Linux has started moving to a model where instead of applying block
queue limits through individual modification functions, a complete
limits structure is built up and applied atomically, either when the
block device or open, or some time afterwards. As of 6.10 this
transition appears only partly completed.
This commit matches that model within OpenZFS in a way that should work
for past and future kernels. We set up a queue limits structure with any
limits that have had their modification functions removed. For newer
kernels that can have limits applied at block device open
(HAVE_BLK_ALLOC_DISK_2ARG), we have a conversion function to turn the
OpenZFS queue limits structure into Linux's queue_limits structure,
which can then be passed in. For older kernels, we provide an
application function that just calls the old functions for each limit in
the structure.
Signed-off-by: Rob Norris <robn@despairlabs.com>
Sponsored-by: https://despairlabs.com/sponsor/
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
zfs_zstd_cache_reap_now is issued every second.
zstd_mempool_reap checks for both pool existence and buffer count, but
that's still 2 func calls which are trivially avoidable.
With clang it even avoids pushing the stack pointer (but still suffers
the mispredict due to a forward jump, not modified in case someone is
using zstd):
<+0>: cmpq $0x0,0x0(%rip) # <zfs_zstd_cache_reap_now+8>
<+8>: je 0x217de4 <zfs_zstd_cache_reap_now+36>
<+10>: push %rbp
<+11>: mov %rsp,%rbp
<+14>: mov 0x0(%rip),%rdi # <zfs_zstd_cache_reap_now+21>
<+21>: call 0x217df0 <zstd_mempool_reap>
<+26>: mov 0x0(%rip),%rdi # <zfs_zstd_cache_reap_now+33>
<+33>: pop %rbp
<+34>: jmp 0x217df0 <zstd_mempool_reap>
<+36>: ret
Preferably the call would not be made to begin with if zstd is not used,
but this retains all the logic confined to zstd code.
Sponsored by: Rubicon Communications, LLC ("Netgate")
Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
Reviewed-by: Allan Jude <allan@klarasystems.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
In the commit of the head_errlog feature we introduced a bug in
dsl_dataset_promote_sync(): we may dereference origin_head and hds, both
dereferencing ddpa after calling promote_sync() on ddpa.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Chunwei Chen <david.chen@nutanix.com>
Reviewed-by: Rob Norris <robn@despairlabs.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#16272Closes#16273
This freeuio() interface was introduced to FreeBSD recently. For now
it simply calls free(), so this change has no effect. However, this
may not always be true, and in CheriBSD this change is required.
Signed-off-by: Mark Johnston <markj@FreeBSD.org>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brooks Davis <brooks.davis@sri.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
You can use the UBSAN_SANITIZE_* Kbuild options to exclude certain
kernel objects from the UBSAN checks. We previously excluded
zap_micro.o with:
UBSAN_SANITIZE_zap_micro.o := n
For some reason that didn't work for the 6.9 kernel, which wants us
to use:
UBSAN_SANITIZE_zfs/zap_micro.o := n
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tony Hutter <hutter2@llnl.gov>
Closes#16278Closes#16330
In several functions, we use a flag variable to track whether
zv_suspend_lock is held. This flag was not getting reset in a
particular case where we need to retry the underlying operation,
resulting in a lock leak. Make sure to update the flag where necessary.
Signed-off-by: Mark Johnston <markj@FreeBSD.org>
Reviewed-by: Allan Jude <allan@klarasystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
The 6.9 kernel behaves differently in how it releases block devices. In
the common case it will async release the device only after the return
to userspace. This is different from the 6.8 and older kernels which
release the block devices synchronously. To get around this, call
add_disk() from a workqueue so that the kernel uses a different
codepath to release our zvols in the way we expect. This stops
zfs_allow_010_pos from hanging.
Fixes: #16089
Signed-off-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Reviewed-by: Rob Norris <rob.norris@klarasystems.com>
Otherwise if zfs is unloaded and reroot is being used it trips over a
stale pointer.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: Rubicon Communications, LLC ("Netgate")
Signed-off-by: Mateusz Guzik <mjguzik@gmail.com>
Closes#16242
Recent UMA changes repurposed the use of UMA_MD_SMALL_ALLOC in a way
that breaks arc_available_memory on -CURRENT. This change
ensures that arc_available_memory uses the new symbol
while maintaining compatibility with older FreeBSD releases.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Bojan Novković <bnovkov@FreeBSD.org>
Closes#16230
We don't build illumos-crypto for FreeBSD.
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#16209
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#16209
Nothing calls it through the KCF interface, so this is all unused.
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#16209
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#16209
sha2_mech_type_t serves double-duty, as the list of MAC providers and
also the algo type for direct callers to SHA2Init. Until we disentangle
that, reorganise it to make the separation more clear. While we're
there, remove the digest mechs we don't use.
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#16209
For whatever reason, we call digest mechanisms directly, not through the
KCF digest provider. So we can remove those entry points entirely.
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#16209
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#16209
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#16209
Still retaining the struture, for now.
