]> git.proxmox.com Git - mirror_zfs.git/log
mirror_zfs.git
6 months agozdb: add missing cleanup for early return
Ameer Hamza [Thu, 9 May 2024 14:31:57 +0000 (19:31 +0500)]
zdb: add missing cleanup for early return

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Don Brady <don.brady@klarasystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Ameer Hamza <ahamza@ixsystems.com>
Closes #16152

6 months agoReplace usage of schedule_timeout with schedule_timeout_interruptible (#16150)
Daniel Perry [Thu, 9 May 2024 14:30:28 +0000 (10:30 -0400)]
Replace usage of schedule_timeout with schedule_timeout_interruptible (#16150)

This commit replaces current usages of schedule_timeout() with
schedule_timeout_interruptible() in code paths that expect the running
task to sleep for a short period of time. When schedule_timeout() is
called without previously calling set_current_state(), the running
task never sleeps because the task state remains in TASK_RUNNING.

By calling schedule_timeout_interruptible() to set the task state to
TASK_INTERRUPTIBLE before calling schedule_timeout() we achieve the
intended/desired behavior of putting the task to sleep for the
specified timeout.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Daniel Perry <dtperry@amazon.com>
Closes #16150

6 months agoDisable high priority ZIO threads on FreeBSD and Linux
Alexander Motin [Fri, 3 May 2024 16:53:34 +0000 (12:53 -0400)]
Disable high priority ZIO threads on FreeBSD and Linux

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

6 months agovdev_disk: disable flushes if device does not support it
Rob N [Thu, 2 May 2024 22:18:35 +0000 (08:18 +1000)]
vdev_disk: disable flushes if device does not support it

If the underlying device doesn't have a write-back cache, the kernel
will just return a successful response. This doesn't hurt anything, but
it's extra work on the IO taskqs that are unnecessary. So, detect this
when we open the device for the first time.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #16148

6 months agoImprove write issue taskqs utilization
Alexander Motin [Wed, 1 May 2024 18:07:20 +0000 (14:07 -0400)]
Improve write issue taskqs utilization

- 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

6 months agoSlightly improve dnode hash
Alexander Motin [Wed, 1 May 2024 17:59:32 +0000 (13:59 -0400)]
Slightly improve dnode hash

As I understand just for being less predictable dnode hash includes
8 bits of objset pointer, starting at 6.  But since objset_t is
more than 1KB in size, its allocations are likely aligned to 2KB,
that means 11 lower bits provide no entropy. Just take the 8 bits
starting from 11.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes #16131

6 months agolibspl/assert: use libunwind for backtrace when available
Rob Norris [Tue, 30 Apr 2024 00:37:29 +0000 (10:37 +1000)]
libspl/assert: use libunwind for backtrace when available

libunwind seems to do a better job of resolving a symbols than
backtrace(), and is also useful on platforms that don't have backtrace()
(eg musl). If it's available, use it.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Sponsored-by: https://despairlabs.com/sponsor/
Closes #16140

6 months agolibspl/assert: dump backtrace in assert
Rob Norris [Sat, 27 Apr 2024 11:35:05 +0000 (21:35 +1000)]
libspl/assert: dump backtrace in assert

Adds a check for the backtrace() function. If available, uses it to show
a stack backtrace in the assertion output.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Sponsored-by: https://despairlabs.com/sponsor/
Closes #16140

6 months agolibspl/assert: add lock around assertion output
Rob Norris [Sun, 28 Apr 2024 02:49:58 +0000 (12:49 +1000)]
libspl/assert: add lock around assertion output

If multiple threads trip an assertion at the same moment (quite common),
they can be printing at the same time, and their output gets messy.

This adds a simple lock around the whole thing, to prevent a second task
printing assert output before the first has finished.

Additionally, if libspl_assert_ok is not set, abort() is called without
dropping the lock, so that any other asserting tasks will be killed
before starting any output, rather than only getting part-way through.
This is a tradeoff; it's assumed that multiple threads asserting at the
same moment are likely the same fault in different instances of a
thread, and so there won't be any more useful information from the other
tasks anyway.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Sponsored-by: https://despairlabs.com/sponsor/
Closes #16140

6 months agolibspl/assert: show process/task details in assert output
Rob Norris [Sun, 21 Apr 2024 11:43:53 +0000 (21:43 +1000)]
libspl/assert: show process/task details in assert output

Makes it much easier to see what thing complained.

Getting thread id, program name and thread name vary wildly between
Linux and FreeBSD, so those are set up in macros. pthread_getname_np()
did not appear in musl until very recently, but the same info has always
been available via prctl(PR_GET_NAME), so we use that instead.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Sponsored-by: https://despairlabs.com/sponsor/
Closes #16140

6 months agolibzpool: set thread names
Rob Norris [Sun, 28 Apr 2024 01:03:11 +0000 (11:03 +1000)]
libzpool: set thread names

Arrange for the thread/task name to be set when new threads are created.
This makes them visible in the process table etc.

pthread_setname_np() is generally available in glibc, musl and FreeBSD,
so no test is required.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Sponsored-by: https://despairlabs.com/sponsor/
Closes #16140

6 months agofind_system_library: fix var cleanup when library not found
Rob Norris [Tue, 30 Apr 2024 02:35:30 +0000 (12:35 +1000)]
find_system_library: fix var cleanup when library not found

The "not found" path is attempting to clear SOMELIB_CFLAGS and
SOMELIB_LIBS by resetting them in AC_SUBST(). However, the second arg to
AC_SUBST is expanded in autoconf with `m4_ifvaln([$2], [[$1]=$2])`,
which is defined as "if the first arg is non-empty". The m4 "empty"
construction is [], therefore, the existing AC_SUBST calls never modify
the variables at all.

The effect of this is that leftovers from the library test can leak out.
At least, if a library header is found in the first stage, but the
library itself is not, -lsomelib is added to SOMELIB_LIBS and further
tests done. If that library is not found, SOMELIB_LIBS will not be
cleared.

For most of our library tests this hasn't been a problem, as they're
either always found properly via pkg-config or set directly, or the
calling test immediately aborts configure. For an optional dependency
however, an apparent "partial" result where the header is found but no
corresponding library causes link errors later.

I think a complete fix should probably not be setting SOMELIB_xxx until
the final result is known, but for now, adjusting the AC_SUBST calls to
explictly set the empty shell string (which is not "empty" to m4) at
least restores the intent.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Sponsored-by: https://despairlabs.com/sponsor/
Closes #16140

6 months agozio: try to execute TYPE_NULL ZIOs on the current task
Rob N [Mon, 29 Apr 2024 22:57:32 +0000 (08:57 +1000)]
zio: try to execute TYPE_NULL ZIOs on the current task

Many TYPE_NULL ZIOs are used to provide a sync point for child ZIOs, and
do not do any actual work themselves. However, they are still dispatched
to a dedicated, single-thread taskq, which leads to their execution
being entirely task switch and dequeue overhead for no actual reason.

This commit changes it so that when selecting a parent ZIO to execute,
if the parent is TYPE_NULL and has no done function (that is, no
additional work), it is executed on the same thread. This reduces task
switches and frees up CPU cores for other work.

