[PATCHSET v3] blk-mq: reimplement timeout handling

[PATCHSET v3] blk-mq: reimplement timeout handling

[Tejun Heo]

Hello,

Changes from [v2]

- Possible extended looping around seqcount and u64_stat_sync fixed.

- Misplaced MQ_RQ_IDLE state setting fixed.

- RQF_MQ_TIMEOUT_EXPIRED added to prevent firing the same timeout
  multiple times.

- s/queue_rq_src/srcu/ patch added.

- Other misc changes.

Changes from [v1]

- BLK_EH_RESET_TIMER handling fixed.

- s/request->gstate_seqc/request->gstate_seq/

- READ_ONCE() added to blk_mq_rq_udpate_state().

- Removed left over blk_clear_rq_complete() invocation from
  blk_mq_rq_timed_out().

Currently, blk-mq timeout path synchronizes against the usual
issue/completion path using a complex scheme involving atomic
bitflags, REQ_ATOM_*, memory barriers and subtle memory coherence
rules.  Unfortunatley, it contains quite a few holes.

It's pretty easy to make blk_mq_check_expired() terminate a later
instance of a request.  If we induce 5 sec delay before
time_after_eq() test in blk_mq_check_expired(), shorten the timeout to
2s, and issue back-to-back large IOs, blk-mq starts timing out
requests spuriously pretty quickly.  Nothing actually timed out.  It
just made the call on a recycle instance of a request and then
terminated a later instance long after the original instance finished.
The scenario isn't theoretical either.

This patchset replaces the broken synchronization mechanism with a RCU
and generation number based one.  Please read the patch description of
the second path for more details.

This patchset contains the following six patches.

0001-blk-mq-protect-completion-path-with-RCU.patch
0002-blk-mq-replace-timeout-synchronization-with-a-RCU-an.patch
0003-blk-mq-use-blk_mq_rq_state-instead-of-testing-REQ_AT.patch
0004-blk-mq-make-blk_abort_request-trigger-timeout-path.patch
0005-blk-mq-remove-REQ_ATOM_COMPLETE-usages-from-blk-mq.patch
0006-blk-mq-remove-REQ_ATOM_STARTED.patch
0007-blk-mq-rename-blk_mq_hw_ctx-queue_rq_srcu-to-srcu.patch

and is available in the following git branch.

 git://git.kernel.org/pub/scm/linux/kernel/git/tj/misc.git blk-mq-timeout-v3

diffstat follows.  Thanks.

 block/blk-core.c       |    2 
 block/blk-mq-debugfs.c |    4 
 block/blk-mq.c         |  285 +++++++++++++++++++++++++++++--------------------
 block/blk-mq.h         |   49 ++++++++
 block/blk-timeout.c    |   11 +
 block/blk.h            |    7 -
 include/linux/blk-mq.h |    3 
 include/linux/blkdev.h |   25 ++++
 8 files changed, 257 insertions(+), 129 deletions(-)

--
tejun

[v2] http://lkml.kernel.org/r/20171010155441.753966-1-tj@kernel.org
[v1] http://lkml.kernel.org/r/20171209192525.982030-1-tj@kernel.org

[PATCH 1/7] blk-mq: protect completion path with RCU

[Tejun Heo]

Currently, blk-mq protects only the issue path with RCU.  This patch
puts the completion path under the same RCU protection.  This will be
used to synchronize issue/completion against timeout by later patches,
which will also add the comments.

Signed-off-by: Tejun Heo <tj@kernel.org>
---
 block/blk-mq.c | 16 ++++++++++++++--
 1 file changed, 14 insertions(+), 2 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 1109747..acf4fbb 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -568,11 +568,23 @@ static void __blk_mq_complete_request(struct request *rq)
 void blk_mq_complete_request(struct request *rq)
 {
 	struct request_queue *q = rq->q;
+	struct blk_mq_hw_ctx *hctx = blk_mq_map_queue(q, rq->mq_ctx->cpu);
+	int srcu_idx;
 
 	if (unlikely(blk_should_fake_timeout(q)))
 		return;
-	if (!blk_mark_rq_complete(rq))
-		__blk_mq_complete_request(rq);
+
+	if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
+		rcu_read_lock();
+		if (!blk_mark_rq_complete(rq))
+			__blk_mq_complete_request(rq);
+		rcu_read_unlock();
+	} else {
+		srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
+		if (!blk_mark_rq_complete(rq))
+			__blk_mq_complete_request(rq);
+		srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
+	}
 }
 EXPORT_SYMBOL(blk_mq_complete_request);
 
-- 
2.9.5

Re: [PATCH 1/7] blk-mq: protect completion path with RCU

[Christoph Hellwig]

Why do you need the srcu protection?  The completion path can never
sleep.

If there is a good reason to keep it please add commment, and
make the srcu variant a separate function only used by drivers that
need it instead of adding the conditional.

Re: [PATCH 1/7] blk-mq: protect completion path with RCU

[Tejun Heo]

Hello, Christoph.

On Fri, Dec 29, 2017 at 02:04:18AM -0800, Christoph Hellwig wrote:
> Why do you need the srcu protection?  The completion path can never
> sleep.
> 
> If there is a good reason to keep it please add commment, and
> make the srcu variant a separate function only used by drivers that
> need it instead of adding the conditional.

It's either that or adding protection against both srcu and rcu in the
counterpart.  It's easier / cleaner to just use single rcu construct
for a given queue.  Will add a comment.  Jens already has a patch
which factors out the conditionals.

Thanks.

-- 
tejun

[PATCH 2/7] blk-mq: replace timeout synchronization with a RCU and generation based scheme

[Tejun Heo]

Currently, blk-mq timeout path synchronizes against the usual
issue/completion path using a complex scheme involving atomic
bitflags, REQ_ATOM_*, memory barriers and subtle memory coherence
rules.  Unfortunately, it contains quite a few holes.

There's a complex dancing around REQ_ATOM_STARTED and
REQ_ATOM_COMPLETE between issue/completion and timeout paths; however,
they don't have a synchronization point across request recycle
instances and it isn't clear what the barriers add.
blk_mq_check_expired() can easily read STARTED from N-2'th iteration,
deadline from N-1'th, blk_mark_rq_complete() against Nth instance.

In fact, it's pretty easy to make blk_mq_check_expired() terminate a
later instance of a request.  If we induce 5 sec delay before
time_after_eq() test in blk_mq_check_expired(), shorten the timeout to
2s, and issue back-to-back large IOs, blk-mq starts timing out
requests spuriously pretty quickly.  Nothing actually timed out.  It
just made the call on a recycle instance of a request and then
terminated a later instance long after the original instance finished.
The scenario isn't theoretical either.

This patch replaces the broken synchronization mechanism with a RCU
and generation number based one.

1. Each request has a u64 generation + state value, which can be
   updated only by the request owner.  Whenever a request becomes
   in-flight, the generation number gets bumped up too.  This provides
   the basis for the timeout path to distinguish different recycle
   instances of the request.

   Also, marking a request in-flight and setting its deadline are
   protected with a seqcount so that the timeout path can fetch both
   values coherently.

2. The timeout path fetches the generation, state and deadline.  If
   the verdict is timeout, it records the generation into a dedicated
   request abortion field and does RCU wait.

3. The completion path is also protected by RCU (from the previous
   patch) and checks whether the current generation number and state
   match the abortion field.  If so, it skips completion.

4. The timeout path, after RCU wait, scans requests again and
   terminates the ones whose generation and state still match the ones
   requested for abortion.

   By now, the timeout path knows that either the generation number
   and state changed if it lost the race or the completion will yield
   to it and can safely timeout the request.

While it's more lines of code, it's conceptually simpler, doesn't
depend on direct use of subtle memory ordering or coherence, and
hopefully doesn't terminate the wrong instance.

