From: "jianchao.wang" <jianchao.w.wang@oracle.com> To: Ming Lei <ming.lei@redhat.com> Cc: Jens Axboe <axboe@kernel.dk>, linux-block@vger.kernel.org, Sagi Grimberg <sagi@grimberg.me>, linux-nvme@lists.infradead.org, Keith Busch <keith.busch@intel.com>, Christoph Hellwig <hch@lst.de> Subject: Re: [PATCH 1/2] nvme: pci: simplify timeout handling Date: Fri, 27 Apr 2018 09:37:06 +0800 [thread overview] Message-ID: <325688af-3ae2-49db-3a59-ef3903adcdf6@oracle.com> (raw) In-Reply-To: <20180426155722.GA3597@ming.t460p> On 04/26/2018 11:57 PM, Ming Lei wrote: > Hi Jianchao, > > On Thu, Apr 26, 2018 at 11:07:56PM +0800, jianchao.wang wrote: >> Hi Ming >> >> Thanks for your wonderful solution. :) >> >> On 04/26/2018 08:39 PM, Ming Lei wrote: >>> +/* >>> + * This one is called after queues are quiesced, and no in-fligh timeout >>> + * and nvme interrupt handling. >>> + */ >>> +static void nvme_pci_cancel_request(struct request *req, void *data, >>> + bool reserved) >>> +{ >>> + /* make sure timed-out requests are covered too */ >>> + if (req->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) { >>> + req->aborted_gstate = 0; >>> + req->rq_flags &= ~RQF_MQ_TIMEOUT_EXPIRED; >>> + } >>> + >>> + nvme_cancel_request(req, data, reserved); >>> +} >>> + >>> static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown) >>> { >>> int i; >>> @@ -2223,10 +2316,17 @@ static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown) >>> for (i = dev->ctrl.queue_count - 1; i >= 0; i--) >>> nvme_suspend_queue(&dev->queues[i]); >>> >>> + /* >>> + * safe to sync timeout after queues are quiesced, then all >>> + * requests(include the time-out ones) will be canceled. >>> + */ >>> + nvme_sync_queues(&dev->ctrl); >>> + blk_sync_queue(dev->ctrl.admin_q); >>> + >> Looks like blk_sync_queue cannot drain all the timeout work. >> >> blk_sync_queue >> -> del_timer_sync >> blk_mq_timeout_work >> -> mod_timer >> -> cancel_work_sync >> the timeout work may come back again. >> we may need to force all the in-flight requests to be timed out with blk_abort_request >> > > blk_abort_request() seems over-kill, we could avoid this race simply by > returning EH_NOT_HANDLED if the controller is in-recovery. return EH_NOT_HANDLED maybe not enough. please consider the following scenario. nvme_error_handler -> nvme_dev_disable -> blk_sync_queue //timeout comes again due to the //scenario above blk_mq_timeout_work -> blk_mq_check_expired -> set aborted_gstate -> nvme_pci_cancel_request -> RQF_MQ_TIMEOUT_EXPIRED has not been set -> nvme_cancel_request -> blk_mq_complete_request -> do nothing -> blk_mq_ternimate_expired -> blk_mq_rq_timed_out -> set RQF_MQ_TIMEOUT_EXPIRED -> .timeout return BLK_EH_NOT_HANDLED Then the timeout request is leaked. > >>> nvme_pci_disable(dev); >> >> the interrupt will not come, but there maybe running one. >> a synchronize_sched() here ? > > We may cover this case by moving nvme_suspend_queue() before > nvme_stop_queues(). > > Both two are very good catch, thanks! >
WARNING: multiple messages have this Message-ID (diff)