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 #16134

6 months agovdev probe to slow disk can stall mmp write checker
Don Brady [Mon, 29 Apr 2024 21:35:53 +0000 (15:35 -0600)]
vdev probe to slow disk can stall mmp write checker

Simplify vdev probes in the zio_vdev_io_done context to
avoid holding the spa config lock for a long duration.

Also allow zpool clear if no evidence of another host
is using the pool.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Olaf Faaland <faaland1@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Don Brady <don.brady@klarasystems.com>
Closes #15839

6 months agoFix arcstats for FreeBSD after zfetch support
Ameer Hamza [Mon, 29 Apr 2024 20:28:50 +0000 (01:28 +0500)]
Fix arcstats for FreeBSD after zfetch support

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Ameer Hamza <ahamza@ixsystems.com>
Closes #16141

6 months agoOverflowing refreservation is bad
Rich Ercolani [Mon, 29 Apr 2024 18:32:49 +0000 (14:32 -0400)]
Overflowing refreservation is bad

Someone came to me and pointed out that you could pretty
readily cause the refreservation calculation to exceed
2**64, given the 2**17 multiplier in it, and produce
refreservations wildly less than the actual volsize in cases where
it should have failed.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rich Ercolani <rincebrain@gmail.com>
Closes #15996

6 months agoGCC: Fixes for gcc 14 on Fedora 40
Tony Hutter [Mon, 29 Apr 2024 18:31:50 +0000 (11:31 -0700)]
GCC: Fixes for gcc 14 on Fedora 40

- Workaround dangling pointer in uu_list.c (#16124)
- Fix calloc() transposed arguments in zpool_vdev_os.c
- Make some temp variables unsigned to prevent triggering a
  '-Werror=alloc-size-larger-than' error.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tony Hutter <hutter2@llnl.gov>
Closes #16124
Closes #16125

6 months agoFix updating the zvol_htable when renaming a zvol
Alan Somers [Thu, 25 Apr 2024 21:24:52 +0000 (16:24 -0500)]
Fix updating the zvol_htable when renaming a zvol

When renaming a zvol, insert it into zvol_htable using the new name, not
the old name.  Otherwise some operations won't work.  For example,
"zfs set volsize" while the zvol is open.

Sponsored by: Axcient
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alek Pinchuk <apinchuk@axcient.com>
Signed-off-by: Alan Somers <asomers@FreeBSD.org>
Closes #16127
Closes #16128

6 months agoPython 3.12 deprecated python3-distutils
Brian Behlendorf [Thu, 25 Apr 2024 20:40:09 +0000 (13:40 -0700)]
Python 3.12 deprecated python3-distutils

As for python-3.12 the distutils package has been deprecated.
The latest ax_python_devel.m4 macro from the autoconf archive
has been updated accordingly so let's pull in the new version.

We can also drop the changes made to our customized version
to continue if the development version is not installed since
this functionality has been included upstream.

Reviewed-by: Rich Ercolani <rincebrain@gmail.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #16126
Closes #16129

6 months agoFast Dedup: ZAP Shrinking
Allan Jude [Wed, 24 Apr 2024 21:51:21 +0000 (17:51 -0400)]
Fast Dedup: ZAP Shrinking

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

6 months agoMake more taskq parameters writable
Alexander Motin [Wed, 24 Apr 2024 21:38:48 +0000 (17:38 -0400)]
Make more taskq parameters writable

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

7 months agoL2ARC: Cleanup buffer re-compression
Alexander Motin [Tue, 23 Apr 2024 16:06:00 +0000 (12:06 -0400)]
L2ARC: Cleanup buffer re-compression

When compressed ARC is disabled, we may have to re-compress when
writing into L2ARC.  If doing so we can't fit it into the original
physical size, we should just fail immediately, since even if it
may still fit into allocation size, its checksum will never match.

While there, refactor the code similar to other compression places
without using abd_return_buf_copy().

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes #16038

7 months agozfs-kmod: fix empty rpm requires/conflicts
Todd [Tue, 23 Apr 2024 00:55:41 +0000 (17:55 -0700)]
zfs-kmod: fix empty rpm requires/conflicts

Fix an error in zfs-kmod.spec that causes kmod-zfs packages not to
include the correct RPM requires/conflicts relationships.  With this
change applied, RPM correctly no longer allows kmod-zfs & zfs-dkms
packages to be installed together.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Todd Seidelmann <18294602+seidelma@users.noreply.github.com>
Closes #16121

7 months agoRefactor dbuf_read() for safer decryption
Alexander Motin [Mon, 22 Apr 2024 18:41:03 +0000 (14:41 -0400)]
Refactor dbuf_read() for safer decryption

In dbuf_read_verify_dnode_crypt():
 - We don't need original dbuf locked there. Instead take a lock
on a dnode dbuf, that is actually manipulated.
 - Block decryption for a dnode dbuf if it is currently being
written.  ARC hash lock does not protect anonymous buffers, so
arc_untransform() is unsafe when used on buffers being written,
that may happen in case of encrypted dnode buffers, since they
are not copied by dbuf_dirty()/dbuf_hold_copy().

In dbuf_read():
 - If the buffer is in flight, recheck its compression/encryption
status after it is cached, since it may need arc_untransform().

Tested-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 #16104

7 months agozfs get: add '-t fs' and '-t vol' options
Ryan [Mon, 22 Apr 2024 17:59:31 +0000 (01:59 +0800)]
zfs get: add '-t fs' and '-t vol' options

Make `zfs get` accept `fs` for `filesystem` and `vol` for `volume`.

Reviewed-by: Rob Norris <rob.norris@klarasystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ryan <errornointernet@envs.net>
Closes #16117

7 months agoztest: use ASSERT3P to compare pointers
Brooks Davis [Mon, 22 Apr 2024 17:48:58 +0000 (10:48 -0700)]
ztest: use ASSERT3P to compare pointers

With a sufficiently modern gcc (I saw this with gcc13), gcc complains
when casting pointers to an integer of a different type (even a larger
one).  On 32-bt ASSERT3U does this on 32-bit systems by casting a 32-bit
pointer to uint64_t so use ASSERT3P which uses uintptr_t.

Fixes: 5caeef02fa53 RAID-Z expansion feature
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Brooks Davis <brooks.davis@sri.com>
Closes #16115

7 months agoZTS: user_namespace_004.ksh avoid error in cleanup if unsupported
Seth Troisi [Mon, 22 Apr 2024 17:47:44 +0000 (10:47 -0700)]
ZTS: user_namespace_004.ksh avoid error in cleanup if unsupported

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Seth Troisi <sethtroisi@google.com>
Closes #16114

7 months agoAdd newline to two zpool messages
Seth Troisi [Mon, 22 Apr 2024 17:45:39 +0000 (10:45 -0700)]
Add newline to two zpool messages

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Seth Troisi <sethtroisi@google.com>
Closes #16113

7 months agoParallel pool import
George Wilson [Mon, 22 Apr 2024 16:42:38 +0000 (12:42 -0400)]
Parallel pool import

This commit allow spa_load() to drop the spa_namespace_lock so
that imports can happen concurrently. Prior to dropping the
spa_namespace_lock, the import logic will set the spa_load_thread
value to track the thread which is doing the import.

