[PATCH v3 0/2] xfs: fix buffer delwri queue state race

[PATCH v3 0/2] xfs: fix buffer delwri queue state race

[Brian Foster]

I'll post the xfs_buf_submit[_wait]() refactoring patch separately after
waiting for some feedback in the previous thread and/or doing some
testing. It really has nothing to do with fixing this problem, anyways.

Brian

v3:
- Leave tracepoint in __xfs_buf_submit and kill
  trace_xfs_buf_submit_wait().
- Comment tweaks.
v2: https://marc.info/?l=linux-xfs&m=152888792024447&w=2
- Implement sync buffer I/O for sync delwri queues instead of buffer
  wait list stealing.
v1: https://marc.info/?l=linux-xfs&m=152837528705511&w=2

Brian Foster (2):
  xfs: refactor buffer submission into a common helper
  xfs: use sync buffer I/O for sync delwri queue submission

 fs/xfs/xfs_buf.c   | 163 ++++++++++++++++++++++-----------------------
 fs/xfs/xfs_trace.h |   1 -
 2 files changed, 78 insertions(+), 86 deletions(-)

-- 
2.17.1

[PATCH v3 1/2] xfs: refactor buffer submission into a common helper

[Brian Foster]

Sync and async buffer submission both do generally similar things
with a couple odd exceptions. Refactor the core buffer submission
code into a common helper to isolate buffer submission from
completion handling of synchronous buffer I/O.

This patch does not change behavior. It is a step towards support
for using synchronous buffer I/O via synchronous delwri queue
submission.

Designed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Brian Foster <bfoster@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
---
 fs/xfs/xfs_buf.c   | 85 ++++++++++++++++++++--------------------------
 fs/xfs/xfs_trace.h |  1 -
 2 files changed, 37 insertions(+), 49 deletions(-)

diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c
index e9c058e3761c..7b0f7c79cd62 100644
--- a/fs/xfs/xfs_buf.c
+++ b/fs/xfs/xfs_buf.c
@@ -1458,22 +1458,20 @@ _xfs_buf_ioapply(
  * a call to this function unless the caller holds an additional reference
  * itself.
  */
-void
-xfs_buf_submit(
+static int
+__xfs_buf_submit(
 	struct xfs_buf	*bp)
 {
 	trace_xfs_buf_submit(bp, _RET_IP_);
 
 	ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
-	ASSERT(bp->b_flags & XBF_ASYNC);
 
 	/* on shutdown we stale and complete the buffer immediately */
 	if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
 		xfs_buf_ioerror(bp, -EIO);
 		bp->b_flags &= ~XBF_DONE;
 		xfs_buf_stale(bp);
-		xfs_buf_ioend(bp);
-		return;
+		return -EIO;
 	}
 
 	if (bp->b_flags & XBF_WRITE)
@@ -1482,23 +1480,14 @@ xfs_buf_submit(
 	/* clear the internal error state to avoid spurious errors */
 	bp->b_io_error = 0;
 
-	/*
-	 * The caller's reference is released during I/O completion.
-	 * This occurs some time after the last b_io_remaining reference is
-	 * released, so after we drop our Io reference we have to have some
-	 * other reference to ensure the buffer doesn't go away from underneath
-	 * us. Take a direct reference to ensure we have safe access to the
-	 * buffer until we are finished with it.
-	 */
-	xfs_buf_hold(bp);
-
 	/*
 	 * Set the count to 1 initially, this will stop an I/O completion
 	 * callout which happens before we have started all the I/O from calling
 	 * xfs_buf_ioend too early.
 	 */
 	atomic_set(&bp->b_io_remaining, 1);
-	xfs_buf_ioacct_inc(bp);
+	if (bp->b_flags & XBF_ASYNC)
+		xfs_buf_ioacct_inc(bp);
 	_xfs_buf_ioapply(bp);
 
 	/*
@@ -1507,14 +1496,39 @@ xfs_buf_submit(
 	 * that we don't return to the caller with completion still pending.
 	 */
 	if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
-		if (bp->b_error)
+		if (bp->b_error || !(bp->b_flags & XBF_ASYNC))
 			xfs_buf_ioend(bp);
 		else
 			xfs_buf_ioend_async(bp);
 	}
 
-	xfs_buf_rele(bp);
+	return 0;
+}
+
+void
+xfs_buf_submit(
+	struct xfs_buf	*bp)
+{
+	int		error;
+
+	ASSERT(bp->b_flags & XBF_ASYNC);
+
+	/*
+	 * The caller's reference is released during I/O completion.
+	 * This occurs some time after the last b_io_remaining reference is
+	 * released, so after we drop our Io reference we have to have some
+	 * other reference to ensure the buffer doesn't go away from underneath
+	 * us. Take a direct reference to ensure we have safe access to the
+	 * buffer until we are finished with it.
+	 */
+	xfs_buf_hold(bp);
+
+	error = __xfs_buf_submit(bp);
+	if (error)
+		xfs_buf_ioend(bp);
+
 	/* Note: it is not safe to reference bp now we've dropped our ref */
+	xfs_buf_rele(bp);
 }
 
 /*
@@ -1526,22 +1540,7 @@ xfs_buf_submit_wait(
 {
 	int		error;
 
-	trace_xfs_buf_submit_wait(bp, _RET_IP_);
-
-	ASSERT(!(bp->b_flags & (_XBF_DELWRI_Q | XBF_ASYNC)));
-
-	if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
-		xfs_buf_ioerror(bp, -EIO);
-		xfs_buf_stale(bp);
-		bp->b_flags &= ~XBF_DONE;
-		return -EIO;
-	}
-
-	if (bp->b_flags & XBF_WRITE)
-		xfs_buf_wait_unpin(bp);
-
-	/* clear the internal error state to avoid spurious errors */
-	bp->b_io_error = 0;
+	ASSERT(!(bp->b_flags & XBF_ASYNC));
 
