Hi Ming

On 09/18/2018 03:42 PM, Ming Lei wrote:

On Tue, Sep 18, 2018 at 09:17:12AM +0800, jianchao.wang wrote:

quoted

Hi Ming

On 09/17/2018 08:07 PM, Ming Lei wrote:

quoted

quoted

quoted

This way will delay runtime pm or system suspend until the queue is unfrozen,
but it isn't reasonable.

This interface is for the __scsi_execute only, before we call into function, we should have
get the device resumed synchronously.

I mean when the queue is frozen, it is reasonable to runtime suspend the queue. However,
blk_queue_preempt_enter() is still waiting for queue becoming unfreezing first.

We don't freeze the queue, but set preempt-only mode when suspend the queue. :)

But the queue can be frozen by other paths. Even though it is frozen, the queue
should be allowed to suspend too.

Yes, the race between freeze and preempt-only is a tricky issue.

Thanks
Jianchao