Consumers of spa_lookup() retain the same behavior by blocking
when either a thread is holding the spa_namespace_lock or the
spa_load_thread value is set. This will ensure that critical
concurrent operations cannot take place while a pool is being
imported.

The zpool command is also enhanced to provide multi-threaded support
when invoking zpool import -a.

Lastly, zinject provides a mechanism to insert artificial delays
when importing a pool and new zfs tests are added to verify parallel
import functionality.

Contributions-by: Don Brady <don.brady@klarasystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Wilson <gwilson@delphix.com>
Closes #16093

7 months agoabd_iter_page: rework to handle multipage scatterlists
Rob N [Fri, 19 Apr 2024 23:41:31 +0000 (09:41 +1000)]
abd_iter_page: rework to handle multipage scatterlists

Previously, abd_iter_page() would assume that every scatterlist would
contain a single page (compound or no), because that's all we ever
create in abd_alloc_chunks(). However, scatterlists can contain multiple
pages of arbitrary provenance, and if we get one of those, we'd get all
the math wrong.

This reworks things to handle multiple pages in a scatterlist, by
properly finding the right page within it for the given offset, and
understanding better where the end of the page is and not crossing it.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reported-by: Brian Atkinson <batkinson@lanl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #16108

7 months agoHandle FLUSH errors as "expected"
Alexander Motin [Fri, 19 Apr 2024 23:18:54 +0000 (19:18 -0400)]
Handle FLUSH errors as "expected"

Before #16061 zio_vdev_io_done() was not used for FLUSH requests.
Addition of it triggers reprobe each TXG for vdevs not supporting
them.  Since those errors are often expected, they are normally
handled by individual vdev drivers and should be ignored here.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Rob Norris <rob.norris@klarasystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes #16110

7 months agotests/quota: consistently clear quota property between tests
Rob Norris [Tue, 16 Apr 2024 04:56:35 +0000 (14:56 +1000)]
tests/quota: consistently clear quota property between tests

When run in isolation, quota_005_pos would fail in cleanup because it
would attempt restore the previous quota, which was 0, and so get an
error (because you can't set quota to '0', you have to use 'none').

It worked as part of the quota tag set because the previous tests did
not clean up their quota, so there was always a non-zero quota to return
to.

This adds a simple quota reset function, and has all quota tests run it
at cleanup. For the ones that weren't cleaning up, they now do, and for
quota_005_pos, which was trying to do the right thing, it now just
resets it.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #16097

7 months agotests/quota_005_pos: use a long int for doubling the quota size
Rob Norris [Tue, 16 Apr 2024 05:03:33 +0000 (15:03 +1000)]
tests/quota_005_pos: use a long int for doubling the quota size

When run in isolation, quota_005_pos would see an empty ~300G dataset.
Doubling it's space overflows a int32, which meant it was trying to then
set the quota to a negative value, and would fail.

When run as part of the quota tests, the filesystem appears to have
stuff in it, and so a lower available space, which doesn't overflow, and
so succeeds.

The bare minimum fix seems to be to use a int64 for the available space,
so it can be comfortably doubled. Here it is.

(Also a typo fix and a tiny bit of cleanup).

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #16097

7 months agoAdd zfetch stats in arcstats
Ameer Hamza [Fri, 19 Apr 2024 17:19:12 +0000 (22:19 +0500)]
Add zfetch stats in arcstats

arc_summary also reports zfetch stats but it's inconvenient to monitor
contiguously incrementing numbers. Adding them in arcstats allows us to
observe streams more conveniently.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Ameer Hamza <ahamza@ixsystems.com>
Closes #16094

7 months agoFix: FreeBSD Arm64 does not build currently
Tino Reichardt [Fri, 19 Apr 2024 17:15:38 +0000 (19:15 +0200)]
Fix: FreeBSD Arm64 does not build currently

The define LD_VERSION isn't defined on FreeBSD Arm64 when OpenZFS is
build with the default compiler: clang.
I used only gcc for testing - my fault.

Fast fix as suggested by @mmatuska

Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Martin Matuska <mm@FreeBSD.org>
Signed-off-by: Tino Reichardt <milky-zfs@mcmilk.de>
Closes #16103

7 months agoLinux 6.8 compat: META (#16099)
Tony Hutter [Wed, 17 Apr 2024 16:29:21 +0000 (09:29 -0700)]
Linux 6.8 compat: META (#16099)

Update the META file to reflect compatibility with the 6.8 kernel.

Signed-off-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Rob Norris <rob.norris@klarasystems.com>
7 months agozts: add a debug option to get full test output
Rob N [Tue, 16 Apr 2024 16:13:01 +0000 (02:13 +1000)]
zts: add a debug option to get full test output

The test runner accumulates output from individual tests, then writes it
to the log at the end. If a test hangs or crashes the system half way
through, we get no insight into how it got to where it did.

This adds a -D option for "debug". When set, all test output is written
to stdout.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Akash B <akash-b@hpe.com>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #16096

7 months agoDo no use .cfi_negate_ra_state within the assembly on Arm64
Tino Reichardt [Mon, 15 Apr 2024 20:56:10 +0000 (22:56 +0200)]
Do no use .cfi_negate_ra_state within the assembly on Arm64

Compiling openzfs on aarch64 with gcc-8 and gcc-9 is failing currently.
See issue #14965 for deeper context.

On platforms without pointer authentication, .cfi_negate_ra_state can be
defined to a no-op:
https://sourceware.org/git/?p=binutils-gdb.git;a=blob;f=gdb/aarch64-tdep.c#l1413

I have tested this on Arm64 FreeBSD 13.2 and AlmaLinux-8.

Reviewed-by: Andrew Turner <andrew.turner4@arm.com>
Signed-off-by: Tino Reichardt <milky-zfs@mcmilk.de>
Closes #14965
Closes #15784

7 months agoAdd the BTI elf note to the AArch64 SHA2 assembly
Andrew Turner [Mon, 15 Apr 2024 20:53:39 +0000 (21:53 +0100)]
Add the BTI elf note to the AArch64 SHA2 assembly

On ELF platforms there is a note to specify when an application or
library supports BTI. When linking one of these the linker needs
all input object files to have the note. If not it will not include
it in the output file.

Normally the compiler would generate it, but for assembly files we
need to do it our selves.

Add the note to the aarch64 sha256 and sha512 assembly files.

Tested by building with BTI enabled and using the -zbti-report=error
flag to lld that makes it an error if the note is missing.

Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Andrew Turner <andrew.turner4@arm.com>
Closes #16086

7 months agozinject: "no-op" error injection
Rob N [Mon, 15 Apr 2024 20:52:20 +0000 (06:52 +1000)]
zinject: "no-op" error injection

When injected, this causes the matching IO to appear to succeed, but the
actual work is never submitted to the physical device. This can be used
to simulate a write-back cache servicing a write, but the backing device
has failed and the cache cannot complete the operation in the
background.

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 #16085

7 months agozts: allow running a single test by name only
Rob N [Mon, 15 Apr 2024 20:44:12 +0000 (06:44 +1000)]
zts: allow running a single test by name only

Specifying a single test is kind of a hassle, because the full relative
path under the test suite dir has to be included, but it's not always
clear what that path even is.

