Existing allocation throttling had a goal to improve write speed
by allocating more data to vdevs that are able to write it faster.
But in the process it completely broken the original mechanism,
designed to balance vdev space usage. With severe vdev space use
imbalance it is possible that some with higher use start growing
fragmentation sooner than others and after getting full will stop
any writes at all. Also after vdev addition it might take a very
long time for pool to restore the balance, since the new vdev does
not have any real preference, unless the old one is already much
slower due to fragmentation. Also the old throttling was request-
based, which was unpredictable with block sizes varying from 512B
to 16MB, neither it made much sense in case of I/O aggregation,
when its 32-100 requests could be aggregated into few, leaving
device underutilized, submitting fewer and/or shorter requests,
or in opposite try to queue up to 1.6GB of writes per device.
This change presents a completely new throttling algorithm. Unlike
the request-based old one, this one measures allocation queue in
bytes. It makes possible to integrate with the reworked allocation
quota (aliquot) mechanism, which is also byte-based. Unlike the
original code, balancing the vdevs amounts of free space, this one
balances their free/used space fractions. It should result in a
lower and more uniform fragmentation in a long run.
This algorithm still allows to improve write speed by allocating
more data to faster vdevs, but does it in more controllable way.
On top of space-based allocation quota, it also calculates minimum
queue depth that vdev is allowed to maintain, and respectively the
amount of extra allocations it can receive if it appear faster.
That amount is based on vdev's capacity and space usage, but also
applied only when the pool is busy. This way the code can choose
between faster writes when needed and better vdev balance when not,
with the choice gradually reducing together with the free space.
This change also makes allocation queues per-class, allowing them
to throttle independently and in parallel. Allocations that are
bounced between classes due to allocation errors will be able to
properly throttle in the new class. Allocations that should not
be throttled (ZIL, gang, copies) are not, but may still follow
the rotor and allocation quota mechanism of the class without
disrupting it.
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
The redundant_metadata setting in ZFS allows users to trade resilience
for performance and space savings. This applies to all data and metadata
blocks in zfs, with one exception: gang blocks. Gang blocks currently
just take the copies property of the IO being ganged and, if it's 1,
sets it to 2. This means that we always make at least two copies of a
gang header, which is good for resilience. However, if the users care
more about performance than resilience, their gang blocks will be even
more of a penalty than usual.
We add logic to calculate the number of gang headers copies directly,
and store it as a separate IO property. This is stored in the IO
properties and not calculated when we decide to gang because by that
point we may not have easy access to the relevant information about what
kind of block is being stored. We also check the redundant_metadata
property when doing so, and use that to decide whether to store an extra
copy of the gang headers, compared to the underlying blocks.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Signed-off-by: Paul Dagnelie <paul.dagnelie@klarasystems.com>
Co-authored-by: Paul Dagnelie <paul.dagnelie@klarasystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Sponsored-by: https://despairlabs.com/sponsor/
Signed-off-by: Rob Norris <robn@despairlabs.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Implementation of DDT pruning introduced verification of DVAs in
a block pointer during ddt_lookup() to not by mistake free previous
pruned incarnation of the entry. But when writing a new block in
zio_ddt_write() we might have the DVAs only from override pointer,
which may never have "D" flag to be confused with pruned DDT entry,
and we'll abandon those DVAs if we find a matching entry in DDT.
This fixes deduplication for blocks written via dmu_sync() for
purposes of indirect ZIL write records, that I have tested. And
I suspect it might actually allow deduplication for Direct I/O,
even though in an odd way -- first write block directly and then
delete it later during TXG commit if found duplicate, which part
I haven't tested.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#17120
Since embedded blocks introduction 11 years ago, their writing was
blocked if dedup is enabled. After searching through the modern
code I see no reason for this restriction to exist. Same time
embedded blocks are dramatically cheaper. Even regular write of
so small blocks would likely be cheaper than deduplication, even
if the last is successful, not mentioning otherwise.
