fcmpset can have two kinds of semantics, weak and strong.
For practical purposes, strong semantics means that if fcmpset fails
then the reported current value is always different from the expected
value. Weak semantics means that the reported current value may be the
same as the expected value even though fcmpset failed. That's a so
called "sporadic" failure.
I originally implemented atomic_cas expecting strong semantics, but many
platforms actually have weak one.
Reported by: pkubaj (not confirmed if same issue)
Discussed with: kib, mjg
MFC after: 19 days
X-MFC with: r353340
This patch fixes 2 issues with the DMU free throttle implemented
in dmu_free_long_range(). The first issue is that get_next_chunk()
was calculating the number of L1 blocks the free would dirty
incorrectly. In some cases involving extremely large files, this
code would greatly overestimate the number of affected L1 blocks,
causing excessive calls to txg_wait_open(). This patch corrects
the calculation.
The second issue is that the free throttle uses the total number
of free'd blocks in all (open, quiescing, and syncing) txgs to
determine whether to throttle. This causes large frees (such as
those created by the first issue) to cause 4 txg syncs before
any further frees were allowed to proceed. This patch ensures
that the accounting is done entirely in a per-txg fashion, so
that frees from a given txg don't affect those that immediately
follow it.
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Matthew Ahrens <mahrens@delphix.com>
Signed-off-by: Tom Caputi <tcaputi@datto.com>
zfsonlinux/zfs@f4c594da94
Freeing throttle should account for holes
Deletion throttle currently does not account for holes in a file.
This means that it can activate when it shouldn't.
To fix it we switch the throttle to be based on the number of
L1 blocks we will have to dirty when freeing
Reviewed-by: Tom Caputi <tcaputi@datto.com>
Reviewed-by: Matt Ahrens <mahrens@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Alek Pinchuk <apinchuk@datto.com>
zfsonlinux/zfs@65282ee9e0
Submitted by: Alek Pinchuk <pinchuk.alek@gmail.com>
Reviewed by: allanjude
MFC after: 2 weeks
Sponsored by: Axcient
Differential Revision: https://reviews.freebsd.org/D21895
membar_producer is supposed to be a store-store barrier.
Also, in the code that FreeBSD has ported from illumos membar_producer
is used only with regular stores to regular memory (with respect to
caching).
We do not have an MI primitive for the store-store barrier, so
atomic_thread_fence_rel is the closest we have as it provides
(load | store) -> store barrier.
Previously, membar_producer was an empty function call on all 32-bit
arm-s, 32-bit powerpc, riscv and all mips variants. I think that it was
inadequate.
On other platforms, such as amd64, arm64, i386, powerpc64, sparc64,
membar_producer was implemented using stronger primitives than required
for a store-store barrier with respect to regular memory access.
For example, it used sfence on amd64 and lock-ed nop in i386 (despite TSO).
On powerpc64 we now use recommended lwsync instead of eieio.
On sparc64 FreeBSD uses TSO mode.
On arm64/aarch64 we now use dmb sy instead of dmb ish. Not sure if this
is an improvement, actually.
After this change we can drop opensolaris_atomic.S for aarch64, amd64,
powerpc64 and sparc64 as all required atomic operations have either
direct or light-weight mapping to FreeBSD native atomic operations.
Discussed with: kib
MFC after: 4 weeks
atomic_cas_32 is implemented using atomic_fcmpset_32 on all platforms.
Ditto for atomic_cas_64 and atomic_fcmpset_64 on platforms that have it.
The only exception is sparc64 that provides MD atomic_cas_32 and
atomic_cas_64.
This is slightly inefficient as fcmpset reports whether the operation
updated the target and that information is not needed for cas.
Nevertheless, there is less code to maintain and to add for new platforms.
Also, the operations are done inline now as opposed to function calls before.
atomic_add_64_nv is implemented using atomic_fetchadd_64 on platforms
that provide it.
casptr, cas32, atomic_or_8, atomic_or_8_nv are completely removed as they
have no users.
atomic_mtx that is used to emulate 64-bit atomics on platforms that lack
them is defined only on those platforms.
