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
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
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
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
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
While 763ca47 closes the situation of block cloning creating
unencrypted records in encrypted datasets, existing data still causes
panic on read. Setting zfs_recover bypasses this but at the cost of
potentially ignoring more serious issues.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Chris Peredun <chris.peredun@ixsystems.com>
Closes#15677
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
Block cloning normally creates dirty record without dr_data. But if
the block is read after cloning, it is moved into DB_CACHED state and
receives the data buffer. If after that we call dbuf_unoverride()
to convert the dirty record into normal write, we should give it the
data buffer from dbuf and release one.
Reviewed-by: Kay Pedersen <mail@mkwg.de>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15654Closes#15656
In some cases dbuf_assign_arcbuf() may be called on a block that
was recently cloned. If it happened in current TXG we must undo
the block cloning first, since the only one dirty record per TXG
can't and shouldn't mean both cloning and overwrite same time.
Reviewed-by: Kay Pedersen <mail@mkwg.de>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15653
Previously, dmu_buf_will_clone() would roll back any dirty record, but
would not clean out the modified data nor reset the state before
releasing the lock. That leaves the last-written data in db_data, but
the dbuf in the wrong state.
This is eventually corrected when the dbuf state is made NOFILL, and
dbuf_noread() called (which clears out the old data), but at this point
its too late, because the lock was already dropped with that invalid
state.
Any caller acquiring the lock before the call into
dmu_buf_will_not_fill() can find what appears to be a clean, readable
buffer, and would take the wrong state from it: it should be getting the
data from the cloned block, not from earlier (unwritten) dirty data.
Even after the state was switched to NOFILL, the old data was still not
cleaned out until dbuf_noread(), which is another gap for a caller to
take the lock and read the wrong data.
This commit fixes all this by properly cleaning up the previous state
and then setting the new state before dropping the lock. The
DBUF_VERIFY() calls confirm that the dbuf is in a valid state when the
lock is down.
Sponsored-by: Klara, Inc.
Sponsored-By: OpenDrives Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Pawel Jakub Dawidek <pawel@dawidek.net>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#15566Closes#15526
Entries in the dbuf cache contribute only the size of the dbuf data to
the cache size. Attached "user" data is not counted. This can lead to
the data currently "owned" by the cache consuming more memory accounting
appears to show. In some cases (eg a metadnode data block with all child
dnode_t slots allocated), the actual size can be as much as 3x as what
the cache believes it to be.
This is arguably correct behaviour, as the cache is only tracking the
size of the dbuf data, not even the overhead of the dbuf_t. On the other
hand, in the above case of dnodes, evicting cached metadnode dbufs is
the only current way to reclaim the dnode objects, and can lead to the
situation where the dbuf cache appears to be comfortably within its
target memory window and yet is holding enormous amounts of slab memory
that cannot be reclaimed.
This commit adds a facility for a dbuf user to artificially inflate the
apparent size of the dbuf for caching purposes. This at least allows for
cache tuning to be adjusted to match something closer to the real memory
overhead.
metadnode dbufs carry a >1KiB allocation per dnode in their user data.
This informs the dbuf cache machinery of that fact, allowing it to make
better decisions when evicting dbufs.
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#15511
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
Building module/zfs/dbuf.c for 32-bit targets can result in a warning:
In file included from
/usr/src/sys/contrib/openzfs/include/sys/zfs_context.h:97,
from /usr/src/sys/contrib/openzfs/module/zfs/dbuf.c:32:
/usr/src/sys/contrib/openzfs/module/zfs/dbuf.c: In function
'dmu_buf_will_clone':
/usr/src/sys/contrib/openzfs/lib/libspl/include/assert.h:116:33: error:
cast from pointer to integer of different size
[-Werror=pointer-to-int-cast]
116 | const uint64_t __left = (uint64_t)(LEFT);
\
| ^
/usr/src/sys/contrib/openzfs/lib/libspl/include/assert.h:148:25: note:
in expansion of macro 'VERIFY0'
148 | #define ASSERT0 VERIFY0
| ^~~~~~~
/usr/src/sys/contrib/openzfs/module/zfs/dbuf.c:2704:9: note: in
expansion of macro 'ASSERT0'
2704 | ASSERT0(dbuf_find_dirty_eq(db, tx->tx_txg));
| ^~~~~~~
This is because dbuf_find_dirty_eq() returns a pointer, which if
pointers are 32-bit results in a warning about the cast to uint64_t.
