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
zio_root() has no arguments for ready callback or parent ZIO. Except
one recent case in ZIL code if root ZIOs ever have a parent it is
also a root ZIO. It means we do not need READY pipeline stage for
them, which takes some time to process, but even more time to wait
for the children and be woken by them, and both for no good reason.
The most visible effect of this change is that it avoids one taskq
wakeup per ZIL block written, previously used to run zio_ready()
for lwb_root_zio and skipped now.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15398
As part of some internal gang block testing within Delphix
we hit the assertion removed by this patch. The assertion
was triggered by a ZIO that had two copies and was a gang
block making the following expression equal to 3:
```
MIN(zp->zp_copies + BP_IS_GANG(bp), spa_max_replication(spa))
```
and failing when we expected the above to be equal to
`BP_GET_NDVAS(bp)`.
The assertion is no longer valid since the following commit:
```
commit 14872aaa4f
Author: Matthew Ahrens <matthew.ahrens@delphix.com>
Date: Mon Feb 6 09:37:06 2023 -0800
EIO caused by encryption + recursive gang
```
The above commit changed gang block headers so they can't
have more than 2 copies but the assertion in question from
this PR was never updated.
Reviewed-by: George Wilson <george.wilson@delphix.com>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Closes#15180
The previous patch #14841 appeared to have significant flaw, causing
deadlocks if zl_get_data callback got blocked waiting for TXG sync. I
already handled some of such cases in the original patch, but issue
#14982 shown cases that were impossible to solve in that design.
This patch fixes the problem by postponing log blocks allocation till
the very end, just before the zios issue, leaving nothing blocking after
that point to cause deadlocks. Before that point though any sleeps are
now allowed, not causing sync thread blockage. This require slightly
more complicated lwb state machine to allocate blocks and issue zios
in proper order. But with removal of special early issue workarounds
the new code is much cleaner now, and should even be more efficient.
Since this patch uses null zios between write, I've found that null
zios do not wait for logical children ready status in zio_ready(),
that makes parent write to proceed prematurely, producing incorrect
log blocks. Added ZIO_CHILD_LOGICAL_BIT to zio_wait_for_children()
fixes it.
Reviewed-by: Rob Norris <rob.norris@klarasystems.com>
Reviewed-by: Mark Maybee <mark.maybee@delphix.com>
Reviewed-by: George Wilson <george.wilson@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15122
Fastwrite was introduced many years ago to improve ZIL writes spread
between multiple top-level vdevs by tracking number of allocated but
not written blocks and choosing vdev with smaller count. It suposed
to reduce ZIL knowledge about allocation, but actually made ZIL to
even more actively report allocation code about the allocations,
complicating both ZIL and metaslabs code.
On top of that, it seems ZIO_FLAG_FASTWRITE setting in dmu_sync()
was lost many years ago, that was one of the declared benefits. Plus
introduction of embedded log metaslab class solved another problem
with allocation rotor accounting both normal and log allocations,
since in most cases those are now in different metaslab classes.
After all that, I'd prefer to simplify already too complicated ZIL,
ZIO and metaslab code if the benefit of complexity is not obvious.
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#15107
Since spa_min_alloc may not be a power of 2, unlike ashifts, in the
case of DRAID, we should not select the minimal value among several
vdevs. Rounding to a multiple of it is unlikely to work for other
vdevs. Instead, using the greatest common divisor produces smaller
yet more reasonable results.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Ameer Hamza <ahamza@ixsystems.com>
Closes#15067
- Pack struct zio_prop by 4 bytes from 84 to 80.
- Skip new child ZIO locking while linking to parent. The newly
allocated ZIO is not externally visible yet, so nobody should care.
- Skip io_bp_copy writes when not used (write && non-debug).
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#14985
Those callbacks were introduced many years ago as part of a bigger
patch to smoothen the write throttling within a txg. They allow to
account completion of individual physical writes within a logical
one, improving cases when some of physical writes complete much
sooner than others, gradually opening the write throttle.
Few years after that ZFS got allocation throttling, working on a
level of logical writes and limiting number of writes queued to
vdevs at any point, and so limiting latency distribution between
the physical writes and especially writes of multiple copies.
The addition of scheduling deadline I proposed in #14925 should
further reduce the latency distribution. Grown memory sizes over
the past 10 years should also reduce importance of the smoothing.
