commit 528b16bfc3ae5f11638e71b3b63a81f9999df727 upstream.
On systems with many cores using dm-crypt, heavy spinlock contention in
percpu_counter_compare() can be observed when the page allocation limit
for a given device is reached or close to be reached. This is due
to percpu_counter_compare() taking a spinlock to compute an exact
result on potentially many CPUs at the same time.
Switch to non-exact comparison of allocated and allowed pages by using
the value returned by percpu_counter_read_positive() to avoid taking
the percpu_counter spinlock.
This may over/under estimate the actual number of allocated pages by at
most (batch-1) * num_online_cpus().
Currently, batch is bounded by 32. The system on which this issue was
first observed has 256 CPUs and 512GB of RAM. With a 4k page size, this
change may over/under estimate by 31MB. With ~10G (2%) allowed dm-crypt
allocations, this seems an acceptable error. Certainly preferred over
running into the spinlock contention.
This behavior was reproduced on an EC2 c5.24xlarge instance with 96 CPUs
and 192GB RAM as follows, but can be provoked on systems with less CPUs
as well.
* Disable swap
* Tune vm settings to promote regular writeback
$ echo 50 > /proc/sys/vm/dirty_expire_centisecs
$ echo 25 > /proc/sys/vm/dirty_writeback_centisecs
$ echo $((128 * 1024 * 1024)) > /proc/sys/vm/dirty_background_bytes
* Create 8 dmcrypt devices based on files on a tmpfs
* Create and mount an ext4 filesystem on each crypt devices
* Run stress-ng --hdd 8 within one of above filesystems
Total %system usage collected from sysstat goes to ~35%. Write throughput
on the underlying loop device is ~2GB/s. perf profiling an individual
kworker kcryptd thread shows the following profile, indicating spinlock
contention in percpu_counter_compare():
99.98% 0.00% kworker/u193:46 [kernel.kallsyms] [k] ret_from_fork
|
--ret_from_fork
kthread
worker_thread
|
--99.92%--process_one_work
|
|--80.52%--kcryptd_crypt
| |
| |--62.58%--mempool_alloc
| | |
| | --62.24%--crypt_page_alloc
| | |
| | --61.51%--__percpu_counter_compare
| | |
| | --61.34%--__percpu_counter_sum
| | |
| | |--58.68%--_raw_spin_lock_irqsave
| | | |
| | | --58.30%--native_queued_spin_lock_slowpath
| | |
| | --0.69%--cpumask_next
| | |
| | --0.51%--_find_next_bit
| |
| |--10.61%--crypt_convert
| | |
| | |--6.05%--xts_crypt
...
After applying this patch and running the same test, %system usage is
lowered to ~7% and write throughput on the loop device increases
to ~2.7GB/s. perf report shows mempool_alloc() as ~8% rather than ~62%
in the profile and not hitting the percpu_counter() spinlock anymore.
|--8.15%--mempool_alloc
| |
| |--3.93%--crypt_page_alloc
| | |
| | --3.75%--__alloc_pages
| | |
| | --3.62%--get_page_from_freelist
| | |
| | --3.22%--rmqueue_bulk
| | |
| | --2.59%--_raw_spin_lock
| | |
| | --2.57%--native_queued_spin_lock_slowpath
| |
| --3.05%--_raw_spin_lock_irqsave
| |
| --2.49%--native_queued_spin_lock_slowpath
Suggested-by: DJ Gregor <dj@corelight.com>
Reviewed-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Arne Welzel <arne.welzel@corelight.com>
Fixes: 5059353df86e ("dm crypt: limit the number of allocated pages")
Cc: stable@vger.kernel.org
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 46d4703b1db4c86ab5acb2331b10df999f005e8e upstream.
We are seeing the following warning in raid10_handle_discard.
[ 695.110751] =============================
[ 695.131439] WARNING: suspicious RCU usage
[ 695.151389] 4.18.0-319.el8.x86_64+debug #1 Not tainted
[ 695.174413] -----------------------------
[ 695.192603] drivers/md/raid10.c:1776 suspicious
rcu_dereference_check() usage!
[ 695.225107] other info that might help us debug this:
[ 695.260940] rcu_scheduler_active = 2, debug_locks = 1
[ 695.290157] no locks held by mkfs.xfs/10186.
In the first loop of function raid10_handle_discard. It already
determines which disk need to handle discard request and add the
rdev reference count rdev->nr_pending. So the conf->mirrors will
not change until all bios come back from underlayer disks. It
doesn't need to use rcu_dereference to get rdev.
