FreeBSD: Add zfs_link_create() error handling

Originally Solaris didn't expect errors there, but they may happen
if we fail to add entry into ZAP.  Linux fixed it in #7421, but it
was never fully ported to FreeBSD.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes #13215
Closes #16138

config/zfs-build.m4: add Alpine Linux bash-completion path

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: omni <omni+vagant@hack.org>
Closes #16164

config/zfs-build.m4: sort vendors

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: omni <omni+vagant@hack.org>
Closes #16164

Correct level handling in zstream recompress.

sscanf returns number of items parsed on success and EOF on failure.

Reviewed-by: Adam Moss <c@yotes.com>
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Rob Norris <robn@despairlabs.com>
Signed-off-by: Rich Ercolani <rincebrain@gmail.com>
Closes #16198

dbuf: separate refcount calls for dbuf and dbuf_user

In 92dc4ad83 I updated the dbuf_cache accounting to track the size of
userdata associated with dbufs. This adds the size of the dbuf+userdata
together in a single call to zfs_refcount_add_many(), but sometime
removes them in separate calls to zfs_refcount_remove_many(), if dbuf
and userdata are evicted separately.

What I didn't realise is that when refcount tracking is on,
zfs_refcount_add_many() and zfs_refcount_remove_many() are expected to
be paired, with their second & third args (count & holder) the same on
both sides. Splitting the remove part into two calls means the counts
don't match up, tripping a panic.

This commit fixes that, by always adding and removing the dbuf and
userdata counts separately.

Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reported-by: Mark Johnston <markj@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes #16191

zdb/ztest: send dbgmsg output to stderr

And, make the output fd an arg to zfs_dbgmsg_print(). This is a change
in behaviour, but keeps it consistent with where crash traces go, and
it's easy to argue this is what we want anyway; this is information
about the task, not the actual output of the task.

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 #16181

zfs_dbgmsg_print: make FreeBSD and Linux consistent

FreeBSD was using fprintf(), which might not be signal-safe. Meanwhile,
Linux's locking did not cover the header output. This two quirks are
unrelated, but both have the same response: be like the other one. So
with this commit, both functions are the same except for the names of
their lock and list variables.

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 #16181

backtrace: rework for signal safety

Mostly, try a lot harder to not allocate anything.

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 #16181

libspl: lift backtrace into a separate file

If it's going to be used directly by zdb/ztest, then it sort of doesn't
make sense to carry it with the assert code.

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 #16181

zdb/ztest: use libspl backtrace for crashes

We can show much nicer backtraces these days, lets use them.

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 #16181

zdb: bring crash handling over from ztest

ztest has a very nice ability to show a backtrace when there's an
unexpected crash. zdb is used often enough on corrupted data and can
blow up too, so nice output is useful there too.

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 #16181

spa_taskq_dispatch_ent: simplify arguments

This renames it to spa_taskq_dispatch(), and reduces and simplifies its
arguments based on these observations from its two call sites:

- arg is always the zio, so it can be typed that way, and we don't need
  to provide it twice;
- ent is always &zio->io_tqent, and zio is always provided, so we can
  use it directly;
- the only flag used is TQ_FRONT, which can just be a bool;
- zio != NULL was part of the "use allocator" test, but it never would
  have got that far, because that arg was only set to NULL in the
  reexecute path, which is forced to type CLAIM, so the condition would
  fail at t == WRITE 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 #16151

spa: flatten spa_taskq_dispatch_ent()

It is the only user of spa_taskq_dispatch_select(), so might as well
just carry it directly.

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 #16151

spa: remove spa_taskq_dispatch_sync()

It has no callers anymore.

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 #16151

zfs_ioc_send: use a dedicated taskq thread for send

When stack space is tight, the stream is written to its target on a
separate taskq thread to make sure there's enough stack space to
complete it.

This has always used an IO taskq, but that doesn't really make sense for
it, and moving it onto a regular taskq lets us get rid of
spa_taskq_dispatch_sync(), which is not used anywhere else.

Stream writes may block for a long time depending on what the target is,
and we have no way of discovering this, so we can't risk using the
system taskq, as there may be many tens of sends in progress. Instead,
we create a dedicated taskq thread for each send writer to run on, and
clean it up when it's done.

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 #16151

Better control the thread pool size when mounting datasets

Ever since a10d50f999, ZFS has mounted file systems in parallel when
importing a pool.  It uses a fixed size of 512 for the thread pool.  But
since c183d164aa1, it has also imported pools in parallel.  So the total
number of threads at one time is 513 * npools + 1.  That can easily
exceed the system's limit on the number of threads per process, which
will cause one or more pools to be unable to allocate any worker
threads, forcing them to fallback to slow serial mounting .  To
forestall that, manage the threadpool size in /sbin/zpool, not libzfs.
Use the same size (512), but divided by the number of pools.

This is a backwards-incompatible change to the libzfs abi.

