[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [dm-devel] [PATCH 1/2] block: fix leaks associated with discard request payload



On Tue, 2010-06-29 at 18:28 -0400, Mikulas Patocka wrote:
> 
> On Sun, 27 Jun 2010, James Bottomley wrote:
> 
> > linux-scsi cc added, since it's a SCSI patch.
> > 
> > On Sat, 2010-06-26 at 15:56 -0400, Mike Snitzer wrote:
> > > Fix leaks introduced via "block: don't allocate a payload for discard
> > > request" commit a1d949f5f44.
> > > 
> > > sd_done() is not called for REQ_TYPE_BLOCK_PC commands so cleanup
> > > discard request's payload directly in scsi_finish_command().
> > > 
> > > Also cleanup page allocated for discard payload in
> > > scsi_setup_discard_cmnd's scsi_setup_blk_pc_cmnd error path.
> > > 
> > > Signed-off-by: Mike Snitzer <snitzer redhat com>
> > > ---
> > >  block/blk-core.c       |   23 +++++++++++++++++++++++
> > >  drivers/scsi/scsi.c    |    8 ++++++++
> > >  drivers/scsi/sd.c      |   18 ++++++++----------
> > >  include/linux/blkdev.h |    1 +
> > >  4 files changed, 40 insertions(+), 10 deletions(-)
> > > 
> > > diff --git a/block/blk-core.c b/block/blk-core.c
> > > index 98b4cee..07925aa 100644
> > > --- a/block/blk-core.c
> > > +++ b/block/blk-core.c
> > > @@ -1167,6 +1167,29 @@ void blk_add_request_payload(struct request *rq, struct page *page,
> > >  }
> > >  EXPORT_SYMBOL_GPL(blk_add_request_payload);
> > >  
> > > +/**
> > > + * blk_clear_request_payload - clear a request's payload
> > > + * @rq: request to update
> > > + *
> > > + * The driver needs to take care of freeing the payload itself.
> > > + */
> > > +void blk_clear_request_payload(struct request *rq)
> > > +{
> > > +	struct bio *bio = rq->bio;
> > > +
> > > +	rq->__data_len = rq->resid_len = 0;
> > > +	rq->nr_phys_segments = 0;
> > > +	rq->buffer = NULL;
> > > +
> > > +	bio->bi_size = 0;
> > > +	bio->bi_vcnt = 0;
> > > +	bio->bi_phys_segments = 0;
> > > +
> > > +	bio->bi_io_vec->bv_page = NULL;
> > > +	bio->bi_io_vec->bv_len = 0;
> > > +}
> > > +EXPORT_SYMBOL_GPL(blk_clear_request_payload);
> > > +
> > >  void init_request_from_bio(struct request *req, struct bio *bio)
> > >  {
> > >  	req->cpu = bio->bi_comp_cpu;
> > > diff --git a/drivers/scsi/scsi.c b/drivers/scsi/scsi.c
> > > index ad0ed21..69c7ea4 100644
> > > --- a/drivers/scsi/scsi.c
> > > +++ b/drivers/scsi/scsi.c
> > > @@ -851,6 +851,14 @@ void scsi_finish_command(struct scsi_cmnd *cmd)
> > >  		 */
> > >  		if (good_bytes == old_good_bytes)
> > >  			good_bytes -= scsi_get_resid(cmd);
> > > +	} else if (cmd->request->cmd_flags & REQ_DISCARD) {
> > > +		/*
> > > +		 * If this is a discard request that originated from the kernel
> > > +		 * we need to free our payload here.  Note that we need to check
> > > +		 * the request flag as the normal payload rules apply for
> > > +		 * pass-through UNMAP / WRITE SAME requests.
> > > +		 */
> > > +		__free_page(bio_page(cmd->request->bio));
> > 
> > This is another layering violation:  the page is allocated in the Upper
> > layer and freed in the mid-layer.
> > 
> > I really hate these growing contortions for discard.  They're a clear
> > signal that we haven't implemented it right.
> > 
> > So let's first work out how it should be done.  I really like Tomo's
> > idea of doing discard through the normal REQ_TYPE_FS route, which means
> > we can control the setup in prep and the tear down in done, all confined
> > to the ULD.
> > 
> > Where I think I'm at is partially what Christoph says:  The command
> > transformation belongs in the ULD so that's where the allocation and
> > deallocation should be, and partly what Tomo says in that we should
> > eliminate the special case paths.
> > 
> > The payload vs actual request size should be a red herring if we've got
> > everything correct:  only the ULD cares about the request parameters.
> > Once we've got everything set up, the mid layer and LLD should only care
> > about the parameters in the command, so we can confine the size changing
> > part to the ULD doing the discard.
> > 
> > Could someone take a stab at this and see if it works without layering
> > violations or any other problematic signals?
> > 
> > Thanks,
> > 
> > James
> 
> Well, I think that you overestimate the importance of scsi code too much.

Not, I think, a deadly sin for a SCSI maintainer.

> There is a layering violation in the code. So what --- you either fix the 
> layering violation or let it be there and grind your teeth on it. But in 
> either case, that layering violation won't affect anyone except scsi 
> developers.

A layering violation is a signal of bad design wherever it occurs, so
that wasn't a SCSI centric argument.

> On the other hand, if you say "because we want to avoid layering violation 
> in SCSI, every issuer of discard request must supply an empty page", you 
> create havoc all over the Linux codebase. md, dm, drbd, xvd, virtio --- 
> whatever you think of, will be allocating a dummy page when constructing 
> a discard request.

Since I didn't actually say any of that, I suggest you re-read text you
quoted above.  The phrase "The command transformation belongs in the ULD
so that's where the allocation and deallocation should be" might be a
relevant one to concentrate on.

> If the layering violation spans only scsi code, it can be eventually 
> fixed, but this, much worse "layering violation" that will be spanning all 
> block device midlayers, won't ever be fixed.
> 
> Imagine for example --- a discard request arrivers at a dm-snapshot 
> device. The driver splits it into chunks, remaps each chunk to the 
> physical chunk, submits the requests, the elevator merges adjacent 
> requests and submits fewer bigger requests to the device. Now, if you had 
> to allocate a zeroed page each time you are splitting the request, that 
> would exhaust memory and burn cpu needlessly. You delete a 100MB file? --- 
> fine, allocate a 100MB of zeroed pages.

This is a straw man:  You've tried to portray a position I've never
taken as mine then attack it ... with what is effectively another bogus
argument.

It's not an either/or choice.  I've asked the relevant parties to
combine the approaches and see if a REQ_TYPE_FS path that does the
allocations in the appropriate place, likely the ULD, produces a good
design.

> So I say --- let there be a layering violation in the scsi code, but don't 
> put this problem with a page allocation to all the other bio midlayer 
> developers.

Thanks for explaining that you have nothing to contribute, I'll make
sure you're not on my list of relevant parties.

James



[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]