Re: INFO: task hung in __io_uring_task_cancel

On 1/2/21 9:14 PM, Palash Oswal wrote:

 Hello,

I was running syzkaller and I found the following issue :

Head Commit : b1313fe517ca3703119dcc99ef3bbf75ab42bcfb ( v5.10.4 )
Git Tree : stable
Console Output :
[  242.769080] INFO: task repro:2639 blocked for more than 120 seconds.
[  242.769096]       Not tainted 5.10.4 #8
[  242.769103] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs"
disables this message.
[  242.769112] task:repro           state:D stack:    0 pid: 2639
ppid:  2638 flags:0x00000004
[  242.769126] Call Trace:
[  242.769148]  __schedule+0x28d/0x7e0
[  242.769162]  ? __percpu_counter_sum+0x75/0x90
[  242.769175]  schedule+0x4f/0xc0
[  242.769187]  __io_uring_task_cancel+0xad/0xf0
[  242.769198]  ? wait_woken+0x80/0x80
[  242.769210]  bprm_execve+0x67/0x8a0
[  242.769223]  do_execveat_common+0x1d2/0x220
[  242.769235]  __x64_sys_execveat+0x5d/0x70
[  242.769249]  do_syscall_64+0x38/0x90
[  242.769260]  entry_SYSCALL_64_after_hwframe+0x44/0xa9
[  242.769270] RIP: 0033:0x7f59ce45967d
[  242.769277] RSP: 002b:00007ffd05d10a58 EFLAGS: 00000246 ORIG_RAX:
0000000000000142
[  242.769290] RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007f59ce45967d
[  242.769297] RDX: 0000000000000000 RSI: 0000000020000180 RDI: 00000000ffffffff
[  242.769304] RBP: 00007ffd05d10a70 R08: 0000000000000000 R09: 00007ffd05d10a70
[  242.769311] R10: 0000000000000000 R11: 0000000000000246 R12: 000055a91d37d320
[  242.769318] R13: 0000000000000000 R14: 0000000000000000 R15: 0000000000000000

Can you see if this helps? The reproducer is pretty brutal, it'll fork
thousands of tasks with rings! But should work of course. I think this
one is pretty straight forward, and actually an older issue with the
poll rewaiting.

diff --git a/fs/io_uring.c b/fs/io_uring.c
index ca46f314640b..539de04f9183 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c

@@ -5103,6 +5103,12 @@ static bool io_poll_rewait(struct io_kiocb *req, struct io_poll_iocb *poll)
 {
 	struct io_ring_ctx *ctx = req->ctx;
 
+	/* Never re-wait on poll if the ctx or task is going away */
+	if (percpu_ref_is_dying(&ctx->refs) || req->task->flags & PF_EXITING) {
+		spin_lock_irq(&ctx->completion_lock);
+		return false;
+	}
+
 	if (!req->result && !READ_ONCE(poll->canceled)) {
 		struct poll_table_struct pt = { ._key = poll->events };
 

-- 
Jens Axboe