While this change makes REQ_ATOM_COMPLETE synchronization unnecessary
between issue/complete and timeout paths, REQ_ATOM_COMPLETE isn't
removed yet as it's still used in other places.  Future patches will
move all state tracking to the new mechanism and remove all bitops in
the hot paths.

v2: - Fixed BLK_EH_RESET_TIMER handling as pointed out by Jianchao.
    - s/request->gstate_seqc/request->gstate_seq/ as suggested by Peter.
    - READ_ONCE() added in blk_mq_rq_update_state() as suggested by Peter.

v3: - Fixed possible extended seqcount / u64_stats_sync read looping
      spotted by Peter.
    - MQ_RQ_IDLE was incorrectly being set in complete_request instead
      of free_request.  Fixed.

Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: "jianchao.wang" <jianchao.w.wang@oracle.com>
Cc: Peter Zijlstra <peterz@infradead.org>
---
 block/blk-core.c       |   2 +
 block/blk-mq.c         | 220 +++++++++++++++++++++++++++++++++----------------
 block/blk-mq.h         |  46 +++++++++++
 block/blk-timeout.c    |   2 +-
 block/blk.h            |   6 --
 include/linux/blk-mq.h |   1 +
 include/linux/blkdev.h |  23 ++++++
 7 files changed, 220 insertions(+), 80 deletions(-)

diff --git a/block/blk-core.c b/block/blk-core.c
index b888175..6034623 100644
--- a/block/blk-core.c
+++ b/block/blk-core.c
@@ -126,6 +126,8 @@ void blk_rq_init(struct request_queue *q, struct request *rq)
 	rq->start_time = jiffies;
 	set_start_time_ns(rq);
 	rq->part = NULL;
+	seqcount_init(&rq->gstate_seq);
+	u64_stats_init(&rq->aborted_gstate_sync);
 }
 EXPORT_SYMBOL(blk_rq_init);
 
diff --git a/block/blk-mq.c b/block/blk-mq.c
index acf4fbb..abd5d01 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -483,6 +483,7 @@ void blk_mq_free_request(struct request *rq)
 	if (blk_rq_rl(rq))
 		blk_put_rl(blk_rq_rl(rq));
 
+	blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
 	clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 	clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
 	if (rq->tag != -1)
@@ -530,6 +531,8 @@ static void __blk_mq_complete_request(struct request *rq)
 	bool shared = false;
 	int cpu;
 
+	WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT);
+
 	if (rq->internal_tag != -1)
 		blk_mq_sched_completed_request(rq);
 	if (rq->rq_flags & RQF_STATS) {
@@ -557,6 +560,30 @@ static void __blk_mq_complete_request(struct request *rq)
 	put_cpu();
 }
 
+static void blk_mq_rq_update_aborted_gstate(struct request *rq, u64 gstate)
+{
+	unsigned long flags;
+
+	local_irq_save(flags);
+	u64_stats_update_begin(&rq->aborted_gstate_sync);
+	rq->aborted_gstate = gstate;
+	u64_stats_update_end(&rq->aborted_gstate_sync);
+	local_irq_restore(flags);
+}
+
+static u64 blk_mq_rq_aborted_gstate(struct request *rq)
+{
+	unsigned int start;
+	u64 aborted_gstate;
+
+	do {
+		start = u64_stats_fetch_begin(&rq->aborted_gstate_sync);
+		aborted_gstate = rq->aborted_gstate;
+	} while (u64_stats_fetch_retry(&rq->aborted_gstate_sync, start));
+
+	return aborted_gstate;
+}
+
 /**
  * blk_mq_complete_request - end I/O on a request
  * @rq:		the request being processed
@@ -574,14 +601,21 @@ void blk_mq_complete_request(struct request *rq)
 	if (unlikely(blk_should_fake_timeout(q)))
 		return;
 
+	/*
+	 * If @rq->aborted_gstate equals the current instance, timeout is
+	 * claiming @rq and we lost.  This is synchronized through RCU.
+	 * See blk_mq_timeout_work() for details.
+	 */
 	if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
 		rcu_read_lock();
-		if (!blk_mark_rq_complete(rq))
+		if (blk_mq_rq_aborted_gstate(rq) != rq->gstate &&
+		    !blk_mark_rq_complete(rq))
 			__blk_mq_complete_request(rq);
 		rcu_read_unlock();
 	} else {
 		srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
-		if (!blk_mark_rq_complete(rq))
+		if (blk_mq_rq_aborted_gstate(rq) != rq->gstate &&
+		    !blk_mark_rq_complete(rq))
 			__blk_mq_complete_request(rq);
 		srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
 	}
@@ -608,34 +642,32 @@ void blk_mq_start_request(struct request *rq)
 		wbt_issue(q->rq_wb, &rq->issue_stat);
 	}
 
-	blk_add_timer(rq);
-
+	WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
 	WARN_ON_ONCE(test_bit(REQ_ATOM_STARTED, &rq->atomic_flags));
 
 	/*
-	 * Mark us as started and clear complete. Complete might have been
-	 * set if requeue raced with timeout, which then marked it as
-	 * complete. So be sure to clear complete again when we start
-	 * the request, otherwise we'll ignore the completion event.
+	 * Mark @rq in-flight which also advances the generation number,
+	 * and register for timeout.  Protect with a seqcount to allow the
+	 * timeout path to read both @rq->gstate and @rq->deadline
+	 * coherently.
 	 *
-	 * Ensure that ->deadline is visible before we set STARTED, such that
-	 * blk_mq_check_expired() is guaranteed to observe our ->deadline when
-	 * it observes STARTED.
+	 * This is the only place where a request is marked in-flight.  If
+	 * the timeout path reads an in-flight @rq->gstate, the
+	 * @rq->deadline it reads together under @rq->gstate_seq is
+	 * guaranteed to be the matching one.
 	 */
-	smp_wmb();
+	preempt_disable();
+	write_seqcount_begin(&rq->gstate_seq);
+
+	blk_mq_rq_update_state(rq, MQ_RQ_IN_FLIGHT);
+	blk_add_timer(rq);
+
+	write_seqcount_end(&rq->gstate_seq);
+	preempt_enable();
+
 	set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
-	if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) {
-		/*
-		 * Coherence order guarantees these consecutive stores to a
-		 * single variable propagate in the specified order. Thus the
-		 * clear_bit() is ordered _after_ the set bit. See
-		 * blk_mq_check_expired().
-		 *
-		 * (the bits must be part of the same byte for this to be
-		 * true).
-		 */
+	if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
 		clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
-	}
 
 	if (q->dma_drain_size && blk_rq_bytes(rq)) {
 		/*
@@ -668,6 +700,7 @@ static void __blk_mq_requeue_request(struct request *rq)
 	blk_mq_sched_requeue_request(rq);
 
 	if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
+		blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
 		if (q->dma_drain_size && blk_rq_bytes(rq))
 			rq->nr_phys_segments--;
 	}
@@ -765,6 +798,7 @@ EXPORT_SYMBOL(blk_mq_tag_to_rq);
 struct blk_mq_timeout_data {
 	unsigned long next;
 	unsigned int next_set;
+	unsigned int nr_expired;
 };
 
 void blk_mq_rq_timed_out(struct request *req, bool reserved)
@@ -792,6 +826,12 @@ void blk_mq_rq_timed_out(struct request *req, bool reserved)
 		__blk_mq_complete_request(req);
 		break;
 	case BLK_EH_RESET_TIMER:
+		/*
+		 * As nothing prevents from completion happening while
+		 * ->aborted_gstate is set, this may lead to ignored
+		 * completions and further spurious timeouts.
+		 */
+		blk_mq_rq_update_aborted_gstate(req, 0);
 		blk_add_timer(req);
 		blk_clear_rq_complete(req);
 		break;
@@ -807,50 +847,51 @@ static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 		struct request *rq, void *priv, bool reserved)
 {
 	struct blk_mq_timeout_data *data = priv;
-	unsigned long deadline;
+	unsigned long gstate, deadline;
+	int start;
+
+	might_sleep();
 
 	if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 		return;
 
