While evicting dbufs of a dnode, a marker node is added to the AVL.
The marker node should be inserted in AVL tree ahead of the dbuf its
trying to delete. The blkid and level is used to ensure this. However,
this could go wrong there's another dbufs with the same blkid and level
in DB_EVICTING state but not yet removed from AVL tree. dbuf_compare()
could fail to give the right location or could cause confusion and
trigger ASSERTs.
To ensure that the marker is inserted before the deleting dbuf, use
the pointer value of the original dbuf for comparision.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Co-authored-by: Sanjeev Bagewadi <sanjeev.bagewadi@nutanix.com>
Signed-off-by: Chunwei Chen <david.chen@nutanix.com>
Closes#12482Closes#15643
Over its history this the dirty dnode test has been changed between
checking for a dnodes being on `os_dirty_dnodes` (`dn_dirty_link`) and
`dn_dirty_record`.
de198f2d9 Fix lseek(SEEK_DATA/SEEK_HOLE) mmap consistency
2531ce372 Revert "Report holes when there are only metadata changes"
ec4f9b8f3 Report holes when there are only metadata changes
454365bba Fix dirty check in dmu_offset_next()
66aca2473 SEEK_HOLE should not block on txg_wait_synced()
Also illumos/illumos-gate@c543ec060dillumos/illumos-gate@2bcf0248e9
It turns out both are actually required.
In the case of appending data to a newly created file, the dnode proper
is dirtied (at least to change the blocksize) and dirty records are
added. Thus, a single logical operation is represented by separate
dirty indicators, and must not be separated.
The incorrect dirty check becomes a problem when the first block of a
file is being appended to while another process is calling lseek to skip
holes. There is a small window where the dnode part is undirtied while
there are still dirty records. In this case, `lseek(fd, 0, SEEK_DATA)`
would not know that the file is dirty, and would go to
`dnode_next_offset()`. Since the object has no data blocks yet, it
returns `ESRCH`, indicating no data found, which results in `ENXIO`
being returned to `lseek()`'s caller.
Since coreutils 9.2, `cp` performs sparse copies by default, that is, it
uses `SEEK_DATA` and `SEEK_HOLE` against the source file and attempts to
replicate the holes in the target. When it hits the bug, its initial
search for data fails, and it goes on to call `fallocate()` to create a
hole over the entire destination file.
This has come up more recently as users upgrade their systems, getting
OpenZFS 2.2 as well as a newer coreutils. However, this problem has been
reproduced against 2.1, as well as on FreeBSD 13 and 14.
This change simply updates the dirty check to check both types of dirty.
If there's anything dirty at all, we immediately go to the "wait for
sync" stage, It doesn't really matter after that; both changes are on
disk, so the dirty fields should be correct.
Sponsored-by: Klara, Inc.
Sponsored-by: Wasabi Technology, Inc.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Rich Ercolani <rincebrain@gmail.com>
Signed-off-by: Rob Norris <rob.norris@klarasystems.com>
Closes#15571Closes#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
Currently redaction bookmarks and their associated redaction lists
have a relatively low limit of 36 redaction snapshots. This is imposed
by the number of snapshot GUIDs that fit in the bonus buffer of the
redaction list object. While this is more than enough for most use
cases, there are some limited cases where larger numbers would be
useful to support.
We tweak the redaction list creation code to use a spill block if
the number of redaction snapshots is above the amount that would fit
in the bonus buffer. We also make a small change to allow spill blocks
to be use for types of data besides SA. In order to fully leverage
this logic, we also change the redaction code to use vmem_alloc, to
handle extremely large allocations if needed. Finally, small tweaks
were made to the zfs commands and the test suite.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: Paul Dagnelie <pcd@delphix.com>
Closes#15018
Unlike regular receive, raw receive require destination to have the
same block structure as the source. In case of dnode reclaim this
triggers two special cases, requiring special handling:
- If dn_nlevels == 1, we can change the ibs, but dnode_set_blksz()
should not dirty the data buffer if block size does not change, or
durign receive dbuf_dirty_lightweight() will trigger assertion.
- If dn_nlevels > 1, we just can't change the ibs, dnode_set_blksz()
would fail and receive_object would trigger assertion, so we should
destroy and recreate the dnode from scratch.
