| Message ID | 20240903081653.65613-1-songmuchun@bytedance.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | Fix some starvation problems in block layer | expand |
> On Sep 3, 2024, at 16:16, Muchun Song <songmuchun@bytedance.com> wrote: > > We encounter a problem on our servers where hundreds of UNINTERRUPTED > processes are all waiting in the WBT wait queue. And the IO hung detector > logged so many messages about "blocked for more than 122 seconds". The > call trace is as follows: > > Call Trace: > __schedule+0x959/0xee0 > schedule+0x40/0xb0 > io_schedule+0x12/0x40 > rq_qos_wait+0xaf/0x140 > wbt_wait+0x92/0xc0 > __rq_qos_throttle+0x20/0x30 > blk_mq_make_request+0x12a/0x5c0 > generic_make_request_nocheck+0x172/0x3f0 > submit_bio+0x42/0x1c0 > ... > > The WBT module is used to throttle buffered writeback, which will block > any buffered writeback IO request until the previous inflight IOs have > been completed. So I checked the inflight IO counter. That was one meaning > one IO request was submitted to the downstream interface like block core > layer or device driver (virtio_blk driver in our case). We need to figure > out why the inflight IO is not completed in time. I confirmed that all > the virtio ring buffers of virtio_blk are empty and the hardware dispatch > list had one IO request, so the root cause is not related to the block > device or the virtio_blk driver since the driver has never received that > IO request. > > We know that block core layer could submit IO requests to the driver through > kworker (the callback function is blk_mq_run_work_fn). I thought maybe the > kworker was blocked by some other resources causing the callback to not be > evoked in time. So I checked all the kworkers and workqueues and confirmed > there was no pending work on any kworker or workqueue. > > Integrate all the investigation information, the problem should be in the > block core layer missing a chance to submit that IO request. After > some investigation of code, I found some scenarios which could cause the > problem. Hi Jens Axboe, May I ask if you have any suggestions for those fixes? Or if they could be merged? Muchun, Thanks. > > Changes in v2: > - Collect RB tag from Ming Lei. > - Use barrier-less approach to fix QUEUE_FLAG_QUIESCED ordering problem > suggested by Ming Lei. > - Apply new approach to fix BLK_MQ_S_STOPPED ordering for easier maintenance. > - Add Fixes tag to each patch. > > Muchun Song (3): > block: fix missing dispatching request when queue is started or > unquiesced > block: fix ordering between checking QUEUE_FLAG_QUIESCED and adding > requests > block: fix ordering between checking BLK_MQ_S_STOPPED and adding > requests > > block/blk-mq.c | 55 ++++++++++++++++++++++++++++++++++++++------------ > block/blk-mq.h | 13 ++++++++++++ > 2 files changed, 55 insertions(+), 13 deletions(-) > > -- > 2.20.1 >
We encounter a problem on our servers where hundreds of UNINTERRUPTED processes are all waiting in the WBT wait queue. And the IO hung detector logged so many messages about "blocked for more than 122 seconds". The call trace is as follows: Call Trace: __schedule+0x959/0xee0 schedule+0x40/0xb0 io_schedule+0x12/0x40 rq_qos_wait+0xaf/0x140 wbt_wait+0x92/0xc0 __rq_qos_throttle+0x20/0x30 blk_mq_make_request+0x12a/0x5c0 generic_make_request_nocheck+0x172/0x3f0 submit_bio+0x42/0x1c0 ... The WBT module is used to throttle buffered writeback, which will block any buffered writeback IO request until the previous inflight IOs have been completed. So I checked the inflight IO counter. That was one meaning one IO request was submitted to the downstream interface like block core layer or device driver (virtio_blk driver in our case). We need to figure out why the inflight IO is not completed in time. I confirmed that all the virtio ring buffers of virtio_blk are empty and the hardware dispatch list had one IO request, so the root cause is not related to the block device or the virtio_blk driver since the driver has never received that IO request. We know that block core layer could submit IO requests to the driver through kworker (the callback function is blk_mq_run_work_fn). I thought maybe the kworker was blocked by some other resources causing the callback to not be evoked in time. So I checked all the kworkers and workqueues and confirmed there was no pending work on any kworker or workqueue. Integrate all the investigation information, the problem should be in the block core layer missing a chance to submit that IO request. After some investigation of code, I found some scenarios which could cause the problem. Changes in v2: - Collect RB tag from Ming Lei. - Use barrier-less approach to fix QUEUE_FLAG_QUIESCED ordering problem suggested by Ming Lei. - Apply new approach to fix BLK_MQ_S_STOPPED ordering for easier maintenance. - Add Fixes tag to each patch. Muchun Song (3): block: fix missing dispatching request when queue is started or unquiesced block: fix ordering between checking QUEUE_FLAG_QUIESCED and adding requests block: fix ordering between checking BLK_MQ_S_STOPPED and adding requests block/blk-mq.c | 55 ++++++++++++++++++++++++++++++++++++++------------ block/blk-mq.h | 13 ++++++++++++ 2 files changed, 55 insertions(+), 13 deletions(-)