-	/*
-	 * Ensures that if we see STARTED we must also see our
-	 * up-to-date deadline, see blk_mq_start_request().
-	 */
-	smp_rmb();
-
-	deadline = READ_ONCE(rq->deadline);
+	/* read coherent snapshots of @rq->state_gen and @rq->deadline */
+	while (true) {
+		start = read_seqcount_begin(&rq->gstate_seq);
+		gstate = READ_ONCE(rq->gstate);
+		deadline = rq->deadline;
+		if (!read_seqcount_retry(&rq->gstate_seq, start))
+			break;
+		cond_resched();
+	}
 
-	/*
-	 * The rq being checked may have been freed and reallocated
-	 * out already here, we avoid this race by checking rq->deadline
-	 * and REQ_ATOM_COMPLETE flag together:
-	 *
-	 * - if rq->deadline is observed as new value because of
-	 *   reusing, the rq won't be timed out because of timing.
-	 * - if rq->deadline is observed as previous value,
-	 *   REQ_ATOM_COMPLETE flag won't be cleared in reuse path
-	 *   because we put a barrier between setting rq->deadline
-	 *   and clearing the flag in blk_mq_start_request(), so
-	 *   this rq won't be timed out too.
-	 */
-	if (time_after_eq(jiffies, deadline)) {
-		if (!blk_mark_rq_complete(rq)) {
-			/*
-			 * Again coherence order ensures that consecutive reads
-			 * from the same variable must be in that order. This
-			 * ensures that if we see COMPLETE clear, we must then
-			 * see STARTED set and we'll ignore this timeout.
-			 *
-			 * (There's also the MB implied by the test_and_clear())
-			 */
-			blk_mq_rq_timed_out(rq, reserved);
-		}
+	/* if in-flight && overdue, mark for abortion */
+	if ((gstate & MQ_RQ_STATE_MASK) == MQ_RQ_IN_FLIGHT &&
+	    time_after_eq(jiffies, deadline)) {
+		blk_mq_rq_update_aborted_gstate(rq, gstate);
+		data->nr_expired++;
+		hctx->nr_expired++;
 	} else if (!data->next_set || time_after(data->next, deadline)) {
 		data->next = deadline;
 		data->next_set = 1;
 	}
 }
 
+static void blk_mq_terminate_expired(struct blk_mq_hw_ctx *hctx,
+		struct request *rq, void *priv, bool reserved)
+{
+	/*
+	 * We marked @rq->aborted_gstate and waited for RCU.  If there were
+	 * completions that we lost to, they would have finished and
+	 * updated @rq->gstate by now; otherwise, the completion path is
+	 * now guaranteed to see @rq->aborted_gstate and yield.  If
+	 * @rq->aborted_gstate still matches @rq->gstate, @rq is ours.
+	 */
+	if (READ_ONCE(rq->gstate) == rq->aborted_gstate &&
+	    !blk_mark_rq_complete(rq))
+		blk_mq_rq_timed_out(rq, reserved);
+}
+
 static void blk_mq_timeout_work(struct work_struct *work)
 {
 	struct request_queue *q =
@@ -858,7 +899,9 @@ static void blk_mq_timeout_work(struct work_struct *work)
 	struct blk_mq_timeout_data data = {
 		.next		= 0,
 		.next_set	= 0,
+		.nr_expired	= 0,
 	};
+	struct blk_mq_hw_ctx *hctx;
 	int i;
 
 	/* A deadlock might occur if a request is stuck requiring a
@@ -877,14 +920,40 @@ static void blk_mq_timeout_work(struct work_struct *work)
 	if (!percpu_ref_tryget(&q->q_usage_counter))
 		return;
 
+	/* scan for the expired ones and set their ->aborted_gstate */
 	blk_mq_queue_tag_busy_iter(q, blk_mq_check_expired, &data);
 
+	if (data.nr_expired) {
+		bool has_rcu = false;
+
+		/*
+		 * Wait till everyone sees ->aborted_gstate.  The
+		 * sequential waits for SRCUs aren't ideal.  If this ever
+		 * becomes a problem, we can add per-hw_ctx rcu_head and
+		 * wait in parallel.
+		 */
+		queue_for_each_hw_ctx(q, hctx, i) {
+			if (!hctx->nr_expired)
+				continue;
+
+			if (!(hctx->flags & BLK_MQ_F_BLOCKING))
+				has_rcu = true;
+			else
+				synchronize_srcu(hctx->queue_rq_srcu);
+
+			hctx->nr_expired = 0;
+		}
+		if (has_rcu)
+			synchronize_rcu();
+
+		/* terminate the ones we won */
+		blk_mq_queue_tag_busy_iter(q, blk_mq_terminate_expired, NULL);
+	}
+
 	if (data.next_set) {
 		data.next = blk_rq_timeout(round_jiffies_up(data.next));
 		mod_timer(&q->timeout, data.next);
 	} else {
-		struct blk_mq_hw_ctx *hctx;
-
 		queue_for_each_hw_ctx(q, hctx, i) {
 			/* the hctx may be unmapped, so check it here */
 			if (blk_mq_hw_queue_mapped(hctx))
@@ -1879,6 +1948,22 @@ static size_t order_to_size(unsigned int order)
 	return (size_t)PAGE_SIZE << order;
 }
 
+static int blk_mq_init_request(struct blk_mq_tag_set *set, struct request *rq,
+			       unsigned int hctx_idx, int node)
+{
+	int ret;
+
+	if (set->ops->init_request) {
+		ret = set->ops->init_request(set, rq, hctx_idx, node);
+		if (ret)
+			return ret;
+	}
+
+	seqcount_init(&rq->gstate_seq);
+	u64_stats_init(&rq->aborted_gstate_sync);
+	return 0;
+}
+
 int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
 		     unsigned int hctx_idx, unsigned int depth)
 {
@@ -1940,12 +2025,9 @@ int blk_mq_alloc_rqs(struct blk_mq_tag_set *set, struct blk_mq_tags *tags,
 			struct request *rq = p;
 
 			tags->static_rqs[i] = rq;
-			if (set->ops->init_request) {
-				if (set->ops->init_request(set, rq, hctx_idx,
-						node)) {
-					tags->static_rqs[i] = NULL;
-					goto fail;
-				}
+			if (blk_mq_init_request(set, rq, hctx_idx, node)) {
+				tags->static_rqs[i] = NULL;
+				goto fail;
 			}
 
 			p += rq_size;
@@ -2084,9 +2166,7 @@ static int blk_mq_init_hctx(struct request_queue *q,
 	if (!hctx->fq)
 		goto sched_exit_hctx;
 
-	if (set->ops->init_request &&
-	    set->ops->init_request(set, hctx->fq->flush_rq, hctx_idx,
-				   node))
+	if (blk_mq_init_request(set, hctx->fq->flush_rq, hctx_idx, node))
 		goto free_fq;
 
 	if (hctx->flags & BLK_MQ_F_BLOCKING)
@@ -2980,12 +3060,6 @@ static bool blk_mq_poll(struct request_queue *q, blk_qc_t cookie)
 
 static int __init blk_mq_init(void)
 {
-	/*
-	 * See comment in block/blk.h rq_atomic_flags enum
-	 */
-	BUILD_BUG_ON((REQ_ATOM_STARTED / BITS_PER_BYTE) !=
-			(REQ_ATOM_COMPLETE / BITS_PER_BYTE));
-
 	cpuhp_setup_state_multi(CPUHP_BLK_MQ_DEAD, "block/mq:dead", NULL,
 				blk_mq_hctx_notify_dead);
 	return 0;
diff --git a/block/blk-mq.h b/block/blk-mq.h
index 6c7c3ff..cf01f6f 100644
--- a/block/blk-mq.h
+++ b/block/blk-mq.h
@@ -27,6 +27,19 @@ struct blk_mq_ctx {
 	struct kobject		kobj;
 } ____cacheline_aligned_in_smp;
 
