Re: [PATCH v4 1/8] scsi: ufs: Flush exception event before suspend

[Can Guo]

On 2020-02-04 11:12, Bart Van Assche wrote:
> On 2020-02-02 22:23, Can Guo wrote:
> > On 2020-01-26 11:29, Bart Van Assche wrote:
> > > On 2020-01-22 23:25, Can Guo wrote:
> > > > ÂÂÂÂÂÂÂÂÂÂÂÂ break;
> > > > ÂÂÂÂÂÂÂÂ case UPIU_TRANSACTION_REJECT_UPIU:
> > > > ÂÂÂÂÂÂÂÂÂÂÂÂ /* TODO: handle Reject UPIU Response */
> > > > @@ -5215,7 +5222,14 @@ static void
> > > > ufshcd_exception_event_handler(struct work_struct *work)
> > > >
> > > > Âout:
> > > > ÂÂÂÂ scsi_unblock_requests(hba->host);
> > > > -ÂÂÂ pm_runtime_put_sync(hba->dev);
> > > > +ÂÂÂ /*
> > > > +ÂÂÂÂ * pm_runtime_get_noresume is called while scheduling
> > > > +ÂÂÂÂ * eeh_work to avoid suspend racing with exception work.
> > > > +ÂÂÂÂ * Hence decrement usage counter using pm_runtime_put_noidle
> > > > +ÂÂÂÂ * to allow suspend on completion of exception event handler.
> > > > +ÂÂÂÂ */
> > > > +ÂÂÂ pm_runtime_put_noidle(hba->dev);
> > > > +ÂÂÂ pm_runtime_put(hba->dev);
> > > > ÂÂÂÂ return;
> > > > Â}
> > > >
> > > > @@ -7901,6 +7915,7 @@ static int ufshcd_suspend(struct ufs_hba *hba,
> > > > enum ufs_pm_op pm_op)
> > > > ÂÂÂÂÂÂÂÂÂÂÂÂ goto enable_gating;
> > > > ÂÂÂÂ }
> > > >
> > > > +ÂÂÂ flush_work(&hba->eeh_work);
> > > > ÂÂÂÂ ret = ufshcd_link_state_transition(hba, req_link_state, 1);
> > > > ÂÂÂÂ if (ret)
> > > > ÂÂÂÂÂÂÂÂ goto set_dev_active;
> > >
> > > I think this patch introduces a new race condition, namely the 
> > > following:
> > > - ufshcd_slave_destroy() tests pm_op_in_progress and reads the value
> > > Â zero from that variable.
> > > - ufshcd_suspend() sets hba->pm_op_in_progress to one.
> > > - ufshcd_slave_destroy() calls schedule_work().
> > >
> > > How about fixing this race condition by calling
> > > pm_runtime_get_noresume() before checking pm_op_in_progress and by
> > > reallowing resume if no work is scheduled?
> >
> > If you apply this patch, you will find the change is not in
> > ufshcd_slave_destroy(), but in ufshcd_transfer_rsp_status().
> > So the racing you mentioned above does not exist.
>
> Hi Can,
>
> Apparently I got a function name wrong. Can the following race 
> condition
> happen:
> - ufshcd_transfer_rsp_status() tests pm_op_in_progress and reads the
>   value zero from that variable.
> - ufshcd_suspend() sets hba->pm_op_in_progress to one.
> - ufshcd_suspend() calls flush_work(&hba->eeh_work).
> - ufshcd_transfer_rsp_status() calls schedule_work(&hba->eeh_work).
>
> Thanks,
>
> Bart.

Hi Bart,

The sequence you mentioned is not possible.

In normal cases, before ufshcd_transfer_rsp_status() returns,
ufshcd_suspend() would not be called (unless you intentionally call
ufshcd_suspend() to screw it). Because ufshcd_transfer_rsp_status() is
called from __ufshcd_transfer_req_compl(), which is being used by either
UFS IRQ handler or err handler. Meanwhile, in 
__ufshcd_transfer_req_compl(),
scsi_done() is called only after ufshcd_transfer_rsp_status() returns. 
When
we are here, it means UFS driver is still handling requests/tasks, so 
suspend
would not kick start at this moment, either runtime suspend or system 
suspend.

And this is why below lines work, calling pm_runtime_get_noresume() 
within
ufshcd_transfer_rsp_status() can prevent runtime suspend from happening
after we leave ufshcd_transfer_rsp_status().

+                if (schedule_work(&hba->eeh_work))
+                    pm_runtime_get_noresume(hba->dev);

Thanks,

Can Guo.