OSDN Git Service

blk-wbt: Avoid lock contention and thundering herd issue in wbt_wait
authorAnchal Agarwal <anchalag@amazon.com>
Tue, 7 Aug 2018 20:40:49 +0000 (14:40 -0600)
committerJens Axboe <axboe@kernel.dk>
Tue, 7 Aug 2018 20:40:49 +0000 (14:40 -0600)
commit2887e41b910bb14fd847cf01ab7a5993db989d88
treef0141ca85946f1c38ceb5ccaec5cca64e963e845
parente33e5c85763e8ac1899ec382c1ebc4603d8c52ae
blk-wbt: Avoid lock contention and thundering herd issue in wbt_wait

I am currently running a large bare metal instance (i3.metal)
on EC2 with 72 cores, 512GB of RAM and NVME drives, with a
4.18 kernel. I have a workload that simulates a database
workload and I am running into lockup issues when writeback
throttling is enabled,with the hung task detector also
kicking in.

Crash dumps show that most CPUs (up to 50 of them) are
all trying to get the wbt wait queue lock while trying to add
themselves to it in __wbt_wait (see stack traces below).

[    0.948118] CPU: 45 PID: 0 Comm: swapper/45 Not tainted 4.14.51-62.38.amzn1.x86_64 #1
[    0.948119] Hardware name: Amazon EC2 i3.metal/Not Specified, BIOS 1.0 10/16/2017
[    0.948120] task: ffff883f7878c000 task.stack: ffffc9000c69c000
[    0.948124] RIP: 0010:native_queued_spin_lock_slowpath+0xf8/0x1a0
[    0.948125] RSP: 0018:ffff883f7fcc3dc8 EFLAGS: 00000046
[    0.948126] RAX: 0000000000000000 RBX: ffff887f7709ca68 RCX: ffff883f7fce2a00
[    0.948128] RDX: 000000000000001c RSI: 0000000000740001 RDI: ffff887f7709ca68
[    0.948129] RBP: 0000000000000002 R08: 0000000000b80000 R09: 0000000000000000
[    0.948130] R10: ffff883f7fcc3d78 R11: 000000000de27121 R12: 0000000000000002
[    0.948131] R13: 0000000000000003 R14: 0000000000000000 R15: 0000000000000000
[    0.948132] FS:  0000000000000000(0000) GS:ffff883f7fcc0000(0000) knlGS:0000000000000000
[    0.948134] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[    0.948135] CR2: 000000c424c77000 CR3: 0000000002010005 CR4: 00000000003606e0
[    0.948136] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[    0.948137] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[    0.948138] Call Trace:
[    0.948139]  <IRQ>
[    0.948142]  do_raw_spin_lock+0xad/0xc0
[    0.948145]  _raw_spin_lock_irqsave+0x44/0x4b
[    0.948149]  ? __wake_up_common_lock+0x53/0x90
[    0.948150]  __wake_up_common_lock+0x53/0x90
[    0.948155]  wbt_done+0x7b/0xa0
[    0.948158]  blk_mq_free_request+0xb7/0x110
[    0.948161]  __blk_mq_complete_request+0xcb/0x140
[    0.948166]  nvme_process_cq+0xce/0x1a0 [nvme]
[    0.948169]  nvme_irq+0x23/0x50 [nvme]
[    0.948173]  __handle_irq_event_percpu+0x46/0x300
[    0.948176]  handle_irq_event_percpu+0x20/0x50
[    0.948179]  handle_irq_event+0x34/0x60
[    0.948181]  handle_edge_irq+0x77/0x190
[    0.948185]  handle_irq+0xaf/0x120
[    0.948188]  do_IRQ+0x53/0x110
[    0.948191]  common_interrupt+0x87/0x87
[    0.948192]  </IRQ>
....
[    0.311136] CPU: 4 PID: 9737 Comm: run_linux_amd64 Not tainted 4.14.51-62.38.amzn1.x86_64 #1
[    0.311137] Hardware name: Amazon EC2 i3.metal/Not Specified, BIOS 1.0 10/16/2017
[    0.311138] task: ffff883f6e6a8000 task.stack: ffffc9000f1ec000
[    0.311141] RIP: 0010:native_queued_spin_lock_slowpath+0xf5/0x1a0
[    0.311142] RSP: 0018:ffffc9000f1efa28 EFLAGS: 00000046
[    0.311144] RAX: 0000000000000000 RBX: ffff887f7709ca68 RCX: ffff883f7f722a00
[    0.311145] RDX: 0000000000000035 RSI: 0000000000d80001 RDI: ffff887f7709ca68
[    0.311146] RBP: 0000000000000202 R08: 0000000000140000 R09: 0000000000000000
[    0.311147] R10: ffffc9000f1ef9d8 R11: 000000001a249fa0 R12: ffff887f7709ca68
[    0.311148] R13: ffffc9000f1efad0 R14: 0000000000000000 R15: ffff887f7709ca00
[    0.311149] FS:  000000c423f30090(0000) GS:ffff883f7f700000(0000) knlGS:0000000000000000
[    0.311150] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[    0.311151] CR2: 00007feefcea4000 CR3: 0000007f7016e001 CR4: 00000000003606e0
[    0.311152] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[    0.311153] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[    0.311154] Call Trace:
[    0.311157]  do_raw_spin_lock+0xad/0xc0
[    0.311160]  _raw_spin_lock_irqsave+0x44/0x4b
[    0.311162]  ? prepare_to_wait_exclusive+0x28/0xb0
[    0.311164]  prepare_to_wait_exclusive+0x28/0xb0
[    0.311167]  wbt_wait+0x127/0x330
[    0.311169]  ? finish_wait+0x80/0x80
[    0.311172]  ? generic_make_request+0xda/0x3b0
[    0.311174]  blk_mq_make_request+0xd6/0x7b0
[    0.311176]  ? blk_queue_enter+0x24/0x260
[    0.311178]  ? generic_make_request+0xda/0x3b0
[    0.311181]  generic_make_request+0x10c/0x3b0
[    0.311183]  ? submit_bio+0x5c/0x110
[    0.311185]  submit_bio+0x5c/0x110
[    0.311197]  ? __ext4_journal_stop+0x36/0xa0 [ext4]
[    0.311210]  ext4_io_submit+0x48/0x60 [ext4]
[    0.311222]  ext4_writepages+0x810/0x11f0 [ext4]
[    0.311229]  ? do_writepages+0x3c/0xd0
[    0.311239]  ? ext4_mark_inode_dirty+0x260/0x260 [ext4]
[    0.311240]  do_writepages+0x3c/0xd0
[    0.311243]  ? _raw_spin_unlock+0x24/0x30
[    0.311245]  ? wbc_attach_and_unlock_inode+0x165/0x280
[    0.311248]  ? __filemap_fdatawrite_range+0xa3/0xe0
[    0.311250]  __filemap_fdatawrite_range+0xa3/0xe0
[    0.311253]  file_write_and_wait_range+0x34/0x90
[    0.311264]  ext4_sync_file+0x151/0x500 [ext4]
[    0.311267]  do_fsync+0x38/0x60
[    0.311270]  SyS_fsync+0xc/0x10
[    0.311272]  do_syscall_64+0x6f/0x170
[    0.311274]  entry_SYSCALL_64_after_hwframe+0x42/0xb7