+/*
+ * Bits for request->gstate.  The lower two bits carry MQ_RQ_* state value
+ * and the upper bits the generation number.
+ */
+enum mq_rq_state {
+	MQ_RQ_IDLE		= 0,
+	MQ_RQ_IN_FLIGHT		= 1,
+
+	MQ_RQ_STATE_BITS	= 2,
+	MQ_RQ_STATE_MASK	= (1 << MQ_RQ_STATE_BITS) - 1,
+	MQ_RQ_GEN_INC		= 1 << MQ_RQ_STATE_BITS,
+};
+
 void blk_mq_freeze_queue(struct request_queue *q);
 void blk_mq_free_queue(struct request_queue *q);
 int blk_mq_update_nr_requests(struct request_queue *q, unsigned int nr);
@@ -85,6 +98,39 @@ extern void blk_mq_rq_timed_out(struct request *req, bool reserved);
 
 void blk_mq_release(struct request_queue *q);
 
+/**
+ * blk_mq_rq_state() - read the current MQ_RQ_* state of a request
+ * @rq: target request.
+ */
+static inline int blk_mq_rq_state(struct request *rq)
+{
+	return READ_ONCE(rq->gstate) & MQ_RQ_STATE_MASK;
+}
+
+/**
+ * blk_mq_rq_update_state() - set the current MQ_RQ_* state of a request
+ * @rq: target request.
+ * @state: new state to set.
+ *
+ * Set @rq's state to @state.  The caller is responsible for ensuring that
+ * there are no other updaters.  A request can transition into IN_FLIGHT
+ * only from IDLE and doing so increments the generation number.
+ */
+static inline void blk_mq_rq_update_state(struct request *rq,
+					  enum mq_rq_state state)
+{
+	u64 old_val = READ_ONCE(rq->gstate);
+	u64 new_val = (old_val & ~MQ_RQ_STATE_MASK) | state;
+
+	if (state == MQ_RQ_IN_FLIGHT) {
+		WARN_ON_ONCE((old_val & MQ_RQ_STATE_MASK) != MQ_RQ_IDLE);
+		new_val += MQ_RQ_GEN_INC;
+	}
+
+	/* avoid exposing interim values */
+	WRITE_ONCE(rq->gstate, new_val);
+}
+
 static inline struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
 					   unsigned int cpu)
 {
diff --git a/block/blk-timeout.c b/block/blk-timeout.c
index 764ecf9..6427be7 100644
--- a/block/blk-timeout.c
+++ b/block/blk-timeout.c
@@ -208,7 +208,7 @@ void blk_add_timer(struct request *req)
 	if (!req->timeout)
 		req->timeout = q->rq_timeout;
 
-	WRITE_ONCE(req->deadline, jiffies + req->timeout);
+	req->deadline = jiffies + req->timeout;
 
 	/*
 	 * Only the non-mq case needs to add the request to a protected list.
diff --git a/block/blk.h b/block/blk.h
index 3f14469..9cb2739 100644
--- a/block/blk.h
+++ b/block/blk.h
@@ -123,12 +123,6 @@ void blk_account_io_done(struct request *req);
  * Internal atomic flags for request handling
  */
 enum rq_atomic_flags {
-	/*
-	 * Keep these two bits first - not because we depend on the
-	 * value of them, but we do depend on them being in the same
-	 * byte of storage to ensure ordering on writes. Keeping them
-	 * first will achieve that nicely.
-	 */
 	REQ_ATOM_COMPLETE = 0,
 	REQ_ATOM_STARTED,
 
diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
index 95c9a5c..460798d 100644
--- a/include/linux/blk-mq.h
+++ b/include/linux/blk-mq.h
@@ -51,6 +51,7 @@ struct blk_mq_hw_ctx {
 	unsigned int		queue_num;
 
 	atomic_t		nr_active;
+	unsigned int		nr_expired;
 
 	struct hlist_node	cpuhp_dead;
 	struct kobject		kobj;
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 8089ca1..2d6fd11 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -27,6 +27,8 @@
 #include <linux/percpu-refcount.h>
 #include <linux/scatterlist.h>
 #include <linux/blkzoned.h>
+#include <linux/seqlock.h>
+#include <linux/u64_stats_sync.h>
 
 struct module;
 struct scsi_ioctl_command;
@@ -228,6 +230,27 @@ struct request {
 
 	unsigned short write_hint;
 
+	/*
+	 * On blk-mq, the lower bits of ->gstate carry the MQ_RQ_* state
+	 * value and the upper bits the generation number which is
+	 * monotonically incremented and used to distinguish the reuse
+	 * instances.
+	 *
+	 * ->gstate_seq allows updates to ->gstate and other fields
+	 * (currently ->deadline) during request start to be read
+	 * atomically from the timeout path, so that it can operate on a
+	 * coherent set of information.
+	 */
+	seqcount_t gstate_seq;
+	u64 gstate;
+
+	/*
+	 * ->aborted_gstate is used by the timeout to claim a specific
+	 * recycle instance of this request.  See blk_mq_timeout_work().
+	 */
+	struct u64_stats_sync aborted_gstate_sync;
+	u64 aborted_gstate;
+
 	unsigned long deadline;
 	struct list_head timeout_list;
 
-- 
2.9.5

[PATCH 3/7] blk-mq: use blk_mq_rq_state() instead of testing REQ_ATOM_COMPLETE

[Tejun Heo]

blk_mq_check_inflight() and blk_mq_poll_hybrid_sleep() test
REQ_ATOM_COMPLETE to determine the request state.  Both uses are
speculative and we can test REQ_ATOM_STARTED and blk_mq_rq_state() for
equivalent results.  Replace the tests.  This will allow removing
REQ_ATOM_COMPLETE usages from blk-mq.

Signed-off-by: Tejun Heo <tj@kernel.org>
---
 block/blk-mq.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index abd5d01..643a38d 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -95,8 +95,7 @@ static void blk_mq_check_inflight(struct blk_mq_hw_ctx *hctx,
 {
 	struct mq_inflight *mi = priv;
 
-	if (test_bit(REQ_ATOM_STARTED, &rq->atomic_flags) &&
-	    !test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags)) {
+	if (blk_mq_rq_state(rq) == MQ_RQ_IN_FLIGHT) {
 		/*
 		 * index[0] counts the specific partition that was asked
 		 * for. index[1] counts the ones that are active on the
@@ -2974,7 +2973,8 @@ static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
 
 	hrtimer_init_sleeper(&hs, current);
 	do {
-		if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
+		if (test_bit(REQ_ATOM_STARTED, &rq->atomic_flags) &&
+		    blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
 			break;
 		set_current_state(TASK_UNINTERRUPTIBLE);
 		hrtimer_start_expires(&hs.timer, mode);
-- 
2.9.5

[PATCH 4/7] blk-mq: make blk_abort_request() trigger timeout path

[Tejun Heo]

With issue/complete and timeout paths now using the generation number
and state based synchronization, blk_abort_request() is the only one
which depends on REQ_ATOM_COMPLETE for arbitrating completion.

There's no reason for blk_abort_request() to be a completely separate
path.  This patch makes blk_abort_request() piggyback on the timeout
path instead of trying to terminate the request directly.

This removes the last dependency on REQ_ATOM_COMPLETE in blk-mq.

Note that this makes blk_abort_request() asynchronous - it initiates
abortion but the actual termination will happen after a short while,
even when the caller owns the request.  AFAICS, SCSI and ATA should be
fine with that and I think mtip32xx and dasd should be safe but not
completely sure.  It'd be great if people who know the drivers take a
look.

Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Asai Thambi SP <asamymuthupa@micron.com>
Cc: Stefan Haberland <sth@linux.vnet.ibm.com>
Cc: Jan Hoeppner <hoeppner@linux.vnet.ibm.com>
---
 block/blk-mq.c      | 2 +-
 block/blk-mq.h      | 2 --
 block/blk-timeout.c | 8 ++++----
 3 files changed, 5 insertions(+), 7 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 643a38d..88baa82 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -800,7 +800,7 @@ struct blk_mq_timeout_data {
 	unsigned int nr_expired;
 };
 
-void blk_mq_rq_timed_out(struct request *req, bool reserved)
+static void blk_mq_rq_timed_out(struct request *req, bool reserved)
 {
 	const struct blk_mq_ops *ops = req->q->mq_ops;
 	enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
diff --git a/block/blk-mq.h b/block/blk-mq.h
index cf01f6f..6b2d616 100644
--- a/block/blk-mq.h
+++ b/block/blk-mq.h
@@ -94,8 +94,6 @@ extern int blk_mq_sysfs_register(struct request_queue *q);
 extern void blk_mq_sysfs_unregister(struct request_queue *q);
 extern void blk_mq_hctx_kobj_init(struct blk_mq_hw_ctx *hctx);
 
-extern void blk_mq_rq_timed_out(struct request *req, bool reserved);
-
 void blk_mq_release(struct request_queue *q);
 
 /**
diff --git a/block/blk-timeout.c b/block/blk-timeout.c
index 6427be7..d580af3 100644
--- a/block/blk-timeout.c
+++ b/block/blk-timeout.c
@@ -156,12 +156,12 @@ void blk_timeout_work(struct work_struct *work)
  */
 void blk_abort_request(struct request *req)
 {
-	if (blk_mark_rq_complete(req))
-		return;
-
 	if (req->q->mq_ops) {
-		blk_mq_rq_timed_out(req, false);
+		req->deadline = jiffies;
+		mod_timer(&req->q->timeout, 0);
 	} else {
+		if (blk_mark_rq_complete(req))
+			return;
 		blk_delete_timer(req);
 		blk_rq_timed_out(req);
 	}
-- 
2.9.5

Re: [PATCH 4/7] blk-mq: make blk_abort_request() trigger timeout path

[Christoph Hellwig]

On Sat, Dec 16, 2017 at 04:07:23AM -0800, Tejun Heo wrote:
> Note that this makes blk_abort_request() asynchronous - it initiates
> abortion but the actual termination will happen after a short while,
> even when the caller owns the request.  AFAICS, SCSI and ATA should be
> fine with that and I think mtip32xx and dasd should be safe but not
> completely sure.  It'd be great if people who know the drivers take a
> look.

For that you'll need to CC linux-ide and linux-scsi, and for the
SAS drivers some of the usual suspects that touch the SAS code.

[PATCH 5/7] blk-mq: remove REQ_ATOM_COMPLETE usages from blk-mq

[Tejun Heo]

After the recent updates to use generation number and state based
synchronization, blk-mq no longer depends on REQ_ATOM_COMPLETE except
to avoid firing the same timeout multiple times.

Remove all REQ_ATOM_COMPLETE usages and use a new rq_flags flag
RQF_MQ_TIMEOUT_EXPIRED to avoid firing the same timeout multiple
times.  This removes atomic bitops from hot paths too.

v2: Removed blk_clear_rq_complete() from blk_mq_rq_timed_out().

v3: Added RQF_MQ_TIMEOUT_EXPIRED flag.

Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: "jianchao.wang" <jianchao.w.wang@oracle.com>
---
 block/blk-mq.c         | 18 ++++++++----------
 block/blk-timeout.c    |  1 +
 include/linux/blkdev.h |  2 ++
 3 files changed, 11 insertions(+), 10 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 88baa82..47e722b 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -607,14 +607,12 @@ void blk_mq_complete_request(struct request *rq)
 	 */
 	if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
 		rcu_read_lock();
-		if (blk_mq_rq_aborted_gstate(rq) != rq->gstate &&
-		    !blk_mark_rq_complete(rq))
+		if (blk_mq_rq_aborted_gstate(rq) != rq->gstate)
 			__blk_mq_complete_request(rq);
 		rcu_read_unlock();
 	} else {
 		srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
-		if (blk_mq_rq_aborted_gstate(rq) != rq->gstate &&
-		    !blk_mark_rq_complete(rq))
+		if (blk_mq_rq_aborted_gstate(rq) != rq->gstate)
 			__blk_mq_complete_request(rq);
 		srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
 	}
@@ -665,8 +663,6 @@ void blk_mq_start_request(struct request *rq)
 	preempt_enable();
 
 	set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
-	if (test_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags))
-		clear_bit(REQ_ATOM_COMPLETE, &rq->atomic_flags);
 
 	if (q->dma_drain_size && blk_rq_bytes(rq)) {
 		/*
@@ -817,6 +813,8 @@ static void blk_mq_rq_timed_out(struct request *req, bool reserved)
 	if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
 		return;
 
+	req->rq_flags |= RQF_MQ_TIMEOUT_EXPIRED;
+
 	if (ops->timeout)
 		ret = ops->timeout(req, reserved);
 
@@ -832,7 +830,6 @@ static void blk_mq_rq_timed_out(struct request *req, bool reserved)
 		 */
 		blk_mq_rq_update_aborted_gstate(req, 0);
 		blk_add_timer(req);
-		blk_clear_rq_complete(req);
 		break;
 	case BLK_EH_NOT_HANDLED:
 		break;
@@ -851,7 +848,8 @@ static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 
 	might_sleep();
 
-	if (!test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
+	if ((rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) ||
+	    !test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
 		return;
 
 	/* read coherent snapshots of @rq->state_gen and @rq->deadline */
@@ -886,8 +884,8 @@ static void blk_mq_terminate_expired(struct blk_mq_hw_ctx *hctx,
 	 * now guaranteed to see @rq->aborted_gstate and yield.  If
 	 * @rq->aborted_gstate still matches @rq->gstate, @rq is ours.
 	 */
-	if (READ_ONCE(rq->gstate) == rq->aborted_gstate &&
-	    !blk_mark_rq_complete(rq))
+	if (!(rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) &&
+	    READ_ONCE(rq->gstate) == rq->aborted_gstate)
 		blk_mq_rq_timed_out(rq, reserved);
 }
 
diff --git a/block/blk-timeout.c b/block/blk-timeout.c
index d580af3..25c4ffa 100644
--- a/block/blk-timeout.c
+++ b/block/blk-timeout.c
@@ -209,6 +209,7 @@ void blk_add_timer(struct request *req)
 		req->timeout = q->rq_timeout;
 
 	req->deadline = jiffies + req->timeout;
+	req->rq_flags &= ~RQF_MQ_TIMEOUT_EXPIRED;
 
 	/*
 	 * Only the non-mq case needs to add the request to a protected list.
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 2d6fd11..13186a7 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -123,6 +123,8 @@ typedef __u32 __bitwise req_flags_t;
 /* Look at ->special_vec for the actual data payload instead of the
    bio chain. */
 #define RQF_SPECIAL_PAYLOAD	((__force req_flags_t)(1 << 18))
+/* timeout is expired */
+#define RQF_MQ_TIMEOUT_EXPIRED	((__force req_flags_t)(1 << 19))
 
 /* flags that prevent us from merging requests: */
 #define RQF_NOMERGE_FLAGS \
-- 
2.9.5

Re: [PATCH 5/7] blk-mq: remove REQ_ATOM_COMPLETE usages from blk-mq

[jianchao.wang]