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored by: iXsystems, Inc.
Closes#15039
This reverts commit 4b3133e671.
Users identified this commit as a possible source of data
corruption:
https://github.com/openzfs/zfs/issues/14753
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Tony Hutter <hutter2@llnl.gov>
Issue #14753Closes#14761
031d7c2fe6 did not handle reverse
iteration, such that the original issue theoretically could still occur.
Note that contrary to the claim in the ZFS disk format specification
that a maximum of 6 levels are possible, 9 levels are possible with
recordsize=512 and and indirect block size of 16KB. In this unusual
configuration, span will be 65. The maximum size of span at 70 can be
reached at recordsize=16K and an indirect blocksize of 16KB.
When we are at this indirection level and are traversing backward, the
minimum value is start, but we cannot calculate that with 64-bit
arithmetic, so we avoid the calculation and instead rely on the earlier
statement that did `*offset = start;`.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Reported-by: Coverity (CID-1466214)
Closes#14618
`lseek(SEEK_DATA | SEEK_HOLE)` are only accurate when the on-disk blocks
reflect all writes, i.e. when there are no dirty data blocks. To ensure
this, if the target dnode is dirty, they wait for the open txg to be
synced, so we can call them "stabilizing operations". If they cause
txg_wait_synced often, it can be detrimental to performance.
Typically, a group of files are all modified, and then SEEK_DATA/HOLE
are performed on them. In this case, the first SEEK does a
txg_wait_synced(), and subsequent SEEKs don't need to wait, so
performance is good.
However, if a workload involves an interleaved metadata modification,
the subsequent SEEK may do a txg_wait_synced() unnecessarily. For
example, if we do a `read()` syscall to each file before we do its SEEK.
This applies even with `relatime=on`, when the `read()` is the first
read after the last write. The txg_wait_synced() is unnecessary because
the SEEK operations only care that the structure of the tree of indirect
and data blocks is up to date on disk. They don't care about metadata
like the contents of the bonus or spill blocks. (They also don't care
if an existing data block is modified, but this would be more involved
to filter out.)
This commit changes the behavior of SEEK_DATA/HOLE operations such that
they do not call txg_wait_synced() if there is only a pending change to
the bonus or spill block.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#13368
Issue #14594
Issue #14512
Issue #14009
The default_bs and default_ibs tunables control the default block size
and indirect block size.
So far, default_bs and default_ibs were tunable only on FreeBSD, e.g.,
sysctl vfs.zfs.default_ibs
Remove the FreeBSD-specific sysctl code and expose default_bs and
default_ibs as tunables on both Linux and FreeBSD using
ZFS_MODULE_PARAM.
One of the use cases for changing the values of those tunables is to
lower the indirect block size, which may improve performance of large
directories (as discussed during the OpenZFS Leadership Meeting
on 2022-08-16).
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Richard Yao <richard.yao@alumni.stonybrook.edu>
Signed-off-by: Mateusz Piotrowski <mateusz.piotrowski@klarasystems.com>
Sponsored-by: Wasabi Technology, Inc.
Closes#14293
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
I've noticed that some of those counters are used in hot paths like
dnode_hold_impl(), and results of this change is visible in profiler.
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#14198
In our codebase, `cond_resched() and `schedule()` are Linux kernel
functions that have replaced the OpenSolaris `kpreempt()` functions in
the codebase to such an extent that `kpreempt()` in zfs_context.h was
broken. Nobody noticed because we did not actually use it. The header
had defined `kpreempt()` as `yield()`, which works on OpenSolaris and
Illumos where `sched_yield()` is a wrapper for `yield()`, but that does
not work on any other platform.
The FreeBSD platform specific code implemented shims for these, but the
shim for `schedule()` forced us to wait, which is different than merely
rescheduling to another thread as the original Linux code does, while
the shim for `cond_resched()` had the same definition as its kernel
kpreempt() shim.
After studying this, I have concluded that we should reintroduce the
kpreempt() function in platform independent code with the following
definitions:
- In the Linux kernel:
kpreempt(unused) -> cond_resched()
- In the FreeBSD kernel:
kpreempt(unused) -> kern_yield(PRI_USER)
- In userspace:
kpreempt(unused) -> sched_yield()
In userspace, nothing changes from this cleanup. In the kernels, the
function `fm_fini()` will now call `kern_yield(PRI_USER)` on FreeBSD and
`cond_resched()` on Linux. This is instead of `pause("schedule", 1)` on
FreeBSD and `schedule()` on Linux. This makes our behavior consistent
across platforms.