Instead, use the ASSERT3P() macro, with == and NULL as second and third
arguments, which should work regardless of the target's bitness.
Reviewed-by: Kay Pedersen <mail@mkwg.de>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Dimitry Andric <dimitry@andric.com>
Closes#15224
Block cloning introduced a new state transition from DB_NOFILL to
DB_READ. This occurs when a block is cloned and then read on the
current txg.
In this case, the clone will move the dbuf to DB_NOFILL, and then the
read will be issued for the overidden block pointer. If that read is
still outstanding when it comes time to write, the dbuf will be in
DB_READ, which is not handled by the checks in dbuf_sync_leaf, thus
tripping the assertions.
This updates those checks to allow DB_READ as a valid state iff the
dirty record is for a BRT write and there is a override block pointer.
This is a safe situation because the block already exists, so there's
nothing that could change from underneath the read.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Kay Pedersen <mail@mkwg.de>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Original-patch-by: Kay Pedersen <mail@mkwg.de>
Sponsored-By: OpenDrives Inc.
Sponsored-By: Klara Inc.
Closes#15050
dbuf_undirty() will (correctly) only removed dirty records for the given
(open) txg. If there is a dirty record for an earlier closed txg that
has not been synced out yet, then db_dirty_records will still have
entries on it, tripping the assertion.
Instead, change the assertion to only consider the current txg. To some
extent this is redundant, as its really just saying "did dbuf_undirty()
work?", but it it doesn't hurt and accurately expresses our
expectations.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Kay Pedersen <mail@mkwg.de>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Original-patch-by: Kay Pedersen <mail@mkwg.de>
Sponsored-By: OpenDrives Inc.
Sponsored-By: Klara Inc.
Closes#15050
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
Commit 555ef90 did some general code refactoring for
dmu_buf_will_not_fill() and dmu_buf_will_fill(). However, the db_mtx was
not held when update db->db_state in those code block. The rest of the
dbuf code always holds the db_mtx when updating db_state. This is
important because cv_wait() db_changed is used to check for db_state
changes.
Updating dmu_buf_will_not_fill() and dmu_buf_will_fill() to hold the
db_mtx when updating db_state.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Brian Atkinson <batkinson@lanl.gov>
Closes#14875
Reimplement some of the block cloning vs dbuf logic, mostly to fix
situation where we clone a block and in the same transaction group
we want to partially overwrite the clone.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Pawel Jakub Dawidek <pawel@dawidek.net>
Closes#14825
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
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Pawel Jakub Dawidek <pawel@dawidek.net>
Closes#14806
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
Undirty the dbuf and destroy its buffer when cloning into it.
Coverity ID: CID-1535375
Reported-by: Richard Yao
Reported-by: Benjamin Coddington
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Pawel Jakub Dawidek <pawel@dawidek.net>
Closes#14655
Coverity reported a dereference after a NULL check in dbuf_verify(). If
`dn` is `NULL`, we can just assume that !dn->dn_free_txg, so we change
`!dn->dn_free_txg` to `(dn == NULL || !dn->dn_free_txg)`.
Reviewed-by: Tino Reichardt <milky-zfs@mcmilk.de>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reported-by: Coverity (CID-992298)
Closes#14619
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
The txg_sync thread will see certain buffers in a DR_IN_DMU_SYNC state
when ZIL is writing them out. Then it waits until the state changes, but
has an assertion to check that they were not DR_NOT_OVERRIDDEN. If the
data write failed with an error, ZIL will put it into the
DR_NOT_OVERRIDDEN state. It looks like the code will handle that state
without an issue, so we can just delete the assertion.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@klarasystems.com>
Sponsored-By: Wasabi Technology, Inc.