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#16209
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#16209
These members have directly references to the global variables
exposed by the kernel. They are not going to be changed by this
kernel module.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Zhenlei Huang <zlei@FreeBSD.org>
Closes#16210
We always call it twice with JUSTLOOKING and then FORREAL.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Pawel Jakub Dawidek <pawel@dawidek.net>
Closes#16225
Optionally turn off disk's enclosure slot if an I/O is hung
triggering the deadman.
It's possible for outstanding I/O to a misbehaving SCSI disk to
neither promptly complete or return an error. This can occur due
to retry and recovery actions taken by the SCSI layer, driver, or
disk. When it occurs the pool will be unresponsive even though
there may be sufficient redundancy configured to proceeded without
this single disk.
When a hung I/O is detected by the kmods it will be posted as a
deadman event. By default an I/O is considered to be hung after
5 minutes. This value can be changed with the zfs_deadman_ziotime_ms
module parameter. If ZED_POWER_OFF_ENCLOSURE_SLOT_ON_DEADMAN is set
the disk's enclosure slot will be powered off causing the outstanding
I/O to fail. The ZED will then handle this like a normal disk failure.
By default ZED_POWER_OFF_ENCLOSURE_SLOT_ON_DEADMAN is not set.
As part of this change `zfs_deadman_events_per_second` is added
to control the ratelimitting of deadman events independantly of
delay events. In practice, a single deadman event is sufficient
and more aren't particularly useful.
Alphabetize the zfs_deadman_* entries in zfs.4.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#16226
- Add old eviction for special and dedup metaslab classes. Those
vdevs may be potentially big and fragmented with large metaslabs,
while their asynchronous write pattern is not really different
from normal class. It seems an omission to not evict old metaslabs
from them.
- If we have metaslab preload enabled, which means we are not too
low on memory, do not evict active metaslabs even if they are not
used for some time. Eviction of active metaslabs means we won't
be able to write anything until we load them, that may take some
time, that is straight opposite to metaslab preload goals. For
small systems the memory saving should be less important after
recent reduction in number of allocators and so open metaslabs.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16214
In case of error dmu_buf_fill_done() returns the buffer back into
DB_UNCACHED state. Since during transition from DB_UNCACHED into
DB_FILL state dbuf_noread() allocates an ARC buffer, we must free
it here, otherwise it will be leaked.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Jorgen Lundman <lundman@lundman.net>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15665Closes#15802Closes#16216
At the end of l2arc_evict() fix an assertion in the case that l2ad_hand
+ distance == l2ad_end.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#16202Closes#16207
C99 6.7.8.17 says that when an undesignated initialiser is used, only
the first element of a union is initialised. If the first element is not
the largest within the union, how the remaining space is initialised is
up to the compiler.
GCC extends the initialiser to the entire union, while Clang treats the
remainder as padding, and so initialises according to whatever
automatic/implicit initialisation rules are currently active.
When Linux is compiled with CONFIG_INIT_STACK_ALL_PATTERN,
-ftrivial-auto-var-init=pattern is added to the kernel CFLAGS. This flag
sets the policy for automatic/implicit initialisation of variables on
the stack.
Taken together, this means that when compiling under
CONFIG_INIT_STACK_ALL_PATTERN on Clang, the "zero" initialiser will only
zero the first element in a union, and the rest will be filled with a
pattern. This is significant for aes_ctx_t, which in
aes_encrypt_atomic() and aes_decrypt_atomic() is initialised to zero,
but then used as a gcm_ctx_t, which is the fifth element in the union,
and thus gets pattern initialisation. Later, it's assumed to be zero,
resulting in a hang.
As confusing and undiscoverable as it is, by the spec, we are at fault
when we initialise a structure containing a union with the zero
initializer. As such, this commit replaces these uses with an explicit
memset(0).
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#16135Closes#16206
If a shrink or truncate had recently freed a portion of the ZAP, the
dbuf could still be sitting on the dbuf cache waiting for eviction. If
it is then allocated for a new leaf before it can be evicted, the
zap_leaf_t is still attached as userdata, tripping the VERIFY.
Instead, just check for the userdata, and if we find it, reuse it.
Sponsored-by: Klara, Inc.
Sponsored-by: iXsystems, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#16157.
Closes#16204
On 32-bit platforms long long is generally 64-bits. Sufficiently modern
versions of gcc (13 in my testing) complains when casting a pointer to
an integer of a different width so cast to uintptr_t first to avoid the
warning.
Fixes: c183d164aa Parallel pool import
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Don Brady <don.brady@klarasystems.com>
Signed-off-by: Brooks Davis <brooks.davis@sri.com>
Closes#16203
Even though block cloning is much faster than regular copying,
it is not instantaneous - the file might be large and the recordsize
small. It would be nice to be able to interrupt it with a signal
(e.g., SIGINFO on FreeBSD to see the progress).
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Pawel Jakub Dawidek <pawel@dawidek.net>
Closes#16208
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#13215Closes#16138
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
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
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
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
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
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
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
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
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
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
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#12366Closes#16159
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
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
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
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
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
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
- 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
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
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