 	/*
 	 * For synchronous IO, the IO does not inherit the submitters reference
@@ -1551,20 +1550,9 @@ xfs_buf_submit_wait(
 	 */
 	xfs_buf_hold(bp);
 
-	/*
-	 * Set the count to 1 initially, this will stop an I/O completion
-	 * callout which happens before we have started all the I/O from calling
-	 * xfs_buf_ioend too early.
-	 */
-	atomic_set(&bp->b_io_remaining, 1);
-	_xfs_buf_ioapply(bp);
-
-	/*
-	 * make sure we run completion synchronously if it raced with us and is
-	 * already complete.
-	 */
-	if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
-		xfs_buf_ioend(bp);
+	error = __xfs_buf_submit(bp);
+	if (error)
+		goto out;
 
 	/* wait for completion before gathering the error from the buffer */
 	trace_xfs_buf_iowait(bp, _RET_IP_);
@@ -1572,6 +1560,7 @@ xfs_buf_submit_wait(
 	trace_xfs_buf_iowait_done(bp, _RET_IP_);
 	error = bp->b_error;
 
+out:
 	/*
 	 * all done now, we can release the hold that keeps the buffer
 	 * referenced for the entire IO.
diff --git a/fs/xfs/xfs_trace.h b/fs/xfs/xfs_trace.h
index 972d45d28097..c2d2e513d722 100644
--- a/fs/xfs/xfs_trace.h
+++ b/fs/xfs/xfs_trace.h
@@ -310,7 +310,6 @@ DEFINE_BUF_EVENT(xfs_buf_hold);
 DEFINE_BUF_EVENT(xfs_buf_rele);
 DEFINE_BUF_EVENT(xfs_buf_iodone);
 DEFINE_BUF_EVENT(xfs_buf_submit);
-DEFINE_BUF_EVENT(xfs_buf_submit_wait);
 DEFINE_BUF_EVENT(xfs_buf_lock);
 DEFINE_BUF_EVENT(xfs_buf_lock_done);
 DEFINE_BUF_EVENT(xfs_buf_trylock_fail);
-- 
2.17.1

[PATCH v3 2/2] xfs: use sync buffer I/O for sync delwri queue submission

[Brian Foster]

If a delwri queue occurs of a buffer that sits on a delwri queue
wait list, the queue sets _XBF_DELWRI_Q without changing the state
of ->b_list. This occurs, for example, if another thread beats the
current delwri waiter thread to the buffer lock after I/O
completion. Once the waiter acquires the lock, it removes the buffer
from the wait list and leaves a buffer with _XBF_DELWRI_Q set but
not populated on a list. This results in a lost buffer submission
and in turn can result in assert failures due to _XBF_DELWRI_Q being
set on buffer reclaim or filesystem lockups if the buffer happens to
cover an item in the AIL.

This problem has been reproduced by repeated iterations of xfs/305
on high CPU count (28xcpu) systems with limited memory (~1GB). Dirty
dquot reclaim races with an xfsaild push of a separate dquot backed
by the same buffer such that the buffer sits on the reclaim wait
list at the time xfsaild attempts to queue it. Since the latter
dquot has been flush locked but the underlying buffer not submitted
for I/O, the dquot pins the AIL and causes the filesystem to
livelock.

This race is essentially made possible by the buffer lock cycle
involved with waiting on a synchronous delwri queue submission.
Close the race by using synchronous buffer I/O for respective delwri
queue submission. This means the buffer remains locked across the
I/O and so is inaccessible from other contexts while in the
intermediate wait list state. The sync buffer I/O wait mechanism is
factored into a helper such that sync delwri buffer submission and
serialization are batched operations.

Designed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Brian Foster <bfoster@redhat.com>
---
 fs/xfs/xfs_buf.c | 78 +++++++++++++++++++++++++-----------------------
 1 file changed, 41 insertions(+), 37 deletions(-)

diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c
index 7b0f7c79cd62..b5c0f6949b30 100644
--- a/fs/xfs/xfs_buf.c
+++ b/fs/xfs/xfs_buf.c
@@ -1531,6 +1531,20 @@ xfs_buf_submit(
 	xfs_buf_rele(bp);
 }
 
+/*
+ * Wait for I/O completion of a sync buffer and return the I/O error code.
+ */
+static int
+xfs_buf_iowait(
+	struct xfs_buf	*bp)
+{
+	trace_xfs_buf_iowait(bp, _RET_IP_);
+	wait_for_completion(&bp->b_iowait);
+	trace_xfs_buf_iowait_done(bp, _RET_IP_);
+
+	return bp->b_error;
+}
+
 /*
  * Synchronous buffer IO submission path, read or write.
  */
@@ -1553,12 +1567,7 @@ xfs_buf_submit_wait(
 	error = __xfs_buf_submit(bp);
 	if (error)
 		goto out;
-
-	/* wait for completion before gathering the error from the buffer */
-	trace_xfs_buf_iowait(bp, _RET_IP_);
-	wait_for_completion(&bp->b_iowait);
-	trace_xfs_buf_iowait_done(bp, _RET_IP_);
-	error = bp->b_error;
+	error = xfs_buf_iowait(bp);
 
 out:
 	/*
@@ -1961,16 +1970,11 @@ xfs_buf_cmp(
 }
 
 /*
- * submit buffers for write.
- *
- * When we have a large buffer list, we do not want to hold all the buffers
- * locked while we block on the request queue waiting for IO dispatch. To avoid
- * this problem, we lock and submit buffers in groups of 50, thereby minimising
- * the lock hold times for lists which may contain thousands of objects.
- *
- * To do this, we sort the buffer list before we walk the list to lock and
- * submit buffers, and we plug and unplug around each group of buffers we
- * submit.
+ * Submit buffers for write. If wait_list is specified, the buffers are
+ * submitted using sync I/O and placed on the wait list such that the caller can
+ * iowait each buffer. Otherwise async I/O is used and the buffers are released
+ * at I/O completion time. In either case, buffers remain locked until I/O
+ * completes and the buffer is released from the queue.
  */
 static int
 xfs_buf_delwri_submit_buffers(
@@ -2012,21 +2016,22 @@ xfs_buf_delwri_submit_buffers(
 		trace_xfs_buf_delwri_split(bp, _RET_IP_);
 