Reviewed-by: Allan Jude <allan@klarasystems.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#17113
Now instead of crashing when attempting to read the corrupt block
pointer, ZFS will return ECKSUM, in a stack that looks like this:
```
none:set-error
zfs.ko`arc_read+0x1d82
zfs.ko`dbuf_read+0xa8c
zfs.ko`dmu_buf_hold_array_by_dnode+0x292
zfs.ko`dmu_read_uio_dnode+0x47
zfs.ko`zfs_read+0x2d5
zfs.ko`zfs_freebsd_read+0x7b
kernel`VOP_READ_APV+0xd0
kernel`vn_read+0x20e
kernel`vn_io_fault_doio+0x45
kernel`vn_io_fault1+0x15e
kernel`vn_io_fault+0x150
kernel`dofileread+0x80
kernel`sys_read+0xb7
kernel`amd64_syscall+0x424
kernel`0xffffffff810633cb
```
This patch should hopefully also prevent such corrupt block pointers
from being written to disk in the first place.
And in zdb, don't crash when printing a block pointer with no valid
DVAs. If a block pointer isn't embedded yet doesn't have any valid
DVAs, that's a data corruption bug. zdb should be able to handle the
situation gracefully.
Finally, remove an extra check for gang blocks in SNPRINTF_BLKPTR. This
check, which compares the asizes of two different DVAs within the same
BP, was added by illumos-gate commit b24ab67[^1], and I can't understand
why. It doesn't appear to do anything useful, so remove it.
[^1]: b24ab67627
Fixes #17077
Sponsored by: ConnectWise
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed by: Alek Pinchuk <pinchuk.alek@gmail.com>
Signed-off-by: Alan Somers <asomers@gmail.com>
Closes#17078
In l2arc_evict(), the config lock may be acquired in reverse order
(e.g., first the config lock (writer), then a hash lock) unlike in
arc_read() during scenarios like L2ARC device removal. To avoid
deadlocks, if the attempt to acquire the config lock (reader) fails
in arc_read(), release the hash lock, wait for the config lock, and
retry from the beginning.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Ameer Hamza <ahamza@ixsystems.com>
Closes#17071
The purpose of no-op is to simulate a failure between a device cache and
its permanent store. We still want it to go through the queue and
respond in the same way to everything else.
So, inject "success" as the very last thing, and then move on to
VDEV_IO_DONE to be dequeued and so any followup work can occur.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#17029
Gang blocks have a significant impact on the long and short term
performance of a zpool, but there is not a lot of observability into
whether they're being used. This change adds gang-specific kstats to
ZFS, to better allow users to see whether ganging is happening.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Paul Dagnelie <paul.dagnelie@klarasystems.com>
Closes#17003
As zios are reexecuted after resume from suspension, their ready and
wait states need to be propagated to wait counts on all their parents.
It's possible for those parents to have active children passing through
READY or DONE, which then end up in zio_notify_parent(), take their
parent's lock, and decrement the wait count. Without also taking a lock
here, it's possible for an increment race to occur, which leads to
either there being no references left (tripping the assert in
zio_notify_parent()), or a parent waiting forever for a nonexistent
child to complete.
To protect against this, we simply take the appropriate zio locks in
zio_reexecute() before updating the wait counts.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Allan Jude <allan@klarasystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#17016
It seems there's no good reason for vdev_disk & vdev_geom to explicitly
detect no support for flush and set vdev_nowritecache. Instead, just
signal it by setting the error to ENOTSUP, and let zio_vdev_io_assess()
take care of it in one place.
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#16855
The first time a device returns ENOTSUP in repsonse to a flush request,
we set vdev_nowritecache so we don't issue flushes in the future and
instead just pretend the succeeded. However, we still return an error
for the initial flush, even though we just decided such errors are
meaningless!
So, when setting vdev_nowritecache in response to a flush error, also
reset the error code to assume success.
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#16855
zio_delay_interrupt(), apparently used for fault injection, is executed
in the I/O pipeline. It can cause the calling thread to go to sleep,
which is not allowed on FreeBSD. This happens only for small delays,
though, and there's no apparent reason to avoid deferring to a taskqueue
in that case, as it already does otherwise.
Simply go to sleep unconditionally. This fixes an occasional panic I
see when running the ZTS on FreeBSD. Also remove an unhelpful comment
referencing the non-existent timeout_generic().