This change allows `-t` to take the name of a single test instead of a
full path. If the value has no `/` characters, we search for a file of
that name under the test root, and if found, use that as the full test
path instead.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Akash B <akash-b@hpe.com>
Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #16088

7 months agobdev_discard_supported: understand discard_granularity=0
Rob N [Fri, 12 Apr 2024 16:00:20 +0000 (02:00 +1000)]
bdev_discard_supported: understand discard_granularity=0

Kernel documentation for the discard_granularity property says:

    A discard_granularity of 0 means that the device does not support
    discard functionality.

Some older kernels had drivers (notably loop, but also some USB-SATA
adapters) that would set the QUEUE_FLAG_DISCARD capability flag, but
have discard_granularity=0. Since 5.10 (torvalds/linux@b35fd7422c2f) the
discard entry point blkdev_issue_discard() has had a check for this,
which would immediately reject the call with EOPNOTSUPP, and throw a
scary diagnostic message into the log. See #16068.

Since 6.8, the block layer sets a non-zero default for
discard_granularity (torvalds/linux@3c407dc723bb), and a future kernel
will remove the check entirely[1].

As such, there's no good reason for us to enable discard when
discard_granularity=0. The kernel will never let the request go in
anyway; better that we just disable it so we can report it properly to
the user.

1. https://patchwork.kernel.org/project/linux-block/patch/20240312144826.1045212-2-hch@lst.de/

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 #16068
Closes #16082

7 months agozio: rename ZIO_TYPE_IOCTL to ZIO_TYPE_FLUSH
Rob Norris [Thu, 4 Apr 2024 11:35:00 +0000 (22:35 +1100)]
zio: rename ZIO_TYPE_IOCTL to ZIO_TYPE_FLUSH

The only possible ioctl is a flush, and any other kind of meta-operation
introduced in the future is likely to have different semantics (much
like trim did). So, lets just call it what it is.

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 #16064

7 months agodkio: remove kernel dkio.h compatibility header
Rob Norris [Wed, 10 Apr 2024 06:07:24 +0000 (16:07 +1000)]
dkio: remove kernel dkio.h compatibility header

Without DKIOCFLUSHWRITECACHE, we no longer need the compat header. Note
that we're keeping the userspace SPL compat header, which is used by
libefi.

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 #16064

7 months agozio: remove io_cmd and DKIOCFLUSHWRITECACHE
Rob Norris [Thu, 4 Apr 2024 11:34:54 +0000 (22:34 +1100)]
zio: remove io_cmd and DKIOCFLUSHWRITECACHE

There's no other options, so we can just always assume its a flush.

Includes some light refactoring where a switch statement was doing
control flow that no longer works.

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 #16064

7 months agozio: remove zio_ioctl()
Rob Norris [Thu, 4 Apr 2024 11:34:42 +0000 (22:34 +1100)]
zio: remove zio_ioctl()

It only had one user, zio_flush(), and there are no other vdev ioctls
anyway.

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 #16064

7 months agoAdd support for zfs mount -R <filesystem>
Umer Saleem [Thu, 11 Apr 2024 22:10:24 +0000 (03:10 +0500)]
Add support for zfs mount -R <filesystem>

This commit adds support for mounting a dataset along with all of
it's children with '-R' flag for zfs mount. There can be scenarios
where we want to mount all datasets under one hierarchy instead of
mounting all datasets present on system with '-a' flag.

'-R' flag should work on all root and non-root datasets. Usage
information and man page has been updated for zfs mount. A test
for verifying the behavior for '-R' flag is also added.

Reviewed-by: Ameer Hamza <ahamza@ixsystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Umer Saleem <usaleem@ixsystems.com>
Closes #16015

7 months agoAUTHORS: refresh with recent new contributors
Rob N [Thu, 11 Apr 2024 21:49:57 +0000 (07:49 +1000)]
AUTHORS: refresh with recent new contributors

Sponsored-by: https://despairlabs.com/sponsor/
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Closes #16079

7 months agovdev_disk: fix alignment check when buffer has non-zero starting offset
Rob Norris [Wed, 10 Apr 2024 03:14:13 +0000 (13:14 +1000)]
vdev_disk: fix alignment check when buffer has non-zero starting offset

If a linear buffer spans multiple pages, and the first page has a
non-zero starting offset, the checker would not include the offset, and
so would think there was an alignment gap at the end of the first page,
rather than at the start.

That is, for a 16K buffer spread across five pages with an initial 512B
offset:

    [.XXXXXXX][XXXXXXXX][XXXXXXXX][XXXXXXXX][XXXXXXX.]

It would be interpreted as:

    [XXXXXXX.][XXXXXXXX]...

And be rejected as misaligned.

Since it's already a linear ABD, the "linearising" copy would just reuse
the buffer as-is, and the second check would failing, tripping the
VERIFY in vdev_disk_io_rw().

This commit fixes all this by including the offset in the check for
end-of-page alignment.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #16076

7 months agotests: add test for vdev_disk page alignment check
Rob Norris [Wed, 10 Apr 2024 01:19:50 +0000 (11:19 +1000)]
tests: add test for vdev_disk page alignment check

This provides a test driver and a set of test vectors for the page
alignment check callback function vdev_disk_check_pages_cb().

Because there's no good facility for exposing this function to a
userspace test right now, for now I'm just duplicating the function and
adding commentary to remind people to keep them in sync.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #16076

7 months agoIllumos#16463 zfs_ioc_recv leaks nvlist
Andy Fiddaman [Thu, 11 Apr 2024 21:38:22 +0000 (22:38 +0100)]
Illumos#16463 zfs_ioc_recv leaks nvlist

In https://www.illumos.org/issues/16463 it was observed that
an nvlist was being leaked in zfs_ioc_recv() due a missing
call to nvlist_free for "hidden_args".
For OpenZFS the same issue exists in zfs_ioc_recv_new() and
is addressed by this PR.

This change also properly frees nvlists in the unlikely
event that a call to get_nvlist() fails.

Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Signed-off-by: Andy Fiddaman <illumos@fiddaman.net>
Closes #16077

7 months agoreturn NULL at end of send_progress_thread
Jason Lee [Wed, 10 Apr 2024 22:01:39 +0000 (16:01 -0600)]
return NULL at end of send_progress_thread

Reviewed-by: Rob Norris <robn@despairlabs.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Jason Lee <jasonlee@lanl.gov>
Closes #16074

7 months agoAdd custom debug printing for your asserts
Rich Ercolani [Wed, 10 Apr 2024 20:30:25 +0000 (16:30 -0400)]
Add custom debug printing for your asserts

Being able to print custom debug information on assert trip
seems useful.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Signed-off-by: Rich Ercolani <rincebrain@gmail.com>
Closes #15792

7 months agoconfig/Substfiles.am: restrict to the dedicated list.
Benda Xu [Tue, 9 Apr 2024 23:34:58 +0000 (07:34 +0800)]
config/Substfiles.am: restrict to the dedicated list.

We recover the scope of $(SUBSTFILES) to explicitly control what files
are being generated from the corresponding .in.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Benda Xu <orv@debian.org>
Closes #15980

7 months agoL2ARC: Relax locking during write
Alexander Motin [Tue, 9 Apr 2024 23:23:19 +0000 (19:23 -0400)]
L2ARC: Relax locking during write

Previous code held ARC state sublist lock throughout all L2ARC
write process, which included number of allocations and even ZIO
issues.  Being blocked in any of those places the code could also
block ARC eviction, that could cause OOM activation or even dead-
lock if system is low on memory or one is too fragmented.

