Currently if a user enqueue a work item using schedule_delayed_work() the
used wq is "system_wq" (per-cpu wq) while queue_delayed_work() use
WORK_CPU_UNBOUND (used when a cpu is not specified). The same applies to
schedule_work() that is using system_wq and queue_work(), that makes use
again of WORK_CPU_UNBOUND.
This lack of consistentcy cannot be addressed without refactoring the API.
alloc_workqueue() treats all queues as per-CPU by default, while unbound
workqueues must opt-in via WQ_UNBOUND.
This default is suboptimal: most workloads benefit from unbound queues,
allowing the scheduler to place worker threads where they’re needed and
reducing noise when CPUs are isolated.
This patch continues the effort to refactor worqueue APIs, which has begun
with the change introducing new workqueues and a new alloc_workqueue flag:
commit 128ea9f6cc ("workqueue: Add system_percpu_wq and system_dfl_wq")
commit 930c2ea566 ("workqueue: Add new WQ_PERCPU flag")
This change adds a new WQ_PERCPU flag to explicitly request
alloc_workqueue() to be per-cpu when WQ_UNBOUND has not been specified.
With the introduction of the WQ_PERCPU flag (equivalent to !WQ_UNBOUND),
any alloc_workqueue() caller that doesn’t explicitly specify WQ_UNBOUND
must now use WQ_PERCPU.
Once migration is complete, WQ_UNBOUND can be removed and unbound will
become the implicit default.
Suggested-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Marco Crivellari <marco.crivellari@suse.com>
Signed-off-by: Coly Li <colyli@fnnas.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Since discard code is removed, now the sysfs interface to enable discard
is useless. This patch removes the corresponding sysfs entry, and remove
bool variable 'discard' from struct cache as well.
Signed-off-by: Coly Li <colyli@fnnas.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Bios are embedded into other structures, and at least spare is unhappy
about embedding structures with variable sized arrays. There's no
real need to the array anyway, we can replace it with a helper pointing
to the memory just behind the bio, and with the previous cleanups there
is very few site doing anything special with it.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: John Garry <john.g.garry@oracle.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Just a simpler wrapper around bio_init for callers that want to
initialize a bio with inline bvecs.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: John Garry <john.g.garry@oracle.com>
Reviewed-by: Yu Kuai <yukuai3@huawei.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
This reverts commit 866898efbb.
The generic bottom-up min_heap implementation causes performance
regression in invalidate_buckets_lru(), a hot path in bcache. Before the
cache is fully populated, new_bucket_prio() often returns zero, leading to
many equal comparisons. In such cases, bottom-up sift_down performs up to
2 * log2(n) comparisons, while the original top-down approach completes
with just O() comparisons, resulting in a measurable performance gap.
The performance degradation is further worsened by the non-inlined
min_heap API functions introduced in commit 92a8b224b8 ("lib/min_heap:
introduce non-inline versions of min heap API functions"), adding function
call overhead to this critical path.
As reported by Robert, bcache now suffers from latency spikes, with P100
(max) latency increasing from 600 ms to 2.4 seconds every 5 minutes.
These regressions degrade bcache's effectiveness as a low-latency cache
layer and lead to frequent timeouts and application stalls in production
environments.
This revert aims to restore bcache's original low-latency behavior.
Link: https://lore.kernel.org/lkml/CAJhEC05+0S69z+3+FB2Cd0hD+pCRyWTKLEOsc8BOmH73p1m+KQ@mail.gmail.com
Link: https://lkml.kernel.org/r/20250614202353.1632957-3-visitorckw@gmail.com
Fixes: 866898efbb ("bcache: remove heap-related macros and switch to generic min_heap")
Fixes: 92a8b224b8 ("lib/min_heap: introduce non-inline versions of min heap API functions")
Signed-off-by: Kuan-Wei Chiu <visitorckw@gmail.com>
Reported-by: Robert Pang <robertpang@google.com>
Closes: https://lore.kernel.org/linux-bcache/CAJhEC06F_AtrPgw2-7CvCqZgeStgCtitbD-ryuPpXQA-JG5XXw@mail.gmail.com
Acked-by: Coly Li <colyli@kernel.org>
Cc: Ching-Chun (Jim) Huang <jserv@ccns.ncku.edu.tw>
Cc: Kent Overstreet <kent.overstreet@linux.dev>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
btree_iter is used in two ways: either allocated on the stack with a
fixed size MAX_BSETS, or from a mempool with a dynamic size based on the
specific cache set. Previously, the struct had a fixed-length array of
size MAX_BSETS which was indexed out-of-bounds for the dynamically-sized
iterators, which causes UBSAN to complain.
