On Fri, Jun 01, 2018 at 11:52:45AM -0600, Jens Axboe wrote:
> On 6/1/18 4:10 AM, Tetsuo Handa wrote:
> > Tetsuo Handa wrote:
> >> Since sum of percpu_count did not change after percpu_ref_kill(), this is
> >> not a race condition while folding percpu counter values into atomic counter
> >> value. That is, for some reason, someone who is responsible for calling
> >> percpu_ref_put(&q->q_usage_counter) (presumably via blk_queue_exit()) is
> >> unable to call percpu_ref_put().
> >> But I don't know how to find someone who is failing to call percpu_ref_put()...
> >
> > I found the someone. It was already there in the backtrace...
> >
> > ----------------------------------------
> > [ 62.065852] a.out D 0 4414 4337 0x00000000
> > [ 62.067677] Call Trace:
> > [ 62.068545] __schedule+0x40b/0x860
> > [ 62.069726] schedule+0x31/0x80
> > [ 62.070796] schedule_timeout+0x1c1/0x3c0
> > [ 62.072159] ? __next_timer_interrupt+0xd0/0xd0
> > [ 62.073670] blk_queue_enter+0x218/0x520
> > [ 62.074985] ? remove_wait_queue+0x70/0x70
> > [ 62.076361] generic_make_request+0x3d/0x540
> > [ 62.077785] ? __bio_clone_fast+0x6b/0x80
> > [ 62.079147] ? bio_clone_fast+0x2c/0x70
> > [ 62.080456] blk_queue_split+0x29b/0x560
> > [ 62.081772] ? blk_queue_split+0x29b/0x560
> > [ 62.083162] blk_mq_make_request+0x7c/0x430
> > [ 62.084562] generic_make_request+0x276/0x540
> > [ 62.086034] submit_bio+0x6e/0x140
> > [ 62.087185] ? submit_bio+0x6e/0x140
> > [ 62.088384] ? guard_bio_eod+0x9d/0x1d0
> > [ 62.089681] do_mpage_readpage+0x328/0x730
> > [ 62.091045] ? __add_to_page_cache_locked+0x12e/0x1a0
> > [ 62.092726] mpage_readpages+0x120/0x190
> > [ 62.094034] ? check_disk_change+0x70/0x70
> > [ 62.095454] ? check_disk_change+0x70/0x70
> > [ 62.096849] ? alloc_pages_current+0x65/0xd0
> > [ 62.098277] blkdev_readpages+0x18/0x20
> > [ 62.099568] __do_page_cache_readahead+0x298/0x360
> > [ 62.101157] ondemand_readahead+0x1f6/0x490
> > [ 62.102546] ? ondemand_readahead+0x1f6/0x490
> > [ 62.103995] page_cache_sync_readahead+0x29/0x40
> > [ 62.105539] generic_file_read_iter+0x7d0/0x9d0
> > [ 62.107067] ? futex_wait+0x221/0x240
> > [ 62.108303] ? trace_hardirqs_on+0xd/0x10
> > [ 62.109654] blkdev_read_iter+0x30/0x40
> > [ 62.110954] generic_file_splice_read+0xc5/0x140
> > [ 62.112538] do_splice_to+0x74/0x90
> > [ 62.113726] splice_direct_to_actor+0xa4/0x1f0
> > [ 62.115209] ? generic_pipe_buf_nosteal+0x10/0x10
> > [ 62.116773] do_splice_direct+0x8a/0xb0
> > [ 62.118056] do_sendfile+0x1aa/0x390
> > [ 62.119255] __x64_sys_sendfile64+0x4e/0xc0
> > [ 62.120666] do_syscall_64+0x6e/0x210
> > [ 62.121909] entry_SYSCALL_64_after_hwframe+0x49/0xbe
> > ----------------------------------------
> >
> > The someone is blk_queue_split() from blk_mq_make_request() who depends on an
> > assumption that blk_queue_enter() from recursively called generic_make_request()
> > does not get blocked due to percpu_ref_tryget_live(&q->q_usage_counter) failure.
> >
> > ----------------------------------------
> > generic_make_request(struct bio *bio) {
> > if (blk_queue_enter(q, flags) < 0) { /* <= percpu_ref_tryget_live() succeeds. */
> > if (!blk_queue_dying(q) && (bio->bi_opf & REQ_NOWAIT))
> > bio_wouldblock_error(bio);
> > else
> > bio_io_error(bio);
> > return ret;
> > }
> > (...snipped...)