 		/*
-		 * We do all IO submission async. This means if we need
-		 * to wait for IO completion we need to take an extra
-		 * reference so the buffer is still valid on the other
-		 * side. We need to move the buffer onto the io_list
-		 * at this point so the caller can still access it.
+		 * If we have a wait list, each buffer (and associated delwri
+		 * queue reference) transfers to it and is submitted
+		 * synchronously. Otherwise, drop the buffer from the delwri
+		 * queue and submit async.
 		 */
 		bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_WRITE_FAIL);
-		bp->b_flags |= XBF_WRITE | XBF_ASYNC;
+		bp->b_flags |= XBF_WRITE;
 		if (wait_list) {
-			xfs_buf_hold(bp);
+			bp->b_flags &= ~XBF_ASYNC;
 			list_move_tail(&bp->b_list, wait_list);
-		} else
+			__xfs_buf_submit(bp);
+		} else {
+			bp->b_flags |= XBF_ASYNC;
 			list_del_init(&bp->b_list);
-
-		xfs_buf_submit(bp);
+			xfs_buf_submit(bp);
+		}
 	}
 	blk_finish_plug(&plug);
 
@@ -2073,8 +2078,11 @@ xfs_buf_delwri_submit(
 
 		list_del_init(&bp->b_list);
 
-		/* locking the buffer will wait for async IO completion. */
-		xfs_buf_lock(bp);
+		/*
+		 * Wait on the locked buffer, check for errors and unlock and
+		 * release the delwri queue reference.
+		 */
+		xfs_buf_iowait(bp);
 		error2 = bp->b_error;
 		xfs_buf_relse(bp);
 		if (!error)
@@ -2121,22 +2129,18 @@ xfs_buf_delwri_pushbuf(
 
 	/*
 	 * Delwri submission clears the DELWRI_Q buffer flag and returns with
-	 * the buffer on the wait list with an associated reference. Rather than
+	 * the buffer on the wait list with the original reference. Rather than
 	 * bounce the buffer from a local wait list back to the original list
 	 * after I/O completion, reuse the original list as the wait list.
 	 */
 	xfs_buf_delwri_submit_buffers(&submit_list, buffer_list);
 
 	/*
-	 * The buffer is now under I/O and wait listed as during typical delwri
-	 * submission. Lock the buffer to wait for I/O completion. Rather than
-	 * remove the buffer from the wait list and release the reference, we
-	 * want to return with the buffer queued to the original list. The
-	 * buffer already sits on the original list with a wait list reference,
-	 * however. If we let the queue inherit that wait list reference, all we
-	 * need to do is reset the DELWRI_Q flag.
+	 * The buffer is now locked, under I/O and wait listed on the original
+	 * delwri queue. Wait for I/O completion, restore the DELWRI_Q flag and
+	 * return with the buffer unlocked and on the original queue.
 	 */
-	xfs_buf_lock(bp);
+	xfs_buf_iowait(bp);
 	error = bp->b_error;
 	bp->b_flags |= _XBF_DELWRI_Q;
 	xfs_buf_unlock(bp);
-- 
2.17.1

Re: [PATCH v3 2/2] xfs: use sync buffer I/O for sync delwri queue submission

[Darrick J. Wong]