As a result, platform specific opensolaris_atomic.S files have lost most of
their code. The only exception is i386 where the compat+contrib code
provides 64-bit atomics for userland use. That code assumes availability of
cmpxchg8b instruction. FreeBSD does not have that assumption for i386
userland and does not provide 64-bit atomics. Hopefully, this can and will
be fixed.
MFC after: 3 weeks
As long as we support ZFS on 32-bit platforms we should do this for all
64-bit variables that are modified in a lockless fashion using atomic
operations. Otherwise, there is a risk of a reading a torn value.
Here is a rationale for why I am doing this in dmu_object_alloc_impl:
- it's very recent code
- the code deals with object IDs and a number of objects in a file
system can overflow 32 bits
- incorrect allocation of an object ID may result in hard to debug
problems
- fixing all plain reads of 64-bit atomic variables is not a trivial
undertaking to do in one shot, so I chose to do it incrementally
MFC after: 3 weeks
X-MFC after: r353301, r353176
The compatibility code for the atomic operations in ZFS code is a bit
messy. In some cases the native definitions are directly made
available, in some cases there are emulated operations in
opensolaris_atomic.c and in yet other cases there are atomic operations
implemented in assembly that were obtained from OpenSolaris / illumos.
This commit adds atomic_swap_64 for use with i386 userland.
The code is copied from illumos.
I am not sure why FreeBSD does not provide that operation natively.
Maybe because we try (or pretend) to support processors that did not
have the necessary instructions.
While here I also added atomic_load_64 for the same reasons.
This is original code based on iilumos atomic_swap_64 and FreeBSD
atomic_load_acq_64_i586.
Pointyhat to: avg
MFC after: 1 week
8423 8199 7432 Implement large_dnode pool feature
7432 Large dnode pool feature
8199 multi-threaded dmu_object_alloc()
8423 Implement large_dnode pool feature
10406 large_dnode changes broke zfs recv of legacy stream
llumos/illumos-gate@54811da5ac54811da5achttps://www.illumos.org/issues/8423https://www.illumos.org/issues/8199https://www.illumos.org/issues/7432illumos/illumos-gate@811964cd9f811964cd9fhttps://www.illumos.org/issues/10406
ZoL issues:
Improved dnode allocation #6564
Clean up large dnode code #6262
Fix dnode_hold() freeing dnode behavior #8172
Fix dnode allocation race #6414, #6439
Partial: Raw sends must be able to decrease nlevels #6821, #6864
Remove unnecessary txg syncs from receive_object() Closes#7197
This updates FreeBSD large_dnode code (that was imported from ZoL) to a
version that was committed to illumos. It has some cleanups,
improvements and fixes comparing to what we have in FreeBSD now.
I think that the most significant update is 8199 multi-threaded
dmu_object_alloc().
This commit reverts r351077 that was a revert of r351074 and r351076 and
restores those changes. Required atomic operations should be available
now on all platforms where we build ZFS.
Obtained from: illumos
MFC after: 3 weeks
Previously, the code used a plain store on platforms that lacked
atomic_swap_64 and possibly some other platforms as the condition worked
only if atomic_swap_64 was a macro.
MFC after: 1 week
X-MFC after: r353166, r353167
Some 32-bit platforms do not provide 64-bit atomic operations that ZFS
requires, either in userland or at all. We emulate those operations for
those platforms using a mutex. That is not entirely correct and it's
very efficient. Besides, the loads are plain loads, so torn values are
possible.
Nevertheless, the emulation seems to work for some definition of work.
This change adds atomic_swap_64, which is already used in ZFS code, and
atomic_load_64 that can be used to prevent torn reads.
MFC after: 1 week
The registers in ilumos and FreeBSD have a different number.
In the illumos, last 32-bits register defined is SS an in FreeBSD is GS.
While translating register we should comper it to the highest one.
PR: 240358
Reported by: lwhsu@
MFC after: 2 weeks
The feature is implemented as an extension of the existing
ZFS_IOC_RENAME ioctl. Both the userland and the DSL interfaces support
renaming only a single bookmark at a time. As of now, there is no ZCP
interface to the new functionality. I am going to add it once the DSL
interface passes a test of time.