While the use of physdone callback may still in theory provide
some smoother throttling, there are cases where we simply can not
afford it. Since dirty data accounting is protected by pool-wide
lock, in case of 6-wide RAIDZ, for example, it requires us to take
it 8 times per logical block write, creating huge lock contention.
My tests of this patch show radical reduction of the lock spinning
time on workloads when smaller blocks are written to RAIDZ pools,
when each of the disks receives 8-16KB chunks, but the total rate
reaching 100K+ blocks per second. Same time attempts to measure
any write time fluctuations didn't show anything noticeable.
While there, remove also io_child_count/io_parent_count counters.
They are used only for couple assertions that can be avoided.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#14948
It was a vdev level read cache, designed to aggregate many small
reads by speculatively issuing bigger reads instead and caching
the result. But since it has almost no idea about what is going
on with exception of ZIO_FLAG_DONT_CACHE flag set by higher layers,
it was found to make more harm than good, for which reason it was
disabled for the past 12 years. These days we have much better
instruments to enlarge the I/Os, such as speculative and prescient
prefetches, I/O scheduler, I/O aggregation etc.
Besides just the dead code removal this removes one extra mutex
lock/unlock per write inside vdev_cache_write(), not otherwise
disabled and trying to do some work.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#14953
We only need to know if ZIO has any parent there. We do not care if
it has more than one, but use of zio_unique_parent() == NULL asserts
that.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#14823
If a block pointer is corrupted (but the block containing it checksums
correctly, e.g. due to a bug that overwrites random memory), we can
often detect it before the block is read, with the `zfs_blkptr_verify()`
function, which is used in `arc_read()`, `zio_free()`, etc.
However, such corruption is not typically recoverable. To recover from
it we would need to detect the memory error before the block pointer is
written to disk.
This PR verifies BP's that are contained in indirect blocks and dnodes
before they are written to disk, in `dbuf_write_ready()`. This way,
we'll get a panic before the on-disk data is corrupted. This will help
us to diagnose what's causing the corruption, as well as being much
easier to recover from.
To minimize performance impact, only checks that can be done without
holding the spa_config_lock are performed.
Additionally, when corruption is detected, the raw words of the block
pointer are logged. (Note that `dprintf_bp()` is a no-op by default,
but if enabled it is not safe to use with invalid block pointers.)
Reviewed-by: Rich Ercolani <rincebrain@gmail.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Paul Zuchowski <pzuchowski@datto.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#14817
Address the following bugs in persistent error log:
1) Check nested clones, eg "fs->snap->clone->snap2->clone2".
2) When deleting files containing error blocks in those clones (from
"clone" the example above), do not break the check chain.
3) When deleting files in the originating fs before syncing the errlog
to disk, do not break the check chain. This happens because at the
time of introducing the error block in the error list, we do not have
its birth txg and the head filesystem. If the original file is
deleted before the error list is synced to the error log (which is
when we actually lookup the birth txg and the head filesystem), then
we do not have access to this info anymore and break the check chain.
The most prominent change is related to achieving (3). We expand the
spa_error_entry_t structure to accommodate the newly introduced
zbookmark_err_phys_t structure (containing the birth txg of the error
block).Due to compatibility reasons we cannot remove the
zbookmark_phys_t structure and we also need to place the new structure
after se_avl, so it is not accounted for in avl_find(). Then we modify
spa_log_error() to also provide the birth txg of the error block. With
these changes in place we simplify the previously introduced function
get_head_and_birth_txg() (now named get_head_ds()).
We chose not to follow the same approach for the head filesystem (thus
completely removing get_head_ds()) to avoid introducing new lock
contentions.
The stack sizes of nested functions (as measured by checkstack.pl in the
linux kernel) are:
check_filesystem [zfs]: 272 (was 912)
check_clones [zfs]: 64
We also introduced two new tests covering the above changes.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#14633
Block Cloning allows to manually clone a file (or a subset of its
blocks) into another (or the same) file by just creating additional
references to the data blocks without copying the data itself.
Those references are kept in the Block Reference Tables (BRTs).
