Michael S. Tsirkin
2023-Jun-08 14:46 UTC
virtio-blk: support completion batching for the IRQ path - failure
On Thu, Jun 08, 2023 at 11:07:21PM +0900, Suwan Kim wrote:> On Thu, Jun 8, 2023 at 7:16?PM Roberts, Martin <martin.roberts at intel.com> wrote: > > > > The rq_affinity change does not resolve the issue; just reduces its occurrence rate; I am still seeing hangs with it set to 2. > > > > Martin > > > > > > > > From: Roberts, Martin > > Sent: Wednesday, June 7, 2023 3:46 PM > > To: Suwan Kim <suwan.kim027 at gmail.com> > > Cc: mst at redhat.com; virtualization <virtualization at lists.linux-foundation.org>; linux-block at vger.kernel.org > > Subject: RE: virtio-blk: support completion batching for the IRQ path - failure > > > > > > > > It is the change indicated that breaks it - changing the IRQ handling to batching. > > > > > > > > > > > > > > > > From reports such as, > > > > [PATCH 1/1] blk-mq: added case for cpu offline during send_ipi in rq_complete (kernel.org) > https://lore.kernel.org/lkml/20220929033428.25948-1-mj0123.lee at samsung.com/T/ > > > > [RFC] blk-mq: Don't IPI requests on PREEMPT_RT - Patchwork (linaro.org) > https://patches.linaro.org/project/linux-rt-users/patch/20201023110400.bx3uzsb7xy5jtsea at linutronix.de/ > > > > > > > > I?m thinking the issue has something to do with which CPU the IRQ is running on. > > > > > > > > So, I set, > > > > # echo 2 > /sys/block/vda/queue/rq_affinity > > > > # echo 2 > /sys/block/vdb/queue/rq_affinity > > > > ? > > > > # echo 2 > /sys/block/vdp/queue/rq_affinity > > > > > > > > > > > > and the system (running 16 disks, 4 queues/disk) has not yet hung (running OK for several hours)? > > > > > > > > Martin > > > > Hi Martin, > > Both codes (original code and your simple path) execute > blk_mq_complete_send_ipi() > at blk_mq_complete_request_remote(). So maybe missing request completion > on other vCPU is not the cause... > > The difference between the original code and your simple path is that > the original code calls blk_mq_end_request_batch() at virtblk_done() > to process request at block layer > and your code calls blk_mq_end_request() at virtblk_done() to do same thing. > > The original code : > virtblk_handle_req() first collects all requests from virtqueue in while loop > and pass it to blk_mq_end_request_batch() at once > > Your simple path: > virtblk_handle_req() get single request from virtqueue and pass it to > blk_mq_end_request() and do it again in while loop until there in no request > in virtqueue > > > I think we need to focus on the difference between blk_mq_end_request() > and blk_mq_end_request_batch() > > Regards, > Suwan Kim >Yes but linux release is imminent and regressions are bad. What do you suggest for now? If there's no better idea I'll send a revert patch and we'll see in the next linux version.> > > > > > > > -----Original Message----- > > From: Suwan Kim <suwan.kim027 at gmail.com> > > Sent: Wednesday, June 7, 2023 3:21 PM > > To: Roberts, Martin <martin.roberts at intel.com> > > Cc: mst at redhat.com; virtualization <virtualization at lists.linux-foundation.org>; linux-block at vger.kernel.org > > Subject: Re: virtio-blk: support completion batching for the IRQ path - failure > > > > > > > > On Wed, Jun 7, 2023 at 6:14?PM Roberts, Martin <martin.roberts at intel.com> wrote: > > > > > > > > > > Re: virtio-blk: support completion batching for the IRQ path ? torvalds/linux at 07b679f ? GitHub > > > > > > > > > > Signed-off-by: Suwan Kim suwan.kim027 at gmail.com > > > > > > > > > > Signed-off-by: Michael S. Tsirkin mst at redhat.com > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > This change appears to have broken things? > > > > > > > > > > We now see applications hanging during disk accesses. > > > > > > > > > > e.g. > > > > > > > > > > multi-port virtio-blk device running in h/w (FPGA) > > > > > > > > > > Host running a simple ?fio? test. > > > > > > > > > > [global] > > > > > > > > > > thread=1 > > > > > > > > > > direct=1 > > > > > > > > > > ioengine=libaio > > > > > > > > > > norandommap=1 > > > > > > > > > > group_reporting=1 > > > > > > > > > > bs=4K > > > > > > > > > > rw=read > > > > > > > > > > iodepth=128 > > > > > > > > > > runtime=1 > > > > > > > > > > numjobs=4 > > > > > > > > > > time_based > > > > > > > > > > [job0] > > > > > > > > > > filename=/dev/vda > > > > > > > > > > [job1] > > > > > > > > > > filename=/dev/vdb > > > > > > > > > > [job2] > > > > > > > > > > filename=/dev/vdc > > > > > > > > > > ... > > > > > > > > > > [job15] > > > > > > > > > > filename=/dev/vdp > > > > > > > > > > > > > > > > > > > > i.e. 16 disks; 4 queues per disk; simple burst of 4KB reads > > > > > > > > > > This is repeatedly run in a loop. > > > > > > > > > > > > > > > > > > > > After a few, normally <10 seconds, fio hangs. > > > > > > > > > > With 64 queues (16 disks), failure occurs within a few seconds; with 8 queues (2 disks) it may take ~hour before hanging. > > > > > > > > > > Last message: > > > > > > > > > > fio-3.19 > > > > > > > > > > Starting 8 threads > > > > > > > > > > Jobs: 1 (f=1): [_(7),R(1)][68.3%][eta 03h:11m:06s] > > > > > > > > > > I think this means at the end of the run 1 queue was left incomplete. > > > > > > > > > > > > > > > > > > > > ?diskstats? (run while fio is hung) shows no outstanding transactions. > > > > > > > > > > e.g. > > > > > > > > > > $ cat /proc/diskstats > > > > > > > > > > ... > > > > > > > > > > 252 0 vda 1843140071 0 14745120568 712568645 0 0 0 0 0 3117947 712568645 0 0 0 0 0 0 > > > > > > > > > > 252 16 vdb 1816291511 0 14530332088 704905623 0 0 0 0 0 3117711 704905623 0 0 0 0 0 0 > > > > > > > > > > ... > > > > > > > > > > > > > > > > > > > > Other stats (in the h/w, and added to the virtio-blk driver ([a]virtio_queue_rq(), [b]virtblk_handle_req(), [c]virtblk_request_done()) all agree, and show every request had a completion, and that virtblk_request_done() never gets called. > > > > > > > > > > e.g. > > > > > > > > > > PF= 0 vq=0 1 2 3 > > > > > > > > > > [a]request_count - 839416590 813148916 105586179 84988123 > > > > > > > > > > [b]completion1_count - 839416590 813148916 105586179 84988123 > > > > > > > > > > [c]completion2_count - 0 0 0 0 > > > > > > > > > > > > > > > > > > > > PF= 1 vq=0 1 2 3 > > > > > > > > > > [a]request_count - 823335887 812516140 104582672 75856549 > > > > > > > > > > [b]completion1_count - 823335887 812516140 104582672 75856549 > > > > > > > > > > [c]completion2_count - 0 0 0 0 > > > > > > > > > > > > > > > > > > > > i.e. the issue is after the virtio-blk driver. > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > This change was introduced in kernel 6.3.0. > > > > > > > > > > I am seeing this using 6.3.3. > > > > > > > > > > If I run with an earlier kernel (5.15), it does not occur. > > > > > > > > > > If I make a simple patch to the 6.3.3 virtio-blk driver, to skip the blk_mq_add_to_batch()call, it does not fail. > > > > > > > > > > e.g. > > > > > > > > > > kernel 5.15 ? this is OK > > > > > > > > > > virtio_blk.c,virtblk_done() [irq handler] > > > > > > > > > > if (likely(!blk_should_fake_timeout(req->q))) { > > > > > > > > > > blk_mq_complete_request(req); > > > > > > > > > > } > > > > > > > > > > > > > > > > > > > > kernel 6.3.3 ? this fails > > > > > > > > > > virtio_blk.c,virtblk_handle_req() [irq handler] > > > > > > > > > > if (likely(!blk_should_fake_timeout(req->q))) { > > > > > > > > > > if (!blk_mq_complete_request_remote(req)) { > > > > > > > > > > if (!blk_mq_add_to_batch(req, iob, virtblk_vbr_status(vbr), virtblk_complete_batch)) { > > > > > > > > > > virtblk_request_done(req); //this never gets called... so blk_mq_add_to_batch() must always succeed > > > > > > > > > > } > > > > > > > > > > } > > > > > > > > > > } > > > > > > > > > > > > > > > > > > > > If I do, kernel 6.3.3 ? this is OK > > > > > > > > > > virtio_blk.c,virtblk_handle_req() [irq handler] > > > > > > > > > > if (likely(!blk_should_fake_timeout(req->q))) { > > > > > > > > > > if (!blk_mq_complete_request_remote(req)) { > > > > > > > > > > virtblk_request_done(req); //force this here... > > > > > > > > > > if (!blk_mq_add_to_batch(req, iob, virtblk_vbr_status(vbr), virtblk_complete_batch)) { > > > > > > > > > > virtblk_request_done(req); //this never gets called... so blk_mq_add_to_batch() must always succeed > > > > > > > > > > } > > > > > > > > > > } > > > > > > > > > > } > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > Perhaps you might like to fix/test/revert this change? > > > > > > > > > > Martin > > > > > > > > > > > > > > > > > > Hi Martin, > > > > > > > > There are many changes between 6.3.0 and 6.3.3. > > > > Could you try to find a commit which triggers the io hang? > > > > Is it ok with 6.3.0 kernel or with reverting > > > > "virtio-blk: support completion batching for the IRQ path" commit? > > > > > > > > We need to confirm which commit is causing the error. > > > > > > > > Regards, > > > > Suwan Kim