This change picks up support for zfs_ioc_namecheck_t::ENTITY_NAME that
was added to ZoL as part of Redacted Send/Receive feature by Paul
Dagnelie <pcd@delphix.com>. This is needed to allow a bookmark name in
zc_name.
Discussed with: mahrens
Reviewed by: bcr (man page)
Sponsored by: CyberSecure
Differential Revision: https://reviews.freebsd.org/D21795
Before my ZIL space optimization few years ago 128KB writes were logged
as two 64KB+ records in two 128KB log blocks. After that change it became
~124KB+/4KB+ in two 128KB log blocks to free space in the second block
for another record. Unfortunately in case of 128KB only writes, when space
in the second block remained unused, that change increased write latency by
imbalancing checksum computation time between parallel threads.
This change introduces new 68KB log block size, used for both writes below
67KB and 128KB-sharp writes. Writes of 68-127KB are still using one 128KB
block to not increase processing overhead. Writes above 131KB are still
using full 128KB blocks, since possible saving there is small. Mixed loads
will likely also fall back to previous 128KB, since code uses maximum of
the last 10 requested block sizes.
On a simple 128KB write test with queue depth of 1 this change demonstrates
~15-20% performance improvement.
MFC after: 2 weeks
Sponsored by: iXsystems, Inc.
There is no reason for these routines to be written in assembly. In
the ports of DTrace to other platforms, they are already written in C.
No functional change intended.
MFC after: 1 week
Sponsored by: Netflix
Add a small wrapper around libzfs_core's lzc_send_space() to libzfs so
that every legacy ZFS_IOC_SEND consumer, along with their userland
counterpart estimate_ioctl(), can leverage ZFS_IOC_SEND_SPACE to
request send space estimation.
The legacy functionality in zfs_ioc_send() is left untouched for
compatibility purposes.
Obtained from: ZoL
Obtained from: zfsonlinux/zfs@cf7684bc8d
Author: loli10K <ezomori.nozomu@gmail.com>
MFC after: 2 weeks
It was incorrect with respect to swapping dataset IDs both in the
on-disk ZAP object and the in-memory queue.
In both cases, if only ds1 was already present, then it would be first
replaced with ds2 and then ds2 would be replaced back with ds1. Also,
both cases did not properly handle a situation where both ds1 and ds2
are already queued. A duplicate insertion would be attempted and its
failure would result in a panic.
This change has also been submitted to ZoL as zfsonlinux/zfs@dd262c9
PR: 239566
Reported by: pascal.guitierrez@gmail.com
MFC after: 4 days
Sponsored by: CyberSecure
New primitive is introduced to denote sections can operate locklessly
on aspects of struct mount, but which can also be disabled if necessary.
This provides an opportunity to start scaling common case modifications
while providing stable state of the struct when facing unmount, write
suspendion or other events.
mnt_ref is the first counter to start being managed in this manner with
the intent to make it per-cpu.
Reviewed by: kib, jeff
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D21425
- Remove a dead variable from the amd64 pmap_extract_and_hold().
- Fix grammar in the vm_page_wire man page.
Reported by: alc
Reviewed by: alc, kib
Sponsored by: Netflix
Differential Revision: https://reviews.freebsd.org/D21639
- VM_ALLOC_NOCREAT will grab without creating a page.
- vm_page_grab_valid() will grab and page in if necessary.
- vm_page_busy_acquire() automates some busy acquire loops.
Discussed with: alc, kib, markj
Tested by: pho (part of larger branch)
Sponsored by: Netflix
Differential Revision: https://reviews.freebsd.org/D21546
races with page busy state. The object lock is still used as an interlock
to ensure that the identity stays valid. Most callers should use
vm_page_sleep_if_busy() to handle the locking particulars.