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Mark Johnston <markj@FreeBSD.org>
Closes#16785
Now that we can handle these different alignments, we don't this
workaround.
This reverts commit aefc2da8a5.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#16687
in some cases not linearizing buffers with disk sector crossing a
page boundary. It is fine for hardware, but somehow required by LUKS.
It is not typical for ZFS to produce such buffers, but it may happen
if 6KB block is compressed to 4KB, while still having 2KB alignment.
Banning the 6KB buffers helps vdevs with ashifh=12.
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
This fixes an oversight in the Direct I/O PR. There is nothing that
stops a process from manipulating the contents of a buffer for a
Direct I/O read while the I/O is in flight. This can lead checksum
verify failures. However, the disk contents are still correct, and this
would lead to false reporting of checksum validation failures.
To remedy this, all Direct I/O reads that have a checksum verification
failure are treated as suspicious. In the event a checksum validation
failure occurs for a Direct I/O read, then the I/O request will be
reissued though the ARC. This allows for actual validation to happen and
removes any possibility of the buffer being manipulated after the I/O
has been issued.
Just as with Direct I/O write checksum validation failures, Direct I/O
read checksum validation failures are reported though zpool status -d in
the DIO column. Also the zevent has been updated to have both:
1. dio_verify_wr -> Checksum verification failure for writes
2. dio_verify_rd -> Checksum verification failure for reads.
This allows for determining what I/O operation was the culprit for the
checksum verification failure. All DIO errors are reported only on the
top-level VDEV.
Even though FreeBSD can write protect pages (stable pages) it still has
the same issue as Linux with Direct I/O reads.
This commit updates the following:
1. Propogates checksum failures for reads all the way up to the
top-level VDEV.
2. Reports errors through zpool status -d as DIO.
3. Has two zevents for checksum verify errors with Direct I/O. One for
read and one for write.
4. Updates FreeBSD ABD code to also check for ABD_FLAG_FROM_PAGES and
handle ABD buffer contents validation the same as Linux.
5. Updated manipulate_user_buffer.c to also manipulate a buffer while a
Direct I/O read is taking place.
6. Adds a new ZTS test case dio_read_verify that stress tests the new
code.
7. Updated man pages.
8. Added an IMPLY statement to zio_checksum_verify() to make sure that
Direct I/O reads are not issued as speculative.
9. Removed self healing through mirror, raidz, and dRAID VDEVs for
Direct I/O reads.
This issue was first observed when installing a Windows 11 VM on a ZFS
dataset with the dataset property direct set to always. The zpool
devices would report checksum failures, but running a subsequent zpool
scrub would not repair any data and report no errors.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Brian Atkinson <batkinson@lanl.gov>
Closes#16598
ZLE compressor needs additional bytes to process
d_len argument efficiently.
Don't use BPE_PAYLOAD_SIZE as d_len with it
before we rework zle compressor somehow.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: George Melikov <mail@gmelikov.ru>
Closes#9416
Now default compression is lz4, which can stop
compression process by itself on incompressible data.
If there are additional size checks -
we will only make our compressratio worse.
New usable compression thresholds are:
- less than BPE_PAYLOAD_SIZE (embedded_data feature);
- at least one saved sector.
Old 12.5% threshold is left to minimize affect
on existing user expectations of CPU utilization.
If data wasn't compressed - it will be saved as
ZIO_COMPRESS_OFF, so if we really need to recompress
data without ashift info and check anything -
we can just compress it with zero threshold.
So, we don't need a new feature flag here!
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: George Melikov <mail@gmelikov.ru>
Closes#9416
So that we can get actual benefit from last commit.
Reviewed by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Signed-off-by: Shengqi Chen <harry-chen@outlook.com>
Closes#16131Closes#16483
Adding O_DIRECT support to ZFS to bypass the ARC for writes/reads.
O_DIRECT support in ZFS will always ensure there is coherency between
buffered and O_DIRECT IO requests. This ensures that all IO requests,
whether buffered or direct, will see the same file contents at all
times. Just as in other FS's , O_DIRECT does not imply O_SYNC. While
data is written directly to VDEV disks, metadata will not be synced
until the associated TXG is synced.
