[PATCH RFC] nvmet-tcp: add new workqueue to surpress lockdep warning
Guoqing Jiang
guoqing.jiang at linux.dev
Tue Sep 12 18:51:56 PDT 2023
On 9/12/23 20:24, Sagi Grimberg wrote:
>
>
> 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
The above should be nvmet_tcp_install_queue instead of 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.
Yes, io_work is run on nvmet_tcp_wq, and the work may trigger
flush_workqueue(nvmet_wq)
io_work = nvmet_tcp_io_work
-> 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_tcp_install_queue
-> nvmet_tcp_install_queue
-> flush_workqueue(nvmet_wq)
Also release_work = nvmet_tcp_release_queue_work need to
call cancel_work_sync(&queue->io_work), but release_work is
queued in nvmet_wq. I think this kind of mutual dependency
scenario is complained by lockdep.
> 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.
Please see above.
>> 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.
IMHO work_pending would check whether there is a pending work
item before flush relevant workqueue.
BTW, Hannes's patch can fix this as well, which might be better.
https://lore.kernel.org/linux-nvme/20230810132006.129365-1-hare@suse.de/
Thanks,
Guoqing
More information about the Linux-nvme
mailing list