[Qemu-devel] [PATCH] scsi: always call notifier on async cancellation

[Qemu-devel] [PATCH] scsi: always call notifier on async cancellation

[Paolo Bonzini]

This was found by code inspection.  If the request is cancelled twice,
the notifier is never called on the second cancellation request,
and hence for example a TMF might never finish.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 hw/scsi/scsi-bus.c | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/hw/scsi/scsi-bus.c b/hw/scsi/scsi-bus.c
index 524a998..4c121fe 100644
--- a/hw/scsi/scsi-bus.c
+++ b/hw/scsi/scsi-bus.c
@@ -1759,9 +1759,6 @@ void scsi_req_cancel_async(SCSIRequest *req, Notifier *notifier)
     if (notifier) {
         notifier_list_add(&req->cancel_notifiers, notifier);
     }
-    if (req->io_canceled) {
-        return;
-    }
     scsi_req_ref(req);
     scsi_req_dequeue(req);
     req->io_canceled = true;
-- 
2.5.0

Re: [Qemu-devel] [PATCH] scsi: always call notifier on async cancellation

[Fam Zheng]

On Wed, 12/16 19:33, Paolo Bonzini wrote:
> This was found by code inspection.  If the request is cancelled twice,
> the notifier is never called on the second cancellation request,
> and hence for example a TMF might never finish.
> 
> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
> ---
>  hw/scsi/scsi-bus.c | 3 ---
>  1 file changed, 3 deletions(-)
> 
> diff --git a/hw/scsi/scsi-bus.c b/hw/scsi/scsi-bus.c
> index 524a998..4c121fe 100644
> --- a/hw/scsi/scsi-bus.c
> +++ b/hw/scsi/scsi-bus.c
> @@ -1759,9 +1759,6 @@ void scsi_req_cancel_async(SCSIRequest *req, Notifier *notifier)
>      if (notifier) {
>          notifier_list_add(&req->cancel_notifiers, notifier);
>      }
> -    if (req->io_canceled) {
> -        return;
> -    }
>      scsi_req_ref(req);
>      scsi_req_dequeue(req);
>      req->io_canceled = true;
       if (req->aiocb) {
           blk_aio_cancel_async(req->aiocb);
       } else {
           scsi_req_cancel_complete(req);
       }

A second TMF must be blk_aio_cancel_async case, otherwise the first one would
have already completed the request synchronously in scsi_req_cancel_complete.
With that in mind, I think returning early is not a problem. But I suppose
these are also idempotent so this change is not breaking anything, either.

Fam

Re: [Qemu-devel] [PATCH] scsi: always call notifier on async cancellation

[Paolo Bonzini]

On 17/12/2015 02:15, Fam Zheng wrote:
>> >      if (notifier) {
>> >          notifier_list_add(&req->cancel_notifiers, notifier);
>> >      }
>> > -    if (req->io_canceled) {
>> > -        return;
>> > -    }
>> >      scsi_req_ref(req);
>> >      scsi_req_dequeue(req);
>> >      req->io_canceled = true;
>        if (req->aiocb) {
>            blk_aio_cancel_async(req->aiocb);
>        } else {
>            scsi_req_cancel_complete(req);
>        }
> 
> A second TMF must be blk_aio_cancel_async case, otherwise the first one would
> have already completed the request synchronously in scsi_req_cancel_complete.

Good point.

> With that in mind, I think returning early is not a problem. But I suppose
> these are also idempotent so this change is not breaking anything, either.

Right, the issue is that all these calls are idempotent, but the
notifier may not; that is why I prefer to be safe and ensure that all
notifier additions are matched by a notify.  But you explained well why
this should be safe, I'll add a note to the commit message.

Paolo

Re: [Qemu-devel] [PATCH] scsi: always call notifier on async cancellation

[Fam Zheng]

On Thu, 12/17 09:41, Paolo Bonzini wrote:
> 
> 
> On 17/12/2015 02:15, Fam Zheng wrote:
> >> >      if (notifier) {
> >> >          notifier_list_add(&req->cancel_notifiers, notifier);
> >> >      }
> >> > -    if (req->io_canceled) {
> >> > -        return;
> >> > -    }
> >> >      scsi_req_ref(req);
> >> >      scsi_req_dequeue(req);
> >> >      req->io_canceled = true;
> >        if (req->aiocb) {
> >            blk_aio_cancel_async(req->aiocb);
> >        } else {
> >            scsi_req_cancel_complete(req);
> >        }
> > 
> > A second TMF must be blk_aio_cancel_async case, otherwise the first one would
> > have already completed the request synchronously in scsi_req_cancel_complete.
> 
> Good point.
> 
> > With that in mind, I think returning early is not a problem. But I suppose
> > these are also idempotent so this change is not breaking anything, either.
> 
> Right, the issue is that all these calls are idempotent, but the
> notifier may not; that is why I prefer to be safe and ensure that all
> notifier additions are matched by a notify.  But you explained well why
> this should be safe, I'll add a note to the commit message.
> 

Thanks, please add my

Reviewed-by: Fam Zheng <famz@redhat.com>