Closes#14283
Clang's static analyzer reports that if a `blkid == DMU_SPILL_BLKID` is
passed, then we can have a NULL pointer dereference when either
->dn_have_spill or `DNODE_FLAG_SPILL_BLKPTR` is not set. This should not
happen. We add an `ASSERT()` to suppress reports about NULL pointer
dereferences.
Originally, I wanted to use one or two IMPLY statements on
pre-conditions before the call to `dbuf_findbp()`, but Clang's static
analyzer did not understand it.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14575
by placing the most common use case (no special vdevs) first and avoid
allocating new variables.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#14494Closes#14563
Clang's static analyzer claims that dbuf_hold_copy() will have a NULL
pointer dereference in data->b_data when called by dbuf_hold_impl().
This is impossible because data is dr->dt.dl.dr_data, which is non-NULL
whenever db->db_level == 0, which is always the case whenever
dbuf_hold_impl() calls dbuf_hold_copy(). We add an assertion to suppress
the complaint.
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#14470
Previously the primarycache property was handled only in the dbuf
layer. Since the speculative prefetcher is implemented in the ARC,
it had to be disabled for uncacheable buffers.
This change gives the ARC knowledge about uncacheable buffers
via arc_read() and arc_write(). So when remove_reference() drops
the last reference on the ARC header, it can either immediately destroy
it, or if it is marked as prefetch, put it into a new arc_uncached state.
That state is scanned every second, evicting stale buffers that were
not demand read.
This change also tracks dbufs that were read from the beginning,
but not to the end. It is assumed that such buffers may receive further
reads, and so they are stored in dbuf cache. If a following
reads reaches the end of the buffer, it is immediately evicted.
Otherwise it will follow regular dbuf cache eviction. Since the dbuf
layer does not know actual file sizes, this logic is not applied to
the final buffer of a dnode.
Since uncacheable buffers should no longer stay in the ARC for long,
this patch also tries to optimize I/O by allocating ARC physical
buffers as linear to allow buffer sharing.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: George Wilson <george.wilson@delphix.com>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#14243
We currently compute a 64-bit hash three times, which consumes 0.8% CPU
time on ARC eviction heavy workloads. Caching the 64-bit value in the
dbuf allows us to avoid that overhead.
Sponsored-By: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: Richard Yao <richard.yao@klarasystems.com>
Closes#14251
When ZFS is built with assertions, a prefetch is done on a redacted
blkptr and `dpa->dpa_dnode` is NULL, we will have a NULL pointer
dereference in `dbuf_prefetch_indirect_done()`.
Both Coverity and Clang's Static Analyzer caught this.
Reported-by: Coverity (CID 1524671)
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#14210
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Closes#14199
These were categorized as the following:
* Dead assignment 23
* Dead increment 4
* Dead initialization 6
* Dead nested assignment 18
Most of these are harmless, but since actual issues can hide among them,
we correct them.
That said, there were a few return values that were being ignored that
appeared to merit some correction:
* `destroy_callback()` in `cmd/zfs/zfs_main.c` ignored the error from
`destroy_batched()`. We handle it by returning -1 if there is an
error.
* `zfs_do_upgrade()` in `cmd/zfs/zfs_main.c` ignored the error from
`zfs_for_each()`. We handle it by doing a binary OR of the error
value from the subsequent `zfs_for_each()` call to the existing
value. This is how errors are mostly handled inside `zfs_for_each()`.
The error value here is passed to exit from the zfs command, so doing
a binary or on it is better than what we did previously.
* `get_zap_prop()` in `module/zfs/zcp_get.c` ignored the error from
`dsl_prop_get_ds()` when the property is not of type string. We
return an error when it does. There is a small concern that the
`zfs_get_temporary_prop()` call would handle things, but in the case
that it does not, we would be pushing an uninitialized numval onto
the lua stack. It is expected that `dsl_prop_get_ds()` will succeed
anytime that `zfs_get_temporary_prop()` does, so that not giving it a
chance to fix things is not a problem.
* `draid_merge_impl()` in `tests/zfs-tests/cmd/draid.c` used
`nvlist_add_nvlist()` twice in ways in which errors are expected to
be impossible, so we switch to `fnvlist_add_nvlist()`.