For both O_DIRECT read and write request the offset and request sizes,
at a minimum, must be PAGE_SIZE aligned. In the event they are not,
then EINVAL is returned unless the direct property is set to always (see
below).
For O_DIRECT writes:
The request also must be block aligned (recordsize) or the write
request will take the normal (buffered) write path. In the event that
request is block aligned and a cached copy of the buffer in the ARC,
then it will be discarded from the ARC forcing all further reads to
retrieve the data from disk.
For O_DIRECT reads:
The only alignment restrictions are PAGE_SIZE alignment. In the event
that the requested data is in buffered (in the ARC) it will just be
copied from the ARC into the user buffer.
For both O_DIRECT writes and reads the O_DIRECT flag will be ignored in
the event that file contents are mmap'ed. In this case, all requests
that are at least PAGE_SIZE aligned will just fall back to the buffered
paths. If the request however is not PAGE_SIZE aligned, EINVAL will
be returned as always regardless if the file's contents are mmap'ed.
Since O_DIRECT writes go through the normal ZIO pipeline, the
following operations are supported just as with normal buffered writes:
Checksum
Compression
Encryption
Erasure Coding
There is one caveat for the data integrity of O_DIRECT writes that is
distinct for each of the OS's supported by ZFS.
FreeBSD - FreeBSD is able to place user pages under write protection so
any data in the user buffers and written directly down to the
VDEV disks is guaranteed to not change. There is no concern
with data integrity and O_DIRECT writes.
Linux - Linux is not able to place anonymous user pages under write
protection. Because of this, if the user decides to manipulate
the page contents while the write operation is occurring, data
integrity can not be guaranteed. However, there is a module
parameter `zfs_vdev_direct_write_verify` that controls the
if a O_DIRECT writes that can occur to a top-level VDEV before
a checksum verify is run before the contents of the I/O buffer
are committed to disk. In the event of a checksum verification
failure the write will return EIO. The number of O_DIRECT write
checksum verification errors can be observed by doing
`zpool status -d`, which will list all verification errors that
have occurred on a top-level VDEV. Along with `zpool status`, a
ZED event will be issues as `dio_verify` when a checksum
verification error occurs.
ZVOLs and dedup is not currently supported with Direct I/O.
A new dataset property `direct` has been added with the following 3
allowable values:
disabled - Accepts O_DIRECT flag, but silently ignores it and treats
the request as a buffered IO request.
standard - Follows the alignment restrictions outlined above for
write/read IO requests when the O_DIRECT flag is used.
always - Treats every write/read IO request as though it passed
O_DIRECT and will do O_DIRECT if the alignment restrictions
are met otherwise will redirect through the ARC. This
property will not allow a request to fail.
There is also a module parameter zfs_dio_enabled that can be used to
force all reads and writes through the ARC. By setting this module
parameter to 0, it mimics as if the direct dataset property is set to
disabled.
Reviewed-by: Brian Behlendorf <behlendorf@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Atkinson <batkinson@lanl.gov>
Co-authored-by: Mark Maybee <mark.maybee@delphix.com>
Co-authored-by: Matt Macy <mmacy@FreeBSD.org>
Co-authored-by: Brian Behlendorf <behlendorf@llnl.gov>
Closes#10018
When reviewing logs after a failure, its useful to see where
unsuspend/resume was requested.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Requires the new 'flat' physical data which has the start
time for a class entry.
The amount to prune can be based on a target percentage of
the unique entries or based on the age (i.e., every entry
older than N days).
Sponsored-by: Klara, Inc.
Sponsored-by: iXsystems, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Don Brady <don.brady@klarasystems.com>
Closes#16277
This commit changes the frontend zio_compress_data and
zio_decompress_data APIs to take ABD points instead of buffer pointers.
All callers are updated to match. Any that already have an appropriate
ABD nearby now use it directly, while at the rest we create an one.
Internally, the ABDs are passed through to the provider directly.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Makes it harder to use memory debuggers like valgrind directly, because
they can't see canary overruns.