On Fri, Jun 15, 2018 at 02:05:36PM -0400, Brian Foster wrote:
> If a delwri queue occurs of a buffer that sits on a delwri queue
> wait list, the queue sets _XBF_DELWRI_Q without changing the state
> of ->b_list. This occurs, for example, if another thread beats the
> current delwri waiter thread to the buffer lock after I/O
> completion. Once the waiter acquires the lock, it removes the buffer
> from the wait list and leaves a buffer with _XBF_DELWRI_Q set but
> not populated on a list. This results in a lost buffer submission
> and in turn can result in assert failures due to _XBF_DELWRI_Q being
> set on buffer reclaim or filesystem lockups if the buffer happens to
> cover an item in the AIL.
> 
> This problem has been reproduced by repeated iterations of xfs/305
> on high CPU count (28xcpu) systems with limited memory (~1GB). Dirty
> dquot reclaim races with an xfsaild push of a separate dquot backed
> by the same buffer such that the buffer sits on the reclaim wait
> list at the time xfsaild attempts to queue it. Since the latter
> dquot has been flush locked but the underlying buffer not submitted
> for I/O, the dquot pins the AIL and causes the filesystem to
> livelock.
> 
> This race is essentially made possible by the buffer lock cycle
> involved with waiting on a synchronous delwri queue submission.
> Close the race by using synchronous buffer I/O for respective delwri
> queue submission. This means the buffer remains locked across the
> I/O and so is inaccessible from other contexts while in the
> intermediate wait list state. The sync buffer I/O wait mechanism is
> factored into a helper such that sync delwri buffer submission and
> serialization are batched operations.
> 
> Designed-by: Dave Chinner <dchinner@redhat.com>
> Signed-off-by: Brian Foster <bfoster@redhat.com>
> ---
>  fs/xfs/xfs_buf.c | 78 +++++++++++++++++++++++++-----------------------
>  1 file changed, 41 insertions(+), 37 deletions(-)
> 
> diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c
> index 7b0f7c79cd62..b5c0f6949b30 100644
> --- a/fs/xfs/xfs_buf.c
> +++ b/fs/xfs/xfs_buf.c
> @@ -1531,6 +1531,20 @@ xfs_buf_submit(
>  	xfs_buf_rele(bp);
>  }
>  
> +/*
> + * Wait for I/O completion of a sync buffer and return the I/O error code.
> + */
> +static int
> +xfs_buf_iowait(
> +	struct xfs_buf	*bp)
> +{
> +	trace_xfs_buf_iowait(bp, _RET_IP_);
> +	wait_for_completion(&bp->b_iowait);
> +	trace_xfs_buf_iowait_done(bp, _RET_IP_);
> +
> +	return bp->b_error;
> +}
> +
>  /*
>   * Synchronous buffer IO submission path, read or write.
>   */
> @@ -1553,12 +1567,7 @@ xfs_buf_submit_wait(
>  	error = __xfs_buf_submit(bp);
>  	if (error)
>  		goto out;
> -
> -	/* wait for completion before gathering the error from the buffer */
> -	trace_xfs_buf_iowait(bp, _RET_IP_);
> -	wait_for_completion(&bp->b_iowait);
> -	trace_xfs_buf_iowait_done(bp, _RET_IP_);
> -	error = bp->b_error;
> +	error = xfs_buf_iowait(bp);
>  
>  out:
>  	/*
> @@ -1961,16 +1970,11 @@ xfs_buf_cmp(
>  }
>  
>  /*
> - * submit buffers for write.
> - *
> - * When we have a large buffer list, we do not want to hold all the buffers
> - * locked while we block on the request queue waiting for IO dispatch. To avoid
> - * this problem, we lock and submit buffers in groups of 50, thereby minimising
> - * the lock hold times for lists which may contain thousands of objects.
> - *
> - * To do this, we sort the buffer list before we walk the list to lock and
> - * submit buffers, and we plug and unplug around each group of buffers we
> - * submit.
> + * Submit buffers for write. If wait_list is specified, the buffers are
> + * submitted using sync I/O and placed on the wait list such that the caller can
> + * iowait each buffer. Otherwise async I/O is used and the buffers are released
> + * at I/O completion time. In either case, buffers remain locked until I/O
> + * completes and the buffer is released from the queue.
>   */
>  static int
>  xfs_buf_delwri_submit_buffers(
> @@ -2012,21 +2016,22 @@ xfs_buf_delwri_submit_buffers(
>  		trace_xfs_buf_delwri_split(bp, _RET_IP_);
>  
>  		/*
> -		 * We do all IO submission async. This means if we need
> -		 * to wait for IO completion we need to take an extra
> -		 * reference so the buffer is still valid on the other
> -		 * side. We need to move the buffer onto the io_list
> -		 * at this point so the caller can still access it.
> +		 * If we have a wait list, each buffer (and associated delwri
> +		 * queue reference) transfers to it and is submitted
> +		 * synchronously. Otherwise, drop the buffer from the delwri
> +		 * queue and submit async.
>  		 */
>  		bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_WRITE_FAIL);
> -		bp->b_flags |= XBF_WRITE | XBF_ASYNC;
> +		bp->b_flags |= XBF_WRITE;
>  		if (wait_list) {
> -			xfs_buf_hold(bp);
> +			bp->b_flags &= ~XBF_ASYNC;
>  			list_move_tail(&bp->b_list, wait_list);
> -		} else
> +			__xfs_buf_submit(bp);
> +		} else {
> +			bp->b_flags |= XBF_ASYNC;
>  			list_del_init(&bp->b_list);
> -
> -		xfs_buf_submit(bp);
> +			xfs_buf_submit(bp);
> +		}
>  	}
>  	blk_finish_plug(&plug);
>  
> @@ -2073,8 +2078,11 @@ xfs_buf_delwri_submit(
>  
>  		list_del_init(&bp->b_list);
>  
> -		/* locking the buffer will wait for async IO completion. */
> -		xfs_buf_lock(bp);
> +		/*
> +		 * Wait on the locked buffer, check for errors and unlock and
> +		 * release the delwri queue reference.
> +		 */
> +		xfs_buf_iowait(bp);
>  		error2 = bp->b_error;

error2 = xfs_buf_iowait()?

>  		xfs_buf_relse(bp);
>  		if (!error)
> @@ -2121,22 +2129,18 @@ xfs_buf_delwri_pushbuf(
>  
>  	/*
>  	 * Delwri submission clears the DELWRI_Q buffer flag and returns with
> -	 * the buffer on the wait list with an associated reference. Rather than
> +	 * the buffer on the wait list with the original reference. Rather than
>  	 * bounce the buffer from a local wait list back to the original list
>  	 * after I/O completion, reuse the original list as the wait list.
>  	 */
>  	xfs_buf_delwri_submit_buffers(&submit_list, buffer_list);
>  
>  	/*
> -	 * The buffer is now under I/O and wait listed as during typical delwri
> -	 * submission. Lock the buffer to wait for I/O completion. Rather than
> -	 * remove the buffer from the wait list and release the reference, we
> -	 * want to return with the buffer queued to the original list. The
> -	 * buffer already sits on the original list with a wait list reference,
> -	 * however. If we let the queue inherit that wait list reference, all we
> -	 * need to do is reset the DELWRI_Q flag.
> +	 * The buffer is now locked, under I/O and wait listed on the original
> +	 * delwri queue. Wait for I/O completion, restore the DELWRI_Q flag and
> +	 * return with the buffer unlocked and on the original queue.
>  	 */
> -	xfs_buf_lock(bp);
> +	xfs_buf_iowait(bp);
>  	error = bp->b_error;

error = xfs_buf_iowait()?

Or is there a specific reason you collected the error separately?

