Re: [PATCH] blk-wbt: Avoid lock contention and thundering herd issue in wbt_wait

From: Jens Axboe
Date: Wed Aug 01 2018 - 18:10:12 EST


On 8/1/18 11:06 AM, Anchal Agarwal wrote:
> On Wed, Aug 01, 2018 at 09:14:50AM -0600, Jens Axboe wrote:
>> On 7/31/18 3:34 PM, Anchal Agarwal wrote:
>>> Hi folks,
>>>
>>> This patch modifies commit e34cbd307477a
>>> (blk-wbt: add general throttling mechanism)
>>>
>>> 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.
>>
>> I like the patch, and a few weeks ago we independently discovered that
>> the waitqueue list checking was bogus as well. My only worry is that
>> changes like this can be delicate, meaning that it's easy to introduce
>> stall conditions. What kind of testing did you push this through?
>>
>> --
>> Jens Axboe
>>
> I ran the following tests on both real HW with NVME devices attached
> and emulated NVME too:
>
> 1. The test case I used to reproduce the issue, spawns a bunch of threads
> to concurrently read and write files with random size and content.
> Files are randomly fsync'd. The implementation is a FIFO queue of files.
> When the queue fills the test starts to verify and remove the files. This
> test will fail if there's a read, write, or hash check failure. It tests
> for file corruption when lots of small files are being read and written
> with high concurrency.
>
> 2. Fio for random writes with a root NVME device of 200GB
>
> fio --name=randwrite --ioengine=libaio --iodepth=1 --rw=randwrite --bs=4k
> --direct=0 --size=10G --numjobs=2 --runtime=60 --group_reporting
>
> fio --name=randwrite --ioengine=libaio --iodepth=1 --rw=randwrite --bs=4k
> --direct=0 --size=5G --numjobs=2 --runtime=30 --fsync=64 --group_reporting
>
> I did see an improvement in the bandwidth numbers reported on the patched
> kernel.
>
> Do you have any test case/suite in mind that you would suggest me to
> run to be sure that patch does not introduce any stall conditions?

One thing that is always useful is to run xfstest, do a full run on
the device. If that works, then do another full run, this time limiting
the queue depth of the SCSI device to 1. If both of those pass, then
I'd feel pretty good getting this applied for 4.19.

--
Jens Axboe