From: jianchao.w.wang@oracle.com (jianchao.wang) Subject: [PATCH 1/2] nvme: pci: simplify timeout handling Date: Fri, 27 Apr 2018 09:37:06 +0800 [thread overview] Message-ID: <325688af-3ae2-49db-3a59-ef3903adcdf6@oracle.com> (raw) In-Reply-To: <20180426155722.GA3597@ming.t460p> On 04/26/2018 11:57 PM, Ming Lei wrote: > Hi Jianchao, > > On Thu, Apr 26, 2018@11:07:56PM +0800, jianchao.wang wrote: >> Hi Ming >> >> Thanks for your wonderful solution. :) >> >> On 04/26/2018 08:39 PM, Ming Lei wrote: >>> +/* >>> + * This one is called after queues are quiesced, and no in-fligh timeout >>> + * and nvme interrupt handling. >>> + */ >>> +static void nvme_pci_cancel_request(struct request *req, void *data, >>> + bool reserved) >>> +{ >>> + /* make sure timed-out requests are covered too */ >>> + if (req->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) { >>> + req->aborted_gstate = 0; >>> + req->rq_flags &= ~RQF_MQ_TIMEOUT_EXPIRED; >>> + } >>> + >>> + nvme_cancel_request(req, data, reserved); >>> +} >>> + >>> static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown) >>> { >>> int i; >>> @@ -2223,10 +2316,17 @@ static void nvme_dev_disable(struct nvme_dev *dev, bool shutdown) >>> for (i = dev->ctrl.queue_count - 1; i >= 0; i--) >>> nvme_suspend_queue(&dev->queues[i]); >>> >>> + /* >>> + * safe to sync timeout after queues are quiesced, then all >>> + * requests(include the time-out ones) will be canceled. >>> + */ >>> + nvme_sync_queues(&dev->ctrl); >>> + blk_sync_queue(dev->ctrl.admin_q); >>> + >> Looks like blk_sync_queue cannot drain all the timeout work. >> >> blk_sync_queue >> -> del_timer_sync >> blk_mq_timeout_work >> -> mod_timer >> -> cancel_work_sync >> the timeout work may come back again. >> we may need to force all the in-flight requests to be timed out with blk_abort_request >> > > blk_abort_request() seems over-kill, we could avoid this race simply by > returning EH_NOT_HANDLED if the controller is in-recovery. return EH_NOT_HANDLED maybe not enough. please consider the following scenario. nvme_error_handler -> nvme_dev_disable -> blk_sync_queue //timeout comes again due to the //scenario above blk_mq_timeout_work -> blk_mq_check_expired -> set aborted_gstate -> nvme_pci_cancel_request -> RQF_MQ_TIMEOUT_EXPIRED has not been set -> nvme_cancel_request -> blk_mq_complete_request -> do nothing -> blk_mq_ternimate_expired -> blk_mq_rq_timed_out -> set RQF_MQ_TIMEOUT_EXPIRED -> .timeout return BLK_EH_NOT_HANDLED Then the timeout request is leaked. > >>> nvme_pci_disable(dev); >> >> the interrupt will not come, but there maybe running one. >> a synchronize_sched() here ? > > We may cover this case by moving nvme_suspend_queue() before > nvme_stop_queues(). > > Both two are very good catch, thanks! >