This patch uses the same approach as in bcachefs's sort_iter and splits
the iterator into a btree_iter with a flexible array member and a
btree_iter_stack which embeds a btree_iter as well as a fixed-length
data array.
Cc: stable@vger.kernel.org
Closes: https://bugs.launchpad.net/ubuntu/+source/linux/+bug/2039368
Signed-off-by: Matthew Mirvish <matthew@mm12.xyz>
Signed-off-by: Coly Li <colyli@suse.de>
Link: https://lore.kernel.org/r/20240509011117.2697-3-colyli@suse.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Pull more bcachefs bugfixes from Kent Overstreet:
- bcache & bcachefs were broken with CFI enabled; patch for closures to
fix type punning
- mark erasure coding as extra-experimental; there are incompatible
disk space accounting changes coming for erasure coding, and I'm
still seeing checksum errors in some tests
- several fixes for durability-related issues (durability is a device
specific setting where we can tell bcachefs that data on a given
device should be counted as replicated x times)
- a fix for a rare livelock when a btree node merge then updates a
parent node that is almost full
- fix a race in the device removal path, where dropping a pointer in a
btree node to a device would be clobbered by an in flight btree write
updating the btree node key on completion
- fix one SRCU lock hold time warning in the btree gc code - ther's
still a bunch more of these to fix
- fix a rare race where we'd start copygc before initializing the "are
we rw" percpu refcount; copygc would think we were already ro and die
immediately
* tag 'bcachefs-2023-11-29' of https://evilpiepirate.org/git/bcachefs: (23 commits)
bcachefs: Extra kthread_should_stop() calls for copygc
bcachefs: Convert gc_alloc_start() to for_each_btree_key2()
bcachefs: Fix race between btree writes and metadata drop
bcachefs: move journal seq assertion
bcachefs: -EROFS doesn't count as move_extent_start_fail
bcachefs: trace_move_extent_start_fail() now includes errcode
bcachefs: Fix split_race livelock
bcachefs: Fix bucket data type for stripe buckets
bcachefs: Add missing validation for jset_entry_data_usage
bcachefs: Fix zstd compress workspace size
bcachefs: bpos is misaligned on big endian
bcachefs: Fix ec + durability calculation
bcachefs: Data update path won't accidentaly grow replicas
bcachefs: deallocate_extra_replicas()
bcachefs: Proper refcounting for journal_keys
bcachefs: preserve device path as device name
bcachefs: Fix an endianness conversion
bcachefs: Start gc, copygc, rebalance threads after initing writes ref
bcachefs: Don't stop copygc thread on device resize
bcachefs: Make sure bch2_move_ratelimit() also waits for move_ops
...
Control flow integrity is now checking that type signatures match on
indirect function calls. That breaks closures, which embed a work_struct
in a closure in such a way that a closure_fn may also be used as a
workqueue fn by the underlying closure code.
So we have to change closure fns to take a work_struct as their
argument - but that results in a loss of clarity, as closure fns have
different semantics from normal workqueue functions (they run owning a
ref on the closure, which must be released with continue_at() or
closure_return()).
Thus, this patc introduces CLOSURE_CALLBACK() and closure_type() macros
as suggested by Kees, to smooth things over a bit.