Note that Linux's SPL continues to use `cond_resched()` and
`schedule()`. However, those functions have been removed from both the
FreeBSD code and userspace code.
This should have the benefit of making it slightly easier to port the
code to new platforms by making how things should be mapped less
confusing.
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#13845
If a dnode has a spill pointer, and we use DN_SLOTS_TO_BONUSLEN() then
we will possibly include the spill pointer in the len calculation and it
will be byteswapped. Then dnode_byteswap() will carry on and swap the
spill pointer again. Fix this by using DN_MAX_BONUS_LEN() instead.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: George Amanakis <gamanakis@gmail.com>
Closes#13002Closes#13015
bcopy() has a confusing argument order and is actually a move, not a
copy; they're all deprecated since POSIX.1-2001 and removed in -2008,
and we shim them out to mem*() on Linux anyway
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Ahelenia Ziemiańska <nabijaczleweli@nabijaczleweli.xyz>
Closes#12996
In addition to flushing memory mapped regions when checking holes,
commit de198f2d95 modified the dirty dnode detection logic to check
the dn->dn_dirty_records instead of the dn->dn_dirty_link. Relying
on the dirty record has not be reliable, switch back to the previous
method.
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #11900Closes#12745
When using lseek(2) to report data/holes memory mapped regions of
the file were ignored. This could result in incorrect results.
To handle this zfs_holey_common() was updated to asynchronously
writeback any dirty mmap(2) regions prior to reporting holes.
Additionally, while not strictly required, the dn_struct_rwlock is
now held over the dirty check to prevent the dnode structure from
changing. This ensures that a clean dnode can't be dirtied before
the data/hole is located. The range lock is now also taken to
ensure the call cannot race with zfs_write().
Furthermore, the code was refactored to provide a dnode_is_dirty()
helper function which checks the dnode for any dirty records to
determine its dirtiness.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Reviewed-by: Rich Ercolani <rincebrain@gmail.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #11900Closes#12724
Remove code duplication by moving code responsible for partial block
zeroing to a separate function: dnode_partial_zero().
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Ryan Moeller <ryan@ixsystems.com>
Signed-off-by: Pawel Jakub Dawidek <pawel@dawidek.net>
Closes#12627
It seems nothing ensures that this array is zeroed when a dnode is
freshly allocated, so in principle it retains the values from the
previous allocation. In practice it seems to be the case that the
fields should end up zeroed, but we can zero the field anyway for
consistency.
This was found using KMSAN.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Signed-off-by: Mark Johnston <markj@FreeBSD.org>
Closes#12383
ZFS loves using %llu for uint64_t, but that requires a cast to not
be noisy - which is even done in many, though not all, places.
Also a couple places used %u for uint64_t, which were promoted
to %llu.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Rich Ercolani <rincebrain@gmail.com>
Closes#12233
This commit partially reverts changes to multilists in PR 7968
(multi-threaded spa-sync()) and adds some cache line alignments to
separate read-only multilists and heavily modified refcount's to different
cache lines.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-by: iXsystems, Inc.
Closes#12158
Previous code tried to keep prefetch streams while moving dnode. But
it was at least not updating per-stream zs_fetchback pointers, causing
use-after-free on next access. Instead of that I see much easier and
cleaner to just drop old prefetch state and start new from scratch.
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#11936Closes#11998
The performance of `zfs receive` can be bottlenecked on the CPU consumed
by the `receive_writer` thread, especially when receiving streams with
small compressed block sizes. Much of the CPU is spent creating and
destroying dbuf's and arc buf's, one for each `WRITE` record in the send
stream.
This commit introduces the concept of "lightweight writes", which allows
`zfs receive` to write to the DMU by providing an ABD, and instantiating
only a new type of `dbuf_dirty_record_t`. The dbuf and arc buf for this
"dirty leaf block" are not instantiated.
Because there is no dbuf with the dirty data, this mechanism doesn't
support reading from "lightweight-dirty" blocks (they would see the
on-disk state rather than the dirty data). Since the dedup-receive code
has been removed, `zfs receive` is write-only, so this works fine.