A few notable ones did not merit use of the return value, so we
suppressed it with `(void)`:
* `write_free_diffs()` in `lib/libzfs/libzfs_diff.c` ignored the error
value from `describe_free()`. A look through the commit history
revealed that this was intentional.
* `arc_evict_hdr()` in `module/zfs/arc.c` did not need to use the
returned handle from `arc_hdr_realloc()` because it is already
referenced in lists.
* `spa_vdev_detach()` in `module/zfs/spa.c` has a comment explicitly
saying not to use the error from `vdev_label_init()` because whatever
causes the error could be the reason why a detach is being done.
Unfortunately, I am not presently able to analyze the kernel modules
with Clang's static analyzer, so I could have missed some cases of this.
In cases where reports were present in code that is duplicated between
Linux and FreeBSD, I made a conscious effort to fix the FreeBSD version
too.
After this commit is merged, regressions like dee8934 should become
extremely obvious with Clang's static analyzer since a regression would
appear in the results as the only instance of unused code. That assumes
that Coverity does not catch the issue first.
My local branch with fixes from all of my outstanding non-draft pull
requests shows 118 reports from Clang's static anlayzer after this
patch. That is down by 51 from 169.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Cedric Berger <cedric@precidata.com>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#13986
Various module parameters such as `zfs_arc_max` were originally
`uint64_t` on OpenSolaris/Illumos, but were changed to `unsigned long`
for Linux compatibility because Linux's kernel default module parameter
implementation did not support 64-bit types on 32-bit platforms. This
caused problems when porting OpenZFS to Windows because its LLP64 memory
model made `unsigned long` a 32-bit type on 64-bit, which created the
undesireable situation that parameters that should accept 64-bit values
could not on 64-bit Windows.
Upon inspection, it turns out that the Linux kernel module parameter
interface is extensible, such that we are allowed to define our own
types. Rather than maintaining the original type change via hacks to to
continue shrinking module parameters on 32-bit Linux, we implement
support for 64-bit module parameters on Linux.
After doing a review of all 64-bit kernel parameters (found via the man
page and also proposed changes by Andrew Innes), the kernel module
parameters fell into a few groups:
Parameters that were originally 64-bit on Illumos:
* dbuf_cache_max_bytes
* dbuf_metadata_cache_max_bytes
* l2arc_feed_min_ms
* l2arc_feed_secs
* l2arc_headroom
* l2arc_headroom_boost
* l2arc_write_boost
* l2arc_write_max
* metaslab_aliquot
* metaslab_force_ganging
* zfetch_array_rd_sz
* zfs_arc_max
* zfs_arc_meta_limit
* zfs_arc_meta_min
* zfs_arc_min
* zfs_async_block_max_blocks
* zfs_condense_max_obsolete_bytes
* zfs_condense_min_mapping_bytes
* zfs_deadman_checktime_ms
* zfs_deadman_synctime_ms
* zfs_initialize_chunk_size
* zfs_initialize_value
* zfs_lua_max_instrlimit
* zfs_lua_max_memlimit
* zil_slog_bulk
Parameters that were originally 32-bit on Illumos:
* zfs_per_txg_dirty_frees_percent
Parameters that were originally `ssize_t` on Illumos:
* zfs_immediate_write_sz
Note that `ssize_t` is `int32_t` on 32-bit and `int64_t` on 64-bit. It
has been upgraded to 64-bit.
Parameters that were `long`/`unsigned long` because of Linux/FreeBSD
influence:
* l2arc_rebuild_blocks_min_l2size
* zfs_key_max_salt_uses
* zfs_max_log_walking
* zfs_max_logsm_summary_length
* zfs_metaslab_max_size_cache_sec
* zfs_min_metaslabs_to_flush
* zfs_multihost_interval
* zfs_unflushed_log_block_max
* zfs_unflushed_log_block_min
* zfs_unflushed_log_block_pct
* zfs_unflushed_max_mem_amt
* zfs_unflushed_max_mem_ppm
New parameters that do not exist in Illumos:
* l2arc_trim_ahead
* vdev_file_logical_ashift
* vdev_file_physical_ashift
* zfs_arc_dnode_limit
* zfs_arc_dnode_limit_percent
* zfs_arc_dnode_reduce_percent
* zfs_arc_meta_limit_percent
* zfs_arc_sys_free
* zfs_deadman_ziotime_ms
* zfs_delete_blocks
* zfs_history_output_max
* zfs_livelist_max_entries
* zfs_max_async_dedup_frees
* zfs_max_nvlist_src_size
* zfs_rebuild_max_segment
* zfs_rebuild_vdev_limit
* zfs_unflushed_log_txg_max
* zfs_vdev_max_auto_ashift
* zfs_vdev_min_auto_ashift
* zfs_vnops_read_chunk_size
* zvol_max_discard_blocks
Rather than clutter the lists with commentary, the module parameters
that need comments are repeated below.
