On Sun, 2019-04-28 at 16:14 +-0800, Ming Lei wrote: +AD4 Just like aio/io+AF8-uring, we need to grab 2 refcount for queuing one +AD4 request, one is for submission, another is for completion. +AD4 +AD4 If the request isn't queued from plug code path, the refcount grabbed +AD4 in generic+AF8-make+AF8-request() serves for submission. In theroy, this +AD4 refcount should have been released after the sumission(async run queue) +AD4 is done. blk+AF8-freeze+AF8-queue() works with blk+AF8-sync+AF8-queue() together +AD4 for avoiding race between cleanup queue and IO submission, given async +AD4 run queue activities are canceled because hctx-+AD4-run+AF8-work is scheduled with +AD4 the refcount held, so it is fine to not hold the refcount when +AD4 running the run queue work function for dispatch IO. +AD4 +AD4 However, if request is staggered into plug list, and finally queued +AD4 from plug code path, the refcount in submission side is actually missed. +AD4 And we may start to run queue after queue is removed because the queue's +AD4 kobject refcount isn't guaranteed to be grabbed in flushing plug list +AD4 context, then kernel oops is triggered, see the following race: +AD4 +AD4 blk+AF8-mq+AF8-flush+AF8-plug+AF8-list(): +AD4 blk+AF8-mq+AF8-sched+AF8-insert+AF8-requests() +AD4 insert requests to sw queue or scheduler queue +AD4 blk+AF8-mq+AF8-run+AF8-hw+AF8-queue +AD4 +AD4 Because of concurrent run queue, all requests inserted above may be +AD4 completed before calling the above blk+AF8-mq+AF8-run+AF8-hw+AF8-queue. Then queue can +AD4 be freed during the above blk+AF8-mq+AF8-run+AF8-hw+AF8-queue(). +AD4 +AD4 Fixes the issue by grab .q+AF8-usage+AF8-counter before calling +AD4 blk+AF8-mq+AF8-sched+AF8-insert+AF8-requests() in blk+AF8-mq+AF8-flush+AF8-plug+AF8-list(). This way is +AD4 safe because the queue is absolutely alive before inserting request. +AD4 +AD4 Cc: Dongli Zhang +ADw-dongli.zhang+AEA-oracle.com+AD4 +AD4 Cc: James Smart +ADw-james.smart+AEA-broadcom.com+AD4 +AD4 Cc: Bart Van Assche +ADw-bart.vanassche+AEA-wdc.com+AD4 +AD4 Cc: linux-scsi+AEA-vger.kernel.org, +AD4 Cc: Martin K . Petersen +ADw-martin.petersen+AEA-oracle.com+AD4, +AD4 Cc: Christoph Hellwig +ADw-hch+AEA-lst.de+AD4, +AD4 Cc: James E . J . Bottomley +ADw-jejb+AEA-linux.vnet.ibm.com+AD4, +AD4 Reviewed-by: Bart Van Assche +ADw-bvanassche+AEA-acm.org+AD4 +AD4 Reviewed-by: Johannes Thumshirn +ADw-jthumshirn+AEA-suse.de+AD4 +AD4 Reviewed-by: Hannes Reinecke +ADw-hare+AEA-suse.com+AD4 +AD4 Tested-by: James Smart +ADw-james.smart+AEA-broadcom.com+AD4 +AD4 Signed-off-by: Ming Lei +ADw-ming.lei+AEA-redhat.com+AD4 I added my +ACI-Reviewed-by+ACI to a previous version of this patch but not to this version of this patch. Several +ACI-Reviewed-by+ACI tags probably should be removed. +AD4 diff --git a/block/blk-mq-sched.c b/block/blk-mq-sched.c +AD4 index aa6bc5c02643..dfe83e7935d6 100644 +AD4 --- a/block/blk-mq-sched.c +AD4 +-+-+- b/block/blk-mq-sched.c +AD4 +AEAAQA -414,6 +-414,13 +AEAAQA void blk+AF8-mq+AF8-sched+AF8-insert+AF8-requests(struct blk+AF8-mq+AF8-hw+AF8-ctx +ACo-hctx, +AD4 +AHs +AD4 struct elevator+AF8-queue +ACo-e+ADs +AD4 +AD4 +- /+ACo +AD4 +- +ACo blk+AF8-mq+AF8-sched+AF8-insert+AF8-requests() is called from flush plug +AD4 +- +ACo context only, and hold one usage counter to prevent queue +AD4 +- +ACo from being released. +AD4 +- +ACo-/ +AD4 +- percpu+AF8-ref+AF8-get(+ACY-hctx-+AD4-queue-+AD4-q+AF8-usage+AF8-counter)+ADs +AD4 +- +AD4 e +AD0 hctx-+AD4-queue-+AD4-elevator+ADs +AD4 if (e +ACYAJg e-+AD4-type-+AD4-ops.insert+AF8-requests) +AD4 e-+AD4-type-+AD4-ops.insert+AF8-requests(hctx, list, false)+ADs +AD4 +AEAAQA -432,6 +-439,8 +AEAAQA void blk+AF8-mq+AF8-sched+AF8-insert+AF8-requests(struct blk+AF8-mq+AF8-hw+AF8-ctx +ACo-hctx, +AD4 +AH0 +AD4 +AD4 blk+AF8-mq+AF8-run+AF8-hw+AF8-queue(hctx, run+AF8-queue+AF8-async)+ADs +AD4 +- +AD4 +- percpu+AF8-ref+AF8-put(+ACY-hctx-+AD4-queue-+AD4-q+AF8-usage+AF8-counter)+ADs +AD4 +AH0 I think that 'hctx' can disappear if all requests queued by this function finish just before blk+AF8-mq+AF8-run+AF8-hw+AF8-queue() returns and if the number of hardware queues is changed from another thread. Shouldn't the request queue pointer be stored in a local variable instead of reading hctx-+AD4-queue twice? Bart.
