xfs: fix buffer delwri queue state race

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 <[email protected]>

Signed-off-by Brian Foster <[email protected]>

---

v3:
- Leave tracepoint in __xfs_buf_submit and kill
trace_xfs_buf_submit_wait().

 fs/xfs/xfs_buf.c   | 85 ++++++++++++++++++++--------------------------
 fs/xfs/xfs_trace.h |  1 -
 2 files changed, 37 insertions(+), 49 deletions(-)

fs/xfs/xfs_buf.c +21 -18

index e9c058e3761c..7b0f7c79cd62 100644
@@ -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

Quoted 104 lines · 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 <[email protected]>

Signed-off-by Brian Foster <[email protected]>

---

v3:
- Leave tracepoint in __xfs_buf_submit and kill
trace_xfs_buf_submit_wait().

 fs/xfs/xfs_buf.c   | 85 ++++++++++++++++++++--------------------------
 fs/xfs/xfs_trace.h |  1 -
 2 files changed, 37 insertions(+), 49 deletions(-)

fs/xfs/xfs_buf.c +21 -18

index e9c058e3761c..7b0f7c79cd62 100644
@@ -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

fs/xfs/xfs_buf.c xfs_buf_submit( · line 1529

Quoted 10 lines

fs/xfs/xfs_buf.c +10 -0

+	 * 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);
+

It seems like we could simple throw in a:

	if (!(bp->b_flags & XBF_ASYNC)) {
		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;;
	}

here and get ret rid of the separate xfs_buf_submit_wait function entirely. Or even factor the above out into a nice little helper.

Otherwise this looks fine to me:

Signed-off-by Christoph Hellwig <[email protected]>

--
To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

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 <[email protected]>

Signed-off-by Brian Foster <[email protected]>

---

fs/xfs/xfs_buf.c +4 -5

@@ -2013,21 +2017,22 @@ 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.

Quoted 29 lines · 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.

Quoted 3 lines

Designed-by Dave Chinner <[email protected]>

Signed-off-by Brian Foster <[email protected]>

---

Just something I noticed on a initial brief scan:

Quoted 17 lines

fs/xfs/xfs_buf.c +5 -6

@@ -2013,21 +2017,22 @@ 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;

fs/xfs/xfs_buf.c xfs_buf_delwri_submit_buffers( · line 2030

Quoted 6 lines

fs/xfs/xfs_buf.c +2 -2

 		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);

We lose a buffer submission tracepoint here.

Cheers,

Dave.

Signature · 2 lines

Dave Chinner
[email protected]

Quoted 53 lines · Dave Chinner

Quoted 29 lines · 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.

Quoted 3 lines

Designed-by Dave Chinner <[email protected]>

Signed-off-by Brian Foster <[email protected]>

---

Just something I noticed on a initial brief scan:

Quoted 17 lines

fs/xfs/xfs_buf.c +5 -6

@@ -2013,21 +2017,22 @@ 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;

Quoted 9 lines

fs/xfs/xfs_buf.c xfs_buf_delwri_submit_buffers( · line 2030

Quoted 6 lines

fs/xfs/xfs_buf.c +2 -2

 		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);

We lose a buffer submission tracepoint here.

Yeah, good catch. What do you think about just killing trace_xfs_buf_submit_wait() and pushing trace_xfs_buf_submit() down into the new helper? It looks like we can already distinguish the io type based on ->b_flags.

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 <[email protected]>

Signed-off-by Brian Foster <[email protected]>

---
 fs/xfs/xfs_buf.c | 78 +++++++++++++++++++++++++-----------------------
 1 file changed, 41 insertions(+), 37 deletions(-)

fs/xfs/xfs_buf.c +0 -0

index 112999ddb75e..113ab6426a40 100644
@@ -1530,6 +1530,21 @@ xfs_buf_submit(
 	xfs_buf_rele(bp);

fs/xfs/xfs_buf.c +4 -0

@@ -2013,21 +2017,22 @@ xfs_buf_delwri_submit_buffers(
 		trace_xfs_buf_delwri_split(bp, _RET_IP_);
 
 		/*
+		 * 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);

Quoted 43 lines · 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 <[email protected]>

Signed-off-by Brian Foster <[email protected]>

---
 fs/xfs/xfs_buf.c | 78 +++++++++++++++++++++++++-----------------------
 1 file changed, 41 insertions(+), 37 deletions(-)

fs/xfs/xfs_buf.c +0 -0

index 112999ddb75e..113ab6426a40 100644
@@ -1530,6 +1530,21 @@ xfs_buf_submit(
 	xfs_buf_rele(bp);
 }
 

fs/xfs/xfs_buf.c xfs_buf_submit( · line 1540

Quoted 8 lines

fs/xfs/xfs_buf.c +8 -0

+/*
+ * Wait on a sync buffer.
+ */
+static int
+xfs_buf_iowait(
+	struct xfs_buf	*bp)
+{
+	/* wait for completion before gathering the error from the buffer */

That comments seems to state the obvious based on the function name and top of function comment. I'd just remove it.