A few parameters were defined in Linux/FreeBSD specific code, where the
use of ulong/long is not an issue for portability, so we leave them
alone:
* zfs_delete_blocks
* zfs_key_max_salt_uses
* zvol_max_discard_blocks
The documentation for a few parameters was found to be incorrect:
* zfs_deadman_checktime_ms - incorrectly documented as int
* zfs_delete_blocks - not documented as Linux only
* zfs_history_output_max - incorrectly documented as int
* zfs_vnops_read_chunk_size - incorrectly documented as long
* zvol_max_discard_blocks - incorrectly documented as ulong
The documentation for these has been fixed, alongside the changes to
document the switch to fixed width types.
In addition, several kernel module parameters were percentages or held
ashift values, so being 64-bit never made sense for them. They have been
downgraded to 32-bit:
* vdev_file_logical_ashift
* vdev_file_physical_ashift
* zfs_arc_dnode_limit_percent
* zfs_arc_dnode_reduce_percent
* zfs_arc_meta_limit_percent
* zfs_per_txg_dirty_frees_percent
* zfs_unflushed_log_block_pct
* zfs_vdev_max_auto_ashift
* zfs_vdev_min_auto_ashift
Of special note are `zfs_vdev_max_auto_ashift` and
`zfs_vdev_min_auto_ashift`, which were already defined as `uint64_t`,
and passed to the kernel as `ulong`. This is inherently buggy on big
endian 32-bit Linux, since the values would not be written to the
correct locations. 32-bit FreeBSD was unaffected because its sysctl code
correctly treated this as a `uint64_t`.
Lastly, a code comment suggests that `zfs_arc_sys_free` is
Linux-specific, but there is nothing to indicate to me that it is
Linux-specific. Nothing was done about that.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Jorgen Lundman <lundman@lundman.net>
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Original-patch-by: Andrew Innes <andrew.c12@gmail.com>
Original-patch-by: Jorgen Lundman <lundman@lundman.net>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#13984Closes#14004
Coverity complains about possible bugs involving referencing NULL return
values and division by zero. The division by zero bugs require that a
block pointer be corrupt, either from in-memory corruption, or on-disk
corruption. The NULL return value complaints are only bugs if
assumptions that we make about the state of data structures are wrong.
Some seem impossible to be wrong and thus are false positives, while
others are hard to analyze.
Rather than dismiss these as false positives by assuming we know better,
we add defensive assertions to let us know when our assumptions are
wrong.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#13972
In #13871, zfs_vdev_aggregation_limit_non_rotating and
zfs_vdev_aggregation_limit being signed was pointed out as a possible
reason not to eliminate an unnecessary MAX(unsigned, 0) since the
unsigned value was assigned from them.
There is no reason for these module parameters to be signed and upon
inspection, it was found that there are a number of other module
parameters that are signed, but should not be, so we make them unsigned.
Making them unsigned made it clear that some other variables in the code
should also be unsigned, so we also make those unsigned. This prevents
users from setting negative values that could potentially cause bad
behaviors. It also makes the code slightly easier to understand.
Mostly module parameters that deal with timeouts, limits, bitshifts and
percentages are made unsigned by this. Any that are boolean are left
signed, since whether booleans should be considered signed or unsigned
does not matter.
Making zfs_arc_lotsfree_percent unsigned caused a
`zfs_arc_lotsfree_percent >= 0` check to become redundant, so it was
removed. Removing the check was also necessary to prevent a compiler
error from -Werror=type-limits.