Looks reasonable and I could follow along this time. :)

--D

>  	bp->b_flags |= _XBF_DELWRI_Q;
>  	xfs_buf_unlock(bp);
> -- 
> 2.17.1
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH v3 1/2] xfs: refactor buffer submission into a common helper

[Darrick J. Wong]

On Fri, Jun 15, 2018 at 02:05:35PM -0400, Brian Foster wrote:
> Sync and async buffer submission both do generally similar things
> with a couple odd exceptions. Refactor the core buffer submission
> code into a common helper to isolate buffer submission from
> completion handling of synchronous buffer I/O.
> 
> This patch does not change behavior. It is a step towards support
> for using synchronous buffer I/O via synchronous delwri queue
> submission.
> 
> Designed-by: Dave Chinner <dchinner@redhat.com>
> Signed-off-by: Brian Foster <bfoster@redhat.com>
> Reviewed-by: Christoph Hellwig <hch@lst.de>

Looks ok,
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>

--D

> ---
>  fs/xfs/xfs_buf.c   | 85 ++++++++++++++++++++--------------------------
>  fs/xfs/xfs_trace.h |  1 -
>  2 files changed, 37 insertions(+), 49 deletions(-)
> 
> diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c
> index e9c058e3761c..7b0f7c79cd62 100644
> --- a/fs/xfs/xfs_buf.c
> +++ b/fs/xfs/xfs_buf.c
> @@ -1458,22 +1458,20 @@ _xfs_buf_ioapply(
>   * a call to this function unless the caller holds an additional reference
>   * itself.
>   */
> -void
> -xfs_buf_submit(
> +static int
> +__xfs_buf_submit(
>  	struct xfs_buf	*bp)
>  {
>  	trace_xfs_buf_submit(bp, _RET_IP_);
>  
>  	ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
> -	ASSERT(bp->b_flags & XBF_ASYNC);
>  
>  	/* on shutdown we stale and complete the buffer immediately */
>  	if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
>  		xfs_buf_ioerror(bp, -EIO);
>  		bp->b_flags &= ~XBF_DONE;
>  		xfs_buf_stale(bp);
> -		xfs_buf_ioend(bp);
> -		return;
> +		return -EIO;
>  	}
>  
>  	if (bp->b_flags & XBF_WRITE)
> @@ -1482,23 +1480,14 @@ xfs_buf_submit(
>  	/* clear the internal error state to avoid spurious errors */
>  	bp->b_io_error = 0;
>  
> -	/*
> -	 * The caller's reference is released during I/O completion.
> -	 * This occurs some time after the last b_io_remaining reference is
> -	 * released, so after we drop our Io reference we have to have some
> -	 * other reference to ensure the buffer doesn't go away from underneath
> -	 * us. Take a direct reference to ensure we have safe access to the
> -	 * buffer until we are finished with it.
> -	 */
> -	xfs_buf_hold(bp);
> -
>  	/*
>  	 * Set the count to 1 initially, this will stop an I/O completion
>  	 * callout which happens before we have started all the I/O from calling
>  	 * xfs_buf_ioend too early.
>  	 */
>  	atomic_set(&bp->b_io_remaining, 1);
> -	xfs_buf_ioacct_inc(bp);
> +	if (bp->b_flags & XBF_ASYNC)
> +		xfs_buf_ioacct_inc(bp);
>  	_xfs_buf_ioapply(bp);
>  
>  	/*
> @@ -1507,14 +1496,39 @@ xfs_buf_submit(
>  	 * that we don't return to the caller with completion still pending.
>  	 */
>  	if (atomic_dec_and_test(&bp->b_io_remaining) == 1) {
> -		if (bp->b_error)
> +		if (bp->b_error || !(bp->b_flags & XBF_ASYNC))
>  			xfs_buf_ioend(bp);
>  		else
>  			xfs_buf_ioend_async(bp);
>  	}
>  
> -	xfs_buf_rele(bp);
> +	return 0;
> +}
> +
> +void
> +xfs_buf_submit(
> +	struct xfs_buf	*bp)
> +{
> +	int		error;
> +
> +	ASSERT(bp->b_flags & XBF_ASYNC);
> +
> +	/*
> +	 * The caller's reference is released during I/O completion.
> +	 * This occurs some time after the last b_io_remaining reference is
> +	 * released, so after we drop our Io reference we have to have some
> +	 * other reference to ensure the buffer doesn't go away from underneath
> +	 * us. Take a direct reference to ensure we have safe access to the
> +	 * buffer until we are finished with it.
> +	 */
> +	xfs_buf_hold(bp);
> +
> +	error = __xfs_buf_submit(bp);
> +	if (error)
> +		xfs_buf_ioend(bp);
> +
>  	/* Note: it is not safe to reference bp now we've dropped our ref */
> +	xfs_buf_rele(bp);
>  }
>  
>  /*
> @@ -1526,22 +1540,7 @@ xfs_buf_submit_wait(
>  {
>  	int		error;
>  
> -	trace_xfs_buf_submit_wait(bp, _RET_IP_);
> -
> -	ASSERT(!(bp->b_flags & (_XBF_DELWRI_Q | XBF_ASYNC)));
> -
> -	if (XFS_FORCED_SHUTDOWN(bp->b_target->bt_mount)) {
> -		xfs_buf_ioerror(bp, -EIO);
> -		xfs_buf_stale(bp);
> -		bp->b_flags &= ~XBF_DONE;
> -		return -EIO;
> -	}
> -
> -	if (bp->b_flags & XBF_WRITE)
> -		xfs_buf_wait_unpin(bp);
> -
> -	/* clear the internal error state to avoid spurious errors */
> -	bp->b_io_error = 0;
> +	ASSERT(!(bp->b_flags & XBF_ASYNC));
>  
>  	/*
>  	 * For synchronous IO, the IO does not inherit the submitters reference
> @@ -1551,20 +1550,9 @@ xfs_buf_submit_wait(
>  	 */
>  	xfs_buf_hold(bp);
>  
> -	/*
> -	 * Set the count to 1 initially, this will stop an I/O completion
> -	 * callout which happens before we have started all the I/O from calling
> -	 * xfs_buf_ioend too early.
> -	 */
> -	atomic_set(&bp->b_io_remaining, 1);
> -	_xfs_buf_ioapply(bp);
> -
> -	/*
> -	 * make sure we run completion synchronously if it raced with us and is
> -	 * already complete.
> -	 */
> -	if (atomic_dec_and_test(&bp->b_io_remaining) == 1)
> -		xfs_buf_ioend(bp);
> +	error = __xfs_buf_submit(bp);
> +	if (error)
> +		goto out;
>  
>  	/* wait for completion before gathering the error from the buffer */
>  	trace_xfs_buf_iowait(bp, _RET_IP_);
> @@ -1572,6 +1560,7 @@ xfs_buf_submit_wait(
>  	trace_xfs_buf_iowait_done(bp, _RET_IP_);
>  	error = bp->b_error;
>  
> +out:
>  	/*
>  	 * all done now, we can release the hold that keeps the buffer
>  	 * referenced for the entire IO.
> diff --git a/fs/xfs/xfs_trace.h b/fs/xfs/xfs_trace.h
> index 972d45d28097..c2d2e513d722 100644
> --- a/fs/xfs/xfs_trace.h
> +++ b/fs/xfs/xfs_trace.h
> @@ -310,7 +310,6 @@ DEFINE_BUF_EVENT(xfs_buf_hold);
>  DEFINE_BUF_EVENT(xfs_buf_rele);
>  DEFINE_BUF_EVENT(xfs_buf_iodone);
>  DEFINE_BUF_EVENT(xfs_buf_submit);
> -DEFINE_BUF_EVENT(xfs_buf_submit_wait);
>  DEFINE_BUF_EVENT(xfs_buf_lock);
>  DEFINE_BUF_EVENT(xfs_buf_lock_done);
>  DEFINE_BUF_EVENT(xfs_buf_trylock_fail);
> -- 
> 2.17.1
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH v3 2/2] xfs: use sync buffer I/O for sync delwri queue submission