Hi tejun

On 12/16/2017 08:07 PM, Tejun Heo wrote:
> After the recent updates to use generation number and state based
> synchronization, blk-mq no longer depends on REQ_ATOM_COMPLETE except
> to avoid firing the same timeout multiple times.
> 
> Remove all REQ_ATOM_COMPLETE usages and use a new rq_flags flag
> RQF_MQ_TIMEOUT_EXPIRED to avoid firing the same timeout multiple
> times.  This removes atomic bitops from hot paths too.
> 
> v2: Removed blk_clear_rq_complete() from blk_mq_rq_timed_out().
> 
> v3: Added RQF_MQ_TIMEOUT_EXPIRED flag.
> 
> Signed-off-by: Tejun Heo <tj@kernel.org>
> Cc: "jianchao.wang" <jianchao.w.wang@oracle.com>
> ---
>  block/blk-mq.c         | 18 ++++++++----------
>  block/blk-timeout.c    |  1 +
>  include/linux/blkdev.h |  2 ++
>  3 files changed, 11 insertions(+), 10 deletions(-)
> 
> diff --git a/block/blk-mq.c b/block/blk-mq.c
> index 88baa82..47e722b 100644
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -607,14 +607,12 @@ void blk_mq_complete_request(struct request *rq)
>  	 */
>  	if (!(hctx->flags & BLK_MQ_F_BLOCKING)) {
>  		rcu_read_lock();
> -		if (blk_mq_rq_aborted_gstate(rq) != rq->gstate &&
> -		    !blk_mark_rq_complete(rq))
> +		if (blk_mq_rq_aborted_gstate(rq) != rq->gstate)
>  			__blk_mq_complete_request(rq);
>  		rcu_read_unlock();
>  	} else {
>  		srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
> -		if (blk_mq_rq_aborted_gstate(rq) != rq->gstate &&
> -		    !blk_mark_rq_complete(rq))
> +		if (blk_mq_rq_aborted_gstate(rq) != rq->gstate)
>  			__blk_mq_complete_request(rq);
>  		srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
>  	}
It's worrying that even though the blk_mark_rq_complete() here is intended to synchronize with
timeout path, but it indeed give the blk_mq_complete_request() the capability to exclude with 
itself. Maybe this capability should be reserved.

Thanks
Jianchao

Re: [PATCH 5/7] blk-mq: remove REQ_ATOM_COMPLETE usages from blk-mq

[Tejun Heo]

Hello,

On Thu, Dec 21, 2017 at 11:56:49AM +0800, jianchao.wang wrote:
> It's worrying that even though the blk_mark_rq_complete() here is intended to synchronize with
> timeout path, but it indeed give the blk_mq_complete_request() the capability to exclude with 
> itself. Maybe this capability should be reserved.

Can you explain further how that'd help?  The problem there is that if
you have two competing completions, where one isn't a timeout, there's
nothing synchronizing the reuse of the request.  IOW, the losing one
can easily complete the next recycle instance.  The atomic bitops
might feel like more protection but it's just feels.

Thanks.

-- 
tejun

Re: [PATCH 5/7] blk-mq: remove REQ_ATOM_COMPLETE usages from blk-mq

[jianchao.wang]

Sorry for my non-detailed description. 

On 12/21/2017 09:50 PM, Tejun Heo wrote:
> Hello,
> 
> On Thu, Dec 21, 2017 at 11:56:49AM +0800, jianchao.wang wrote:
>> It's worrying that even though the blk_mark_rq_complete() here is intended to synchronize with
>> timeout path, but it indeed give the blk_mq_complete_request() the capability to exclude with 
There could be scenario where the driver itself stop a request itself with blk_mq_complete_request() or
some other interface that will invoke it, races with the normal completion path where a same request comes.
For example:
a reset could be triggered through sysfs on nvme-rdma
Then the driver will cancel all the reqs, including in-flight ones.
nvme_rdma_reset_ctrl_work()
    nvme_rdma_shutdown_ctrl()
    >>>>
        if (ctrl->ctrl.queue_count > 1) {
            nvme_stop_queues(&ctrl->ctrl); //quiesce the queue
            blk_mq_tagset_busy_iter(&ctrl->tag_set,
                        nvme_cancel_request, &ctrl->ctrl); //invoke blk_mq_complete_request()
            nvme_rdma_destroy_io_queues(ctrl, shutdown);
        }
    >>>>

These operations could race with the normal completion path of in-flight ones.
It should drain all the in-flight ones first here. But there maybe some other
places similar with this.
 
>> itself. Maybe this capability should be reserved.
> 
> Can you explain further how that'd help?  The problem there is that if
> you have two competing completions, where one isn't a timeout, there's
> nothing synchronizing the reuse of the request.  IOW, the losing on
> can easily complete the next recycle instance.  The atomic bitops
> might feel like more protection but it's just feels.

In above case, the request may simultaneously enter requeue and end path.

Thanks
Jianchao

Re: [PATCH 5/7] blk-mq: remove REQ_ATOM_COMPLETE usages from blk-mq

[Tejun Heo]

Hello, Jianchao.

On Fri, Dec 22, 2017 at 12:02:20PM +0800, jianchao.wang wrote:
> > On Thu, Dec 21, 2017 at 11:56:49AM +0800, jianchao.wang wrote:
> >> It's worrying that even though the blk_mark_rq_complete() here is
> >> intended to synchronize with timeout path, but it indeed give the
> >> blk_mq_complete_request() the capability to exclude with
>
> There could be scenario where the driver itself stop a request
> itself with blk_mq_complete_request() or some other interface that
> will invoke it, races with the normal completion path where a same
> request comes.

But what'd prevent the completion reinitializing the request and then
the actual completion path coming in and completing the request again?

> For example:
> a reset could be triggered through sysfs on nvme-rdma
> Then the driver will cancel all the reqs, including in-flight ones.
> nvme_rdma_reset_ctrl_work()
>     nvme_rdma_shutdown_ctrl()
>     >>>>
>         if (ctrl->ctrl.queue_count > 1) {
>             nvme_stop_queues(&ctrl->ctrl); //quiesce the queue
>             blk_mq_tagset_busy_iter(&ctrl->tag_set,
>                         nvme_cancel_request, &ctrl->ctrl); //invoke blk_mq_complete_request()
>             nvme_rdma_destroy_io_queues(ctrl, shutdown);
>         }
>     >>>>
> 
> These operations could race with the normal completion path of in-flight ones.
> It should drain all the in-flight ones first here. But there maybe some other
> places similar with this.

If there are any such places, they should be using an interface which
is propelry synchronized like blk_abort_request(), which btw is what
libata already does.  Otherwise, it's racy with or without these
patches.

Thanks.

-- 
tejun

Re: [PATCH 5/7] blk-mq: remove REQ_ATOM_COMPLETE usages from blk-mq

[jianchao.wang]

Hi tejun

Many thanks for you kindly response.

On 01/09/2018 01:27 AM, Tejun Heo wrote:
> Hello, Jianchao.
> 
> On Fri, Dec 22, 2017 at 12:02:20PM +0800, jianchao.wang wrote:
>>> On Thu, Dec 21, 2017 at 11:56:49AM +0800, jianchao.wang wrote:
>>>> It's worrying that even though the blk_mark_rq_complete() here is
>>>> intended to synchronize with timeout path, but it indeed give the
>>>> blk_mq_complete_request() the capability to exclude with
>>
>> There could be scenario where the driver itself stop a request
>> itself with blk_mq_complete_request() or some other interface that
>> will invoke it, races with the normal completion path where a same
>> request comes.
> 
> But what'd prevent the completion reinitializing the request and then
> the actual completion path coming in and completing the request again?
> 
blk_mark_rq_complete() will gate and ensure there will be only one 
__blk_mq_complete_request() to be invoked.

>> For example:
>> a reset could be triggered through sysfs on nvme-rdma
>> Then the driver will cancel all the reqs, including in-flight ones.
>> nvme_rdma_reset_ctrl_work()
>>     nvme_rdma_shutdown_ctrl()
>>     >>>>
>>         if (ctrl->ctrl.queue_count > 1) {
>>             nvme_stop_queues(&ctrl->ctrl); //quiesce the queue
>>             blk_mq_tagset_busy_iter(&ctrl->tag_set,
>>                         nvme_cancel_request, &ctrl->ctrl); //invoke blk_mq_complete_request()
>>             nvme_rdma_destroy_io_queues(ctrl, shutdown);
>>         }
>>     >>>>
>>
>> These operations could race with the normal completion path of in-flight ones.
>> It should drain all the in-flight ones first here. But there maybe some other
>> places similar with this.
> 
> If there are any such places, they should be using an interface which
> is propelry synchronized like blk_abort_request(), which btw is what
> libata already does.  Otherwise, it's racy with or without these
> patches.
Yes, it is that.

