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#16027Closes#16033
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
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
If a zvol has more than 15 partitions, the minor device number exhausts
the slot count reserved for partitions next to the zvol itself. As a
result, the minor number cannot be used to determine the partition
number for the higher partition, and doing so results in wrong named
symlinks being generated by udev.
Since the partition number is encoded in the block device name anyway,
let's just extract it from there instead.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Signed-off-by: Fabian Grünbichler <f.gruenbichler@proxmox.com>
Closes#15904Closes#15970
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
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>
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@b35fd7422c) 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@3c407dc723), 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.
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
(cherry picked from commit b181b2e604)
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
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
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
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
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
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
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
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
- 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#15994Closes#16007
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
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
Since brt_pending_apply() is running in syncing context, no other
brt_pending_tree accesses are possible for the TXG. We don't need
to acquire brt_pending_lock here.
Reviewed-by: Pawel Jakub Dawidek <pawel@dawidek.net>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Reviewed-by: Rob Norris <robn@despairlabs.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15955
Before this change ZAP called dnode_hold() for almost every block
access, that was clearly visible in profiler under heavy load, such
as BRT. This patch makes it always hold the dnode reference between
zap_lockdir() and zap_unlockdir(). It allows to avoid most of dnode
operations between those. It also adds several new _by_dnode() APIs
to ZAP and uses them in BRT code. Also adds dmu_prefetch_by_dnode()
variant and uses it in the ZAP code.
After this there remains only one call to dmu_buf_dnode_enter(),
which seems to be unneeded. So remove the call and the functions.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15951
If there is a pending entry for this block, then we've already
issued BRT prefetch for it within this TXG, so don't do it again.
BRT vdev lookup and following zap_prefetch_uint64() call can be
pretty expensive and should be avoided when not necessary.
Reviewed-by: Pawel Jakub Dawidek <pawel@dawidek.net>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15941
- Remove custom zap_memset(), use regular memset().
- Use PANIC() instead of opaque cmn_err(CE_PANIC).
- Provide entry parameter to zap_leaf_rehash_entry().
- Reduce branching in zap_leaf_array_create() inner loop.
- Remove signedness where it should not be.
Should be no function changes.
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15976
It does not look important how exactly brt_pending_tree is sorted.
When cloning large file, it is quite likely that all of its blocks
have identical physical birth times, so comparing them first does
not provide useful entropy, while accesses additional cache line.
In most cases combination of vdev and offset provides unique result
and physical birth time comparison is not even needed. Meanwhile,
when traversing the tree inside brt_pending_apply(), it can be
beneficial for dbuf cache and CPU cache hits to group processing
by vdev and so by the per-VDEV BRT ZAPs.
Reviewed-by: Rob Norris <robn@despairlabs.com>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15954
Before this change resume token was updated only on data receive.
Usually it is enough to resume replication without much overlap.
But we've got a report of a curios case, where replication source
was traversed with recursive grep, which through enabled atime
modified every object without modifying any data. It produced
several gigabytes of replication traffic without a single data
write and so without a single resume point.
While the resume token was not designed to resume from an object,
I've found that the send implementation always sends object before
any data. So by requesting resume from offset 0 we are effectively
resuming from the object, followed (or not) by the data at offset
0, just as we need it.
Reviewed-by: Allan Jude <allan@klarasystems.com>
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15927
This changes taskq_thread_should_stop() to limit maximum exit rate
for idle threads to one per 5 seconds. I believe the previous one
was broken, not allowing any thread exits for tasks arriving more
than one at a time and so completing while others are running.
Also while there:
- Remove taskq_thread_spawn() calls on task allocation errors.
- Remove extra taskq_thread_should_stop() call.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Rich Ercolani <rincebrain@gmail.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15873
- Split dmu_prefetch_dnode() from dmu_prefetch() into a separate
function. It is quite inconvenient to read the code where len = 0
means dnode prefetch instead indirect/data prefetch. One function
doing both has no benefits, since the code paths are independent.
- Improve dmu_prefetch() handling of long block ranges. Instead
of limiting L0 data length to prefetch for to dmu_prefetch_max,
make dmu_prefetch_max limit the actual amount of prefetch at the
specified level, and, if there is more, prefetch all the rest at
higher indirection level. It should improve random access times
within the prefetched range of any length, reducing importance of
specific dmu_prefetch_max value.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15076
While picking parts from #14909 I've missed Linux tracing specific
ones, that went unnoticed in default configurations, but breaks the
build in some.
Reviewed-by: Ameer Hamza <ahamza@ixsystems.com>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15730
Track history in context of bursts, not individual log blocks. It
allows to not blow away all the history by single large burst of
many block, and same time allows optimizations covering multiple
blocks in a burst and even predicted following burst. For each
burst account its optimal block size and minimal first block size.
