[PATCH] nvme: fix deadlock between reset and scan

[yaoma]

I have previously tried the method that you proposed, and it could solve 
the deadlock issue. My worry is that if an I/O timeout occurs during the 
scan, it will trigger a reset. However, the reset will wait for the scan 
to end, which could introduce a new risk of deadlock.
I agree with the suggestion made by Sagi Grimberg that this approach 
does not introduce new problems.

On 2023/11/29 02:00, Keith Busch wrote:
> On Tue, Nov 28, 2023 at 12:13:59PM +0200, Sagi Grimberg wrote:
>>
>>
>> On 11/28/23 08:22, yaoma wrote:
>>> Hi Keith Busch
>>>
>>> Thanks for your reply.
>>>
>>> The idea to avoid such a deadlock between nvme_reset and nvme_scan is to
>>> ensure that no namespace can be added to ctrl->namespaces after
>>> nvme_start_freeze has already been called. We can achieve this goal by
>>> assessing the ctrl->state after we have already acquired the
>>> ctrl->namespaces_rwsem lock, to decide whether to add the namespace to
>>> the list or not.
>>> 1. After we determine that ctrl->state is LIVE, it may be immediately
>>> changed to another state. However, since we have already acquired the
>>> lock, other tasks cannot access ctrl->namespace, so we can still safely
>>> add the namespace to the list. After acquiring the lock,
>>> nvme_start_freeze will freeze all ns->q in the list, including any newly
>>> added namespaces.
>>> 2. Before the completion of nvme_reset, ctrl->state will not be changed
>>> to LIVE, so we will not add any more namespaces to the list. All ns->q
>>> in the list is frozen, so nvme_wait_freeze can exit normally.
>>
>> I agree with the analysis, there is a hole between start_freeze and
>> freeze_wait that a scan may add a ns to the ctrl ns list.
>>
>> However the fix should be to mark the ctrl with say NVME_CTRL_FROZEN
>> flag set in nvme_freeze_start and cleared in nvme_unfreeze (similar
>> to what we did with quiesce). Then the scan can check it before adding
>> the new namespace (under the namespaces_rwsem).
> 
> Could we just make sure that scan_work isn't running? If we reset a live
> controller, then we're not depending on reset_work to unblock scan_work,
> and can let scan_work end gracefully. The scan_work can't be rescheduled
> again while in the resetting state.
> 
> ---
> diff --git a/drivers/nvme/host/pci.c b/drivers/nvme/host/pci.c
> index fad4cccce745c..5d6305475bad5 100644
> --- a/drivers/nvme/host/pci.c
> +++ b/drivers/nvme/host/pci.c
> @@ -2701,8 +2701,10 @@ static void nvme_reset_work(struct work_struct *work)
>           * If we're called to reset a live controller first shut it down before
>           * moving on.
>           */
> -       if (dev->ctrl.ctrl_config & NVME_CC_ENABLE)
> +       if (dev->ctrl.ctrl_config & NVME_CC_ENABLE) {
> +               flush_work(&dev->ctrl.scan_work);
>                  nvme_dev_disable(dev, false);
> +       }
>          nvme_sync_queues(&dev->ctrl);
> 
>          mutex_lock(&dev->shutdown_lock);
> --