Because there are no arc bufs for the received data, the received data
is no longer cached in the ARC.
Testing a receive of a stream with average compressed block size of 4KB,
this commit improves performance by 50%, while also reducing CPU usage
by 50% of a CPU. On a per-block basis, CPU consumed by receive_writer()
and dbuf_evict() is now 1/7th (14%) of what it was.
Baseline: 450MB/s, CPU in receive_writer() 40% + dbuf_evict() 35%
New: 670MB/s, CPU in receive_writer() 17% + dbuf_evict() 0%
The code is also restructured in a few ways:
Added a `dr_dnode` field to the dbuf_dirty_record_t. This simplifies
some existing code that no longer needs `DB_DNODE_ENTER()` and related
routines. The new field is needed by the lightweight-type dirty record.
To ensure that the `dr_dnode` field remains valid until the dirty record
is freed, we have to ensure that the `dnode_move()` doesn't relocate the
dnode_t. To do this we keep a hold on the dnode until it's zio's have
completed. This is already done by the user-accounting code
(`userquota_updates_task()`), this commit extends that so that it always
keeps the dnode hold until zio completion (see `dnode_rele_task()`).
`dn_dirty_txg` was previously zeroed when the dnode was synced. This
was not necessary, since its meaning can be "when was this dnode last
dirtied". This change simplifies the new `dnode_rele_task()` code.
Removed some dead code related to `DRR_WRITE_BYREF` (dedup receive).
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: George Wilson <gwilson@delphix.com>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#11105
Prefetching of dnodes in dbuf_read() can cause significant mutex
contention for some workloads and isn't very helpful. This is
because we already get 32 dnodes for each block read, and when
iterating over a directory we prefetch the dnodes in the directory.
Disable this prefetching to prevent the lock contention.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Submitted-by: Adam Moss <c@yotes.com>
Submitted-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: Adam Moss <c@yotes.com>
Closes#10877Closes#10953
Reviewed-by: Ryan Moeller <freqlabs@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Pavel Snajdr <snajpa@snajpa.net>
Closes#10879
Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#10744
dnode_special_close() waits for the refcount of dn_holds to go to zero
without holding the dn_mtx. dnode_rele_and_unlock() does the final
remove to dn_holds with dn_mtx being held:
refs = zfs_refcount_remove(&dn->dn_holds, tag);
mutex_exit(&dn->dn_mtx);
So, there is a race condition after the remove until dn_mtx is
dropped. During that time, dnode_destroy() can get called, which ends
up in dnode_dest() calling mutex_destroy() and a panic since the lock
is still held.
This change adds a condvar to wait for the final dnode_rele_and_unlock()
to release the dn_mtx before calling dnode_destroy().
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: John Poduska <jpoduska@datto.com>
Closes#7814Closes#10101
* Add dedicated donde_set_dirtyctx routine.
* Add empty dirty record on destroy assertion.
* Make much more extensive use of the SET_ERROR macro.
Reviewed-by: Will Andrews <wca@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes#9924
Since AVL already has embedded element counter, use dn_dbufs_count
only for dbufs not counted there (bonus buffers) and just add them.
This removes two atomics per dbuf life cycle.
According to profiler it reduces time spent by dbuf_destroy() inside
bottlenecked dbuf_evict_thread() from 13.36% to 9.20% of the core.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matt Ahrens <matt@delphix.com>
Signed-off-by: Alexander Motin <mav@FreeBSD.org>
Sponsored-By: iXsystems, Inc.
Closes#9949
Additionally pull in state machine comments about
upcoming async cow work.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matt Ahrens <matt@delphix.com>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes#9902
The dnp argument can only be set to NULL when the DNODE_DRY_RUN flag
is set. In which case, an early return path will be executed and a
NULL pointer dereference at the given location is impossible. Add
an additional ASSERT to silence the cppcheck warning and document
that dbp must never be NULL at the point in the function.
[module/zfs/dnode.c:1566]: (warning) Possible null pointer deref: dnp
Reviewed-by: Tony Hutter <hutter2@llnl.gov>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#9732
Remove the ASSERTV macro and handle suppressing unused
compiler warnings for variables only in ASSERTs using the
__attribute__((unused)) compiler annotation. The annotation
is understood by both gcc and clang.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Jorgen Lundman <lundman@lundman.net>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes#9671
This change modifies some of the infrastructure for enabling the use of
the DTRACE_PROBE* macros, such that we can use tehm in the "spl" module.