The whole design of block cloning is documented in module/zfs/brt.c.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Christian Schwarz <christian.schwarz@nutanix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Rich Ercolani <rincebrain@gmail.com>
Signed-off-by: Pawel Jakub Dawidek <pawel@dawidek.net>
Closes#13392
Hole detection in the zio compression code allows us to
opportunistically skip compression on holes. We can go a step further
by not doing memory allocations on holes either.
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Richard Yao <richard.yao@klarasystems.com>
Sponsored-by: Wasabi Technology, Inc.
Closes#14500
Clang's static analyzer correctly identified a NULL pointer dereference
in zio_ready() when ZIO_FLAG_NODATA has been set on a zio that is
missing a block pointer. The NULL pointer dereference occurs because we
have logic intended to disable ZIO_FLAG_NODATA when it has been set on a
gang block.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14469
Encrypted blocks can not have 3 DVAs, because they use the space of the
3rd DVA for the IV+salt. zio_write_gang_block() takes this into
account, setting `gbh_copies` to no more than 2 in this case. Gang
members BP's do not have the X (encrypted) bit set (nor do they have the
DMU level and type fields set), because encryption is not handled at
this level. The gang block is reassembled, and then encryption (and
compression) are handled.
To check if this gang block is encrypted, the code in
zio_write_gang_block() checks `pio->io_bp`. This is normally fine,
because the block that's being ganged is typically the encrypted BP.
The problem is that if there is "recursive ganging", where a gang member
is itself a gang block, then when zio_write_gang_block() is called to
create a gang block for a gang member, `pio->io_bp` is the gang member's
BP, which doesn't have the X bit set, so the number of DVA's is not
restricted to 2. It should instead be looking at the the "gang leader",
i.e. the top-level gang block, to determine how many DVA's can be used,
to avoid a "NDVA's inversion" (where a child has more DVA's than its
parent).
gang leader BP: X (encrypted) bit set, 2 DVA's, IV+salt in 3rd DVA's
space:
```
DVA[0]=<1:...:100400> DVA[1]=<0:...:100400> salt=... iv=...
[L0 ZFS plain file] fletcher4 uncompressed encrypted LE
gang unique double size=100000L/100000P birth=... fill=1 cksum=...
```
leader's GBH contains a BP with gang bit set and 3 DVA's:
```
DVA[0]=<1:...:55600> DVA[1]=<0:...:55600>
[L0 unallocated] fletcher4 uncompressed unencrypted LE
contiguous unique double size=55600L/55600P birth=... fill=0 cksum=...
DVA[0]=<1:...:55600> DVA[1]=<0:...:55600>
[L0 unallocated] fletcher4 uncompressed unencrypted LE
contiguous unique double size=55600L/55600P birth=... fill=0 cksum=...
DVA[0]=<1:...:55600> DVA[1]=<0:...:55600> DVA[2]=<1:...:200>
[L0 unallocated] fletcher4 uncompressed unencrypted LE
gang unique double size=55400L/55400P birth=... fill=0 cksum=...
```
On nondebug bits, having the 3rd DVA in the gang block works for the
most part, because it's true that all 3 DVA's are available in the gang
member BP (in the GBH). However, for accounting purposes, gang block
DVA's ASIZE include all the space allocated below them, i.e. the
512-byte gang block header (GBH) as well as the gang members below that.
We see that above where the gang leader BP is 1MB logical (and after
compression: 0x`100000P`), but the ASIZE of each DVA is 2 sectors (1KB)
more than 1MB (0x`100400`).
Since thre are 3 copies of a block below it, we increment the ATIME of
the 3rd DVA of the gang leader by the space used by the 3rd DVA of the
child (1 sector, in this case). But there isn't really a 3rd DVA of the
parent; the salt is stored in place of the 3rd DVA's ASIZE.
So when zio_write_gang_member_ready() increments the parent's BP's
`DVA[2]`'s ASIZE, it's actually incrementing the parent's salt. When we
later try to read the encrypted recursively-ganged block, the salt
doesn't match what we used to write it, so MAC verification fails and we
get an EIO.