Reviewed by: alc, kib, markj
Sponsored by: Netflix
Differential Revision: https://reviews.freebsd.org/D21255
There are several mechanisms by which a vm_page reference is held,
preventing the page from being freed back to the page allocator. In
particular, holding the page's object lock is sufficient to prevent the
page from being freed; holding the busy lock or a wiring is sufficent as
well. These references are protected by the page lock, which must
therefore be acquired for many per-page operations. This results in
false sharing since the page locks are external to the vm_page
structures themselves and each lock protects multiple structures.
Transition to using an atomically updated per-page reference counter.
The object's reference is counted using a flag bit in the counter. A
second flag bit is used to atomically block new references via
pmap_extract_and_hold() while removing managed mappings of a page.
Thus, the reference count of a page is guaranteed not to increase if the
page is unbusied, unmapped, and the object's write lock is held. As
a consequence of this, the page lock no longer protects a page's
identity; operations which move pages between objects are now
synchronized solely by the objects' locks.
The vm_page_wire() and vm_page_unwire() KPIs are changed. The former
requires that either the object lock or the busy lock is held. The
latter no longer has a return value and may free the page if it releases
the last reference to that page. vm_page_unwire_noq() behaves the same
as before; the caller is responsible for checking its return value and
freeing or enqueuing the page as appropriate. vm_page_wire_mapped() is
introduced for use in pmap_extract_and_hold(). It fails if the page is
concurrently being unmapped, typically triggering a fallback to the
fault handler. vm_page_wire() no longer requires the page lock and
vm_page_unwire() now internally acquires the page lock when releasing
the last wiring of a page (since the page lock still protects a page's
queue state). In particular, synchronization details are no longer
leaked into the caller.
The change excises the page lock from several frequently executed code
paths. In particular, vm_object_terminate() no longer bounces between
page locks as it releases an object's pages, and direct I/O and
sendfile(SF_NOCACHE) completions no longer require the page lock. In
these latter cases we now get linear scalability in the common scenario
where different threads are operating on different files.
__FreeBSD_version is bumped. The DRM ports have been updated to
accomodate the KPI changes.
Reviewed by: jeff (earlier version)
Tested by: gallatin (earlier version), pho
Sponsored by: Netflix
Differential Revision: https://reviews.freebsd.org/D20486
This fixes a hole in the situation where the resume state is left from
receiving a new dataset and, so, the state is set on the dataset itself
(as opposed to %recv child).
Additionally, distinguish incremental and resume streams in error
messages.
This was also committed to ZoL:
zfsonlinux/zfs@ebeb6f23bf
MFC after: 2 weeks
Sponsored by: CyberSecure
vnodes have 2 reference counts - holdcnt to keep the vnode itself from getting
freed and usecount to denote it is actively used.
Previously all operations bumping usecount would also bump holdcnt, which is
not necessary. We can detect if usecount is already > 1 (in which case holdcnt
is also > 1) and utilize it to avoid bumping holdcnt on our own. This saves
on atomic ops.
Reviewed by: kib
Tested by: pho (previous version)
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D21471
The page daemon periodically invokes uma_reclaim() to reclaim cached
items from each zone when the system is under memory pressure. This
is important since the size of these caches is unbounded by default.
However it also results in bursts of high latency when allocating from
heavily used zones as threads miss in the per-CPU caches and must
access the keg in order to allocate new items.
With r340405 we maintain an estimate of each zone's usage of its
(per-NUMA domain) cache of full buckets. Start making use of this
estimate to avoid reclaiming the entire cache when under memory
pressure. In particular, introduce TRIM, DRAIN and DRAIN_CPU
verbs for uma_reclaim() and uma_zone_reclaim(). When trimming, only
items in excess of the estimate are reclaimed. Draining a zone
reclaims all of the cached full buckets (the previous behaviour of
uma_reclaim()), and may further drain the per-CPU caches in extreme
cases.
Now, when under memory pressure, the page daemon will trim zones
rather than draining them. As a result, heavily used zones do not incur
bursts of bucket cache misses following reclamation, but large, unused
caches will be reclaimed as before.