Currently, when the DTRACE_PROBE* macros are used, they get expanded to
create new functions, and these dynamically generated functions become
part of the "zfs" module.
Since the "spl" module does not depend on the "zfs" module, the use of
DTRACE_PROBE* in the "spl" module would result in undefined symbols
being used in the "spl" module. Specifically, DTRACE_PROBE* would turn
into a function call, and the function being called would be a symbol
only contained in the "zfs" module; which results in a linker and/or
runtime error.
Thus, this change adds the necessary logic to the "spl" module, to
mirror the tracing functionality available to the "zfs" module. After
this change, we'll have a "trace_zfs.h" header file which defines the
probes available only to the "zfs" module, and a "trace_spl.h" header
file which defines the probes available only to the "spl" module.
Reviewed by: Brad Lewis <brad.lewis@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Prakash Surya <prakash.surya@delphix.com>
Closes#9525
This assert makes non portable assumptions about the state of memory
returned by the memory allocator.
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Reviewed-by: Jorgen Lundman <lundman@lundman.net>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes#9506
This patch implements a new tree structure for ZFS, and uses it to
store range trees more efficiently.
The new structure is approximately a B-tree, though there are some
small differences from the usual characterizations. The tree has core
nodes and leaf nodes; each contain data elements, which the elements
in the core nodes acting as separators between its children. The
difference between core and leaf nodes is that the core nodes have an
array of children, while leaf nodes don't. Every node in the tree may
be only partially full; in most cases, they are all at least 50% full
(in terms of element count) except for the root node, which can be
less full. Underfull nodes will steal from their neighbors or merge to
remain full enough, while overfull nodes will split in two. The data
elements are contained in tree-controlled buffers; they are copied
into these on insertion, and overwritten on deletion. This means that
the elements are not independently allocated, which reduces overhead,
but also means they can't be shared between trees (and also that
pointers to them are only valid until a side-effectful tree operation
occurs). The overhead varies based on how dense the tree is, but is
usually on the order of about 50% of the element size; the per-node
overheads are very small, and so don't make a significant difference.
The trees can accept arbitrary records; they accept a size and a
comparator to allow them to be used for a variety of purposes.
The new trees replace the AVL trees used in the range trees today.
Currently, the range_seg_t structure contains three 8 byte integers
of payload and two 24 byte avl_tree_node_ts to handle its storage in
both an offset-sorted tree and a size-sorted tree (total size: 64
bytes). In the new model, the range seg structures are usually two 4
byte integers, but a separate one needs to exist for the size-sorted
and offset-sorted tree. Between the raw size, the 50% overhead, and
the double storage, the new btrees are expected to use 8*1.5*2 = 24
bytes per record, or 33.3% as much memory as the AVL trees (this is
for the purposes of storing metaslab range trees; for other purposes,
like scrubs, they use ~50% as much memory).
We reduced the size of the payload in the range segments by teaching
range trees about starting offsets and shifts; since metaslabs have a
fixed starting offset, and they all operate in terms of disk sectors,
we can store the ranges using 4-byte integers as long as the size of
the metaslab divided by the sector size is less than 2^32. For 512-byte
sectors, this is a 2^41 (or 2TB) metaslab, which with the default
settings corresponds to a 256PB disk. 4k sector disks can handle
metaslabs up to 2^46 bytes, or 2^63 byte disks. Since we do not
anticipate disks of this size in the near future, there should be
almost no cases where metaslabs need 64-byte integers to store their
ranges. We do still have the capability to store 64-byte integer ranges
to account for cases where we are storing per-vdev (or per-dnode) trees,
which could reasonably go above the limits discussed. We also do not
store fill information in the compact version of the node, since it
is only used for sorted scrub.
We also optimized the metaslab loading process in various other ways
to offset some inefficiencies in the btree model. While individual
operations (find, insert, remove_from) are faster for the btree than
they are for the avl tree, remove usually requires a find operation,
while in the AVL tree model the element itself suffices. Some clever
changes actually caused an overall speedup in metaslab loading; we use
approximately 40% less cpu to load metaslabs in our tests on Illumos.