[Brian Foster]

On Mon, Jun 18, 2018 at 10:21:22PM -0700, Darrick J. Wong wrote:
> On Fri, Jun 15, 2018 at 02:05:36PM -0400, Brian Foster wrote:
> > If a delwri queue occurs of a buffer that sits on a delwri queue
> > wait list, the queue sets _XBF_DELWRI_Q without changing the state
> > of ->b_list. This occurs, for example, if another thread beats the
> > current delwri waiter thread to the buffer lock after I/O
> > completion. Once the waiter acquires the lock, it removes the buffer
> > from the wait list and leaves a buffer with _XBF_DELWRI_Q set but
> > not populated on a list. This results in a lost buffer submission
> > and in turn can result in assert failures due to _XBF_DELWRI_Q being
> > set on buffer reclaim or filesystem lockups if the buffer happens to
> > cover an item in the AIL.
> > 
> > This problem has been reproduced by repeated iterations of xfs/305
> > on high CPU count (28xcpu) systems with limited memory (~1GB). Dirty
> > dquot reclaim races with an xfsaild push of a separate dquot backed
> > by the same buffer such that the buffer sits on the reclaim wait
> > list at the time xfsaild attempts to queue it. Since the latter
> > dquot has been flush locked but the underlying buffer not submitted
> > for I/O, the dquot pins the AIL and causes the filesystem to
> > livelock.
> > 
> > This race is essentially made possible by the buffer lock cycle
> > involved with waiting on a synchronous delwri queue submission.
> > Close the race by using synchronous buffer I/O for respective delwri
> > queue submission. This means the buffer remains locked across the
> > I/O and so is inaccessible from other contexts while in the
> > intermediate wait list state. The sync buffer I/O wait mechanism is
> > factored into a helper such that sync delwri buffer submission and
> > serialization are batched operations.
> > 
> > Designed-by: Dave Chinner <dchinner@redhat.com>
> > Signed-off-by: Brian Foster <bfoster@redhat.com>
> > ---
> >  fs/xfs/xfs_buf.c | 78 +++++++++++++++++++++++++-----------------------
> >  1 file changed, 41 insertions(+), 37 deletions(-)
> > 
> > diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c
> > index 7b0f7c79cd62..b5c0f6949b30 100644
> > --- a/fs/xfs/xfs_buf.c
> > +++ b/fs/xfs/xfs_buf.c
> > @@ -1531,6 +1531,20 @@ xfs_buf_submit(
> >  	xfs_buf_rele(bp);
> >  }
> >  
> > +/*
> > + * Wait for I/O completion of a sync buffer and return the I/O error code.
> > + */
> > +static int
> > +xfs_buf_iowait(
> > +	struct xfs_buf	*bp)
> > +{
> > +	trace_xfs_buf_iowait(bp, _RET_IP_);
> > +	wait_for_completion(&bp->b_iowait);
> > +	trace_xfs_buf_iowait_done(bp, _RET_IP_);
> > +
> > +	return bp->b_error;
> > +}
> > +
> >  /*
> >   * Synchronous buffer IO submission path, read or write.
> >   */
> > @@ -1553,12 +1567,7 @@ xfs_buf_submit_wait(
> >  	error = __xfs_buf_submit(bp);
> >  	if (error)
> >  		goto out;
> > -
> > -	/* wait for completion before gathering the error from the buffer */
> > -	trace_xfs_buf_iowait(bp, _RET_IP_);
> > -	wait_for_completion(&bp->b_iowait);
> > -	trace_xfs_buf_iowait_done(bp, _RET_IP_);
> > -	error = bp->b_error;
> > +	error = xfs_buf_iowait(bp);
> >  
> >  out:
> >  	/*
> > @@ -1961,16 +1970,11 @@ xfs_buf_cmp(
> >  }
> >  
> >  /*
> > - * submit buffers for write.
> > - *
> > - * When we have a large buffer list, we do not want to hold all the buffers
> > - * locked while we block on the request queue waiting for IO dispatch. To avoid
> > - * this problem, we lock and submit buffers in groups of 50, thereby minimising
> > - * the lock hold times for lists which may contain thousands of objects.
> > - *
> > - * To do this, we sort the buffer list before we walk the list to lock and
> > - * submit buffers, and we plug and unplug around each group of buffers we
> > - * submit.
> > + * Submit buffers for write. If wait_list is specified, the buffers are
> > + * submitted using sync I/O and placed on the wait list such that the caller can
> > + * iowait each buffer. Otherwise async I/O is used and the buffers are released
> > + * at I/O completion time. In either case, buffers remain locked until I/O
> > + * completes and the buffer is released from the queue.
> >   */
> >  static int
> >  xfs_buf_delwri_submit_buffers(
> > @@ -2012,21 +2016,22 @@ xfs_buf_delwri_submit_buffers(
> >  		trace_xfs_buf_delwri_split(bp, _RET_IP_);
> >  
> >  		/*
> > -		 * We do all IO submission async. This means if we need
> > -		 * to wait for IO completion we need to take an extra
> > -		 * reference so the buffer is still valid on the other
> > -		 * side. We need to move the buffer onto the io_list
> > -		 * at this point so the caller can still access it.
> > +		 * If we have a wait list, each buffer (and associated delwri
> > +		 * queue reference) transfers to it and is submitted
> > +		 * synchronously. Otherwise, drop the buffer from the delwri
> > +		 * queue and submit async.
> >  		 */
> >  		bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_WRITE_FAIL);
> > -		bp->b_flags |= XBF_WRITE | XBF_ASYNC;
> > +		bp->b_flags |= XBF_WRITE;
> >  		if (wait_list) {
> > -			xfs_buf_hold(bp);
> > +			bp->b_flags &= ~XBF_ASYNC;
> >  			list_move_tail(&bp->b_list, wait_list);
> > -		} else
> > +			__xfs_buf_submit(bp);
> > +		} else {
> > +			bp->b_flags |= XBF_ASYNC;
> >  			list_del_init(&bp->b_list);
> > -
> > -		xfs_buf_submit(bp);
> > +			xfs_buf_submit(bp);
> > +		}
> >  	}
> >  	blk_finish_plug(&plug);
> >  
> > @@ -2073,8 +2078,11 @@ xfs_buf_delwri_submit(
> >  
> >  		list_del_init(&bp->b_list);
> >  
> > -		/* locking the buffer will wait for async IO completion. */
> > -		xfs_buf_lock(bp);
> > +		/*
> > +		 * Wait on the locked buffer, check for errors and unlock and
> > +		 * release the delwri queue reference.
> > +		 */
> > +		xfs_buf_iowait(bp);
> >  		error2 = bp->b_error;
> 
> error2 = xfs_buf_iowait()?
> 
> >  		xfs_buf_relse(bp);
> >  		if (!error)
> > @@ -2121,22 +2129,18 @@ xfs_buf_delwri_pushbuf(
> >  
> >  	/*
> >  	 * Delwri submission clears the DELWRI_Q buffer flag and returns with
> > -	 * the buffer on the wait list with an associated reference. Rather than
> > +	 * the buffer on the wait list with the original reference. Rather than
> >  	 * bounce the buffer from a local wait list back to the original list
> >  	 * after I/O completion, reuse the original list as the wait list.
> >  	 */
> >  	xfs_buf_delwri_submit_buffers(&submit_list, buffer_list);
> >  
> >  	/*
> > -	 * The buffer is now under I/O and wait listed as during typical delwri
> > -	 * submission. Lock the buffer to wait for I/O completion. Rather than
> > -	 * remove the buffer from the wait list and release the reference, we
> > -	 * want to return with the buffer queued to the original list. The
> > -	 * buffer already sits on the original list with a wait list reference,
> > -	 * however. If we let the queue inherit that wait list reference, all we
> > -	 * need to do is reset the DELWRI_Q flag.
> > +	 * The buffer is now locked, under I/O and wait listed on the original
> > +	 * delwri queue. Wait for I/O completion, restore the DELWRI_Q flag and
> > +	 * return with the buffer unlocked and on the original queue.
> >  	 */
> > -	xfs_buf_lock(bp);
> > +	xfs_buf_iowait(bp);
> >  	error = bp->b_error;
> 
> error = xfs_buf_iowait()?
> 
> Or is there a specific reason you collected the error separately?
> 