Thanks for you kindly response again.
Jianchao

Re: [PATCH 5/7] blk-mq: remove REQ_ATOM_COMPLETE usages from blk-mq

[Tejun Heo]

Hello,

On Tue, Jan 09, 2018 at 11:08:04AM +0800, jianchao.wang wrote:
> > But what'd prevent the completion reinitializing the request and then
> > the actual completion path coming in and completing the request again?
>
> blk_mark_rq_complete() will gate and ensure there will be only one 
> __blk_mq_complete_request() to be invoked.

Yeah, but then the complete flag will be cleared once completion is
done and the request is reinitialized.

Thanks.

-- 
tejun

Re: [PATCH 5/7] blk-mq: remove REQ_ATOM_COMPLETE usages from blk-mq

[jianchao.wang]

Hi tejun

Many thanks for your kindly response.

On 01/09/2018 11:37 AM, Tejun Heo wrote:
> Hello,
> 
> On Tue, Jan 09, 2018 at 11:08:04AM +0800, jianchao.wang wrote:
>>> But what'd prevent the completion reinitializing the request and then
>>> the actual completion path coming in and completing the request again?
>>
>> blk_mark_rq_complete() will gate and ensure there will be only one 
>> __blk_mq_complete_request() to be invoked.
> 
> Yeah, but then the complete flag will be cleared once completion is
> done and the request is reinitialized.

Yes, it is. What I mean is not to reserve blk_mark_rq_complete and  REQ_ATOM_COMPLETE
usages in blk-mq but the side-effect after blk_mq_complete_request cannot exclude with itself.
As you said, the scene is racy and should be modified. :)

Thanks
Jianchao 

[PATCH 6/7] blk-mq: remove REQ_ATOM_STARTED

[Tejun Heo]

After the recent updates to use generation number and state based
synchronization, we can easily replace REQ_ATOM_STARTED usages by
adding an extra state to distinguish completed but not yet freed
state.

Add MQ_RQ_COMPLETE and replace REQ_ATOM_STARTED usages with
blk_mq_rq_state() tests.  REQ_ATOM_STARTED no longer has any users
left and is removed.

Signed-off-by: Tejun Heo <tj@kernel.org>
---
 block/blk-mq-debugfs.c |  4 +---
 block/blk-mq.c         | 37 ++++++++-----------------------------
 block/blk-mq.h         |  1 +
 block/blk.h            |  1 -
 4 files changed, 10 insertions(+), 33 deletions(-)

diff --git a/block/blk-mq-debugfs.c b/block/blk-mq-debugfs.c
index b56a4f3..8adc837 100644
--- a/block/blk-mq-debugfs.c
+++ b/block/blk-mq-debugfs.c
@@ -271,7 +271,6 @@ static const char *const cmd_flag_name[] = {
 #define RQF_NAME(name) [ilog2((__force u32)RQF_##name)] = #name
 static const char *const rqf_name[] = {
 	RQF_NAME(SORTED),
-	RQF_NAME(STARTED),
 	RQF_NAME(QUEUED),
 	RQF_NAME(SOFTBARRIER),
 	RQF_NAME(FLUSH_SEQ),
@@ -295,7 +294,6 @@ static const char *const rqf_name[] = {
 #define RQAF_NAME(name) [REQ_ATOM_##name] = #name
 static const char *const rqaf_name[] = {
 	RQAF_NAME(COMPLETE),
-	RQAF_NAME(STARTED),
 	RQAF_NAME(POLL_SLEPT),
 };
 #undef RQAF_NAME
@@ -409,7 +407,7 @@ static void hctx_show_busy_rq(struct request *rq, void *data, bool reserved)
 	const struct show_busy_params *params = data;
 
 	if (blk_mq_map_queue(rq->q, rq->mq_ctx->cpu) == params->hctx &&
-	    test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
+	    blk_mq_rq_state(rq) != MQ_RQ_IDLE)
 		__blk_mq_debugfs_rq_show(params->m,
 					 list_entry_rq(&rq->queuelist));
 }
diff --git a/block/blk-mq.c b/block/blk-mq.c
index 47e722b..724d340 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -483,7 +483,6 @@ void blk_mq_free_request(struct request *rq)
 		blk_put_rl(blk_rq_rl(rq));
 
 	blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
-	clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
 	clear_bit(REQ_ATOM_POLL_SLEPT, &rq->atomic_flags);
 	if (rq->tag != -1)
 		blk_mq_put_tag(hctx, hctx->tags, ctx, rq->tag);
@@ -531,6 +530,7 @@ static void __blk_mq_complete_request(struct request *rq)
 	int cpu;
 
 	WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT);
+	blk_mq_rq_update_state(rq, MQ_RQ_COMPLETE);
 
 	if (rq->internal_tag != -1)
 		blk_mq_sched_completed_request(rq);
@@ -621,7 +621,7 @@ EXPORT_SYMBOL(blk_mq_complete_request);
 
 int blk_mq_request_started(struct request *rq)
 {
-	return test_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
+	return blk_mq_rq_state(rq) != MQ_RQ_IDLE;
 }
 EXPORT_SYMBOL_GPL(blk_mq_request_started);
 
@@ -640,7 +640,6 @@ void blk_mq_start_request(struct request *rq)
 	}
 
 	WARN_ON_ONCE(blk_mq_rq_state(rq) != MQ_RQ_IDLE);
-	WARN_ON_ONCE(test_bit(REQ_ATOM_STARTED, &rq->atomic_flags));
 
 	/*
 	 * Mark @rq in-flight which also advances the generation number,
@@ -662,8 +661,6 @@ void blk_mq_start_request(struct request *rq)
 	write_seqcount_end(&rq->gstate_seq);
 	preempt_enable();
 
-	set_bit(REQ_ATOM_STARTED, &rq->atomic_flags);
-
 	if (q->dma_drain_size && blk_rq_bytes(rq)) {
 		/*
 		 * Make sure space for the drain appears.  We know we can do
@@ -676,13 +673,9 @@ void blk_mq_start_request(struct request *rq)
 EXPORT_SYMBOL(blk_mq_start_request);
 
 /*
- * When we reach here because queue is busy, REQ_ATOM_COMPLETE
- * flag isn't set yet, so there may be race with timeout handler,
- * but given rq->deadline is just set in .queue_rq() under
- * this situation, the race won't be possible in reality because
- * rq->timeout should be set as big enough to cover the window
- * between blk_mq_start_request() called from .queue_rq() and
- * clearing REQ_ATOM_STARTED here.
+ * When we reach here because queue is busy, it's safe to change the state
+ * to IDLE without checking @rq->aborted_gstate because we should still be
+ * holding the RCU read lock and thus protected against timeout.
  */
 static void __blk_mq_requeue_request(struct request *rq)
 {
@@ -694,7 +687,7 @@ static void __blk_mq_requeue_request(struct request *rq)
 	wbt_requeue(q->rq_wb, &rq->issue_stat);
 	blk_mq_sched_requeue_request(rq);
 
-	if (test_and_clear_bit(REQ_ATOM_STARTED, &rq->atomic_flags)) {
+	if (blk_mq_rq_state(rq) != MQ_RQ_IDLE) {
 		blk_mq_rq_update_state(rq, MQ_RQ_IDLE);
 		if (q->dma_drain_size && blk_rq_bytes(rq))
 			rq->nr_phys_segments--;
@@ -801,18 +794,6 @@ static void blk_mq_rq_timed_out(struct request *req, bool reserved)
 	const struct blk_mq_ops *ops = req->q->mq_ops;
 	enum blk_eh_timer_return ret = BLK_EH_RESET_TIMER;
 