Reviewed by: jeff
Tested by: pho (an earlier version)
MFC after: 2 months
Sponsored by: Netflix
Differential Revision: https://reviews.freebsd.org/D16667
Current implementation of vnode_create_vobject() and
vnode_destroy_vobject() is written so that it prepared to handle the
vm object destruction for live vnode. Practically, no filesystems use
this, except for some remnants that were present in UFS till today.
One of the consequences of that model is that each filesystem must
call vnode_destroy_vobject() in VOP_RECLAIM() or earlier, as result
all of them get rid of the v_object in reclaim.
Move the call to vnode_destroy_vobject() to vgonel() before
VOP_RECLAIM(). This makes v_object stable: either the object is NULL,
or it is valid vm object till the vnode reclamation. Remove code from
vnode_create_vobject() to handle races with the parallel destruction.
Reviewed by: markj
Tested by: pho
Sponsored by: The FreeBSD Foundation
Differential revision: https://reviews.freebsd.org/D21412
Previously, the permissions were checked on the pool which was obviously
incorrect.
After this change, zfs_check_userprops() only validates the properties
without any permission checks. The permissions are checked individually
for each snapshotted dataset.
This was also committed to ZoL: zfsonlinux/zfs@e6203d2
Reported by: CyberSecure
MFC after: 1 week
Sponsored by: CyberSecure
If vn_lock() failed, then the function returned the error but the vnode
obtained via zfs_zget() was never released.
MFC after: 10 days
Sponsored by: Panzura
illumos/illumos-gate@811964cd9f811964cd9fhttps://www.illumos.org/issues/10406
The large dnode changes from 8423 caused problems in zfs recv for a legacy
stream. This manifests when attempting to mount the received stream, but the
problem is in the receive code. We missed the following commit from ZoL which
fixes this.
commit da2feb42fb
Author: Tom Caputi <tcaputi@datto.com>
Date: Thu Jun 28 17:55:11 2018 -0400
Fix 'zfs recv' of non large_dnode send streams
Currently, there is a bug where older send streams without the
DMU_BACKUP_FEATURE_LARGE_DNODE flag are not handled correctly.
The code in receive_object() fails to handle cases where
drro->drr_dn_slots is set to 0, which is always the case when the
sending code does not support this feature flag. This patch fixes
the issue by ensuring that that a value of 0 is treated as
DNODE_MIN_SLOTS.
Author: Tom Caputi <tcaputi@datto.com>
MFC after: 3 weeks
X-MFC after: r351074
8423 8199 7432 Implement large_dnode pool feature
8423 Implement large_dnode pool feature
8199 multi-threaded dmu_object_alloc()
7432 Large dnode pool feature
llumos/illumos-gate@54811da5ac54811da5achttps://www.illumos.org/issues/8423https://www.illumos.org/issues/8199https://www.illumos.org/issues/7432
ZoL issues:
Improved dnode allocation #6564
Clean up large dnode code #6262
Fix dnode_hold() freeing dnode behavior #8172
Fix dnode allocation race #6414, #6439
Partial: Raw sends must be able to decrease nlevels #6821, #6864
Remove unnecessary txg syncs from receive_object() Closes#7197
This updates FreeBSD large_dnode code (that was imported from ZoL) to a version
that was committed to illumos. It has some cleanups, improvements and fixes
comparing to what we have in FreeBSD now. I think that the most significant
update is 8199 multi-threaded dmu_object_alloc().
Obtained from: illumos
MFC after: 3 weeks
illumos/illumos-gate@892586e8a1892586e8a1https://www.illumos.org/issues/6585
In any pool without the extensible dataset feature flag already enabled,
creating a dataset with dedup set to use one of the new checksums would result
in the following panic as soon as any data was added:
panic[cpu0]/thread=ffffff0006761c40: feature_get_refcount(spa, feature,
&refcount) != 48 (0x30 != 0x30), file: ../../common/fs/zfs/zfeature.c line 390
ffffff0006761830 fffffffffba8fbdd ()
ffffff0006761890 zfs:feature_do_action+11a ()
ffffff00067618c0 zfs:spa_feature_incr+1e ()
ffffff0006761920 zfs:dmu_object_zapify+b7 ()
ffffff00067619b0 zfs:dsl_dataset_activate_feature+97 ()
ffffff0006761a20 zfs:dsl_dataset_sync+ba ()
ffffff0006761ab0 zfs:dsl_pool_sync+153 ()
ffffff0006761b70 zfs:spa_sync+26e ()
ffffff0006761c20 zfs:txg_sync_thread+227 ()
ffffff0006761c30 unix:thread_start+8 ()
Inspection showed that feature->fi_feature was 7, which is the value of
SPA_FEATURE_EXTENSIBLE_DATASET in the spa_feature enum.
