Skip to contentSkip to conversation

hung_task: fix missing hung task detection for kthread in TASK_WAKEKILL state

PATCH In review Yafang Shao 1 participant 2 messages Dec 23, 2024

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 hung_task: fix missing hung task detection for kthread in TASK_WAKEKILL state

We recently encountered an XFS deadlock issue, which is a known problem resolved in the upstream kernel [0]. During the analysis of this issue, I observed that a kernel thread in the TASK_WAKEKILL state could not be detected as a hung task by the hung_task detector. The details are as follows:

Using the following command, I identified nine tasks stuck in the D state:

$ ps -eLo state,comm,tid,wchan  | grep ^D
D java            4177339 xfs_buf_lock
D kworker/93:3+xf 3025535 xfs_buf_lock
D kworker/87:0+xf 3426612 xfs_extent_busy_flush
D kworker/85:0+xf 3479378 xfs_buf_lock
D kworker/91:1+xf 3584478 xfs_buf_lock
D kworker/80:3+xf 3655680 xfs_buf_lock
D kworker/89:0+xf 3671691 xfs_buf_lock
D kworker/84:1+xf 3708397 xfs_buf_lock
D kworker/81:1+xf 4005763 xfs_buf_lock

However, the hung_task detector only reported eight of these tasks:

[3108840.650652] INFO: task java:4177339 blocked for more than 247779 seconds.
[3108840.654197] INFO: task kworker/93:3:3025535 blocked for more than 248427 seconds.
[3108840.657711] INFO: task kworker/85:0:3479378 blocked for more than 247836 seconds.
[3108840.661483] INFO: task kworker/91:1:3584478 blocked for more than 249638 seconds.
[3108840.664871] INFO: task kworker/80:3:3655680 blocked for more than 249638 seconds.
[3108840.668495] INFO: task kworker/89:0:3671691 blocked for more than 249047 seconds.
[3108840.672418] INFO: task kworker/84:1:3708397 blocked for more than 247836 seconds.
[3108840.676175] INFO: task kworker/81:1:4005763 blocked for more than 247836 seconds.

Task 3426612, although in the D state, was not reported as a hung task.

I confirmed that task 3426612 remained in the D (disk sleep) state and experienced no context switches over a long period:

$ cat /proc/3426612/status | grep -E "State:|ctxt_switches:";   \
  sleep 60; echo "----"; \
  cat /proc/3426612/status | grep -E "State:|ctxt_switches:"
State:  D (disk sleep)
voluntary_ctxt_switches:        7516
nonvoluntary_ctxt_switches:     0
----
State:  D (disk sleep)
voluntary_ctxt_switches:        7516
nonvoluntary_ctxt_switches:     0

The system's hung_task detector settings were configured as follows:

kernel.hung_task_timeout_secs = 28
kernel.hung_task_warnings = -1

The issue lies in the handling of task state in the XFS code. Specifically, the thread in question (3426612) was set to the TASK_KILLABLE state in xfs_extent_busy_flush():

  xfs_extent_busy_flush
    prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE);

When a task is in the TASK_WAKEKILL state (a subset of TASK_KILLABLE), the hung_task detector ignores it, as it assumes such tasks can be terminated. However, in this case, the kernel thread cannot be killed, meaning it effectively becomes a hung task.

To address this issue, the hung_task detector should report the kthreads in the TASK_WAKEKILL state.

---
 kernel/hung_task.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
kernel/hung_task.c +2 -1
index c18717189f32..ed63fd84ce2e 100644
@@ -220,8 +220,9 @@ static void check_hung_uninterruptible_tasks(unsigned long timeout)
 		 */
 		state = READ_ONCE(t->__state);
 		if ((state & TASK_UNINTERRUPTIBLE) &&
+		    (t->flags & PF_KTHREAD ||
 		    !(state & TASK_WAKEKILL) &&
-		    !(state & TASK_NOLOAD))
+		    !(state & TASK_NOLOAD)))
 			check_hung_task(t, timeout);
 	}
  unlock:
Signature · 1 line
2.43.5
#1 lore raw
Quoted 90 lines · Yafang Shao