Another memory and performance optimization was achieved by changing
what is stored in the size-sorted trees. When a disk is heavily
fragmented, the df algorithm used by default in ZFS will almost always
find a number of small regions in its initial cursor-based search; it
will usually only fall back to the size-sorted tree to find larger
regions. If we increase the size of the cursor-based search slightly,
and don't store segments that are smaller than a tunable size floor
in the size-sorted tree, we can further cut memory usage down to
below 20% of what the AVL trees store. This also results in further
reductions in CPU time spent loading metaslabs.
The 16KiB size floor was chosen because it results in substantial memory
usage reduction while not usually resulting in situations where we can't
find an appropriate chunk with the cursor and are forced to use an
oversized chunk from the size-sorted tree. In addition, even if we do
have to use an oversized chunk from the size-sorted tree, the chunk
would be too small to use for ZIL allocations, so it isn't as big of a
loss as it might otherwise be. And often, more small allocations will
follow the initial one, and the cursor search will now find the
remainder of the chunk we didn't use all of and use it for subsequent
allocations. Practical testing has shown little or no change in
fragmentation as a result of this change.
If the size-sorted tree becomes empty while the offset sorted one still
has entries, it will load all the entries from the offset sorted tree
and disregard the size floor until it is unloaded again. This operation
occurs rarely with the default setting, only on incredibly thoroughly
fragmented pools.
There are some other small changes to zdb to teach it to handle btrees,
but nothing major.
Reviewed-by: George Wilson <gwilson@delphix.com>
Reviewed-by: Matt Ahrens <matt@delphix.com>
Reviewed by: Sebastien Roy seb@delphix.com
Reviewed-by: Igor Kozhukhov <igor@dilos.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Paul Dagnelie <pcd@delphix.com>
Closes#9181
Move Linux specific tracing headers and source to platform directories
and update the build system.
Reviewed-by: Allan Jude <allanjude@freebsd.org>
Reviewed-by: Ryan Moeller <ryan@ixsystems.com>
Reviewed by: Brad Lewis <brad.lewis@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matt Macy <mmacy@FreeBSD.org>
Closes#9290
Reviewed-by: Matt Ahrens <matt@delphix.com>
Reviewed-by: Ryan Moeller <ryan@ixsystems.com>
Reviewed-by: Richard Laager <rlaager@wiktel.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Andrea Gelmini <andrea.gelmini@gelma.net>
Closes#9240
If TX_REMOVE is followed by TX_CREATE on the same object id, we need to
make sure the object removal is completely finished before creation. The
current implementation relies on dnode_hold_impl with
DNODE_MUST_BE_ALLOCATED returning ENOENT. While this check seems to work
fine before, in current version it does not guarantee the object removal
is completed.
We fix this by checking if DNODE_MUST_BE_FREE returns successful
instead. Also add test and remove dead code in dnode_hold_impl.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tom Caputi <tcaputi@datto.com>
Signed-off-by: Chunwei Chen <david.chen@nutanix.com>
Closes#7151Closes#8910Closes#9123Closes#9145
This patch introduces an assertion that can catch pitfalls in
development where there is a mismatch between the size of
reads and writes between a *_phys structure and its respective
in-core structure when bonus buffers are used.
This debugging-aid should be complementary to the verification
done by ztest in ztest_verify_dnode_bt().
A side to this patch is that we now clear out any extra bytes
past a bonus buffer's new size when the buffer is shrinking.
Reviewed-by: Matt Ahrens <matt@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tom Caputi <tcaputi@datto.com>
Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com>
Closes#8348
The rwlock implementation on linux does not perform as well as mutexes.
We can realize a performance benefit by replacing the zf_rwlock with a
mutex. Local microbenchmarks show ~50% improvement, and over NFS we see
~5% improvement on several of the ZFS Performance Tests cases,
especially randwrite and seq_write.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Tony Nguyen <tony.nguyen@delphix.com>
Reviewed-by: Olaf Faaland <faaland1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes#9062
External consumers such as Lustre require access to the dnode
interfaces in order to correctly manipulate dnodes.
