On Fri, Jan 12, 2024 at 11:20:04AM +0100, Jan Kara wrote: > On Fri 12-01-24 17:27:48, Kemeng Shi wrote: > > > > > > on 1/11/2024 11:54 PM, Ming Lei wrote: > > > In blk_mq_mark_tag_wait(), __add_wait_queue() may be re-ordered > > > with the following blk_mq_get_driver_tag() in case of getting driver > > > tag failure. > > > > > > Then in __sbitmap_queue_wake_up(), waitqueue_active() may not observe > > > the added waiter in blk_mq_mark_tag_wait() and wake up nothing, meantime > > > blk_mq_mark_tag_wait() can't get driver tag successfully. > > > > > > This issue can be reproduced by running the following test in loop, and > > > fio hang can be observed in < 30min when running it on my test VM > > > in laptop. > > > > > > modprobe -r scsi_debug > > > modprobe scsi_debug delay=0 dev_size_mb=4096 max_queue=1 host_max_queue=1 submit_queues=4 > > > dev=`ls -d /sys/bus/pseudo/drivers/scsi_debug/adapter*/host*/target*/*/block/* | head -1 | xargs basename` > > > fio --filename=/dev/"$dev" --direct=1 --rw=randrw --bs=4k --iodepth=1 \ > > > --runtime=100 --numjobs=40 --time_based --name=test \ > > > --ioengine=libaio > > > > > > Fix the issue by adding one explicit barrier in blk_mq_mark_tag_wait(), which > > > is just fine in case of running out of tag. > > > > > > Apply the same pattern in blk_mq_get_tag() which should have same risk. > > > > > > Reported-by: Changhui Zhong <czhong@xxxxxxxxxx> > > > Signed-off-by: Ming Lei <ming.lei@xxxxxxxxxx> > > > --- > > > BTW, Changhui is planning to upstream the test case to blktests. > > > > > > block/blk-mq-tag.c | 19 +++++++++++++++++++ > > > block/blk-mq.c | 16 ++++++++++++++++ > > > 2 files changed, 35 insertions(+) > > > > > > diff --git a/block/blk-mq-tag.c b/block/blk-mq-tag.c > > > index cc57e2dd9a0b..29f77cae8eb2 100644 > > > --- a/block/blk-mq-tag.c > > > +++ b/block/blk-mq-tag.c > > > @@ -179,6 +179,25 @@ unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data) > > > > > > sbitmap_prepare_to_wait(bt, ws, &wait, TASK_UNINTERRUPTIBLE); > > > > > > + /* > > > + * Add one explicit barrier since __blk_mq_get_tag() may not > > > + * imply barrier in case of failure. > > > + * > > > + * Order adding us to wait queue and the following allocating > > > + * tag in __blk_mq_get_tag(). > > > + * > > > + * The pair is the one implied in sbitmap_queue_wake_up() > > > + * which orders clearing sbitmap tag bits and > > > + * waitqueue_active() in __sbitmap_queue_wake_up(), since > > > + * waitqueue_active() is lockless > > > + * > > > + * Otherwise, re-order of adding wait queue and getting tag > > > + * may cause __sbitmap_queue_wake_up() to wake up nothing > > > + * because the waitqueue_active() may not observe us in wait > > > + * queue. > > > + */ > > > + smp_mb(); > > > + > > Hi Ming, thanks for the fix. I'm not sure if we should explicitly imply > > a memory barrier here as prepare_to_wait variants normally imply a general > > memory barrier (see section "SLEEP AND WAKE-UP FUNCTIONS " in [1]). > > Wish this helps! > > Indeed, good spotting with the ordering bug Ming! I agree with Kemeng > though that set_current_state() called from sbitmap_prepare_to_wait() is > guaranteed to contain a memory barrier and thus reads from > __blk_mq_get_tag() are guaranteed to be ordered properly wrt addition into > the waitqueue. > > So only blk_mq_mark_tag_wait() is vulnerable to the problem you have > spotted AFAICT. Indeed, I will remove the one in blk_mq_get_tag() in V2. thanks, Ming