Suggested-by: Kees Cook <keescook@chromium.org>
Cc: Coly Li <colyli@suse.de>
Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
We get a kernel crash about "unable to handle kernel paging request":
```dmesg
[368033.032005] BUG: unable to handle kernel paging request at ffffffffad9ae4b5
[368033.032007] PGD fc3a0d067 P4D fc3a0d067 PUD fc3a0e063 PMD 8000000fc38000e1
[368033.032012] Oops: 0003 [#1] SMP PTI
[368033.032015] CPU: 23 PID: 55090 Comm: bch_dirtcnt[0] Kdump: loaded Tainted: G OE --------- - - 4.18.0-147.5.1.es8_24.x86_64 #1
[368033.032017] Hardware name: Tsinghua Tongfang THTF Chaoqiang Server/072T6D, BIOS 2.4.3 01/17/2017
[368033.032027] RIP: 0010:native_queued_spin_lock_slowpath+0x183/0x1d0
[368033.032029] Code: 8b 02 48 85 c0 74 f6 48 89 c1 eb d0 c1 e9 12 83 e0
03 83 e9 01 48 c1 e0 05 48 63 c9 48 05 c0 3d 02 00 48 03 04 cd 60 68 93
ad <48> 89 10 8b 42 08 85 c0 75 09 f3 90 8b 42 08 85 c0 74 f7 48 8b 02
[368033.032031] RSP: 0018:ffffbb48852abe00 EFLAGS: 00010082
[368033.032032] RAX: ffffffffad9ae4b5 RBX: 0000000000000246 RCX: 0000000000003bf3
[368033.032033] RDX: ffff97b0ff8e3dc0 RSI: 0000000000600000 RDI: ffffbb4884743c68
[368033.032034] RBP: 0000000000000001 R08: 0000000000000000 R09: 000007ffffffffff
[368033.032035] R10: ffffbb486bb01000 R11: 0000000000000001 R12: ffffffffc068da70
[368033.032036] R13: 0000000000000003 R14: 0000000000000000 R15: 0000000000000000
[368033.032038] FS: 0000000000000000(0000) GS:ffff97b0ff8c0000(0000) knlGS:0000000000000000
[368033.032039] CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[368033.032040] CR2: ffffffffad9ae4b5 CR3: 0000000fc3a0a002 CR4: 00000000003626e0
[368033.032042] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[368033.032043] bcache: bch_cached_dev_attach() Caching rbd479 as bcache462 on set 8cff3c36-4a76-4242-afaa-7630206bc70b
[368033.032045] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[368033.032046] Call Trace:
[368033.032054] _raw_spin_lock_irqsave+0x32/0x40
[368033.032061] __wake_up_common_lock+0x63/0xc0
[368033.032073] ? bch_ptr_invalid+0x10/0x10 [bcache]
[368033.033502] bch_dirty_init_thread+0x14c/0x160 [bcache]
[368033.033511] ? read_dirty_submit+0x60/0x60 [bcache]
[368033.033516] kthread+0x112/0x130
[368033.033520] ? kthread_flush_work_fn+0x10/0x10
[368033.034505] ret_from_fork+0x35/0x40
```
The crash occurred when call wake_up(&state->wait), and then we want
to look at the value in the state. However, bch_sectors_dirty_init()
is not found in the stack of any task. Since state is allocated on
the stack, we guess that bch_sectors_dirty_init() has exited, causing
bch_dirty_init_thread() to be unable to handle kernel paging request.
In order to verify this idea, we added some printing information during
wake_up(&state->wait). We find that "wake up" is printed twice, however
we only expect the last thread to wake up once.
```dmesg
[ 994.641004] alcache: bch_dirty_init_thread() wake up
[ 994.641018] alcache: bch_dirty_init_thread() wake up
[ 994.641523] alcache: bch_sectors_dirty_init() init exit
```
There is a race. If bch_sectors_dirty_init() exits after the first wake
up, the second wake up will trigger this bug("unable to handle kernel
paging request").
Proceed as follows:
bch_sectors_dirty_init
kthread_run ==============> bch_dirty_init_thread(bch_dirtcnt[0])
... ...
atomic_inc(&state.started) ...
... ...
atomic_read(&state.enough) ...
... atomic_set(&state->enough, 1)
kthread_run ======================================================> bch_dirty_init_thread(bch_dirtcnt[1])
... atomic_dec_and_test(&state->started) ...
atomic_inc(&state.started) ... ...
... wake_up(&state->wait) ...
atomic_read(&state.enough) atomic_dec_and_test(&state->started)
... ...
wait_event(state.wait, atomic_read(&state.started) == 0) ...
return ...
wake_up(&state->wait)
We believe it is very common to wake up twice if there is no dirty, but
crash is an extremely low probability event. It's hard for us to reproduce
this issue. We attached and detached continuously for a week, with a total
of more than one million attaches and only one crash.
Putting atomic_inc(&state.started) before kthread_run() can avoid waking
up twice.
