[PATCH] dm_crypt: Issue all previous bios before remapping barrier bio

[PATCH] dm_crypt: Issue all previous bios before remapping barrier bio

[Nikanth Karthikesan]

Before remapping a barrier bio, make sure that there are no un-issued
pending I/O in the workqueue. Otherwise a dm_flush would simply add
a barrier request to the underlying device, without issuing the bio's
queued in the dm_crypt device.

Signed-off-by: Nikanth Karthikesan <knikanth@suse.de>

--

diff --git a/drivers/md/dm-crypt.c b/drivers/md/dm-crypt.c
index a936372..be68e70 100644
--- a/drivers/md/dm-crypt.c
+++ b/drivers/md/dm-crypt.c
@@ -1252,6 +1252,7 @@ static int crypt_map(struct dm_target *ti, struct bio *bio,
 
 	if (unlikely(bio_empty_barrier(bio))) {
 		cc = ti->private;
+		flush_workqueue(cc->io_queue);
 		bio->bi_bdev = cc->dev->bdev;
 		return DM_MAPIO_REMAPPED;
 	}

Re: [PATCH] dm_crypt: Issue all previous bios before remapping barrier bio

[Milan Broz]

On 02/01/2010 04:19 PM, Nikanth Karthikesan wrote:
> Before remapping a barrier bio, make sure that there are no un-issued
> pending I/O in the workqueue. Otherwise a dm_flush would simply add
> a barrier request to the underlying device, without issuing the bio's
> queued in the dm_crypt device.

This should not be needed. Core dm code waits for all bios submitted before
barrier requests.

Mikulas, could you verify it?

Milan

Re: [PATCH] dm_crypt: Issue all previous bios before remapping barrier bio

[Mikulas Patocka]

On Mon, 1 Feb 2010, Milan Broz wrote:

> On 02/01/2010 04:19 PM, Nikanth Karthikesan wrote:
> > Before remapping a barrier bio, make sure that there are no un-issued
> > pending I/O in the workqueue. Otherwise a dm_flush would simply add
> > a barrier request to the underlying device, without issuing the bio's
> > queued in the dm_crypt device.
> 
> This should not be needed. Core dm code waits for all bios submitted before
> barrier requests.
> 
> Mikulas, could you verify it?
> 
> Milan

Yes, you are right, the patch is not needed.

When the target receives an empty barrier request, there are no other data 
requests in flight.

BTW. The target may receive more empty barrier requests simultaneously, if 
it asked for it by setting ti->num_flush_requests to more than one. But 
barrier and non-barrier requests are never mixed.

Mikulas

Re: [PATCH] dm_crypt: Issue all previous bios before remapping barrier bio

[Nikanth Karthikesan]

On Tuesday 02 February 2010 18:17:24 Mikulas Patocka wrote:
> On Mon, 1 Feb 2010, Milan Broz wrote:
> > On 02/01/2010 04:19 PM, Nikanth Karthikesan wrote:
> > > Before remapping a barrier bio, make sure that there are no un-issued
> > > pending I/O in the workqueue. Otherwise a dm_flush would simply add
> > > a barrier request to the underlying device, without issuing the bio's
> > > queued in the dm_crypt device.
> >
> > This should not be needed. Core dm code waits for all bios submitted
> > before barrier requests.
> >
> > Mikulas, could you verify it?
> >
> > Milan
> 
> Yes, you are right, the patch is not needed.
> 
> When the target receives an empty barrier request, there are no other data
> requests in flight.
> 
> BTW. The target may receive more empty barrier requests simultaneously, if
> it asked for it by setting ti->num_flush_requests to more than one. But
> barrier and non-barrier requests are never mixed.
> 

Correct. Sorry, I was a bit hasty in sending that patch. It is not required. 
Thanks for the patient explanation.

Thanks
Nikanth