```
zio_encrypt(): encrypted 515/2/0/403 salt: 25 25 bb 9d ad d6 cd 89
zio_decrypt(): decrypting 515/2/0/403 salt: 26 25 bb 9d ad d6 cd 89
```
This commit addresses the problem by not increasing the number of copies
of the GBH beyond 2 (even for non-encrypted blocks). This simplifies
the logic while maintaining the ability to traverse all metadata
(including gang blocks) even if one copy is lost. (Note that 3 copies
of the GBH will still be created if requested, e.g. for `copies=3` or
MOS blocks.) Additionally, the code that increments the parent's DVA's
ASIZE is made to check the parent DVA's NDVAS even on nondebug bits. So
if there's a similar bug in the future, it will cause a panic when
trying to write, rather than corrupting the parent BP and causing an
error when reading.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Co-authored-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Caused-by: #14356Closes#14440Closes#14413
Encrypted blocks can have up to 2 DVA's, as the third DVA is reserved
for the salt+IV. However, dmu_write_policy() allows non-encrypted
blocks (e.g. DMU_OT_OBJSET) inside encrypted datasets to request and
allocate 3 DVA's, since they don't need a salt+IV (they are merely
authenicated).
However, if such a block becomes a gang block, the gang code incorrectly
limits the gang block header to 2 DVA's. This leads to a "NDVAs
inversion", where a parent block (the gang block header) has less DVA's
than its children (the gang members), causing an assertion failure in
zio_write_gang_member_ready().
This commit addresses the problem by only restricting the gang block
header to 2 DVA's if the block is actually encrypted (and thus its gang
block members can have at most 2 DVA's).
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#14250Closes#14356
When doing a device removal on a pool with gang blocks, the zio pipeline
can deadlock when trying to free blocks from a device which is being
removed with a stack similar to this:
0xffff8ab9a13a1740 UNINTERRUPTIBLE 4
__schedule+0x2e5
__schedule+0x2e5
schedule+0x33
schedule_preempt_disabled+0xe
__mutex_lock.isra.12+0x2a7
__mutex_lock.isra.12+0x2a7
__mutex_lock_slowpath+0x13
mutex_lock+0x2c
free_from_removing_vdev+0x61
metaslab_free_impl+0xd6
metaslab_free_dva+0x5e
metaslab_free+0x196
zio_free_sync+0xe4
zio_free_gang+0x38
zio_gang_tree_issue+0x42
zio_gang_tree_issue+0xa2
zio_gang_issue+0x6d
zio_execute+0x94
zio_execute+0x94
taskq_thread+0x23b
kthread+0x120
ret_from_fork+0x1f
Since there are gang blocks we have to read the gang members as part of
the free. This can be seen with a zio dependency tree that looks like
this:
sdb> echo 0xffff900c24f8a700 | zio -rc | zio
ADDRESS TYPE STAGE WAITER
0xffff900c24f8a700 NULL CHECKSUM_VERIFY 0xffff900ddfd31740
0xffff900c24f8c920 FREE GANG_ASSEMBLE -
0xffff900d93d435a0 READ DONE
In the illustration above we are processing frees but because of gang
block we have to read the constituents blocks. Once we finish the READ
in the zio pipeline we will execute the parent. In this case the parent
is a FREE but the zio taskq is a READ and we continue to process the
pipeline leading to the stack above. In the stack above, we are blocked
waiting for the svr_lock so as a result a READ interrupt taskq thread
is now consumed. Eventually, all of the READ taskq threads end up
blocked and we're unable to complete any read requests.
In zio_notify_parent there is an optimization to continue to use
the taskq thread to exectue the parent's pipeline. To resolve the
deadlock above, we only allow this optimization if the parent's
zio type matches the child which just completed.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: George Wilson <gwilson@delphix.com>
External-issue: DLPX-80130
Closes#14236
After a device has been removed, any nopwrites for blocks on that
indirect vdev should be ignored and a new block should be allocated. The
original code attempted to handle this but used the wrong block pointer
when checking for indirect vdevs and failed to check all DVAs.
This change corrects both of these issues and modifies the test case
to ensure that it properly tests nopwrites with device removal.
Reviewed-by: Prakash Surya <prakash.surya@delphix.com>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Wilson <gwilson@delphix.com>
Closes#14235
The checksum error counter is incremented after reporting to ZED. This
leads ZED to receiving a checksum error report with 0 checksum errors.
To avoid this, bump the checksum error counter before reporting to ZED.
Sponsored-by: Seagate Technology LLC
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rob Wing <rob.wing@klarasystems.com>
Closes#14190