Testing shows that the panic can be prevented by explicitly setting extensible
dataset as a dependency for the sha512, edonr, and skein feature flags.
Alternatively, the new checksums code could possibly be changed to obviate the
need for the dependency.
Reviewed by: Matthew Ahrens <mahrens@delphix.com>
Reviewed by: Richard Laager <rlaager@wiktel.com>
Approved by: Robert Mustacchi <rm@joyent.com>
Author: ilovezfs <ilovezfs@icloud.com>
Note that FreeBSD does not support ednor yet.
MFC after: 2 weeks
The race was introduced in r337669, the large dnode feature import from
ZoL. The problem was debugged by ZoL developers and then,
independently, on FreeBSD.
The fix is an early proposal by Brian Behlendorf:
50f32ed74e
This fix never went into ZoL. A larger change that was committed later
included a different solution because of the re-worked code.
Ideally, we want to revert this fix and re-synchronize FreeBSD large
dnode code with that in illumos (or newer ZoL). illumos has a later
import of the feature from ZoL that does not have the bug.
PR: 236480
Obtained from: Brian Behlendorf <behlendorf1@llnl.gov>
Submitted by: ncrogers@gmail.com (patch adaptation)
Reported by: ncrogers@gmail.com
Tested by: ncrogers@gmail.com,
Dennis Noordsij <dennis.noordsij@alumni.helsinki.fi>,
Julien Cigar <julien@perdition.city>
MFC after: 10 days
We should support removing vdev from boot pool. Update loader zfs reader
to support com.delphix:removing.
Reviewed by: allanjude
MFC after: 2 weeks
Differential Revision: https://reviews.freebsd.org/D18901
with an eventual goal to convert all legacl zlib callers to the new zlib
version:
* Move generic zlib shims that are not specific to zlib 1.0.4 to
sys/dev/zlib.
* Connect new zlib (1.2.11) to the zlib kernel module, currently built
with Z_SOLO.
* Prefix the legacy zlib (1.0.4) with 'zlib104_' namespace.
* Convert sys/opencrypto/cryptodeflate.c to use new zlib.
* Remove bundled zlib 1.2.3 from ZFS and adapt it to new zlib and make
it depend on the zlib module.
* Fix Z_SOLO build of new zlib.
PR: 229763
Submitted by: Yoshihiro Ota <ota j email ne jp>
Reviewed by: markm (sys/dev/zlib/zlib_kmod.c)
Relnotes: yes
Differential Revision: https://reviews.freebsd.org/D19706
This ioctl is used when a breakpoint is encountered while disassembling
a symbol in the target process. Since only one DTrace consumer can
toggle or enumerate fasttrap probes from a given process at time, this
ioctl does not appear to be used in practice.
The hold_count and wire_count fields of struct vm_page are separate
reference counters with similar semantics. The remaining essential
differences are that holds are not counted as a reference with respect
to LRU, and holds have an implicit free-on-last unhold semantic whereas
vm_page_unwire() callers must explicitly determine whether to free the
page once the last reference to the page is released.
This change removes the KPIs which directly manipulate hold_count.
Functions such as vm_fault_quick_hold_pages() now return wired pages
instead. Since r328977 the overhead of maintaining LRU for wired pages
is lower, and in many cases vm_fault_quick_hold_pages() callers would
swap holds for wirings on the returned pages anyway, so with this change
we remove a number of page lock acquisitions.
No functional change is intended. __FreeBSD_version is bumped.