Quoted 16 lines

fs/xfs/xfs_buf.c +8 -0

@@ -2013,21 +2017,22 @@ xfs_buf_delwri_submit_buffers(
 		trace_xfs_buf_delwri_split(bp, _RET_IP_);
 
 		/*
+		 * 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;
 		if (wait_list) {
+			bp->b_flags &= ~XBF_ASYNC;
 			list_move_tail(&bp->b_list, wait_list);
+			__xfs_buf_submit(bp);
+		} else {

fs/xfs/xfs_buf.c xfs_buf_delwri_submit_buffers( · line 2035

Quoted 4 lines

fs/xfs/xfs_buf.c +3 -0

+			bp->b_flags |= XBF_ASYNC;
 			list_del_init(&bp->b_list);
+			xfs_buf_submit(bp);
+		}

Ok, that breaks my idea of just checking XBF_ASYNC in the previous reply. But we could still do that with an explicit flag instead of the duplication.

--
To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

Not totally sure I follow... do you essentially mean to rename xfs_buf_submit_wait() -> xfs_buf_submit() and make the iowait conditional on !XBF_ASYNC and absence of some new "sync_nowait" parameter to the function? Could you clarify how you envision the updated xfs_buf_submit() function signature to look?

If I'm following correctly, that seems fairly reasonable at first thought. This is a separate patch though (refactoring the interface vs. refactoring the implementation to fix a bug).

Quoted 6 lines · Brian Foster

Not totally sure I follow... do you essentially mean to rename xfs_buf_submit_wait() -> xfs_buf_submit() and make the iowait conditional on !XBF_ASYNC and absence of some new "sync_nowait" parameter to the function? Could you clarify how you envision the updated xfs_buf_submit() function signature to look?

Quoted 3 lines

If I'm following correctly, that seems fairly reasonable at first thought. This is a separate patch though (refactoring the interface vs. refactoring the implementation to fix a bug).

Well. I'd suggest something like the patch below for patch 1:

Quoted 7 lines · Christoph Hellwig

Quoted 6 lines · Brian Foster

Not totally sure I follow... do you essentially mean to rename xfs_buf_submit_wait() -> xfs_buf_submit() and make the iowait conditional on !XBF_ASYNC and absence of some new "sync_nowait" parameter to the function? Could you clarify how you envision the updated xfs_buf_submit() function signature to look?

Quoted 6 lines

Quoted 3 lines

If I'm following correctly, that seems fairly reasonable at first thought. This is a separate patch though (refactoring the interface vs. refactoring the implementation to fix a bug).

Well. I'd suggest something like the patch below for patch 1:

Ok, codewise I don't have much of a preference, but I don't think it's worth redoing the regression testing and lowmem testing and whatnot just to change how the guts are refactored here. What's the endgame? I came up with the following on top of patch 2. Compile tested only, and I can refold the _common() helper back into the caller and invert the nowait logic or whatnot..

Brian

--- 8< ---

fs/xfs/xfs_buf.c +23 -49

index 311ca301b7fd..89d8cedf2828 100644
@@ -757,11 +757,7 @@ _xfs_buf_read(
 	bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD);
 	bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
 
-	if (flags & XBF_ASYNC) {
-		xfs_buf_submit(bp);
-		return 0;
-	}
-	return xfs_buf_submit_wait(bp);
+	return xfs_buf_submit(bp);
 }
 
 xfs_buf_t *

@@ -846,7 +842,7 @@ xfs_buf_read_uncached(
 	bp->b_flags |= XBF_READ;
 	bp->b_ops = ops;
 
-	xfs_buf_submit_wait(bp);
+	xfs_buf_submit(bp);
 	if (bp->b_error) {
 		int	error = bp->b_error;
 		xfs_buf_relse(bp);

@@ -1249,7 +1245,7 @@ xfs_bwrite(
 	bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q |
 			 XBF_WRITE_FAIL | XBF_DONE);
 