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#16253
Traditional dedup keeps a separate ddt_phys_t "type" for each possible
count of DVAs (that is, copies=) parameter. Each of these are tracked
independently of each other, and have their own set of DVAs. This leads
to an (admittedly rare) situation where you can create as many as six
copies of the data, by changing the copies= parameter between copying.
This is both a waste of storage on disk, but also a waste of space in
the stored DDT entries, since there never needs to be more than three
DVAs to handle all possible values of copies=.
This commit adds a new FDT feature, DDT_FLAG_FLAT. When active, only the
first ddt_phys_t is used. Each time a block is written with the dedup
bit set, this single phys is checked to see if it has enough DVAs to
fulfill the request. If it does, the block is filled with the saved DVAs
as normal. If not, an adjusted write is issued to create as many extra
copies as are needed to fulfill the request, which are then saved into
the entry too.
Because a single phys is no longer an all-or-nothing, but can be
transitioning from fewer to more DVAs, the write path now has to keep a
copy of the previous "known good" DVA set so we can revert to it in case
an error occurs. zio_ddt_write() has been restructured and heavily
commented to make it much easier to see what's happening.
Backwards compatibility is maintained simply by allocating four
ddt_phys_t when the DDT_FLAG_FLAT flag is not set, and updating the phys
selection macros to check the flag. In the old arrangement, each number
of copies gets a whole phys, so it will always have either zero or all
necessary DVAs filled, with no in-between, so the old behaviour
naturally falls out of the new code.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Co-authored-by: Don Brady <don.brady@klarasystems.com>
Sponsored-by: Klara, Inc.
Sponsored-by: iXsystems, Inc.
Closes#15893
This slims down the in-memory entry to as small as it can be. The
IO-related parts are made into a separate entry, since they're
relatively rarely needed.
The variable allocation for dde_phys is to support the upcoming flat
format.
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: iXsystems, Inc.
Closes#15893
The "flat phys" feature will use only a single phys slot for all
entries, which means the old "single", "double" etc naming now makes no
sense, and more importantly, means that choosing the right slot for a
given block pointer will depend on how many slots are in use for a given
DDT.
This removes the old names, and adds accessor macros to decouple
specific phys array indexes from any particular meaning.
(These macros look strange in isolation, mainly in the way they take the
ddt_t* as an arg but don't use it. This is mostly a separate commit to
introduce the concept to the reader before the "flat phys" commit
extends 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: iXsystems, Inc.
Closes#15893
The upcoming dedup features break the long held assumption that all
blocks on disk with a 'D' dedup bit will always be present in the DDT,
or will have the same set of DVA allocations on disk as in the DDT.
If the DDT is no longer a complete picture of all the dedup blocks that
will be and should be on disk, then it does us no good to walk and prime
it up front, since it won't necessarily match up with every block we'll
see anyway.
Instead, we rework things here to be more like the BRT checks. When we
see a dedup'd block, we look it up in the DDT, consume a refcount, and
for the second-or-later instances, count them as duplicates.
The DDT and BRT are moved ahead of the space accounting. This will
become important for the "flat" feature, which may need to count a
modified version of the block.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Co-authored-by: Allan Jude <allan@klarasystems.com>
Co-authored-by: Don Brady <don.brady@klarasystems.com>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Sponsored-by: Klara, Inc.
Sponsored-by: iXsystems, Inc.
Closes#15892
It's now the caller's responsibility do special handling for holes if
that's something it wants.
This also makes zio_compress_data() and zio_decompress_data() properly
the inverse of each other.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Reviewed-by: Jason Lee <jasonlee@lanl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#16326
- Skip config lock enter/exit for embedded blocks. They have no
DVAs, so there is nothing to check under the lock.
- Skip CHECKSUM check and properly check PSIZE for embedded blocks.
- Add static branch predictions for unlikely conditions.
- Do not verify DVAs for blocks already in ARC. ARC hit already
"verified" the first (often the only) DVA, and it does not worth to
enter/exit config lock for nothing.