Cc: stable@vger.kernel.org
Fixes: d30588b2731f ('md/raid10: improve raid10 discard request')
Signed-off-by: Xiao Ni <xni@redhat.com>
Acked-by: Guoqing Jiang <guoqing.jiang@linux.dev>
Signed-off-by: Song Liu <songliubraving@fb.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
commit 6607cd319b6b91bff94e90f798a61c031650b514 upstream. We can't split write behind bio with more than BIO_MAX_VECS sectors, otherwise the below call trace was triggered because we could allocate oversized write behind bio later. [ 8.097936] bvec_alloc+0x90/0xc0 [ 8.098934] bio_alloc_bioset+0x1b3/0x260 [ 8.099959] raid1_make_request+0x9ce/0xc50 [raid1] [ 8.100988] ? __bio_clone_fast+0xa8/0xe0 [ 8.102008] md_handle_request+0x158/0x1d0 [md_mod] [ 8.103050] md_submit_bio+0xcd/0x110 [md_mod] [ 8.104084] submit_bio_noacct+0x139/0x530 [ 8.105127] submit_bio+0x78/0x1d0 [ 8.106163] ext4_io_submit+0x48/0x60 [ext4] [ 8.107242] ext4_writepages+0x652/0x1170 [ext4] [ 8.108300] ? do_writepages+0x41/0x100 [ 8.109338] ? __ext4_mark_inode_dirty+0x240/0x240 [ext4] [ 8.110406] do_writepages+0x41/0x100 [ 8.111450] __filemap_fdatawrite_range+0xc5/0x100 [ 8.112513] file_write_and_wait_range+0x61/0xb0 [ 8.113564] ext4_sync_file+0x73/0x370 [ext4] [ 8.114607] __x64_sys_fsync+0x33/0x60 [ 8.115635] do_syscall_64+0x33/0x40 [ 8.116670] entry_SYSCALL_64_after_hwframe+0x44/0xae Thanks for the comment from Christoph. [1]. https://bugs.archlinux.org/task/70992 Cc: stable@vger.kernel.org # v5.12+ Reported-by: Jens Stutte <jens@chianterastutte.eu> Tested-by: Jens Stutte <jens@chianterastutte.eu> Reviewed-by: Christoph Hellwig <hch@lst.de> Signed-off-by: Guoqing Jiang <jiangguoqing@kylinos.cn> Signed-off-by: Song Liu <songliubraving@fb.com> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
[ Upstream commit 224b0683228c5f332f9cee615d85e75e9a347170 ] Except for the IDA none of the allocations in bcache_device_init is unwound on error, fix that. Signed-off-by: Christoph Hellwig <hch@lst.de> Acked-by: Coly Li <colyli@suse.de> Link: https://lore.kernel.org/r/20210809064028.1198327-7-hch@lst.de Signed-off-by: Jens Axboe <axboe@kernel.dk> Signed-off-by: Sasha Levin <sashal@kernel.org>
Pull block fixes from Jens Axboe: "A few minor fixes: - Fix ldm kernel-doc warning (Bart) - Fix adding offset twice for DMA address in n64cart (Christoph) - Fix use-after-free in dasd path handling (Stefan) - Order kyber insert trace correctly (Vincent) - raid1 errored write handling fix (Wei) - Fix blk-iolatency queue get failure handling (Yu)" * tag 'block-5.14-2021-08-07' of git://git.kernel.dk/linux-block: kyber: make trace_block_rq call consistent with documentation block/partitions/ldm.c: Fix a kernel-doc warning blk-iolatency: error out if blk_get_queue() failed in iolatency_set_limit() n64cart: fix the dma address in n64cart_do_bvec s390/dasd: fix use after free in dasd path handling md/raid10: properly indicate failure when ending a failed write request
* 'md-fixes' of https://git.kernel.org/pub/scm/linux/kernel/git/song/md: md/raid10: properly indicate failure when ending a failed write request
Similar to [1], this patch fixes the same bug in raid10. Also cleanup the
comments.
[1] commit 2417b9869b81 ("md/raid1: properly indicate failure when ending
a failed write request")
Cc: stable@vger.kernel.org
Fixes: 7cee6d4e6035 ("md/raid10: end bio when the device faulty")
Signed-off-by: Wei Shuyu <wsy@dogben.com>
Acked-by: Guoqing Jiang <jiangguoqing@kylinos.cn>
Signed-off-by: Song Liu <song@kernel.org>
Pull device mapper updates from Mike Snitzer: - Various DM persistent-data library improvements and fixes that benefit both the DM thinp and cache targets. - A few small DM kcopyd efficiency improvements. - Significant zoned related block core, DM core and DM zoned target changes that culminate with adding zoned append emulation (which is required to properly fix DM crypt's zoned support). - Various DM writecache target changes that improve efficiency. Adds an optional "metadata_only" feature that only promotes bios flagged with REQ_META. But the most significant improvement is writecache's ability to pause writeback, for a confiurable time, if/when the working set is larger than the cache (and the cache is full) -- this ensures performance is no worse than the slower origin device. * tag 'for-5.14/dm-changes' of git://git.kernel.org/pub/scm/linux/kernel/git/device-mapper/linux-dm: (35 commits) dm writecache: make writeback pause configurable dm writecache: pause writeback if cache full and origin being written directly dm io tracker: factor out IO tracker dm btree remove: assign new_root only when removal succeeds dm zone: fix dm_revalidate_zones() memory allocation dm ps io affinity: remove redundant continue statement dm writecache: add optional "metadata_only" parameter dm writecache: add "cleaner" and "max_age" to Documentation dm writecache: write at least 4k when committing dm writecache: flush origin device when writing and cache is full dm writecache: have ssd writeback wait if the kcopyd workqueue is busy dm writecache: use list_move instead of list_del/list_add in writecache_writeback() dm writecache: commit just one block, not a full page dm writecache: remove unused gfp_t argument from wc_add_block() dm crypt: Fix zoned block device support dm: introduce zone append emulation dm: rearrange core declarations for extended use from dm-zone.c block: introduce BIO_ZONE_WRITE_LOCKED bio flag block: introduce bio zone helpers block: improve handling of all zones reset operation ...
