在 2021/2/19 上午3:15, Pavel Begunkov 写道:
On 18/02/2021 17:16, Hao Xu wrote:
在 2021/1/25 下午12:31, Jens Axboe 写道:
On 1/23/21 2:40 AM, Hao Xu wrote:
Abaci reported the following warning:
[ 97.862205] ============================================
[ 97.863400] WARNING: possible recursive locking detected
[ 97.864640] 5.11.0-rc4+ #12 Not tainted
[ 97.865537] --------------------------------------------
[ 97.866748] a.out/2890 is trying to acquire lock:
[ 97.867829] ffff8881046763e8 (&ctx->uring_lock){+.+.}-{3:3}, at:
io_wq_submit_work+0x155/0x240
[ 97.869735]
[ 97.869735] but task is already holding lock:
[ 97.871033] ffff88810dfe0be8 (&ctx->uring_lock){+.+.}-{3:3}, at:
__x64_sys_io_uring_enter+0x3f0/0x5b0
[ 97.873074]
[ 97.873074] other info that might help us debug this:
[ 97.874520] Possible unsafe locking scenario:
[ 97.874520]
[ 97.875845] CPU0
[ 97.876440] ----
[ 97.877048] lock(&ctx->uring_lock);
[ 97.877961] lock(&ctx->uring_lock);
[ 97.878881]
[ 97.878881] *** DEADLOCK ***
[ 97.878881]
[ 97.880341] May be due to missing lock nesting notation
[ 97.880341]
[ 97.881952] 1 lock held by a.out/2890:
[ 97.882873] #0: ffff88810dfe0be8 (&ctx->uring_lock){+.+.}-{3:3}, at:
__x64_sys_io_uring_enter+0x3f0/0x5b0
[ 97.885108]
[ 97.885108] stack backtrace:
[ 97.886209] CPU: 0 PID: 2890 Comm: a.out Not tainted 5.11.0-rc4+ #12
[ 97.887683] Hardware name: Alibaba Cloud Alibaba Cloud ECS, BIOS
rel-1.7.5-0-ge51488c-20140602_164612-nilsson.home.kraxel.org 04/01/2014
[ 97.890457] Call Trace:
[ 97.891121] dump_stack+0xac/0xe3
[ 97.891972] __lock_acquire+0xab6/0x13a0
[ 97.892940] lock_acquire+0x2c3/0x390
[ 97.893853] ? io_wq_submit_work+0x155/0x240
[ 97.894894] __mutex_lock+0xae/0x9f0
[ 97.895785] ? io_wq_submit_work+0x155/0x240
[ 97.896816] ? __lock_acquire+0x782/0x13a0
[ 97.897817] ? io_wq_submit_work+0x155/0x240
[ 97.898867] ? io_wq_submit_work+0x155/0x240
[ 97.899916] ? _raw_spin_unlock_irqrestore+0x2d/0x40
[ 97.901101] io_wq_submit_work+0x155/0x240
[ 97.902112] io_wq_cancel_cb+0x162/0x490
[ 97.903084] ? io_uring_get_socket+0x40/0x40
[ 97.904126] io_async_find_and_cancel+0x3b/0x140
[ 97.905247] io_issue_sqe+0x86d/0x13e0
[ 97.906186] ? __lock_acquire+0x782/0x13a0
[ 97.907195] ? __io_queue_sqe+0x10b/0x550
[ 97.908175] ? lock_acquire+0x2c3/0x390
[ 97.909122] __io_queue_sqe+0x10b/0x550
[ 97.910080] ? io_req_prep+0xd8/0x1090
[ 97.911044] ? mark_held_locks+0x5a/0x80
[ 97.912042] ? mark_held_locks+0x5a/0x80
[ 97.913014] ? io_queue_sqe+0x235/0x470
[ 97.913971] io_queue_sqe+0x235/0x470
[ 97.914894] io_submit_sqes+0xcce/0xf10
[ 97.915842] ? xa_store+0x3b/0x50
[ 97.916683] ? __x64_sys_io_uring_enter+0x3f0/0x5b0
[ 97.917872] __x64_sys_io_uring_enter+0x3fb/0x5b0
[ 97.918995] ? lockdep_hardirqs_on_prepare+0xde/0x180
[ 97.920204] ? syscall_enter_from_user_mode+0x26/0x70
[ 97.921424] do_syscall_64+0x2d/0x40
[ 97.922329] entry_SYSCALL_64_after_hwframe+0x44/0xa9