Fix it by dropping the lock as soon as we see a block eligible
for L2ARC writing and pick it up later using earlier inserted
marker.  While there, also reduce scope of hash lock, moving
ZIO allocation and other operations not requiring header access
out of it.  All operations requiring header access move under
hash lock, since L2_WRITING flag does not prevent header eviction
only transition to arc_l2c_only state with L1 header.

To be able to manipulate sublist lock and marker as needed add few
more multilist functions and modify one.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes #16040

7 months agoSmall fix to prefetch ranges aggregation
Alexander Motin [Tue, 9 Apr 2024 23:14:04 +0000 (19:14 -0400)]
Small fix to prefetch ranges aggregation

When after #16022 adding new range we aggregate more than two
existing ranges, that should be very rare, only if several streams
overlap, we may need to zero not the last range, but some earlier.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes #16072

7 months agoetc/init.d: decide which variant to use at build time.
Benda Xu [Mon, 8 Apr 2024 23:52:24 +0000 (07:52 +0800)]
etc/init.d: decide which variant to use at build time.

Let Debian use the sysv-rc variant of the script, even when OpenRC is
installed. Unlike on Gentoo, OpenRC on Debian consumes both the
sysv-rc scripts and OpenRC ones. ZFS initscripts on Debian should be
the sysv-rc version to provide most compatibility and to integrate
with the rest of initscripts for dependency tracking.

Restrict the substitution in the Makefile to the dedicated list.

This construct is inspired by Mo Zhou's detection of the execution
shell and follows the strategy of Peter in 6ef28c526ba7.

As of 2024, the initscripts are mostly relevant on Debian, Gentoo and
their derivatives.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Benda Xu <orv@debian.org>
Issue #8063
Issue #8204
Issue #8359
Closes #15977

7 months agoFix locale-specific time
Maxim Filimonov [Mon, 8 Apr 2024 22:37:41 +0000 (02:37 +0400)]
Fix locale-specific time

In `zpool status -t`, scrub date/time is reported using the C locale,
while trim time is reported using the current one. This is inconsistent.
This patch fixes that.

Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Maxim Filimonov <che@bein.link>
Closes #15878
Closes #15879

7 months agoRemove db_state DB_NOFILL checks from syncing context
Alexander Motin [Mon, 8 Apr 2024 22:23:43 +0000 (18:23 -0400)]
Remove db_state DB_NOFILL checks from syncing context

Syncing context should not depend on current state of dbuf, which
could already change several times in later transaction groups,
but rely solely on dirty record for the transaction group being
synced. Some of the checks seem already impossible, while instead
of others I think we should better check for absence of data in
the specific dirty record rather than DB_NOFILL.

Reviewed-by: Robert Evans <evansr@google.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes #16057

7 months agoSpeculative prefetch for reordered requests
Alexander Motin [Mon, 8 Apr 2024 22:13:27 +0000 (18:13 -0400)]
Speculative prefetch for reordered requests

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

7 months agoFix read errors race after block cloning
Alexander Motin [Mon, 8 Apr 2024 19:03:18 +0000 (15:03 -0400)]
Fix read errors race after block cloning

Investigating read errors triggering panic fixed in #16042 I've
found that we have a race in a sync process between the moment
dirty record for cloned block is removed and the moment dbuf is
destroyed.  If dmu_buf_hold_array_by_dnode() take a hold on a
cloned dbuf before it is synced/destroyed, then dbuf_read_impl()
may see it still in DB_NOFILL state, but without the dirty record.
Such case is not an error, but equivalent to DB_UNCACHED, since
the dbuf block pointer is already updated by dbuf_write_ready().
Unfortunately it is impossible to safely change the dbuf state
to DB_UNCACHED there, since there may already be another cloning
in progress, that dropped dbuf lock before creating a new dirty
record, protected only by the range lock.

Reviewed-by: Rob Norris <robn@despairlabs.com>
Reviewed-by: Robert Evans <evansr@google.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes #16052

7 months agozinject: inject device errors into ioctls
Rob N [Mon, 8 Apr 2024 18:59:04 +0000 (04:59 +1000)]
zinject: inject device errors into ioctls

Adds 'ioctl' as a valid IO type for device error injection, so we can
simulate a flush error (which OpenZFS currently ignores, but that's by
the by).

To support this, adding ZIO_STAGE_VDEV_IO_DONE to ZIO_IOCTL_PIPELINE,
since that's where device error injection happens. This needs a small
exclusion to avoid the vdev_queue, since flushes are not queued, and I'm
assuming that the various failure responses are still reasonable for
flush failures (probes, media change, etc). This seems reasonable to me,
as a flush failure is not unlike a write failure in this regard, however
this may be too aggressive or subtle to assume in just this change.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #16061

7 months agovdev_disk: ensure trim errors are returned immediately
Rob N [Mon, 8 Apr 2024 18:50:24 +0000 (04:50 +1000)]
vdev_disk: ensure trim errors are returned immediately

After 06e25f9c4, the discard issuing code was organised such that if
requesting an async discard or secure erase failed before the IO was
issued (that is, calling __blkdev_issue_discard() returned an error),
the failed zio would never be executed, resulting in txg_sync hanging
forever waiting for IO to finish.

This commit fixes that by immediately executing a failed zio on error.
To handle the successful synchronous op case, we fake an async op by,
when not using an asynchronous submission method, queuing the successful
result zio as part of the discard handler.

Since it was hard to understand the differences between discard and
secure erase, and sync and async, across different kernel versions, I've
commented and reorganised the code a bit to try and make everything more
contained and linear.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Ameer Hamza <ahamza@ixsystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #16070

7 months agozvol_os: fix compile with blk-mq on Linux 4.x
Rob N [Mon, 8 Apr 2024 18:38:49 +0000 (04:38 +1000)]
zvol_os: fix compile with blk-mq on Linux 4.x

99741bde5 accesses a cached blk-mq hardware context through the mq_hctx
field of struct request. However, this field did not exist until 5.0.
Before that, the private function blk_mq_map_queue() was used to dig it
out of broader queue context. This commit detects this situation, and
handles it with a poor-man's simulation of that function.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Ameer Hamza <ahamza@ixsystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #16069

7 months agozvol_os: fix build on Linux <3.13
Rob N [Mon, 8 Apr 2024 17:13:27 +0000 (03:13 +1000)]
zvol_os: fix build on Linux <3.13

99741bde5 introduced zvol_num_taskqs, but put it behind the HAVE_BLK_MQ
define, preventing builds on versions of Linux that don't have it
(<3.13, incl EL7).

Nothing about it seems dependent on blk-mq, so this just moves it out
from behind that define and so fixes the build.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Ameer Hamza <ahamza@ixsystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #16062

7 months agozvol: use multiple taskq
Ameer Hamza [Thu, 4 Apr 2024 01:21:25 +0000 (06:21 +0500)]
zvol: use multiple taskq

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

7 months agoFix panics when truncating/deleting files
Pavel Snajdr [Thu, 4 Apr 2024 01:09:19 +0000 (03:09 +0200)]
Fix panics when truncating/deleting files

There's an union in dbuf_dirty_record_t; dr_brtwrite could evaluate
to B_TRUE if the dirty record is of another type than dl. Adding
more explicit dr type check before trying to access dr_brtwrite.