Several end of line comments had to be moved to their own lines because
replacing int with uint_t caused us to exceed the 80 character limit
enforced by cstyle.pl.
The following were kept signed because they are passed to
taskq_create(), which expects signed values and modifying the
OpenSolaris/Illumos DDI is out of scope of this patch:
* metaslab_load_pct
* zfs_sync_taskq_batch_pct
* zfs_zil_clean_taskq_nthr_pct
* zfs_zil_clean_taskq_minalloc
* zfs_zil_clean_taskq_maxalloc
* zfs_arc_prune_task_threads
Also, negative values in those parameters was found to be harmless.
The following were left signed because either negative values make
sense, or more analysis was needed to determine whether negative values
should be disallowed:
* zfs_metaslab_switch_threshold
* zfs_pd_bytes_max
* zfs_livelist_min_percent_shared
zfs_multihost_history was made static to be consistent with other
parameters.
A number of module parameters were marked as signed, but in reality
referenced unsigned variables. upgrade_errlog_limit is one of the
numerous examples. In the case of zfs_vdev_async_read_max_active, it was
already uint32_t, but zdb had an extern int declaration for it.
Interestingly, the documentation in zfs.4 was right for
upgrade_errlog_limit despite the module parameter being wrongly marked,
while the documentation for zfs_vdev_async_read_max_active (and friends)
was wrong. It was also wrong for zstd_abort_size, which was unsigned,
but was documented as signed.
Also, the documentation in zfs.4 incorrectly described the following
parameters as ulong when they were int:
* zfs_arc_meta_adjust_restarts
* zfs_override_estimate_recordsize
They are now uint_t as of this patch and thus the man page has been
updated to describe them as uint.
dbuf_state_index was left alone since it does nothing and perhaps should
be removed in another patch.
If any module parameters were missed, they were not found by `grep -r
'ZFS_MODULE_PARAM' | grep ', INT'`. I did find a few that grep missed,
but only because they were in files that had hits.
This patch intentionally did not attempt to address whether some of
these module parameters should be elevated to 64-bit parameters, because
the length of a long on 32-bit is 32-bit.
Lastly, it was pointed out during review that uint_t is a better match
for these variables than uint32_t because FreeBSD kernel parameter
definitions are designed for uint_t, whose bit width can change in
future memory models. As a result, we change the existing parameters
that are uint32_t to use uint_t.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Neal Gompa <ngompa@datto.com>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#13875
Incorrectly sizing the array of hash locks used to protect the
dbuf hash table can lead to contention and reduce performance.
We could unconditionally allocate a larger array for the locks
but it's wasteful, particularly for a low-memory system.
Instead, dynamically allocate the array of locks and scale
it based on total system memory.
Additionally, add a new `dbuf_mutex_cache_shift` module option
which can be used to override the hash lock array size. This is
disabled by default (dbuf_mutex_hash_shift=0) and can only be
set at module load time. The minimum target array size is set
to 8192, this matches the current constant value.
Note that the count of the dbuf hash table and count of the
mutex array were added to the /proc/spl/kstat/zfs/dbufstats
kstat.
Finally, this change removes the _KERNEL conditional checks.
These were not required since for the user space build there
is no difference between the kmem and vmem interfaces.
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#13928
This reverts commit 34dbc618f5. While this
change resolved the lock contention observed for certain workloads, it
inadventantly reduced the maximum hash inserts/removes per second. This
appears to be due to the slightly higher acquisition cost of a rwlock vs
a mutex.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Coverity complains about a "use-after-free" bug in
`dbuf_prefetch_indirect_done()` because we use a pointer value after
freeing its buffer. The pointer is used for refcounting in ARC (as the
reference holder). There is a theoretical situation where the pointer
would be reused in a way that causes the refcounting to collide, so we
change the order in which we call arc_buf_destroy() and
dbuf_prefetch_fini() to match the rest of the function. This prevents
the theoretical situation from being a possibility.
Also, we have a few return statements with a value, despite this being a
void function. We clean those up while we are making changes here.