Use that statistics from the last 8 bursts to predict first block
size of the next burst.
Remove predefined set of block sizes. Allocate any size we see fit,
multiple of 4KB, as required by ZIL now. With compression enabled
by default, ZFS already writes pretty random block sizes, so this
should not surprise space allocator any more.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15635
- Generalize vdev_nowritecache handling by traversing through the
VDEV tree and skipping children ZIOs where not supported.
- Remove intermediate zio_null() in case of several VDEV children.
- Remove children handling from zio_ioctl(). There are no other
use cases for this code beside DKIOCFLUSHWRITECACHED, and would there
be, I doubt they would so straightforward apply to all VDEV children.
Comparing to removed previous optimization this should improve cases
of redundant ZILs/SLOGs.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Wilson <george.wilson@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15515
... by checking that previous block is fully written and flushed.
It allows to skip commit delays since we can give up on aggregation
in that case. This removes zil_min_commit_timeout parameter, since
for single-threaded workloads it is not needed at all, while on very
fast devices even some multi-threaded workloads may get detected as
single-threaded and still bypass the wait. To give multi-threaded
workloads more aggregation chances increase zfs_commit_timeout_pct
from 5 to 10%, as they should suffer less from additional latency.
Also single-threaded workloads detection allows in perspective better
prediction of the next block size.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Prakash Surya <prakash.surya@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15381
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
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
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
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.
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
(cherry picked from commit 1bf649cb0a)
After 08fd5ccc3, 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.
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
(cherry picked from commit ba9f587a77)
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>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
(cherry picked from commit 917ff75e95)
1) Make mmap flushes synchronous. Linux may skip flushing dirty pages
already in writeback unless data-integrity sync is requested.
2) Change zfs_putpage to use TXG_WAIT. Otherwise dirty pages may be
skipped due to DMU pushing back on TX assign.
3) Add missing mmap flush when doing block cloning.
4) While here, pass errors from putpage to writepage/writepages.
This change fixes corruption edge cases, but unfortunately adds
synchronous ZIL flushes for dirty mmap pages to llseek and bclone
operations. It may be possible to avoid these sync writes later
but would need more tricky refactoring of the writeback code.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Robert Evans <evansr@google.com>
Closes#15933Closes#16019
We don't want to change to brand-new code in the middle of a stable
series, but we want it available to test for people running into page
splitting issues.
This commits make zfs_vdev_disk_classic=1 the default, and updates the
documentation to better explain what's going on.
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
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#15533Closes#15588
(cherry picked from commit c6be6ce175)
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#15533Closes#15588
(cherry picked from commit 72fd834c47)
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#15533Closes#15588
(cherry picked from commit df2169d141)
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#15533Closes#15588
(cherry picked from commit 06a196020e)
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#15533Closes#15588
(cherry picked from commit c4a13ba483)
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#15533Closes#15588
(cherry picked from commit 867178ae1d)
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#15533Closes#15588
(cherry picked from commit f3b85d706b)
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#15533Closes#15588
(cherry picked from commit 390b448726)
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#15533Closes#15588
(cherry picked from commit df04efe321)
Linux 6.8 removes generic_copy_file_range(), which had been reduced to a
simple wrapper around splice_copy_file_range(). Detect that function
directly and use it if generic_ is not available.
Sponsored-by: https://despairlabs.com/sponsor/
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <robn@despairlabs.com>
Closes#15930Closes#15931
(cherry picked from commit ef08a4d406)
When ZFS overwrites a whole block, it does not bother to read the
old content from disk. It is a good optimization, but if the buffer
fill fails due to page fault or something else, the buffer ends up
corrupted, neither keeping old content, nor getting the new one.
On FreeBSD this is additionally complicated by page faults being
blocked by VFS layer, always returning EFAULT on attempt to write
from mmap()'ed but not yet cached address range. Normally it is
not a big problem, since after original failure VFS will retry the
write after reading the required data. The problem becomes worse
in specific case when somebody tries to write into a file its own
mmap()'ed content from the same location. In that situation the
only copy of the data is getting corrupted on the page fault and
the following retries only fixate the status quo. Block cloning
makes this issue easier to reproduce, since it does not read the
old data, unlike traditional file copy, that may work by chance.
This patch provides the fill status to dmu_buf_fill_done(), that
in case of error can destroy the corrupted buffer as if no write
happened. One more complication in case of block cloning is that
if error is possible during fill, dmu_buf_will_fill() must read
the data via fall-back to dmu_buf_will_dirty(). It is required
to allow in case of error restoring the buffer to a state after
the cloning, not not before it, that would happen if we just call
dbuf_undirty().
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#15665
