[PATCH RFC] nvmet-tcp: add new workqueue to surpress lockdep warning

Sagi Grimberg sagi at grimberg.me
Tue Sep 12 05:24:33 PDT 2023



On 7/26/23 17:29, Guoqing Jiang wrote:
> During the test of nvme-tcp, lockdep complains when discover local
> nvme tcp device.
> 
> [   87.699136] ======================================================
> [   87.699137] WARNING: possible circular locking dependency detected
> [   87.699138] 6.5.0-rc3+ #16 Tainted: G            E
> [   87.699139] ------------------------------------------------------
> [   87.699140] kworker/0:4H/1522 is trying to acquire lock:
> [   87.699141] ffff93c4df45f538 ((wq_completion)nvmet-wq){+.+.}-{0:0}, at: __flush_workqueue+0x99/0x4f0
> [   87.699147]
>                 but task is already holding lock:
> [   87.699148] ffffafb40272fe40 ((work_completion)(&queue->io_work)){+.+.}-{0:0}, at: process_one_work+0x236/0x590
> [   87.699151]
>                 which lock already depends on the new lock.
> [   87.699152]
>                 the existing dependency chain (in reverse order) is:
> [   87.699153]
>                 -> #2 ((work_completion)(&queue->io_work)){+.+.}-{0:0}:
> [   87.699155]        __flush_work+0x7a/0x5c0
> [   87.699157]        __cancel_work_timer+0x155/0x1e0
> [   87.699158]        cancel_work_sync+0x10/0x20
> [   87.699160]        nvmet_tcp_release_queue_work+0xcf/0x490 [nvmet_tcp]
> [   87.699163]        process_one_work+0x2bd/0x590
> [   87.699165]        worker_thread+0x52/0x3f0
> [   87.699166]        kthread+0x109/0x140
> [   87.699168]        ret_from_fork+0x46/0x70
> [   87.699170]        ret_from_fork_asm+0x1b/0x30
> [   87.699172]
>                 -> #1 ((work_completion)(&queue->release_work)){+.+.}-{0:0}:
> [   87.699174]        process_one_work+0x28c/0x590
> [   87.699175]        worker_thread+0x52/0x3f0
> [   87.699177]        kthread+0x109/0x140
> [   87.699177]        ret_from_fork+0x46/0x70
> [   87.699179]        ret_from_fork_asm+0x1b/0x30
> [   87.699180]
>                 -> #0 ((wq_completion)nvmet-wq){+.+.}-{0:0}:
> [   87.699182]        __lock_acquire+0x1523/0x2590
> [   87.699184]        lock_acquire+0xd6/0x2f0
> [   87.699185]        __flush_workqueue+0xc5/0x4f0
> [   87.699187]        nvmet_tcp_install_queue+0x30/0x160 [nvmet_tcp]
> [   87.699189]        nvmet_install_queue+0xbf/0x200 [nvmet]
> [   87.699196]        nvmet_execute_admin_connect+0x18b/0x2f0 [nvmet]
> [   87.699200]        nvmet_tcp_io_work+0x7e3/0x850 [nvmet_tcp]
> [   87.699203]        process_one_work+0x2bd/0x590
> [   87.699204]        worker_thread+0x52/0x3f0
> [   87.699206]        kthread+0x109/0x140
> [   87.699207]        ret_from_fork+0x46/0x70
> [   87.699208]        ret_from_fork_asm+0x1b/0x30
> [   87.699209]
>                 other info that might help us debug this:
> [   87.699210] Chain exists of:
>                   (wq_completion)nvmet-wq --> (work_completion)(&queue->release_work) --> (work_completion)(&queue->io_work)
> [   87.699212]  Possible unsafe locking scenario:
> [   87.699213]        CPU0                    CPU1
> [   87.699214]        ----                    ----
> [   87.699214]   lock((work_completion)(&queue->io_work));
> [   87.699215]                                lock((work_completion)(&queue->release_work));
> [   87.699217]                                lock((work_completion)(&queue->io_work));
> [   87.699218]   lock((wq_completion)nvmet-wq);
>                          -> need to hold release_work since queue_work(nvmet_wq, &queue->release_work)
> [   87.699219]
>                  *** DEADLOCK ***
> [   87.699220] 2 locks held by kworker/0:4H/1522:
> [   87.699221]  #0: ffff93c4df45f338 ((wq_completion)nvmet_tcp_wq){+.+.}-{0:0}, at: process_one_work+0x236/0x590
> [   87.699224]  #1: ffffafb40272fe40 ((work_completion)(&queue->io_work)){+.+.}-{0:0}, at: process_one_work+0x236/0x590
> [   87.699227]
>                 stack backtrace:
> [   87.699229] CPU: 0 PID: 1522 Comm: kworker/0:4H Tainted: G            E      6.5.0-rc3+ #16
> [   87.699230] Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS rel-1.16.0-0-gd239552c-rebuilt.opensuse.org 04/01/2014
> [   87.699231] Workqueue: nvmet_tcp_wq nvmet_tcp_io_work [nvmet_tcp]
> 
> The above happens because nvmet_tcp_io_work can trigger below path
> 
> 	-> nvmet_tcp_try_recv
> 	 -> nvmet_tcp_try_recv_one
> 	  -> nvmet_tcp_try_recv_data
> 	   -> nvmet_tcp_execute_request
> 	    -> cmd->req.execute = nvmet_execute_admin_connect
> 	     -> nvmet_install_queue
> 	      -> ctrl->ops->install_queue = nvmet_install_queue
> 	       -> nvmet_tcp_install_queue
> 	        -> flush_workqueue(nvmet_wq)
> 
> And release_work (nvmet_tcp_release_queue_work) is also queued in
> nvmet_wq, which need to flush io_work (nvmet_tcp_io_work) due to
> cancel_work_sync(&queue->io_work).