Reviewed-by: James Simmons <uja.ornl@yahoo.com>
Reviewed-by: Olaf Faaland <faaland1@llnl.gov>
Reviewed-by: George Melikov <mail@gmelikov.ru>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #8994Closes#9027
Currently, sequential async write workloads spend a lot of time
contending on the dn_struct_rwlock. This lock is responsible for
protecting the entire block tree below it; this naturally results
in some serialization during heavy write workloads. This can be
resolved by having per-dbuf locking, which will allow multiple
writers in the same object at the same time.
We introduce a new rwlock, the db_rwlock. This lock is responsible
for protecting the contents of the dbuf that it is a part of; when
reading a block pointer from a dbuf, you hold the lock as a reader.
When writing data to a dbuf, you hold it as a writer. This allows
multiple threads to write to different parts of a file at the same
time.
Reviewed by: Brad Lewis <brad.lewis@delphix.com>
Reviewed by: Matt Ahrens matt@delphix.com
Reviewed by: George Wilson george.wilson@delphix.com
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Paul Dagnelie <pcd@delphix.com>
External-issue: DLPX-52564
External-issue: DLPX-53085
External-issue: DLPX-57384
Closes#8946
The db_dirtycnt of an EVICTING dbuf is always 0. However, it still
appears in the dn_dbufs tree. If we call dnode_dirty_l1range on a
range that contains an EVICTING dbuf, we will attempt to mark it dirty
(which will fail because it's EVICTING, resulting in a new dbuf being
created and dirtied). Later, in ZFS_DEBUG mode, we assert that all the
dbufs in the range are dirty. If the EVICTING dbuf is still present,
this will trip the assertion erroneously.
Reviewed-by: Matt Ahrens <mahrens@delphix.com>
Reviewed-by: Richard Elling <Richard.Elling@RichardElling.com>
Reviewed-by: Sara Hartse <sara.hartse@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Paul Dagnelie <pcd@delphix.com>
Closes#8745
When receiving a DRR_OBJECT record the receive_object() function
needs to determine how to handle a spill block associated with the
object. It may need to be removed or kept depending on how the
object was modified at the source.
This determination is currently accomplished using a heuristic which
takes in to account the DRR_OBJECT record and the existing object
properties. This is a problem because there isn't quite enough
information available to do the right thing under all circumstances.
For example, when only the block size changes the spill block is
removed when it should be kept.
What's needed to resolve this is an additional flag in the DRR_OBJECT
which indicates if the object being received references a spill block.
The DRR_OBJECT_SPILL flag was added for this purpose. When set then
the object references a spill block and it must be kept. Either
it is update to date, or it will be replaced by a subsequent DRR_SPILL
record. Conversely, if the object being received doesn't reference
a spill block then any existing spill block should always be removed.
Since previous versions of ZFS do not understand this new flag
additional DRR_SPILL records will be inserted in to the stream.
This has the advantage of being fully backward compatible. Existing
ZFS systems receiving this stream will recreate the spill block if
it was incorrectly removed. Updated ZFS versions will correctly
ignore the additional spill blocks which can be identified by
checking for the DRR_SPILL_UNMODIFIED flag.
The small downside to this approach is that is may increase the size
of the stream and of the received snapshot on previous versions of
ZFS. Additionally, when receiving streams generated by previous
unpatched versions of ZFS spill blocks may still be lost.
OpenZFS-issue: https://www.illumos.org/issues/9952
FreeBSD-issue: https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=233277
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Matt Ahrens <mahrens@delphix.com>
Reviewed-by: Tom Caputi <tcaputi@datto.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#8668
Fix style issue for 'tx->tx_txg&TXG_MASK'. There should be white
space around the '&' character. Split the dnode_reallocate() ASSERT
to make it more readable to clearly separate the checks.
Reviewed-by: George Melikov <mail@gmelikov.ru>
Reviewed-by: Tom Caputi <tcaputi@datto.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Closes#8606
This partially reverts commit 5dbf8b4ed. This change resolved
the issues observed with truncated files in raw sends. However,
the required changes to dnode_allocate() introduced a regression
for non-raw streams which needs to be understood.
The additional debugging improvements from the original patch
were not reverted.
Reviewed-by: Tom Caputi <tcaputi@datto.com>
Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
Issue #7378
Issue #8528
Issue #8540
Issue #8565Close#8584