Nope, just an oversight. Will fix..

Brian

> Looks reasonable and I could follow along this time. :)
> 
> --D
> 
> >  	bp->b_flags |= _XBF_DELWRI_Q;
> >  	xfs_buf_unlock(bp);
> > -- 
> > 2.17.1
> > 
> > --
> > To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
> > the body of a message to majordomo@vger.kernel.org
> > More majordomo info at  http://vger.kernel.org/majordomo-info.html
> --
> To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH v4] xfs: use sync buffer I/O for sync delwri queue submission

[Brian Foster]

If a delwri queue occurs of a buffer that sits on a delwri queue
wait list, the queue sets _XBF_DELWRI_Q without changing the state
of ->b_list. This occurs, for example, if another thread beats the
current delwri waiter thread to the buffer lock after I/O
completion. Once the waiter acquires the lock, it removes the buffer
from the wait list and leaves a buffer with _XBF_DELWRI_Q set but
not populated on a list. This results in a lost buffer submission
and in turn can result in assert failures due to _XBF_DELWRI_Q being
set on buffer reclaim or filesystem lockups if the buffer happens to
cover an item in the AIL.

This problem has been reproduced by repeated iterations of xfs/305
on high CPU count (28xcpu) systems with limited memory (~1GB). Dirty
dquot reclaim races with an xfsaild push of a separate dquot backed
by the same buffer such that the buffer sits on the reclaim wait
list at the time xfsaild attempts to queue it. Since the latter
dquot has been flush locked but the underlying buffer not submitted
for I/O, the dquot pins the AIL and causes the filesystem to
livelock.

This race is essentially made possible by the buffer lock cycle
involved with waiting on a synchronous delwri queue submission.
Close the race by using synchronous buffer I/O for respective delwri
queue submission. This means the buffer remains locked across the
I/O and so is inaccessible from other contexts while in the
intermediate wait list state. The sync buffer I/O wait mechanism is
factored into a helper such that sync delwri buffer submission and
serialization are batched operations.