We recently encountered an XFS deadlock issue, which is a known problem resolved in the upstream kernel [0]. During the analysis of this issue, I observed that a kernel thread in the TASK_WAKEKILL state could not be detected as a hung task by the hung_task detector. The details are as follows:

Using the following command, I identified nine tasks stuck in the D state:

$ ps -eLo state,comm,tid,wchan  | grep ^D
D java            4177339 xfs_buf_lock
D kworker/93:3+xf 3025535 xfs_buf_lock
D kworker/87:0+xf 3426612 xfs_extent_busy_flush
D kworker/85:0+xf 3479378 xfs_buf_lock
D kworker/91:1+xf 3584478 xfs_buf_lock
D kworker/80:3+xf 3655680 xfs_buf_lock
D kworker/89:0+xf 3671691 xfs_buf_lock
D kworker/84:1+xf 3708397 xfs_buf_lock
D kworker/81:1+xf 4005763 xfs_buf_lock

However, the hung_task detector only reported eight of these tasks:

[3108840.650652] INFO: task java:4177339 blocked for more than 247779 seconds.
[3108840.654197] INFO: task kworker/93:3:3025535 blocked for more than 248427 seconds.
[3108840.657711] INFO: task kworker/85:0:3479378 blocked for more than 247836 seconds.
[3108840.661483] INFO: task kworker/91:1:3584478 blocked for more than 249638 seconds.
[3108840.664871] INFO: task kworker/80:3:3655680 blocked for more than 249638 seconds.
[3108840.668495] INFO: task kworker/89:0:3671691 blocked for more than 249047 seconds.
[3108840.672418] INFO: task kworker/84:1:3708397 blocked for more than 247836 seconds.
[3108840.676175] INFO: task kworker/81:1:4005763 blocked for more than 247836 seconds.

Task 3426612, although in the D state, was not reported as a hung task.

I confirmed that task 3426612 remained in the D (disk sleep) state and experienced no context switches over a long period:

$ cat /proc/3426612/status | grep -E "State:|ctxt_switches:";   \
  sleep 60; echo "----"; \
  cat /proc/3426612/status | grep -E "State:|ctxt_switches:"
State:  D (disk sleep)
voluntary_ctxt_switches:        7516
nonvoluntary_ctxt_switches:     0
----
State:  D (disk sleep)
voluntary_ctxt_switches:        7516
nonvoluntary_ctxt_switches:     0

The system's hung_task detector settings were configured as follows:

kernel.hung_task_timeout_secs = 28
kernel.hung_task_warnings = -1

The issue lies in the handling of task state in the XFS code. Specifically, the thread in question (3426612) was set to the TASK_KILLABLE state in xfs_extent_busy_flush():

  xfs_extent_busy_flush
    prepare_to_wait(&pag->pagb_wait, &wait, TASK_KILLABLE);

When a task is in the TASK_WAKEKILL state (a subset of TASK_KILLABLE), the hung_task detector ignores it, as it assumes such tasks can be terminated. However, in this case, the kernel thread cannot be killed, meaning it effectively becomes a hung task.

To address this issue, the hung_task detector should report the kthreads in the TASK_WAKEKILL state.

---
 kernel/hung_task.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
kernel/hung_task.c +2 -2
index c18717189f32..ed63fd84ce2e 100644
@@ -220,8 +220,9 @@ static void check_hung_uninterruptible_tasks(unsigned long timeout)
                 */
                state = READ_ONCE(t->__state);
                if ((state & TASK_UNINTERRUPTIBLE) &&
+                   (t->flags & PF_KTHREAD ||
                    !(state & TASK_WAKEKILL) &&
-                   !(state & TASK_NOLOAD))
+                   !(state & TASK_NOLOAD)))
                        check_hung_task(t, timeout);
        }
  unlock:
--
Quoted 2 lines

2.43.5

Please disregard this. There may be multiple hung tasks in the TASK_IDLE state. I will send a new one.

--
Regards
Yafang