Re: [PATCH] blk-wbt: Fix io starvation in wbt_rqw_done()

Hi,

On Fri, Aug 1, 2025 at 1:13 AM Yu Kuai [off-list ref] wrote:

Hi,

在 2025/7/31 23:40, Yizhou Tang 写道:

quoted

Hi Julian,

On Thu, Jul 31, 2025 at 8:33 PM Julian Sun [off-list ref] wrote:

quoted

Recently, we encountered the following hungtask:

INFO: task kworker/11:2:2981147 blocked for more than 6266 seconds
"echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
kworker/11:2    D    0 2981147      2 0x80004000
Workqueue: cgroup_destroy css_free_rwork_fn
Call Trace:
  __schedule+0x934/0xe10
  schedule+0x40/0xb0
  wb_wait_for_completion+0x52/0x80

I don’t see __wbt_wait() or rq_qos_wait() here, so I suspect this call
stack is not directly related to wbt.

quoted

  ? finish_wait+0x80/0x80
  mem_cgroup_css_free+0x3a/0x1b0
  css_free_rwork_fn+0x42/0x380
  process_one_work+0x1a2/0x360
  worker_thread+0x30/0x390
  ? create_worker+0x1a0/0x1a0
  kthread+0x110/0x130
  ? __kthread_cancel_work+0x40/0x40
  ret_from_fork+0x1f/0x30

This is writeback cgroup is waiting for writeback to be done, if you
figured out
they are throttled by wbt, you need to explain clearly, and it's very
important to
provide evidence to support your analysis. However, the following
analysis is
a mess :(

Thanks for the detailed review.
Yes, the description is a bit confusing. I will take a more detailed
look at the on-site information.

quoted

quoted

This is because the writeback thread has been continuously and repeatedly
throttled by wbt, but at the same time, the writes of another thread
proceed quite smoothly.
After debugging, I believe it is caused by the following reasons.

When thread A is blocked by wbt, the I/O issued by thread B will
use a deeper queue depth(rwb->rq_depth.max_depth) because it
meets the conditions of wb_recent_wait(), thus allowing thread B's
I/O to be issued smoothly and resulting in the inflight I/O of wbt
remaining relatively high.

However, when I/O completes, due to the high inflight I/O of wbt,
the condition "limit - inflight >= rwb->wb_background / 2"
in wbt_rqw_done() cannot be satisfied, causing thread A's I/O
to remain unable to be woken up.

 From your description above, it seems you're suggesting that if A is
throttled by wbt, then a writer B on the same device could
continuously starve A.
This situation is not possible — please refer to rq_qos_wait(): if A
is already sleeping, then when B calls wq_has_sleeper(), it will
detect A’s presence, meaning B will also be throttled.

Yes, there are three rq_wait in wbt, and each one is FIFO. It will be
possible
if  A is backgroup, and B is swap.

quoted

Thanks,
Yi

quoted

Some on-site information:

quoted

quoted

quoted

rwb.rq_depth.max_depth

(unsigned int)48

quoted

quoted

quoted

rqw.inflight.counter.value_()

44

quoted

quoted

quoted

rqw.inflight.counter.value_()

35

quoted

quoted

quoted

prog['jiffies'] - rwb.rqos.q.backing_dev_info.last_bdp_sleep

(unsigned long)3

quoted

quoted

quoted

prog['jiffies'] - rwb.rqos.q.backing_dev_info.last_bdp_sleep

(unsigned long)2

quoted

quoted

quoted

prog['jiffies'] - rwb.rqos.q.backing_dev_info.last_bdp_sleep

(unsigned long)20

quoted

quoted

quoted

prog['jiffies'] - rwb.rqos.q.backing_dev_info.last_bdp_sleep

(unsigned long)12

cat wb_normal
24
cat wb_background
12

To fix this issue, we can use max_depth in wbt_rqw_done(), so that
the handling of wb_recent_wait by wbt_rqw_done() and get_limit()
will also be consistent, which is more reasonable.

Are you able to reproduce this problem, and give this patch a test before
you send it?

Thanks,
Kuai

quoted

quoted

Signed-off-by: Julian Sun <redacted>
Fixes: e34cbd307477 ("blk-wbt: add general throttling mechanism")
---
  block/blk-wbt.c | 2 ++
  1 file changed, 2 insertions(+)

diff --git a/block/blk-wbt.c b/block/blk-wbt.c
index a50d4cd55f41..d6a2782d442f 100644
--- a/block/blk-wbt.c
+++ b/block/blk-wbt.c

@@ -210,6 +210,8 @@ static void wbt_rqw_done(struct rq_wb *rwb, struct rq_wait *rqw,
         else if (blk_queue_write_cache(rwb->rqos.disk->queue) &&
                  !wb_recent_wait(rwb))
                 limit = 0;
+       else if (wb_recent_wait(rwb))
+               limit = rwb->rq_depth.max_depth;
         else
                 limit = rwb->wb_normal;

--

2.20.1

Thanks,
-- 
Julian Sun [off-list ref]