Designed-by: Dave Chinner <dchinner@redhat.com>
Signed-off-by: Brian Foster <bfoster@redhat.com>
---

v4:
- Condense a couple buffer error checks to use the xfs_buf_iowait()
  return value.

 fs/xfs/xfs_buf.c | 80 +++++++++++++++++++++++++-----------------------
 1 file changed, 41 insertions(+), 39 deletions(-)

diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c
index 7b0f7c79cd62..ef234847b4e6 100644
--- a/fs/xfs/xfs_buf.c
+++ b/fs/xfs/xfs_buf.c
@@ -1531,6 +1531,20 @@ xfs_buf_submit(
 	xfs_buf_rele(bp);
 }
 
+/*
+ * Wait for I/O completion of a sync buffer and return the I/O error code.
+ */
+static int
+xfs_buf_iowait(
+	struct xfs_buf	*bp)
+{
+	trace_xfs_buf_iowait(bp, _RET_IP_);
+	wait_for_completion(&bp->b_iowait);
+	trace_xfs_buf_iowait_done(bp, _RET_IP_);
+
+	return bp->b_error;
+}
+
 /*
  * Synchronous buffer IO submission path, read or write.
  */
@@ -1553,12 +1567,7 @@ xfs_buf_submit_wait(
 	error = __xfs_buf_submit(bp);
 	if (error)
 		goto out;
-
-	/* wait for completion before gathering the error from the buffer */
-	trace_xfs_buf_iowait(bp, _RET_IP_);
-	wait_for_completion(&bp->b_iowait);
-	trace_xfs_buf_iowait_done(bp, _RET_IP_);
-	error = bp->b_error;
+	error = xfs_buf_iowait(bp);
 
 out:
 	/*
@@ -1961,16 +1970,11 @@ xfs_buf_cmp(
 }
 
 /*
- * submit buffers for write.
- *
- * When we have a large buffer list, we do not want to hold all the buffers
- * locked while we block on the request queue waiting for IO dispatch. To avoid
- * this problem, we lock and submit buffers in groups of 50, thereby minimising
- * the lock hold times for lists which may contain thousands of objects.
- *
- * To do this, we sort the buffer list before we walk the list to lock and
- * submit buffers, and we plug and unplug around each group of buffers we
- * submit.
+ * Submit buffers for write. If wait_list is specified, the buffers are
+ * submitted using sync I/O and placed on the wait list such that the caller can
+ * iowait each buffer. Otherwise async I/O is used and the buffers are released
+ * at I/O completion time. In either case, buffers remain locked until I/O
+ * completes and the buffer is released from the queue.
  */
 static int
 xfs_buf_delwri_submit_buffers(
@@ -2012,21 +2016,22 @@ xfs_buf_delwri_submit_buffers(
 		trace_xfs_buf_delwri_split(bp, _RET_IP_);
 
 		/*
-		 * We do all IO submission async. This means if we need
-		 * to wait for IO completion we need to take an extra
-		 * reference so the buffer is still valid on the other
-		 * side. We need to move the buffer onto the io_list
-		 * at this point so the caller can still access it.
+		 * If we have a wait list, each buffer (and associated delwri
+		 * queue reference) transfers to it and is submitted
+		 * synchronously. Otherwise, drop the buffer from the delwri
+		 * queue and submit async.
 		 */
 		bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_WRITE_FAIL);
-		bp->b_flags |= XBF_WRITE | XBF_ASYNC;
+		bp->b_flags |= XBF_WRITE;
 		if (wait_list) {
-			xfs_buf_hold(bp);
+			bp->b_flags &= ~XBF_ASYNC;
 			list_move_tail(&bp->b_list, wait_list);
-		} else
+			__xfs_buf_submit(bp);
+		} else {
+			bp->b_flags |= XBF_ASYNC;
 			list_del_init(&bp->b_list);
-
-		xfs_buf_submit(bp);
+			xfs_buf_submit(bp);
+		}
 	}
 	blk_finish_plug(&plug);
 
@@ -2073,9 +2078,11 @@ xfs_buf_delwri_submit(
 
 		list_del_init(&bp->b_list);
 
-		/* locking the buffer will wait for async IO completion. */
-		xfs_buf_lock(bp);
-		error2 = bp->b_error;
+		/*
+		 * Wait on the locked buffer, check for errors and unlock and
+		 * release the delwri queue reference.
+		 */
+		error2 = xfs_buf_iowait(bp);
 		xfs_buf_relse(bp);
 		if (!error)
 			error = error2;
@@ -2121,23 +2128,18 @@ xfs_buf_delwri_pushbuf(
 
 	/*
 	 * Delwri submission clears the DELWRI_Q buffer flag and returns with
-	 * the buffer on the wait list with an associated reference. Rather than
+	 * the buffer on the wait list with the original reference. Rather than
 	 * bounce the buffer from a local wait list back to the original list
 	 * after I/O completion, reuse the original list as the wait list.
 	 */
 	xfs_buf_delwri_submit_buffers(&submit_list, buffer_list);
 
 	/*
-	 * The buffer is now under I/O and wait listed as during typical delwri
-	 * submission. Lock the buffer to wait for I/O completion. Rather than
-	 * remove the buffer from the wait list and release the reference, we
-	 * want to return with the buffer queued to the original list. The
-	 * buffer already sits on the original list with a wait list reference,
-	 * however. If we let the queue inherit that wait list reference, all we
-	 * need to do is reset the DELWRI_Q flag.
+	 * The buffer is now locked, under I/O and wait listed on the original
+	 * delwri queue. Wait for I/O completion, restore the DELWRI_Q flag and
+	 * return with the buffer unlocked and on the original queue.
 	 */
-	xfs_buf_lock(bp);
-	error = bp->b_error;
+	error = xfs_buf_iowait(bp);
 	bp->b_flags |= _XBF_DELWRI_Q;
 	xfs_buf_unlock(bp);
 
-- 
2.17.1