Fixes two similar panics:

[ 1373.806119] VERIFY0(db->db_level) failed (0 == 1)
[ 1373.807232] PANIC at dbuf.c:2549:dbuf_undirty()
[ 1373.814979]  dump_stack_lvl+0x71/0x90
[ 1373.815799]  spl_panic+0xd3/0x100 [spl]
[ 1373.827709]  dbuf_undirty+0x62a/0x970 [zfs]
[ 1373.829204]  dmu_buf_will_dirty_impl+0x1e9/0x5b0 [zfs]
[ 1373.831010]  dnode_free_range+0x532/0x1220 [zfs]
[ 1373.833922]  dmu_free_long_range+0x4e0/0x930 [zfs]
[ 1373.835277]  zfs_trunc+0x75/0x1e0 [zfs]
[ 1373.837958]  zfs_freesp+0x9b/0x470 [zfs]
[ 1373.847236]  zfs_setattr+0x161a/0x3500 [zfs]
[ 1373.855267]  zpl_setattr+0x125/0x320 [zfs]
[ 1373.856725]  notify_change+0x1ee/0x4a0
[ 1373.859207]  do_truncate+0x7f/0xd0
[ 1373.859968]  do_sys_ftruncate+0x28e/0x2e0
[ 1373.860962]  do_syscall_64+0x38/0x90
[ 1373.861751]  entry_SYSCALL_64_after_hwframe+0x6e/0xd8

[ 1822.381337] VERIFY0(db->db_level) failed (0 == 1)
[ 1822.382376] PANIC at dbuf.c:2549:dbuf_undirty()
[ 1822.389232]  dump_stack_lvl+0x71/0x90
[ 1822.389920]  spl_panic+0xd3/0x100 [spl]
[ 1822.399567]  dbuf_undirty+0x62a/0x970 [zfs]
[ 1822.400583]  dmu_buf_will_dirty_impl+0x1e9/0x5b0 [zfs]
[ 1822.401752]  dnode_free_range+0x532/0x1220 [zfs]
[ 1822.402841]  dmu_object_free+0x74/0x120 [zfs]
[ 1822.403869]  zfs_znode_delete+0x75/0x120 [zfs]
[ 1822.404906]  zfs_rmnode+0x3f6/0x7f0 [zfs]
[ 1822.405870]  zfs_inactive+0xa3/0x610 [zfs]
[ 1822.407803]  zpl_evict_inode+0x3e/0x90 [zfs]
[ 1822.408831]  evict+0xc1/0x1c0
[ 1822.409387]  do_unlinkat+0x147/0x300
[ 1822.410060]  __x64_sys_unlinkat+0x33/0x60
[ 1822.410802]  do_syscall_64+0x38/0x90
[ 1822.411458]  entry_SYSCALL_64_after_hwframe+0x6e/0xd8

Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Pavel Snajdr <snajpa@snajpa.net>
Closes #15983

7 months agoman: move zfs_prepare_disk.8 to nodist_man_MANS
Shengqi Chen [Thu, 4 Apr 2024 01:04:15 +0000 (09:04 +0800)]
man: move zfs_prepare_disk.8 to nodist_man_MANS

The commit b53077a added zfs_prepare_disk.8 to the wrong list
dist_man_MANS, in which @zfsexecdir@ will not be properly substituted.
This leads to wrong path in the manpage in generated release tarballs.

Reported-by: Benda Xu <orv@debian.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Shengqi Chen <harry-chen@outlook.com>
Closes #15979

7 months agovdev props comment and manpage should include zfsd and FreeBSD mentions
Alek P [Thu, 4 Apr 2024 00:56:34 +0000 (20:56 -0400)]
vdev props comment and manpage should include zfsd and FreeBSD mentions

Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Rob Norris <robn@despairlabs.com>
Reviewed-by: Allan Jude <allan@klarasystems.com>
Signed-off-by: Alek Pinchuk <apinchuk@axcient.com>
Closes #15968

7 months agozap_leaf: make l_hash[] variable length to silence UBSAN
Rob N [Wed, 3 Apr 2024 23:38:18 +0000 (10:38 +1100)]
zap_leaf: make l_hash[] variable length to silence UBSAN

When UBSAN is active and OpenZFS is a debug build, the l_hash assert at
the bottom of zap_open_leaf() causes UBSAN to complain.

This follows the example in 786641dcf to shut it up.

Sponsored-by: https://despairlabs.com/sponsor/
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Closes #15964

7 months agoGive a better message from 'zpool get' with invalid pool name
Paul Dagnelie [Wed, 3 Apr 2024 23:34:46 +0000 (16:34 -0700)]
Give a better message from 'zpool get' with invalid pool name

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Don Brady <don.brady@klarasystems.com>
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Signed-off-by: Paul Dagnelie <pcd@delphix.com>
Closes #15942

7 months agotests: simple zinject disk fault arg check
Rob Norris [Wed, 27 Mar 2024 05:15:48 +0000 (16:15 +1100)]
tests: simple zinject disk fault arg check

Just making sure the valid values for disk faults are accepted.
Obviously we can do a lot more, but this will do to get us started.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #15953

7 months agozinject: show more device fault fields
Rob Norris [Thu, 29 Feb 2024 23:38:41 +0000 (10:38 +1100)]
zinject: show more device fault fields

Once there's a few different kinds injected, its pretty hard to see them
otherwise.

So, lets show IO type, error type and frequency fields in the table too.

Since we now have to convert from error code to pretty string, refactor
the error names into a table and add lookup functions.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #15953

7 months agoMakefile.bsd: sort and cleanup source file list
Rob N [Wed, 3 Apr 2024 22:49:22 +0000 (09:49 +1100)]
Makefile.bsd: sort and cleanup source file list

All files now in their correct sections, and all sections match on-disk
dir layout, and all sorted.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #15943

7 months agoLinux 6.9 compat: blk_alloc_disk() now takes two args
Rob Norris [Wed, 27 Mar 2024 00:24:57 +0000 (11:24 +1100)]
Linux 6.9 compat: blk_alloc_disk() now takes two args

There's an extra nullable arg for queue limits. Detect it, and set it to
NULL. Similar change for blk_mq_alloc_disk(), now three args, same
treatment.

Error return now has error encoded in the return, so detect with
IS_ERR() and explicitly NULL our own return.

Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Sponsored-by: https://despairlabs.com/sponsor/
Closes #16027
Closes #16033

7 months agoLinux 6.9 compat: bdev handles are now struct file
Rob Norris [Tue, 26 Mar 2024 23:07:50 +0000 (10:07 +1100)]
Linux 6.9 compat: bdev handles are now struct file

bdev_open_by_path() is replaced by bdev_file_open_by_path(), which
returns a plain old struct file*. Release function is gone entirely; the
regular file release function fput() will take care of the bdev
specifics.

Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Sponsored-by: https://despairlabs.com/sponsor/
Closes #16027
Closes #16033

7 months agovdev_disk: don't touch vbio after its handed off to the kernel
Rob N [Wed, 3 Apr 2024 22:17:07 +0000 (09:17 +1100)]
vdev_disk: don't touch vbio after its handed off to the kernel

After IO is unplugged, it may complete immediately and vbio_completion
be called on interrupt context. That may interrupt or deschedule our
task. If its the last bio, the vbio will be freed. Then, we get
rescheduled, and try to write to freed memory through vbio->.

This patch just removes the the cleanup, and the corresponding assert.
These were leftovers from a previous iteration of vbio_submit() and were
always "belt and suspenders" ops anyway, never strictly required.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc
Reported-by: Rich Ercolani <rincebrain@gmail.com>
Reviewed-by: Laurențiu Nicola <lnicola@dend.ro>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: George Wilson <george.wilson@delphix.com>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #16045
Closes #16050
Closes #16049

7 months agoxdr: header cleanup
Rob N [Wed, 3 Apr 2024 22:13:27 +0000 (09:13 +1100)]
xdr: header cleanup

#16047 notes that include/os/freebsd/spl/rpc/xdr.h carried an
(apparently) incompatible license. While looking into it, it seems that
this file is actually unnecessary these days - FreeBSD's kernel XDR has
XDR_CONTROL, xdrmem_control and XDR_GET_BYTES_AVAIL, while userspace has
XDR_CONTROL and xdrmem_control, and our implementation of
XDR_GET_BYTES_AVAIL for libspl works nicely with it. So this removes
that file outright.

To keep the includes in nvpair.c tidy, I've made a few small adjustments
to the Linux headers. By definition, rpc/types.h provides bool_t and is
included before rpc/xdr.h, so I've created rpc/types.h for Linux. This
isn't necessary for userspace; both FreeBSD native and tirpc on Linux
already have these headers set up correctly.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Sponsored-by: https://despairlabs.com/sponsor/
Closes #16047
Closes #16051

7 months agoImprove dbuf_read() error reporting
Alexander Motin [Wed, 3 Apr 2024 22:04:26 +0000 (18:04 -0400)]
Improve dbuf_read() error reporting

Previous code reported non-ZIO errors only via return value, but
not via parent ZIO.  It could cause NULL-dereference panics due
to dmu_buf_hold_array_by_dnode() ignoring the return value,
relying solely on parent ZIO status.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ameer Hamza <ahamza@ixsystems.com>
Reported by: Ameer Hamza <ahamza@ixsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes #16042

7 months agoLinux 5.18+ compat: Detect filemap_range_has_page
Robert Evans [Sat, 30 Mar 2024 00:11:52 +0000 (20:11 -0400)]
Linux 5.18+ compat: Detect filemap_range_has_page

In v5.18 `filemap_range_has_page` moved to `pagemap.h`

`pagemap.h` has been around since 3.10 so just include both

Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Rob Norris <robn@despairlabs.com>
Signed-off-by: Robert Evans <evansr@google.com>
Closes #16034

7 months agoFix buffer underflow if sysfs file is empty
Robert Evans [Fri, 29 Mar 2024 21:59:23 +0000 (17:59 -0400)]
Fix buffer underflow if sysfs file is empty

Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Jason Lee <jasonlee@lanl.gov>
Signed-off-by: Robert Evans <evansr@google.com>
Closes #16028
Closes #16035

7 months agovdev_disk: clean up spa/bdev mode conversion
Rob N [Fri, 29 Mar 2024 21:51:33 +0000 (08:51 +1100)]
vdev_disk: clean up spa/bdev mode conversion

43e8f6e37 introduced a subtle API misuse, in that it passed the output
from vdev_bdev_mode() back into itself. Fortunately, the
SPA_MODE_(READ|WRITE) bit values exactly map to the FMODE_(READ|WRITE) &
BLK_OPEN_(READ|WRITE) bit values, so it didn't result in a bug, but it
was hard to read and understand, so I cleaned it up.

In doing so, I noticed that the only call to vdev_bdev_mode() without
the "exclusive" flag set was in that misuse, and actually, we never do a
non-exclusive blkdev_get_by_path(). So I've just made exclusive be
always-on.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Allan Jude <allan@klarasystems.com>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #15995

7 months agozvols: prevent overflow of minor device numbers
Fabian-Gruenbichler [Fri, 29 Mar 2024 21:37:40 +0000 (22:37 +0100)]
zvols: prevent overflow of minor device numbers

currently, the linux kernel allows 2^20 minor devices per major device
number.  ZFS reserves blocks of 2^4 minors per zvol: 1 for the zvol
itself, the other 15 for the first partitions of that zvol. as a result,
only 2^16 such blocks are available for use.

there are no checks in place to avoid overflowing into the major device
number when more than 2^16 zvols are allocated (with volmode=dev or
default). instead of ignoring this limit, which comes with all sorts of
weird knock-on effects, detect this situation and simply fail allocating
the zvol block device early on.

without this safeguard, the kernel will reject the attempt to create an
already existing block device, but ZFS doesn't handle this error and
gets confused about which zvol occupies which minor slot, potentially
resulting in kernel NULL derefs and other issues later on.

Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Fabian Grünbichler <f.gruenbichler@proxmox.com>
Closes #16006

7 months agoAdd ashift validation when adding devices to a pool
George Wilson [Fri, 29 Mar 2024 19:15:56 +0000 (15:15 -0400)]
Add ashift validation when adding devices to a pool

Currently, zpool add allows users to add top-level vdevs that have
different ashifts but doing so prevents users from being able to
perform a top-level vdev removal. Often times consumers may not realize
that they have mismatched ashifts until the top-level removal fails.

This feature adds ashift validation to the zpool add command and will
fail the operation if the sector size of the specified vdev does not
match the existing pool. This behavior can be disabled by using the -f
flag. In addition, new flags have been added to provide fine-grained
control to disable specific checks. These flags
are:

--allow-in-use
--allow-ashift-mismatch
--allow-replicaton-mismatch

The force flag will disable all of these checks.

Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Mark Maybee <mmaybee@delphix.com>
Signed-off-by: George Wilson <gwilson@delphix.com>
Closes #15509

7 months agoZTS: fix flakiness in cp_files_002_pos
Robert Evans [Wed, 27 Mar 2024 21:59:16 +0000 (17:59 -0400)]
ZTS: fix flakiness in cp_files_002_pos

Fix RANDOM to not return zero.

Overwriting with `dd ... count=0` does not test anything.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Allan Jude <allan@klarasystems.com>
Signed-off-by: Robert Evans <evansr@google.com>
Closes #16029

7 months agoBRT: Check pool clone stats in more tests
Alexander Motin [Tue, 19 Mar 2024 17:08:05 +0000 (13:08 -0400)]
BRT: Check pool clone stats in more tests

This should allow to catch some leaks, if those happen.

While there fix some cosmetic issues.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes #16007

7 months agoBRT: Fix tests to work on non-empty pools
Alexander Motin [Tue, 19 Mar 2024 16:25:14 +0000 (12:25 -0400)]
BRT: Fix tests to work on non-empty pools

It should not normally happen, but if it does, better to not fail
everything for no good reason, or it may be hard to debug.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes #16007

7 months agoBRT: Fix holes cloning.
Alexander Motin [Mon, 18 Mar 2024 18:19:53 +0000 (14:19 -0400)]
BRT: Fix holes cloning.