Fixes: b144e45fc5 ("bcache: make bch_sectors_dirty_init() to be multithreaded")
Signed-off-by: Mingzhe Zou <mingzhe.zou@easystack.cn>
Cc: <stable@vger.kernel.org>
Signed-off-by: Coly Li <colyli@suse.de>
Link: https://lore.kernel.org/r/20231120052503.6122-8-colyli@suse.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We found that after long run, the dirty_data of the bcache device
will have errors. This error cannot be eliminated unless re-register.
We also found that reattach after detach, this error can accumulate.
In bch_sectors_dirty_init(), all inode <= d->id keys will be recounted
again. This is wrong, we only need to count the keys of the current
device.
Fixes: b144e45fc5 ("bcache: make bch_sectors_dirty_init() to be multithreaded")
Signed-off-by: Mingzhe Zou <mingzhe.zou@easystack.cn>
Cc: <stable@vger.kernel.org>
Signed-off-by: Coly Li <colyli@suse.de>
Link: https://lore.kernel.org/r/20231120052503.6122-6-colyli@suse.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Variable cur_idx is being initialized with a value that is never read,
it is being re-assigned later in a while-loop. Remove the redundant
assignment. Cleans up clang scan build warning:
drivers/md/bcache/writeback.c:916:2: warning: Value stored to 'cur_idx'
is never read [deadcode.DeadStores]
Signed-off-by: Colin Ian King <colin.i.king@gmail.com>
Reviewed-by: Coly Li <colyli@suse.de>
Signed-off-by: Coly Li <colyli@suse.de>
Link: https://lore.kernel.org/r/20231120052503.6122-4-colyli@suse.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
We get a kernel crash about "list_add corruption. next->prev should be
prev (ffff9c801bc01210), but was ffff9c77b688237c.
(next=ffffae586d8afe68)."
crash> struct list_head 0xffff9c801bc01210
struct list_head {
next = 0xffffae586d8afe68,
prev = 0xffffae586d8afe68
}
crash> struct list_head 0xffff9c77b688237c
struct list_head {
next = 0x0,
prev = 0x0
}
crash> struct list_head 0xffffae586d8afe68
struct list_head struct: invalid kernel virtual address: ffffae586d8afe68 type: "gdb_readmem_callback"
Cannot access memory at address 0xffffae586d8afe68
[230469.019492] Call Trace:
[230469.032041] prepare_to_wait+0x8a/0xb0
[230469.044363] ? bch_btree_keys_free+0x6c/0xc0 [escache]
[230469.056533] mca_cannibalize_lock+0x72/0x90 [escache]
[230469.068788] mca_alloc+0x2ae/0x450 [escache]
[230469.080790] bch_btree_node_get+0x136/0x2d0 [escache]
[230469.092681] bch_btree_check_thread+0x1e1/0x260 [escache]
[230469.104382] ? finish_wait+0x80/0x80
[230469.115884] ? bch_btree_check_recurse+0x1a0/0x1a0 [escache]
[230469.127259] kthread+0x112/0x130
[230469.138448] ? kthread_flush_work_fn+0x10/0x10
[230469.149477] ret_from_fork+0x35/0x40
bch_btree_check_thread() and bch_dirty_init_thread() may call
mca_cannibalize() to cannibalize other cached btree nodes. Only one thread
can do it at a time, so the op of other threads will be added to the
btree_cache_wait list.
We must call finish_wait() to remove op from btree_cache_wait before free
it's memory address. Otherwise, the list will be damaged. Also should call
bch_cannibalize_unlock() to release the btree_cache_alloc_lock and wake_up
other waiters.
Fixes: 8e7102273f ("bcache: make bch_btree_check() to be multithreaded")
Fixes: b144e45fc5 ("bcache: make bch_sectors_dirty_init() to be multithreaded")
Cc: stable@vger.kernel.org
Signed-off-by: Mingzhe Zou <mingzhe.zou@easystack.cn>
Signed-off-by: Coly Li <colyli@suse.de>
Link: https://lore.kernel.org/r/20230615121223.22502-7-colyli@suse.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Inside set_at_max_writeback_rate() the calculation in following if()
check is wrong,
if (atomic_inc_return(&c->idle_counter) <
atomic_read(&c->attached_dev_nr) * 6)
Because each attached backing device has its own writeback thread
running and increasing c->idle_counter, the counter increates much
faster than expected. The correct calculation should be,
(counter / dev_nr) < dev_nr * 6
which equals to,
counter < dev_nr * dev_nr * 6
This patch fixes the above mistake with correct calculation, and helper
routine idle_counter_exceeded() is added to make code be more clear.