Some profiles show me up to 3% of CPU saving from this change.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Allan Jude <allan@klarasystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#16387
Linux provides SLAB_RECLAIM_ACCOUNT and __GFP_RECLAIMABLE flags to
mark memory allocations that can be freed via shinker calls. It
should allow kernel to tune and group such allocations for lower
memory fragmentation and better reclamation under pressure.
This patch marks as reclaimable most of ARC memory, directly
evictable via ZFS shrinker, plus also dnode/znode/sa memory,
indirectly evictable via kernel's superblock shrinker.
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Allan Jude <allan@klarasystems.com>
spa_preferred_class() selects a storage class based on (among other
things) the DMU object type. This only works for old-style object types
that match only one specific kind of thing. For DMU_OTN_ types we need
another way to signal the storage class.
This commit allows the object type to be overridden in the IO policy for
the purposes of choosing a storage class. It then adds the ability to
set the storage type on a dnode hold, such that all writes generated
under that hold will get it.
This method has two shortcomings:
- it would be better if we could "name" a set of storage class
preferences rather than it being implied by the object type.
- it would be better if this info were stored in the dnode on disk.
In the absence of those things, this seems like the smallest possible
change.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Sponsored-by: Klara, Inc.
Sponsored-by: iXsystems, Inc.
Closes#15894
Rather than picking out specific values out of the properties, just pass
the entire zio in, to make it easier in the future to use more of that
info to decide on the storage class.
I would have rathered just pass io_prop in, but having spa.h include
zio.h gets a bit tricky.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Sponsored-by: Klara, Inc.
Sponsored-by: iXsystems, Inc.
Closes#15894
This adds two new pool properties:
- dedup_table_size, the total size of all DDTs on the pool; and
- dedup_table_quota, the maximum possible size of all DDTs in the pool
When set, quota will be enforced by checking when a new entry is about
to be created. If the pool is over its dedup quota, the entry won't be
created, and the corresponding write will be converted to a regular
non-dedup write. Note that existing entries can be updated (ie their
refcounts changed), as that reuses the space rather than requiring more.
dedup_table_quota can be set to 'auto', which will set it based on the
size of the devices backing the "dedup" allocation device. This makes it
possible to limit the DDTs to the size of a dedup vdev only, such that
when the device fills, no new blocks are deduplicated.
Sponsored-by: iXsystems, Inc.
Sponsored-By: Klara Inc.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Signed-off-by: Don Brady <don.brady@klarasystems.com>
Co-authored-by: Don Brady <don.brady@klarasystems.com>
Co-authored-by: Rob Wing <rob.wing@klarasystems.com>
Co-authored-by: Sean Eric Fagan <sean.fagan@klarasystems.com>
Closes#15889
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
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
- 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
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
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
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
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
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
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
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
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
There exist a couple of macros that are used to update the blkptr birth
times but they can often be confusing. For example, the
BP_PHYSICAL_BIRTH() macro will provide either the physical birth time
if it is set or else return back the logical birth time. The
complement to this macro is BP_SET_BIRTH() which will set the logical
birth time and set the physical birth time if they are not the same.
Consumers may get confused when they are trying to get the physical
birth time and use the BP_PHYSICAL_BIRTH() macro only to find out that
the logical birth time is what is actually returned.
This change cleans up these macros and makes them symmetrical. The same
functionally is preserved but the name is changed. Instead of calling
BP_PHYSICAL_BIRTH(), consumer can now call BP_GET_BIRTH(). In
additional to cleaning up this naming conventions, two new sets of
macros are introduced -- BP_[SET|GET]_LOGICAL_BIRTH() and
BP_[SET|GET]_PHYSICAL_BIRTH. These new macros allow the consumer to
get and set the specific birth time.
As part of the cleanup, the unused GRID macros have been removed and
that portion of the blkptr are currently unused.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Signed-off-by: George Wilson <gwilson@delphix.com>
Closes#15962
Since we use a limited set of kmem caches, quite often we have unused
memory after the end of the buffer. Put there up to a 512-byte canary
when built with debug to detect buffer overflows at the free time.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15553
- 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
As part of transaction group commit, dsl_pool_sync() sequentially calls
dsl_dataset_sync() for each dirty dataset, which subsequently calls
dmu_objset_sync(). dmu_objset_sync() in turn uses up to 75% of CPU
cores to run sync_dnodes_task() in taskq threads to sync the dirty
dnodes (files).