-	/*
-	 * We know that complete is set at this point. If STARTED isn't set
-	 * anymore, then the request isn't active and the "timeout" should
-	 * just be ignored. This can happen due to the bitflag ordering.
-	 * Timeout first checks if STARTED is set, and if it is, assumes
-	 * the request is active. But if we race with completion, then
-	 * both flags will get cleared. So check here again, and ignore
-	 * a timeout event with a request that isn't active.
-	 */
-	if (!test_bit(REQ_ATOM_STARTED, &req->atomic_flags))
-		return;
-
 	req->rq_flags |= RQF_MQ_TIMEOUT_EXPIRED;
 
 	if (ops->timeout)
@@ -848,8 +829,7 @@ static void blk_mq_check_expired(struct blk_mq_hw_ctx *hctx,
 
 	might_sleep();
 
-	if ((rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED) ||
-	    !test_bit(REQ_ATOM_STARTED, &rq->atomic_flags))
+	if (rq->rq_flags & RQF_MQ_TIMEOUT_EXPIRED)
 		return;
 
 	/* read coherent snapshots of @rq->state_gen and @rq->deadline */
@@ -2971,8 +2951,7 @@ static bool blk_mq_poll_hybrid_sleep(struct request_queue *q,
 
 	hrtimer_init_sleeper(&hs, current);
 	do {
-		if (test_bit(REQ_ATOM_STARTED, &rq->atomic_flags) &&
-		    blk_mq_rq_state(rq) != MQ_RQ_IN_FLIGHT)
+		if (blk_mq_rq_state(rq) == MQ_RQ_COMPLETE)
 			break;
 		set_current_state(TASK_UNINTERRUPTIBLE);
 		hrtimer_start_expires(&hs.timer, mode);
diff --git a/block/blk-mq.h b/block/blk-mq.h
index 6b2d616..8591a54 100644
--- a/block/blk-mq.h
+++ b/block/blk-mq.h
@@ -34,6 +34,7 @@ struct blk_mq_ctx {
 enum mq_rq_state {
 	MQ_RQ_IDLE		= 0,
 	MQ_RQ_IN_FLIGHT		= 1,
+	MQ_RQ_COMPLETE		= 2,
 
 	MQ_RQ_STATE_BITS	= 2,
 	MQ_RQ_STATE_MASK	= (1 << MQ_RQ_STATE_BITS) - 1,
diff --git a/block/blk.h b/block/blk.h
index 9cb2739..a68dbe3 100644
--- a/block/blk.h
+++ b/block/blk.h
@@ -124,7 +124,6 @@ void blk_account_io_done(struct request *req);
  */
 enum rq_atomic_flags {
 	REQ_ATOM_COMPLETE = 0,
-	REQ_ATOM_STARTED,
 
 	REQ_ATOM_POLL_SLEPT,
 };
-- 
2.9.5

[PATCH 7/7] blk-mq: rename blk_mq_hw_ctx->queue_rq_srcu to ->srcu

[Tejun Heo]

The RCU protection has been expanded to cover both queueing and
completion paths making ->queue_rq_srcu a misnomer.  Rename it to
->srcu as suggested by Bart.

Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Bart Van Assche <Bart.VanAssche@wdc.com>
---
 block/blk-mq.c         | 22 +++++++++++-----------
 include/linux/blk-mq.h |  2 +-
 2 files changed, 12 insertions(+), 12 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index 724d340..872d4ee 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -219,7 +219,7 @@ void blk_mq_quiesce_queue(struct request_queue *q)
 
 	queue_for_each_hw_ctx(q, hctx, i) {
 		if (hctx->flags & BLK_MQ_F_BLOCKING)
-			synchronize_srcu(hctx->queue_rq_srcu);
+			synchronize_srcu(hctx->srcu);
 		else
 			rcu = true;
 	}
@@ -611,10 +611,10 @@ void blk_mq_complete_request(struct request *rq)
 			__blk_mq_complete_request(rq);
 		rcu_read_unlock();
 	} else {
-		srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
+		srcu_idx = srcu_read_lock(hctx->srcu);
 		if (blk_mq_rq_aborted_gstate(rq) != rq->gstate)
 			__blk_mq_complete_request(rq);
-		srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
+		srcu_read_unlock(hctx->srcu, srcu_idx);
 	}
 }
 EXPORT_SYMBOL(blk_mq_complete_request);
@@ -916,7 +916,7 @@ static void blk_mq_timeout_work(struct work_struct *work)
 			if (!(hctx->flags & BLK_MQ_F_BLOCKING))
 				has_rcu = true;
 			else
-				synchronize_srcu(hctx->queue_rq_srcu);
+				synchronize_srcu(hctx->srcu);
 
 			hctx->nr_expired = 0;
 		}
@@ -1279,9 +1279,9 @@ static void __blk_mq_run_hw_queue(struct blk_mq_hw_ctx *hctx)
 	} else {
 		might_sleep();
 
-		srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
+		srcu_idx = srcu_read_lock(hctx->srcu);
 		blk_mq_sched_dispatch_requests(hctx);
-		srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
+		srcu_read_unlock(hctx->srcu, srcu_idx);
 	}
 }
 
@@ -1718,9 +1718,9 @@ static void blk_mq_try_issue_directly(struct blk_mq_hw_ctx *hctx,
 
 		might_sleep();
 
-		srcu_idx = srcu_read_lock(hctx->queue_rq_srcu);
+		srcu_idx = srcu_read_lock(hctx->srcu);
 		__blk_mq_try_issue_directly(hctx, rq, cookie, true);
-		srcu_read_unlock(hctx->queue_rq_srcu, srcu_idx);
+		srcu_read_unlock(hctx->srcu, srcu_idx);
 	}
 }
 
@@ -2074,7 +2074,7 @@ static void blk_mq_exit_hctx(struct request_queue *q,
 		set->ops->exit_hctx(hctx, hctx_idx);
 
 	if (hctx->flags & BLK_MQ_F_BLOCKING)
-		cleanup_srcu_struct(hctx->queue_rq_srcu);
+		cleanup_srcu_struct(hctx->srcu);
 
 	blk_mq_remove_cpuhp(hctx);
 	blk_free_flush_queue(hctx->fq);
@@ -2147,7 +2147,7 @@ static int blk_mq_init_hctx(struct request_queue *q,
 		goto free_fq;
 
 	if (hctx->flags & BLK_MQ_F_BLOCKING)
-		init_srcu_struct(hctx->queue_rq_srcu);
+		init_srcu_struct(hctx->srcu);
 
 	blk_mq_debugfs_register_hctx(q, hctx);
 
@@ -2436,7 +2436,7 @@ static int blk_mq_hw_ctx_size(struct blk_mq_tag_set *tag_set)
 {
 	int hw_ctx_size = sizeof(struct blk_mq_hw_ctx);
 
-	BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, queue_rq_srcu),
+	BUILD_BUG_ON(ALIGN(offsetof(struct blk_mq_hw_ctx, srcu),
 			   __alignof__(struct blk_mq_hw_ctx)) !=
 		     sizeof(struct blk_mq_hw_ctx));
 
diff --git a/include/linux/blk-mq.h b/include/linux/blk-mq.h
index 460798d..8efcf49 100644
--- a/include/linux/blk-mq.h
+++ b/include/linux/blk-mq.h
@@ -66,7 +66,7 @@ struct blk_mq_hw_ctx {
 #endif
 
 	/* Must be the last member - see also blk_mq_hw_ctx_size(). */
-	struct srcu_struct	queue_rq_srcu[0];
+	struct srcu_struct	srcu[0];
 };
 
 struct blk_mq_tag_set {
-- 
2.9.5

Re: [PATCHSET v3] blk-mq: reimplement timeout handling

[Christoph Hellwig]

This seems to miss the linux-block list once again.  Please include
it in the next resend.

Re: [PATCHSET v3] blk-mq: reimplement timeout handling

[Tejun Heo]

On Fri, Dec 29, 2017 at 02:02:39AM -0800, Christoph Hellwig wrote:
> This seems to miss the linux-block list once again.  Please include
> it in the next resend.

Sorry about that.  Copy/pasted from the older thread without thinking.

Thanks.

-- 
tejun