> > ret = q->make_request_fn(q, bio);
> > (...snipped...)
> > if (q)
> > blk_queue_exit(q);
> > }
> > ----------------------------------------
> >
> > where q->make_request_fn == blk_mq_make_request which does
> >
> > ----------------------------------------
> > blk_mq_make_request(struct request_queue *q, struct bio *bio) {
> > blk_queue_split(q, &bio);
> > }
> >
> > blk_queue_split(struct request_queue *q, struct bio **bio) {
> > generic_make_request(*bio); /* <= percpu_ref_tryget_live() fails and waits until atomic_read(&q->mq_freeze_depth) becomes 0. */
> > }
> > ----------------------------------------
> >
> > and meanwhile atomic_inc_return(&q->mq_freeze_depth) and
> > percpu_ref_kill() are called by blk_freeze_queue_start()...
> >
> > Now, it is up to you about how to fix this race problem.
>
> Ahh, nicely spotted! One idea would be the one below. For this case,
> we're recursing, so we can either do a non-block queue enter, or we
> can just do a live enter.
>
> diff --git a/block/blk-core.c b/block/blk-core.c
> index 85909b431eb0..7de12e0dcc3d 100644
> --- a/block/blk-core.c
> +++ b/block/blk-core.c
> @@ -2392,7 +2392,9 @@ blk_qc_t generic_make_request(struct bio *bio)
>
> if (bio->bi_opf & REQ_NOWAIT)
> flags = BLK_MQ_REQ_NOWAIT;
> - if (blk_queue_enter(q, flags) < 0) {
> + if (bio->bi_opf & REQ_QUEUE_ENTERED)
> + blk_queue_enter_live(q);
> + else if (blk_queue_enter(q, flags) < 0) {
> if (!blk_queue_dying(q) && (bio->bi_opf & REQ_NOWAIT))
> bio_wouldblock_error(bio);
> else
> diff --git a/block/blk-merge.c b/block/blk-merge.c
> index 782940c65d8a..e1063e8f7eda 100644
> --- a/block/blk-merge.c
> +++ b/block/blk-merge.c
> @@ -210,6 +210,16 @@ void blk_queue_split(struct request_queue *q, struct bio **bio)
> /* there isn't chance to merge the splitted bio */
> split->bi_opf |= REQ_NOMERGE;
>
> + /*
> + * Since we're recursing into make_request here, ensure
> + * that we mark this bio as already having entered the queue.
> + * If not, and the queue is going away, we can get stuck
> + * forever on waiting for the queue reference to drop. But
> + * that will never happen, as we're already holding a
> + * reference to it.
> + */
> + (*bio)->bi_opf |= REQ_QUEUE_ENTERED;
> +
> bio_chain(split, *bio);
> trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
> generic_make_request(*bio);
> diff --git a/include/linux/blk_types.h b/include/linux/blk_types.h
> index 17b18b91ebac..a6bc789b83e7 100644
> --- a/include/linux/blk_types.h
> +++ b/include/linux/blk_types.h
> @@ -282,6 +282,8 @@ enum req_flag_bits {
> /* command specific flags for REQ_OP_WRITE_ZEROES: */
> __REQ_NOUNMAP, /* do not free blocks when zeroing */
>
> + __REQ_QUEUE_ENTERED, /* queue already entered for this request */
> +
> /* for driver use */
> __REQ_DRV,
>
> @@ -302,9 +304,8 @@ enum req_flag_bits {
> #define REQ_RAHEAD (1ULL << __REQ_RAHEAD)
> #define REQ_BACKGROUND (1ULL << __REQ_BACKGROUND)
> #define REQ_NOWAIT (1ULL << __REQ_NOWAIT)
> -
> #define REQ_NOUNMAP (1ULL << __REQ_NOUNMAP)
> -
> +#define REQ_QUEUE_ENTERED (1ULL << __REQ_QUEUE_ENTERED)
> #define REQ_DRV (1ULL << __REQ_DRV)
>
> #define REQ_FAILFAST_MASK \
bio clone inherits .bi_opf, so this way may not work for cloned bio.
Given most of bio_split() is run in .make_request_fn(), holding the
queue ref may not be needed in case of bio splitting, so another
candidate fix might be to introduce one extra parameter of 'need_queue_ref'
to generic_make_request().
Thanks,
Ming
