caught this.
Reported-by: Coverity (CID 1524671)
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14210
range is always deferenced before it reaches this check, such that the
kmem_zalloc() call is never executed.
A previously version of this had erronously also pruned the
`range->eos_marker = B_TRUE` line, but it must be set whenever we
encounter an error or are cancelled early.
Coverity incorrectly complained about a potential NULL pointer
dereference because of this.
Reported-by: Coverity (CID 1524550)
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14210
There was the series from me a year ago which fixed most of the
callback vs implementation prototype mismatches. It was based on
running the CFI-enabled kernel (in permissive mode -- warning
instead of panic) and performing a full ZTS cycle, and then fixing
all of the problems caught by CFI.
Now, Clang 16-dev has new warning flag, -Wcast-function-type-strict,
which detect such mismatches at compile-time. It allows to find the
remaining issues missed by the first series.
There are only two of them left: one for the
secpolicy_vnode_setattr() callback and one for taskq_dispatch().
The fix is easy, since they are not used anywhere else.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Lobakin <alobakin@pm.me>
Closes#14207
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#14199
I've noticed that some of those counters are used in hot paths like
dnode_hold_impl(), and results of this change is visible in profiler.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#14198
atomic_dec_32() should be a bit lighter than atomic_dec_32_nv().
Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Signed-off-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#14200
If the attached disk already contains a vdev GUID, it
means the disk is not clean. In such a scenario, the
physical path would be a match that makes the disk
faulted when trying to online it. So, we would only
want to proceed if either GUID matches with the last
attached disk or the disk is in a clean state.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Ameer Hamza <ahamza@ixsystems.com>
Closes#14181