There are two problems:
1. Each ZVOL in a pool is a separate dataset/objset having a single
dnode. This means the objsets are synchronized serially, which
leads to a bottleneck of ~330K blocks written per second per pool.
2. In the case of multiple dirty dnodes/files on a dataset/objset on a
big system they will be sync'd in parallel taskq threads. However,
it is inefficient to to use 75% of CPU cores of a big system to do
that, because of (a) bottlenecks on a single write issue taskq, and
(b) allocation throttling. In addition, if not for the allocation
throttling sorting write requests by bookmarks (logical address),
writes for different files may reach space allocators interleaved,
leading to unwanted fragmentation.
The solution to both problems is to always sync no more and (if
possible) no fewer dnodes at the same time than there are allocators
the pool.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Edmund Nadolski <edmund.nadolski@ixsystems.com>
Closes#15197
We should not always use PAGESIZE alignment for caches bigger than
it and SPA_MINBLOCKSIZE otherwise. Doing that caches for 5, 6, 7,
10 and 14KB rounded up to 8, 12 and 16KB respectively make no sense.
Instead specify as alignment the biggest power-of-2 divisor. This
way 2KB and 6KB caches are both aligned to 2KB, while 4KB and 8KB
are aligned to 4KB.
Reduce number of caches to half-power of 2 instead of quarter-power
of 2. This removes caches difficult for underlying allocators to
fit into page-granular slabs, such as: 2.5, 3.5, 5, 7, 10KB, etc.
Since these caches are mostly used for transient allocations like
ZIOs and small DBUF cache it does not worth being too aggressive.
Due to the above alignment issue some of those caches were not
working properly any way. 6KB cache now finally has a chance to
work right, placing 2 buffers into 3 pages, that makes sense.
Remove explicit alignment in Linux user-space case. I don't think
it should be needed any more with the above fixes.
As result on FreeBSD instead of such numbers of pages per slab:
vm.uma.zio_buf_comb_16384.keg.ppera: 4
vm.uma.zio_buf_comb_14336.keg.ppera: 4
vm.uma.zio_buf_comb_12288.keg.ppera: 3
vm.uma.zio_buf_comb_10240.keg.ppera: 3
vm.uma.zio_buf_comb_8192.keg.ppera: 2
vm.uma.zio_buf_comb_7168.keg.ppera: 2
vm.uma.zio_buf_comb_6144.keg.ppera: 2 <= Broken
vm.uma.zio_buf_comb_5120.keg.ppera: 2
vm.uma.zio_buf_comb_4096.keg.ppera: 1
vm.uma.zio_buf_comb_3584.keg.ppera: 7 <= Hard to free
vm.uma.zio_buf_comb_3072.keg.ppera: 3
vm.uma.zio_buf_comb_2560.keg.ppera: 2
vm.uma.zio_buf_comb_2048.keg.ppera: 1
vm.uma.zio_buf_comb_1536.keg.ppera: 2
vm.uma.zio_buf_comb_1024.keg.ppera: 1
vm.uma.zio_buf_comb_512.keg.ppera: 1
I am now getting such:
vm.uma.zio_buf_comb_16384.keg.ppera: 4
vm.uma.zio_buf_comb_12288.keg.ppera: 3
vm.uma.zio_buf_comb_8192.keg.ppera: 2
vm.uma.zio_buf_comb_6144.keg.ppera: 3 <= Fixed, 2 in 3 pages
vm.uma.zio_buf_comb_4096.keg.ppera: 1
vm.uma.zio_buf_comb_3072.keg.ppera: 3
vm.uma.zio_buf_comb_2048.keg.ppera: 1
vm.uma.zio_buf_comb_1536.keg.ppera: 2
vm.uma.zio_buf_comb_1024.keg.ppera: 1
vm.uma.zio_buf_comb_512.keg.ppera: 1
Reviewed-by: Allan Jude <allan@klarasystems.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15452