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Neal Gompa <ngompa@datto.com>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Closes#13869
Move the use of the db pointer after it is freed. It's only used as
a tag so a dereference would never occur, but there's no reason we
can't invert the order to resolve the warning.
module/zfs/dbuf.c: In function 'dbuf_destroy':
module/zfs/dbuf.c:2953:17: error:
pointer 'db' may be used after 'free' [-Werror=use-after-free]
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#13528Closes#13575
Move the use of the private pointer after it is freed. It's only
used as a tag so a dereference would never occur, but there's no
harm in inverting the order to resolve the warning.
module/zfs/dbuf.c: In function 'dbuf_issue_final_prefetch_done':
module/zfs/dbuf.c:3204:17: error:
pointer 'private' may be used after 'free' [-Werror=use-after-free]
Reviewed-by: Ryan Moeller <ryan@iXsystems.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#13528Closes#13575
- Make prefetch distance adaptive: up to 4MB prefetch doubles for
every, hit same as before, but after that it grows by 1/8 every time
the prefetch read does not complete in time to satisfy the demand.
My tests show that 4MB is sufficient for wide NVMe pool to saturate
single reader thread at 2.5GB/s, while new 64MB maximum allows the
same thread to reach 1.5GB/s on wide HDD pool. Further distance
increase may increase speed even more, but less dramatic and with
higher latency.
- Allow early reuse of inactive prefetch streams: streams that never
saw hits can be reused immediately if there is a demand, while others
can be reused after 1s of inactivity, starting with the oldest. After
2s of inactivity streams are deleted to free resources same as before.
This allows by several times increase strided read performance on HDD
pool in presence of simultaneous random reads, previously filling the
zfetch_max_streams limit for seconds and so blocking most of prefetch.
- Always issue intermediate indirect block reads with SYNC priority.
Each of those reads if delayed for longer may delay up to 1024 other
block prefetches, that may be not good for wide pools.
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#13452
clang-15 emits the following error message for functions without
a prototype:
fs/zfs/os/linux/spl/spl-kmem-cache.c:1423:27: error:
a function declaration without a prototype is deprecated
in all versions of C [-Werror,-Wstrict-prototypes]
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Aidan Harris <me@aidanharr.is>
Closes#13421
Holding a dbuf is a common operation which can become highly contended
in dbuf_find() when acquiring the dbuf hash mutex. This is particularly
true on Linux when reading/writing volumes since by default up to 32
threads from the zvol_taskq may be taking a hold of the same dbuf.
This should also be observable on FreeBSD as long as there are enough
processes accessing the volume concurrently.
This is further aggregrated by the fact that only the block id will
be unique when calculating the dbuf hash for a single volume. The
objset id, object id, and level will be the same for data blocks.
This has been observed to result in a somehwat less than uniform hash
distribution and a longer than expected max hash chain depth (~20)
on a large memory system (256 GB) using volumes.
This commit improves the siutation by switching the hash mutex to
an rwlock to allow concurrent lookups, and increasing DBUF_RWLOCKS
from 2048 to 8192 to further reduce the odds of a hash collision.
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#13405
Parts of the Linux kernel build system struggle with _Noreturn. This
results in the following warnings when building on RHEL 8.5, and likely
other environments. Switch to using the __attribute__((noreturn)).
warning: objtool: dbuf_free_range()+0x2b8:
return with modified stack frame
warning: objtool: dbuf_free_range()+0x0:
stack state mismatch: cfa1=7+40 cfa2=7+8
...
WARNING: EXPORT symbol "arc_buf_size" [zfs.ko] version generation
failed, symbol will not be versioned.
WARNING: EXPORT symbol "spa_open" [zfs.ko] version generation
failed, symbol will not be versioned.
...
Additionally, __thread_exit() has been renamed spl_thread_exit() and
made a static inline function. This was needed because the kernel
will generate a warning for symbols which are __attribute__((noreturn))
and then exported with EXPORT_SYMBOL.
While we could continue to use _Noreturn in user space I've also
switched it to __attribute__((noreturn)) purely for consistency
throughout the code base.
Reviewed-by: Ryan Moeller <freqlabs@FreeBSD.org>
Reviewed-by: Brian Atkinson <batkinson@lanl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#13238