[ 97.923538] RIP: 0033:0x7f0b62601239
[ 97.924437] Code: 01 00 48 81 c4 80 00 00 00 e9 f1 fe ff ff 0f 1f 00
48 89 f8 48 89 f7 48 89 d6 48 89 ca 4d 89 c2 4d 89 c8 4c 8b 4c 24 08 0f
05 <48> 3d 01 f0 ff ff 73 01 c3 48 8b 0d 27 ec 2c 00 f7 d8 64 89 01
48
[ 97.928628] RSP: 002b:00007f0b62cc4d28 EFLAGS: 00000246 ORIG_RAX:
00000000000001aa
[ 97.930422] RAX: ffffffffffffffda RBX: 0000000000000000 RCX:
00007f0b62601239
[ 97.932073] RDX: 0000000000000000 RSI: 0000000000006cf6 RDI:
0000000000000005
[ 97.933710] RBP: 00007f0b62cc4e20 R08: 0000000000000000 R09:
0000000000000000
[ 97.935369] R10: 0000000000000000 R11: 0000000000000246 R12:
0000000000000000
[ 97.937008] R13: 0000000000021000 R14: 0000000000000000 R15:
00007f0b62cc5700
This is caused by try to hold uring_lock in io_wq_submit_work() without
checking if we are in io-wq thread context or not. It can be in original
context when io_wq_submit_work() is called from IORING_OP_ASYNC_CANCEL
code path, where we already held uring_lock.
Looks like another fallout of the split CLOSE handling. I've got the
right fixes pending for 5.12:
https://git.kernel.dk/cgit/linux-block/commit/?h=for-5.12/io_uring&id=6bb0079ef3420041886afe1bcd8e7a87e08992e1
(and the prep patch before that in the tree). But that won't really
help us for 5.11 and earlier, though we probably should just queue
those two patches for 5.11 and get them into stable. I really don't
like the below patch, though it should fix it. But the root cause
is really the weird open cancelation...
Hi Jens,
I've repro-ed this issue on branch for-5.12/io_uring-2021-02-17
which contains the patch you give, the issue still exists.
I think this one is not an async close specifical problem.
The rootcause is we try to run an iowq work in the original
context(queue an iowq work, then async cancel it).
If you mean cancellation executed from task_work or inline (during
submission), then yes, I agree.
Yea, that's what I mean.
Can you try a diff below?
Tested, it works well, thanks Pavel.
diff --git a/fs/io_uring.c b/fs/io_uring.c
index 2fdfe5fa00b0..8dab07f42b34 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -2337,7 +2337,9 @@ static void io_req_task_cancel(struct callback_head *cb)
struct io_kiocb *req = container_of(cb, struct io_kiocb, task_work);
struct io_ring_ctx *ctx = req->ctx;
+ mutex_lock(&ctx->uring_lock);
__io_req_task_cancel(req, -ECANCELED);
+ mutex_unlock(&ctx->uring_lock);
percpu_ref_put(&ctx->refs);
}
@@ -6426,8 +6428,13 @@ static void io_wq_submit_work(struct io_wq_work *work)
if (timeout)
io_queue_linked_timeout(timeout);
- if (work->flags & IO_WQ_WORK_CANCEL)
- ret = -ECANCELED;
+ if (work->flags & IO_WQ_WORK_CANCEL) {
+ /* io-wq is going to take down one */
+ refcount_inc(&req->refs);
+ percpu_ref_get(&req->ctx->refs);
+ io_req_task_work_add_fallback(req, io_req_task_cancel);
+ return;
+ }
if (!ret) {
do {
