From mboxrd@z Thu Jan 1 00:00:00 1970 From: Kent Overstreet Subject: Re: [PATCH for-next] dm: fix missing bi_remaining accounting Date: Mon, 4 Nov 2013 12:12:04 -0800 Message-ID: <20131104201204.GA27733@kmo> References: <20131101135931.GA22903@redhat.com> <5273C2B8.5030608@kernel.dk> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Return-path: Content-Disposition: inline In-Reply-To: Sender: linux-kernel-owner@vger.kernel.org To: Mikulas Patocka Cc: Jens Axboe , Mike Snitzer , dm-devel@redhat.com, linux-kernel@vger.kernel.org, Alasdair Kergon List-Id: dm-devel.ids On Mon, Nov 04, 2013 at 10:06:00AM -0500, Mikulas Patocka wrote: > > > On Fri, 1 Nov 2013, Jens Axboe wrote: > > > On 11/01/2013 07:59 AM, Mike Snitzer wrote: > > > Add the missing bi_remaining increment, required by the block layer's > > > new bio-chaining code, to both the verity and old snapshot DM targets. > > > > > > Otherwise users will hit the bi_remaining <= 0 BUG_ON in bio_endio(). > > > > Thanks Mike, added to the mix. > > > > -- > > Jens Axboe > > Hi > > This improves a little bit on the previous patch, by replacing costly > atomic_inc with cheap atomic_set. IMO, this is a bad idea; the behaviour with this patch does _not_ match the naming of bio_endio_nodec(), and the performance difference should be well in the noise anyways because we're touching a cacheline we already have in cache and won't be contended. The fact that it's currently safe is accidental, I could see this easily tripping people up and being a pain in the ass to debug in the future. > > > From: Mikulas Patocka > > dm: change atomic_inc to atomic_set(1) > > There are places in dm where we save bi_endio and bi_private, set them to > target's routine, submit the bio, from the target's bi_endio routine we > restore bi_endio and bi_private and end the bio with bi_endio. > > This causes underflow of bi_remaining, so we must restore bi_remaining > before ending the bio from the target bi_endio routine. > > The code uses atomic_inc for restoration of bi_remaining. This patch > changes it to atomic_set(1) to avoid an interlocked instruction. In the > target's bi_endio routine we are sure that bi_remaining is zero > (otherwise, the bi_endio routine wouldn't be called) and there are no > concurrent users of the bio, so we can replace atomic_inc with > atomic_set(1).