Pull block driver updates from Jens Axboe:
"Pretty calm round, mostly just NVMe and a bit of MD:
- NVMe updates (via Christoph)
- improve the APST configuration algorithm (Alexey Bogoslavsky)
- look for StorageD3Enable on companion ACPI device
(Mario Limonciello)
- allow selecting the network interface for TCP connections
(Martin Belanger)
- misc cleanups (Amit Engel, Chaitanya Kulkarni, Colin Ian King,
Christoph)
- move the ACPI StorageD3 code to drivers/acpi/ and add quirks
for certain AMD CPUs (Mario Limonciello)
- zoned device support for nvmet (Chaitanya Kulkarni)
- fix the rules for changing the serial number in nvmet
(Noam Gottlieb)
- various small fixes and cleanups (Dan Carpenter, JK Kim,
Chaitanya Kulkarni, Hannes Reinecke, Wesley Sheng, Geert
Uytterhoeven, Daniel Wagner)
- MD updates (Via Song)
- iostats rewrite (Guoqing Jiang)
- raid5 lock contention optimization (Gal Ofri)
- Fall through warning fix (Gustavo)
- Misc fixes (Gustavo, Jiapeng)"
* tag 'for-5.14/drivers-2021-06-29' of git://git.kernel.dk/linux-block: (78 commits)
nvmet: use NVMET_MAX_NAMESPACES to set nn value
loop: Fix missing discard support when using LOOP_CONFIGURE
nvme.h: add missing nvme_lba_range_type endianness annotations
nvme: remove zeroout memset call for struct
nvme-pci: remove zeroout memset call for struct
nvmet: remove zeroout memset call for struct
nvmet: add ZBD over ZNS backend support
nvmet: add Command Set Identifier support
nvmet: add nvmet_req_bio put helper for backends
nvmet: add req cns error complete helper
block: export blk_next_bio()
nvmet: remove local variable
nvmet: use nvme status value directly
nvmet: use u32 type for the local variable nsid
nvmet: use u32 for nvmet_subsys max_nsid
nvmet: use req->cmd directly in file-ns fast path
nvmet: use req->cmd directly in bdev-ns fast path
nvmet: make ver stable once connection established
nvmet: allow mn change if subsys not discovered
nvmet: make sn stable once connection was established
...
Pull core block updates from Jens Axboe: - disk events cleanup (Christoph) - gendisk and request queue allocation simplifications (Christoph) - bdev_disk_changed cleanups (Christoph) - IO priority improvements (Bart) - Chained bio completion trace fix (Edward) - blk-wbt fixes (Jan) - blk-wbt enable/disable fix (Zhang) - Scheduler dispatch improvements (Jan, Ming) - Shared tagset scheduler improvements (John) - BFQ updates (Paolo, Luca, Pietro) - BFQ lock inversion fix (Jan) - Documentation improvements (Kir) - CLONE_IO block cgroup fix (Tejun) - Remove of ancient and deprecated block dump feature (zhangyi) - Discard merge fix (Ming) - Misc fixes or followup fixes (Colin, Damien, Dan, Long, Max, Thomas, Yang) * tag 'for-5.14/block-2021-06-29' of git://git.kernel.dk/linux-block: (129 commits) block: fix discard request merge block/mq-deadline: Remove a WARN_ON_ONCE() call blk-mq: update hctx->dispatch_busy in case of real scheduler blk: Fix lock inversion between ioc lock and bfqd lock bfq: Remove merged request already in bfq_requests_merged() block: pass a gendisk to bdev_disk_changed block: move bdev_disk_changed block: add the events* attributes to disk_attrs block: move the disk events code to a separate file block: fix trace completion for chained bio block/partitions/msdos: Fix typo inidicator -> indicator block, bfq: reset waker pointer with shared queues block, bfq: check waker only for queues with no in-flight I/O block, bfq: avoid delayed merge of async queues block, bfq: boost throughput by extending queue-merging times block, bfq: consider also creation time in delayed stable merge block, bfq: fix delayed stable merge check block, bfq: let also stably merged queues enjoy weight raising blk-wbt: make sure throttle is enabled properly blk-wbt: introduce a new disable state to prevent false positive by rwb_enabled() ...