Reported-by: Mingzhe Zou <mingzhe.zou@easystack.cn>
Signed-off-by: Coly Li <colyli@suse.de>
Acked-by: Mingzhe Zou <mingzhe.zou@easystack.cn>
Link: https://lore.kernel.org/r/20220919161647.81238-6-colyli@suse.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
The kworker routine update_writeback_rate() is schedued to update the
writeback rate in every 5 seconds by default. Before calling
__update_writeback_rate() to do real job, semaphore dc->writeback_lock
should be held by the kworker routine.
At the same time, bcache writeback thread routine bch_writeback_thread()
also needs to hold dc->writeback_lock before flushing dirty data back
into the backing device. If the dirty data set is large, it might be
very long time for bch_writeback_thread() to scan all dirty buckets and
releases dc->writeback_lock. In such case update_writeback_rate() can be
starved for long enough time so that kernel reports a soft lockup warn-
ing started like:
watchdog: BUG: soft lockup - CPU#246 stuck for 23s! [kworker/246:31:179713]
Such soft lockup condition is unnecessary, because after the writeback
thread finishes its job and releases dc->writeback_lock, the kworker
update_writeback_rate() may continue to work and everything is fine
indeed.
This patch avoids the unnecessary soft lockup by the following method,
- Add new member to struct cached_dev
- dc->rate_update_retry (0 by default)
- In update_writeback_rate() call down_read_trylock(&dc->writeback_lock)
firstly, if it fails then lock contention happens.
- If dc->rate_update_retry <= BCH_WBRATE_UPDATE_MAX_SKIPS (15), doesn't
acquire the lock and reschedules the kworker for next try.
- If dc->rate_update_retry > BCH_WBRATE_UPDATE_MAX_SKIPS, no retry
anymore and call down_read(&dc->writeback_lock) to wait for the lock.
By the above method, at worst case update_writeback_rate() may retry for
1+ minutes before blocking on dc->writeback_lock by calling down_read().
For a 4TB cache device with 1TB dirty data, 90%+ of the unnecessary soft
lockup warning message can be avoided.
When retrying to acquire dc->writeback_lock in update_writeback_rate(),
of course the writeback rate cannot be updated. It is fair, because when
the kworker is blocked on the lock contention of dc->writeback_lock, the
writeback rate cannot be updated neither.
This change follows Jens Axboe's suggestion to a more clear and simple
version.
Signed-off-by: Coly Li <colyli@suse.de>
Link: https://lore.kernel.org/r/20220528124550.32834-2-colyli@suse.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
After making bch_sectors_dirty_init() being multithreaded, the existing
incremental dirty sector counting in bch_root_node_dirty_init() doesn't
release btree occupation after iterating 500000 (INIT_KEYS_EACH_TIME)
bkeys. Because a read lock is added on btree root node to prevent the
btree to be split during the dirty sectors counting, other I/O requester
has no chance to gain the write lock even restart bcache_btree().
That is to say, the incremental dirty sectors counting is incompatible
to the multhreaded bch_sectors_dirty_init(). We have to choose one and
drop another one.
In my testing, with 512 bytes random writes, I generate 1.2T dirty data
and a btree with 400K nodes. With single thread and incremental dirty
sectors counting, it takes 30+ minites to register the backing device.
And with multithreaded dirty sectors counting, the backing device
registration can be accomplished within 2 minutes.
The 30+ minutes V.S. 2- minutes difference makes me decide to keep
multithreaded bch_sectors_dirty_init() and drop the incremental dirty
sectors counting. This is what this patch does.
But INIT_KEYS_EACH_TIME is kept, in sectors_dirty_init_fn() the CPU
will be released by cond_resched() after every INIT_KEYS_EACH_TIME keys
iterated. This is to avoid the watchdog reports a bogus soft lockup
warning.
Fixes: b144e45fc5 ("bcache: make bch_sectors_dirty_init() to be multithreaded")
Signed-off-by: Coly Li <colyli@suse.de>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/20220524102336.10684-4-colyli@suse.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Commit b144e45fc5 ("bcache: make bch_sectors_dirty_init() to be
multithreaded") makes bch_sectors_dirty_init() to be much faster
when counting dirty sectors by iterating all dirty keys in the btree.