Sponsored by: Axcient
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Wilson <gwilson@delphix.com>
Signed-off-by: Alan Somers <asomers@FreeBSD.org>
Closes #16178

libzfs: Fix mounting datasets under thread limit pressure

During parallel zpool import, /sbin/zpool will create a separate thread
pool for each pool, used to mount that pool's datasets.  If the total
thread count exceed's the system's limit on threads per process, then
tpool_dispatch may fail.  If it does, directly execute the mount
operation instead.

Sponsored by: Axcient
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Wilson <gwilson@delphix.com>
Signed-off-by: Alan Somers <asomers@FreeBSD.org>
Closes #16178
Fixes #16172

tpool_dispatch: fail if it cannot start at least 1 worker.

Sponsored by: Axcient
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Wilson <gwilson@delphix.com>
Signed-off-by: Alan Somers <asomers@FreeBSD.org>
Closes #16178

Simplified the scope of the namespace lock

If we wait until after we check for no spa references to drop the
namespace lock, then we know that spa consumers will need to call
spa_lookup() and end up waiting on the spa_namespace_cv until we
finish.  This narrows the external checks to spa_lookup and we no
longer need to worry about the spa_vdev_enter case.

Sponsored-By: Klara Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Wilson <gwilson@delphix.com>
Signed-off-by: Don Brady <don.brady@klarasystems.com>
Closes #16153

Add support for parallel pool exports

Changed spa_export_common() such that it no longer holds the
spa_namespace_lock for the entire duration and instead sets
spa_export_thread to indicate an import is in progress on the
spa.  This allows for an export to a diffent pool to proceed
in parallel while an export is still processing potentially
long operations like spa_unload_log_sm_flush_all().

Calls like spa_lookup() and spa_vdev_enter() that rely on
the spa_namespace_lock to serialize them against a concurrent
export, now wait for any in-progress export thread to complete
before proceeding.

The 'zpool import -a' sub-command also provides multi-threaded
support, using a thread pool to submit the exports in parallel.

Sponsored-By: Klara Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Wilson <gwilson@delphix.com>
Signed-off-by: Don Brady <don.brady@klarasystems.com>
Closes #16153

Linux: disable lockdep for a couple of locks

When running a debug kernel with lockdep enabled there
are several locks which report false positives.  Set
MUTEX_NOLOCKDEP/RW_NOLOCKDEP to disable these warnings.

Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes #16188

ZAP: Fix leaf references on zap_expand_leaf() errors

Depending on kind of error zap_expand_leaf() may return with or
without valid leaf reference held.  Make sure it returns NULL if
due to error it has no leaf to return.  Make its callers to check
the returned leaf pointer, and release the leaf if it is not NULL.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes #12366
Closes #16159

Replace P2ALIGN with P2ALIGN_TYPED and delete P2ALIGN.

In P2ALIGN, the result would be incorrect when align is unsigned
integer and x is larger than max value of the type of align.
In that case, -(align) would be a positive integer, which means
high bits would be zero and finally stay zero after '&' when
align is converted to a larger integer type.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Youzhong Yang <yyang@mathworks.com>
Signed-off-by: Qiuhao Chen <chenqiuhao1997@gmail.com>
Closes #15940

libspl_assert: always link -lpthread on FreeBSD

The pthread_* functions are in -lpthread on FreeBSD. Some of them are
implicitly linked through libc, but on FreeBSD 13 at least
pthread_getname_np() is not. Just be explicit, since -lpthread is the
documented interface anyway.

Sponsored-by: https://despairlabs.com/sponsor/
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Closes #16168

Unbreak FreeBSD cross-build on MacOS broken in 051460b8b

MacOS used FreeBSD-compatible getprogname() and pthread_getname_np().
But pthread_getthreadid_np() does not exist on MacOS. This implements
libspl_gettid() using pthread_threadid_np() to get the thread id
of the current thread.

Tested with FreeBSD GitHub actions
freebsd-src/.github/workflows/cross-bootstrap-tools.yml

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Rob Norris <rob.norris@klarasystems.com>
Signed-off-by: Martin Matuska <mm@FreeBSD.org>
Closes #16167

Fix ZIL clone records for legacy holes

Previous code overengineered cloned range calculation by using
BP_GET_LSIZE(). The problem is that legacy holes don't have the
logical size, so result will be wrong.  But we also don't need
to look on every block size, since they all must be identical.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes #16165

Fix scn_queue races on very old pools

Code for pools before version 11 uses dmu_objset_find_dp() to scan
for children datasets/clones.  It calls enqueue_clones_cb() and
enqueue_cb() callbacks in parallel from multiple taskq threads.
It ends up bad for scan_ds_queue_insert(), corrupting scn_queue
AVL-tree.  Fix it by introducing a mutex to protect those two
scan_ds_queue_insert() calls.  All other calls are done from the
sync thread and so serialized.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes #16162

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

Add newline to two zpool messages

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

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

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

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

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

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

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

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

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>

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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