[dm-devel] scsi: address leak in the error path of discard page allocation

[James Bottomley]

On Thu, 2010-07-01 at 16:15 -0400, Mike Snitzer wrote:
> On Thu, Jul 01 2010 at  9:03am -0400,
> Mike Snitzer <snitzer at redhat.com> wrote:
> 
> > On Thu, Jul 01 2010 at  6:49am -0400,
> > FUJITA Tomonori <fujita.tomonori at lab.ntt.co.jp> wrote:
> > 
> > > This fixes discard page leak by using q->unprep_rq_fn facility.
> > > 
> > > q->unprep_rq_fn is called when all the data buffer (req->bio and
> > > scsi_data_buffer) in the request is freed.
> > > 
> > > sd_unprep() uses rq->buffer to free discard page allocated in
> > > sd_prepare_discard().
> > > 
> > > Signed-off-by: FUJITA Tomonori <fujita.tomonori at lab.ntt.co.jp>
> > 
> > Thanks for sorting this out Tomo, all 3 patches work great!
> > 
> > BTW, there is one remaining (rare) leak in the allocation path.
> > 
> > The following patch serves to fix it but I'm not sure if there is a more
> > elegant way to address this.
> 
> I've continued to look at this to arrive at alternative implementation.
> Here is a summary of the problem:
> 
> A 'scsi_setup_discard_cmnd' return other than BLKPREP_OK will not cause
> a discard request to get completely stripped down ('blk_finish_request'
> isn't calling 'blk_unprep_request' because REQ_DONTPREP is not set by
> 'scsi_prep_return' for none BLKPREP_OK return).  Therefore the discard
> request's page will _not_ get cleaned up.
> 
> Aside from code inspection, I confirmed this by adding some test code to
> force a one-time initial BLKPREP_DEFER return from
> 'scsi_setup_discard_cmnd'.
> 
> > An alternative would be to check if the page is already allocated
> > (before allocating the page in scsi_setup_discard_cmnd)?
> 
> Unfortunatey this "alternative" won't work because it completely ignores
> the case where BLKPREP_KILL is returned from scsi_setup_discard_cmnd'.
>  
> > Please advise, thanks.
> 
> In short, I'm not too happy that the following patch doesn't allow for
> centralized cleanup of the discard request's page (via sd_unprep_fn).
> But in order to do that we'd likely have to:
> 1) relax blk_finish_request's REQ_DONTPREP constraint
> 2) add other weird conditionals within blk_unprep_request because
>    the discard request wasn't _really_ prepared?
> 
> So given this I'm inclined to stick with the following patch.
> 
> Jens and/or James, what do you think?

The rules are pretty clear:  Unprep is only called if the request gets
prepped ... that means you have to return BLKPREP_OK.  Defer or kill
assume there's no teardown to do, so the allocation (if it took place)
must be reversed before returning them

James