 - When reading L0 block pointers handle buffers without ones and
without dirty records as a holes.  Those appear when dnode size
was increased, but the end was never written, so there are no new
indirection levels to store the pointers.  It makes no sense to
return EAGAIN here, since sync won't create new indirection levels
until there will be actual writes.
 - When cloning blocks set destination hole logical birth time
to the current TXG.  Otherwise if we are cloning over existing
data, newly created holes may not be properly replicated later.
Use BP_SET_BIRTH() when possible to not replicate its logic.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes #15994
Closes #16007

7 months agoBRT: Skip getting length in brt_entry_lookup()
Alexander Motin [Tue, 26 Mar 2024 00:13:45 +0000 (20:13 -0400)]
BRT: Skip getting length in brt_entry_lookup()

Unlike DDT, where ZAP values may have different lengths due to
compression, all BRT entries are identical 8-byte counters.  It
does not make sense to first fetch the length only to assert it.
zap_lookup_uint64() is specifically designed to work with counters
of different size and should return error if something odd found.
Calling it straight allows to save some measurable CPU time.

Reviewed-by: Pawel Jakub Dawidek <pawel@dawidek.net>
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 #15950

7 months agoabd_iter_page: don't use compound heads on Linux <4.5
Rob Norris [Wed, 13 Mar 2024 23:57:30 +0000 (10:57 +1100)]
abd_iter_page: don't use compound heads on Linux <4.5

Before 4.5 (specifically, torvalds/linux@ddc58f2), head and tail pages
in a compound page were refcounted separately. This means that using the
head page without taking a reference to it could see it cleaned up later
before we're finished with it. Specifically, bio_add_page() would take a
reference, and drop its reference after the bio completion callback
returns.

If the zio is executed immediately from the completion callback, this is
usually ok, as any data is referenced through the tail page referenced
by the ABD, and so becomes "live" that way. If there's a delay in zio
execution (high load, error injection), then the head page can be freed,
along with any dirty flags or other indicators that the underlying
memory is used. Later, when the zio completes and that memory is
accessed, its either unmapped and an unhandled fault takes down the
entire system, or it is mapped and we end up messing around in someone
else's memory. Both of these are very bad.

The solution on these older kernels is to take a reference to the head
page when we use it, and release it when we're done. There's not really
a sensible way under our current structure to do this; the "best" would
be to keep a list of head page references in the ABD, and release them
when the ABD is freed.

Since this additional overhead is totally unnecessary on 4.5+, where
head and tail pages share refcounts, I've opted to simply not use the
compound head in ABD page iteration there. This is theoretically less
efficient (though cleaning up head page references would add overhead),
but its safe, and we still get the other benefits of not mapping pages
before adding them to a bio and not mis-splitting pages.

There doesn't appear to be an obvious symbol name or config option we
can match on to discover this behaviour in configure (and the mm/page
APIs have changed a lot since then anyway), so I've gone with a simple
version check.

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 #15533
Closes #15588

7 months agovdev_disk: use bio_chain() to submit multiple BIOs
Rob Norris [Wed, 21 Feb 2024 00:07:21 +0000 (11:07 +1100)]
vdev_disk: use bio_chain() to submit multiple BIOs

Simplifies our code a lot, so we don't have to wait for each and
reassemble them.

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 #15533
Closes #15588

7 months agovdev_disk: add module parameter to select BIO submission method
Rob Norris [Tue, 9 Jan 2024 02:28:57 +0000 (13:28 +1100)]
vdev_disk: add module parameter to select BIO submission method

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 #15533
Closes #15588

7 months agovdev_disk: rewrite BIO filling machinery to avoid split pages
Rob Norris [Tue, 18 Jul 2023 01:11:29 +0000 (11:11 +1000)]
vdev_disk: rewrite BIO filling machinery to avoid split pages

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 #15533
Closes #15588

7 months agovdev_disk: make read/write IO function configurable
Rob Norris [Tue, 9 Jan 2024 01:29:19 +0000 (12:29 +1100)]
vdev_disk: make read/write IO function configurable

This is just setting up for the next couple of commits, which will add a
new IO function and a parameter to select it.

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 #15533
Closes #15588

7 months agovdev_disk: reorganise vdev_disk_io_start
Rob Norris [Tue, 9 Jan 2024 01:23:30 +0000 (12:23 +1100)]
vdev_disk: reorganise vdev_disk_io_start

Light reshuffle to make it a bit more linear to read and get rid of a
bunch of args that aren't needed in all cases.

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 #15533
Closes #15588

7 months agovdev_disk: rename existing functions to vdev_classic_*
Rob Norris [Tue, 9 Jan 2024 01:12:56 +0000 (12:12 +1100)]
vdev_disk: rename existing functions to vdev_classic_*

This is just renaming the existing functions we're about to replace and
grouping them together to make the next commits easier to follow.

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 #15533
Closes #15588

7 months agoabd: add page iterator
Rob Norris [Mon, 11 Dec 2023 05:05:54 +0000 (16:05 +1100)]
abd: add page iterator

The regular ABD iterators yield data buffers, so they have to map and
unmap pages into kernel memory. If the caller only wants to count
chunks, or can use page pointers directly, then the map/unmap is just
unnecessary overhead.

This adds adb_iterate_page_func, which yields unmapped struct page
instead.

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 #15533
Closes #15588

7 months agolinux 5.4 compat: page_size()
Rob Norris [Mon, 13 Nov 2023 06:55:29 +0000 (17:55 +1100)]
linux 5.4 compat: page_size()

Before 5.4 we have to do a little math.

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 #15533
Closes #15588

7 months agoBRT: Make BRT block sizes configurable
Alexander Motin [Mon, 25 Mar 2024 22:02:38 +0000 (18:02 -0400)]
BRT: Make BRT block sizes configurable

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

7 months agoProvide macros for setting and getting blkptr birth times
George Wilson [Mon, 25 Mar 2024 22:01:54 +0000 (18:01 -0400)]
Provide macros for setting and getting blkptr birth times

There exist a couple of macros that are used to update the blkptr birth
times but they can often be confusing. For example, the
BP_PHYSICAL_BIRTH() macro will provide either the physical birth time
if it is set or else return back the logical birth time. The
complement to this macro is BP_SET_BIRTH() which will set the logical
birth time and set the physical birth time if they are not the same.
Consumers may get confused when they are trying to get the physical
birth time and use the BP_PHYSICAL_BIRTH() macro only to find out that
the logical birth time is what is actually returned.

This change cleans up these macros and makes them symmetrical. The same
functionally is preserved but the name is changed. Instead of calling
BP_PHYSICAL_BIRTH(), consumer can now call BP_GET_BIRTH(). In
additional to cleaning up this naming conventions, two new sets of
macros are introduced -- BP_[SET|GET]_LOGICAL_BIRTH() and
BP_[SET|GET]_PHYSICAL_BIRTH.  These new macros allow the consumer to
get and set the specific birth time.

As part of the cleanup, the unused GRID macros have been removed and
that portion of the blkptr are currently unused.

Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Signed-off-by: George Wilson <gwilson@delphix.com>
Closes #15962