In the original patch, wbt_done is waking up all the exclusive
processes in the wait queue, which can cause a thundering herd
if there is a large number of writer threads in the queue. The
original intention of the code seems to be to wake up one thread
only however, it uses wake_up_all() in __wbt_done(), and then
uses the following check in __wbt_wait to have only one thread
actually get out of the wait loop:

if (waitqueue_active(&rqw->wait) &&
            rqw->wait.head.next != &wait->entry)
                return false;

The problem with this is that the wait entry in wbt_wait is
define with DEFINE_WAIT, which uses the autoremove wakeup function.
That means that the above check is invalid - the wait entry will
have been removed from the queue already by the time we hit the
check in the loop.

Secondly, auto-removing the wait entries also means that the wait
queue essentially gets reordered "randomly" (e.g. threads re-add
themselves in the order they got to run after being woken up).
Additionally, new requests entering wbt_wait might overtake requests
that were queued earlier, because the wait queue will be
(temporarily) empty after the wake_up_all, so the waitqueue_active
check will not stop them. This can cause certain threads to starve
under high load.

The fix is to leave the woken up requests in the queue and remove
them in finish_wait() once the current thread breaks out of the
wait loop in __wbt_wait. This will ensure new requests always
end up at the back of the queue, and they won't overtake requests
that are already in the wait queue. With that change, the loop
in wbt_wait is also in line with many other wait loops in the kernel.
Waking up just one thread drastically reduces lock contention, as
does moving the wait queue add/remove out of the loop.

A significant drop in lockdep's lock contention numbers is seen when
running the test application on the patched kernel.

Signed-off-by: Anchal Agarwal <anchalag@amazon.com>
Signed-off-by: Frank van der Linden <fllinden@amazon.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
block/blk-wbt.c