I'm not sure I understand the resolution here. io_work does not
run on nvmet_wq, but on nvmet_tcp_wq. What does separating another
workqueue give here?

> 
> We can surpress the lockdep warning by checking if the relevant work
> is pending. So the simplest might be just add the checking before
> flush_workqueue(nvmet_wq). However, there are other works are also
> queued on the same queue, I am not sure if we should flush other
> works unconditionally, so a new dedicated workqueue is added.
> 
> Signed-off-by: Guoqing Jiang <guoqing.jiang at linux.dev>
> ---
>   drivers/nvme/target/tcp.c | 14 +++++++++++++-
>   1 file changed, 13 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/nvme/target/tcp.c b/drivers/nvme/target/tcp.c
> index 868aa4de2e4c..ac611cb299a8 100644
> --- a/drivers/nvme/target/tcp.c
> +++ b/drivers/nvme/target/tcp.c
> @@ -189,6 +189,7 @@ static LIST_HEAD(nvmet_tcp_queue_list);
>   static DEFINE_MUTEX(nvmet_tcp_queue_mutex);
>   
>   static struct workqueue_struct *nvmet_tcp_wq;
> +static struct workqueue_struct *nvmet_tcp_release_wq;
>   static const struct nvmet_fabrics_ops nvmet_tcp_ops;
>   static void nvmet_tcp_free_cmd(struct nvmet_tcp_cmd *c);
>   static void nvmet_tcp_free_cmd_buffers(struct nvmet_tcp_cmd *cmd);
> @@ -1288,7 +1289,7 @@ static void nvmet_tcp_schedule_release_queue(struct nvmet_tcp_queue *queue)
>   	spin_lock(&queue->state_lock);
>   	if (queue->state != NVMET_TCP_Q_DISCONNECTING) {
>   		queue->state = NVMET_TCP_Q_DISCONNECTING;
> -		queue_work(nvmet_wq, &queue->release_work);
> +		queue_work(nvmet_tcp_release_wq, &queue->release_work);
>   	}
>   	spin_unlock(&queue->state_lock);
>   }
> @@ -1847,6 +1848,8 @@ static u16 nvmet_tcp_install_queue(struct nvmet_sq *sq)
>   	if (sq->qid == 0) {
>   		/* Let inflight controller teardown complete */
>   		flush_workqueue(nvmet_wq);
> +		if (work_pending(&queue->release_work))
> +			flush_workqueue(nvmet_tcp_release_wq);

This is effectively just never flushes anything. when we install
the queue it's own release_work never really runs. So what your
patch effectively does is just to remove the flush altogether.

The workqueue separation has nothing to do with anything afaict.

Am I missing something here?



More information about the Linux-nvme mailing list