next prev parent reply other threads:[~2018-04-27 1:37 UTC|newest] Thread overview: 72+ messages / expand[flat|nested] mbox.gz Atom feed top 2018-04-26 12:39 [PATCH 0/2] nvme: pci: fix & improve timeout handling Ming Lei 2018-04-26 12:39 ` Ming Lei 2018-04-26 12:39 ` [PATCH 1/2] nvme: pci: simplify " Ming Lei 2018-04-26 12:39 ` Ming Lei 2018-04-26 15:07 ` jianchao.wang 2018-04-26 15:07 ` jianchao.wang 2018-04-26 15:57 ` Ming Lei 2018-04-26 15:57 ` Ming Lei 2018-04-26 16:16 ` Ming Lei 2018-04-26 16:16 ` Ming Lei 2018-04-27 1:37 ` jianchao.wang [this message] 2018-04-27 1:37 ` jianchao.wang 2018-04-27 14:57 ` Ming Lei 2018-04-27 14:57 ` Ming Lei 2018-04-28 14:00 ` jianchao.wang 2018-04-28 14:00 ` jianchao.wang 2018-04-28 21:57 ` Ming Lei 2018-04-28 21:57 ` Ming Lei 2018-04-28 22:27 ` Ming Lei 2018-04-28 22:27 ` Ming Lei 2018-04-29 1:36 ` Ming Lei 2018-04-29 1:36 ` Ming Lei 2018-04-29 2:21 ` jianchao.wang 2018-04-29 2:21 ` jianchao.wang 2018-04-29 14:13 ` Ming Lei 2018-04-29 14:13 ` Ming Lei 2018-04-27 17:51 ` Keith Busch 2018-04-27 17:51 ` Keith Busch 2018-04-28 3:50 ` Ming Lei 2018-04-28 3:50 ` Ming Lei 2018-04-28 13:35 ` Keith Busch 2018-04-28 13:35 ` Keith Busch 2018-04-28 14:31 ` jianchao.wang 2018-04-28 14:31 ` jianchao.wang 2018-04-28 21:39 ` Ming Lei 2018-04-28 21:39 ` Ming Lei 2018-04-30 19:52 ` Keith Busch 2018-04-30 19:52 ` Keith Busch 2018-04-30 23:14 ` Ming Lei 2018-04-30 23:14 ` Ming Lei 2018-05-08 15:30 ` Keith Busch 2018-05-08 15:30 ` Keith Busch 2018-05-10 20:52 ` Ming Lei 2018-05-10 20:52 ` Ming Lei 2018-05-10 21:05 ` Keith Busch 2018-05-10 21:05 ` Keith Busch 2018-05-10 21:10 ` Ming Lei 2018-05-10 21:10 ` Ming Lei 2018-05-10 21:18 ` Keith Busch 2018-05-10 21:18 ` Keith Busch 2018-05-10 21:24 ` Ming Lei 2018-05-10 21:24 ` Ming Lei 2018-05-10 21:44 ` Keith Busch 2018-05-10 21:44 ` Keith Busch 2018-05-10 21:50 ` Ming Lei 2018-05-10 21:50 ` Ming Lei 2018-05-10 21:53 ` Ming Lei 2018-05-10 21:53 ` Ming Lei 2018-05-10 22:03 ` Ming Lei 2018-05-10 22:03 ` Ming Lei 2018-05-10 22:43 ` Keith Busch 2018-05-10 22:43 ` Keith Busch 2018-05-11 0:14 ` Ming Lei 2018-05-11 0:14 ` Ming Lei 2018-05-11 2:10 ` Ming Lei 2018-05-11 2:10 ` Ming Lei 2018-04-26 12:39 ` [PATCH 2/2] nvme: pci: guarantee EH can make progress Ming Lei 2018-04-26 12:39 ` Ming Lei 2018-04-26 16:24 ` Keith Busch 2018-04-26 16:24 ` Keith Busch 2018-04-28 3:28 ` Ming Lei 2018-04-28 3:28 ` Ming Lei
Reply instructions: You may reply publicly to this message via plain-text email using any one of the following methods: * Save the following mbox file, import it into your mail client, and reply-to-all from there: mbox Avoid top-posting and favor interleaved quoting: https://en.wikipedia.org/wiki/Posting_style#Interleaved_style * Reply using the --to, --cc, and --in-reply-to switches of git-send-email(1): git send-email \ --in-reply-to=325688af-3ae2-49db-3a59-ef3903adcdf6@oracle.com \ --to=jianchao.w.wang@oracle.com \ --cc=axboe@kernel.dk \ --cc=hch@lst.de \ --cc=keith.busch@intel.com \ --cc=linux-block@vger.kernel.org \ --cc=linux-nvme@lists.infradead.org \ --cc=ming.lei@redhat.com \ --cc=sagi@grimberg.me \ /path/to/YOUR_REPLY https://kernel.org/pub/software/scm/git/docs/git-send-email.html * If your mail client supports setting the In-Reply-To header via mailto: links, try the mailto: linkBe sure your reply has a Subject: header at the top and a blank line before the message body.
This is an external index of several public inboxes, see mirroring instructions on how to clone and mirror all data and code used by this external index.