Reviewed by: alc, kib
Discussed with: jeff
Discussed with: jhb, np (cxgbe)
Tested by: pho (previous version)
Sponsored by: Netflix
Differential Revision: https://reviews.freebsd.org/D19247
DMU sync code calls taskq_dispatch() for each sublist of os_dirty_dnodes
and os_synced_dnodes. Since the number of sublists by default is equal
to number of CPUs, it will dispatch equal, potentially large, number of
tasks, waking up many CPUs to handle them, even if only one or few of
sublists actually have any work to do.
This change adds check for empty sublists to avoid this.
For busy ARC situation when arc_size close to arc_c is desired. But
then it is quite likely that aggsum_compare(&arc_size, arc_c) will need
to flush per-CPU buckets to find exact comparison result. Doing that
often in a hot path penalizes whole idea of aggsum usage there, since it
replaces few simple atomic additions with dozens of lock acquisitions.
Replacing aggsum_compare() with aggsum_upper_bound() in code increasing
arc_p when ARC is growing (arc_size < arc_c) according to PMC profiles
allows to save ~5% of CPU time in aggsum code during sequential write
to 12 ZVOLs with 16KB block size on large dual-socket system.
I suppose there some minor arc_p behavior change due to lower precision
of the new code, but I don't think it is a big deal, since it should
affect only very small window in time (aggsum buckets are flushed every
second) and in ARC size (buckets are limited to 10 average ARC blocks
per CPU).
MFC after: 2 weeks
Sponsored by: iXsystems, Inc.
It is too generous to collect in production debug traces that can only
be read with kernel debugger. Illumos includes special code in their
mdb debugger to read it, we don't.
MFC after: 1 week
Sponsored by: iXsystems, Inc.
Memory copy is too heavy operation to do under the congested lock.
Moving it out reduces congestion by many times to almost invisible.
Since the original zio removed from the queue, and the child zio is
not executed yet, I don't see why would the copy need protection.
My guess it just remained like this from the time when lock was not
dropped here, which was added later to fix lock ordering issue.
Multi-threaded sequential write tests with both HDD and SSD pools
with ZVOL block sizes of 4KB, 16KB, 64KB and 128KB all show major
reduction of lock congestion, saving from 15% to 35% of CPU time
and increasing throughput from 10% to 40%.
Reviewed by: ahrens, behlendorf, ryao
MFC after: 2 weeks
Sponsored by: iXsystems, Inc.
When ARC size is very small, aggsum_lower_bound(&arc_size) may return
negative values, that due to unsigned comparison caused delays, waiting
for arc_adjust() to "fix" it by calling aggsum_value(&arc_size). Use
of signed comparison there fixes the problem.
MFC after: 2 weeks
Sponsored by: iXsystems, Inc.
While formally it is not necessary, but the sooner it start, the sooner it
finish, and supposedly less disturbing for workload it will be.
MFC after: 2 weeks
Before r305323 (MFV r302991: 6950 ARC should cache compressed data)
arc_read() code did this for access to a ghost buffer:
arc_adapt() (from arc_get_data_buf())
arc_access(hdr, hash_lock)
I.e., we first checked access to the MFU ghost/MRU ghost buffer and
adapt MFU/MRU sizes (in arc_adapt()) and next move buffer from the ghost
state to regular.
After r305323 the sequence is different:
arc_access(hdr, hash_lock);
arc_hdr_alloc_pabd(hdr);
I.e., we first move the buffer from the ghost state in arc_access() and
then we check access to buffer in ghost state (in arc_hdr_alloc_pabd()
-> arc_get_data_abd() -> arc_get_data_impl() -> arc_adapt()). This is
incorrect: arc_adapt() never see access to the ghost buffer because
arc_access() already migrated the buffer from the ghost state to
regular.
So, the fix is to restore a call to arc_adapt() before arc_access() and
to suppress the call to arc_adapt() after arc_access().
Submitted by: Slawa Olhovchenkov <slw@zxy.spb.ru>
MFC after: 2 weeks
Sponsored by: Integros [integros.com]
Differential Revision: https://reviews.freebsd.org/D19094