But it isn't in ideal shape yet, still can be improved.
This patch does the following changes to improve current parallel dirty
keys iteration on the btree,
- Add read lock to root node when multiple threads iterating the btree,
to prevent the root node gets split by I/Os from other registered
bcache devices.
- Remove local variable "char name[32]" and generate kernel thread name
string directly when calling kthread_run().
- Allocate "struct bch_dirty_init_state state" directly on stack and
avoid the unnecessary dynamic memory allocation for it.
- Decrease BCH_DIRTY_INIT_THRD_MAX from 64 to 12 which is enough indeed.
- Increase &state->started to count created kernel thread after it
succeeds to create.
- When wait for all dirty key counting threads to finish, use
wait_event() to replace wait_event_interruptible().
With the above changes, the code is more clear, and some potential error
conditions are avoided.
Fixes: b144e45fc5 ("bcache: make bch_sectors_dirty_init() to be multithreaded")
Signed-off-by: Coly Li <colyli@suse.de>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/20220524102336.10684-3-colyli@suse.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
When multiple threads to check btree nodes in parallel, the main
thread wait for all threads to stop or CACHE_SET_IO_DISABLE flag:
wait_event_interruptible(check_state->wait,
atomic_read(&check_state->started) == 0 ||
test_bit(CACHE_SET_IO_DISABLE, &c->flags));
However, the bch_btree_node_read and bch_btree_node_read_done
maybe call bch_cache_set_error, then the CACHE_SET_IO_DISABLE
will be set. If the flag already set, the main thread return
error. At the same time, maybe some threads still running and
read NULL pointer, the kernel will crash.
This patch change the event wait condition, the main thread must
wait for all threads to stop.
Fixes: 8e7102273f ("bcache: make bch_btree_check() to be multithreaded")
Signed-off-by: Mingzhe Zou <mingzhe.zou@easystack.cn>
Cc: stable@vger.kernel.org # v5.7+
Signed-off-by: Coly Li <colyli@suse.de>
When attaching a cached device (a.k.a backing device) to a cache
device, bch_sectors_dirty_init() is called to count dirty sectors
and stripes (see what bcache_dev_sectors_dirty_add() does) on the
cache device.
When bcache_dev_sectors_dirty_add() is called, set_bit(stripe,
d->full_dirty_stripes) or clear_bit(stripe, d->full_dirty_stripes)
operation will always be performed. In full_dirty_stripes, each 1bit
represents stripe_size (8192) sectors (512B), so 1bit=4MB (8192*512),
and each CPU cache line=64B=512bit=2048MB. When 20 threads process
a cached disk with 100G dirty data, a single thread processes about
23M at a time, and 20 threads total 460M. These full_dirty_stripes
bits corresponding to the 460M data is likely to fall in the same CPU
cache line. When one of these threads performs a set_bit or clear_bit
operation, the same CPU cache line of other threads will become invalid
and must read the full_dirty_stripes from the main memory again. Compared
with single thread, the time of a bcache_dev_sectors_dirty_add()
call is increased by about 50 times in our test (100G dirty data,
20 threads, bcache_dev_sectors_dirty_add() is called more than
20 million times).
This patch tries to test_bit before set_bit or clear_bit operation.
Therefore, a lot of force set and clear operations will be avoided,
and most of bcache_dev_sectors_dirty_add() calls will only read CPU
cache line.
Signed-off-by: Mingzhe Zou <mingzhe.zou@easystack.cn>
Signed-off-by: Coly Li <colyli@suse.de>
Pass the block_device that we plan to use this bio for and the
operation to bio_init to optimize the assignment. A NULL block_device
can be passed, both for the passthrough case on a raw request_queue and
to temporarily avoid refactoring some nasty code.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Chaitanya Kulkarni <kch@nvidia.com>
Link: https://lore.kernel.org/r/20220124091107.642561-19-hch@lst.de
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Marco CrivellariandJens Axboe
Coly LiandJens Axboe
Christoph HellwigandJens Axboe
Kuan-Wei ChiuandAndrew Morton
John GarryandJens Axboe
Matthew MirvishandJens Axboe
Linus Torvalds
Kent Overstreet
Mingzhe ZouandJens Axboe
Colin Ian KingandJens Axboe
Coly LiandJens Axboe
Li LeiandJens Axboe