-	error = xfs_buf_submit_wait(bp);
+	error = xfs_buf_submit(bp);
 	if (error) {
 		xfs_force_shutdown(bp->b_target->bt_mount,
 				   SHUTDOWN_META_IO_ERROR);

@@ -1459,7 +1455,7 @@ _xfs_buf_ioapply(
  * itself.
  */
 static int
-__xfs_buf_submit(
+__xfs_buf_submit_common(
 	struct xfs_buf	*bp)
 {
 	trace_xfs_buf_submit(bp, _RET_IP_);

@@ -1505,32 +1501,6 @@ __xfs_buf_submit(
 	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);
-}
-
 /*
  * Wait for I/O completion of a sync buffer and return the I/O error code.
  */

@@ -1549,30 +1519,33 @@ xfs_buf_iowait(
  * Synchronous buffer IO submission path, read or write.
  */
 int
-xfs_buf_submit_wait(
-	struct xfs_buf	*bp)
+__xfs_buf_submit(
+	struct xfs_buf	*bp,
+	bool		sync_nowait)
 {
 	int		error;
 
-	ASSERT(!(bp->b_flags & XBF_ASYNC));
-
 	/*
-	 * For synchronous IO, the IO does not inherit the submitters reference
-	 * count, nor the buffer lock. Hence we cannot release the reference we
-	 * are about to take until we've waited for all IO completion to occur,
-	 * including any xfs_buf_ioend_async() work that may be pending.
+	 * Grab a reference so the buffer does not go away underneath us. For
+	 * async buffers, I/O completion drops the callers reference, which
+	 * could occur before submission returns.
 	 */
 	xfs_buf_hold(bp);
 
-	error = __xfs_buf_submit(bp);
-	if (error)
+	error = __xfs_buf_submit_common(bp);
+	if (error) {
+		if (bp->b_flags & XBF_ASYNC)
+			xfs_buf_ioend(bp);
 		goto out;
-	error = xfs_buf_iowait(bp);
+	}
 
+	if (!(bp->b_flags & XBF_ASYNC) && !sync_nowait)
+		error = xfs_buf_iowait(bp);
 out:
 	/*
-	 * all done now, we can release the hold that keeps the buffer
-	 * referenced for the entire IO.
+	 * Release the hold that keeps the buffer referenced for the entire
+	 * I/O. Note that if the buffer is async, it is not safe to reference
+	 * after this release.
 	 */
 	xfs_buf_rele(bp);
 	return error;

@@ -1984,6 +1957,7 @@ xfs_buf_delwri_submit_buffers(
 	LIST_HEAD		(submit_list);
 	int			pinned = 0;
 	struct blk_plug		plug;
+	bool			nowait = false;
 
 	list_sort(NULL, buffer_list, xfs_buf_cmp);
 

@@ -2025,12 +1999,12 @@ xfs_buf_delwri_submit_buffers(
 		if (wait_list) {
 			bp->b_flags &= ~XBF_ASYNC;
 			list_move_tail(&bp->b_list, wait_list);
-			__xfs_buf_submit(bp);
+			nowait = true;
 		} else {
 			bp->b_flags |= XBF_ASYNC;
 			list_del_init(&bp->b_list);
-			xfs_buf_submit(bp);
 		}
+		__xfs_buf_submit(bp, nowait);
 	}
 	blk_finish_plug(&plug);
 

fs/xfs/xfs_buf.h +7 -2

index d24dbd4dac39..9bae5c201003 100644
@@ -298,8 +298,13 @@ extern void __xfs_buf_ioerror(struct xfs_buf *bp, int error,
 		xfs_failaddr_t failaddr);
 #define xfs_buf_ioerror(bp, err) __xfs_buf_ioerror((bp), (err), __this_address)
 extern void xfs_buf_ioerror_alert(struct xfs_buf *, const char *func);
-extern void xfs_buf_submit(struct xfs_buf *bp);
-extern int xfs_buf_submit_wait(struct xfs_buf *bp);
+
+extern int __xfs_buf_submit(struct xfs_buf *bp, bool);
+static inline int xfs_buf_submit(struct xfs_buf *bp)
+{
+	return __xfs_buf_submit(bp, false);
+}
+
 extern void xfs_buf_iomove(xfs_buf_t *, size_t, size_t, void *,
 				xfs_buf_rw_t);
 #define xfs_buf_zero(bp, off, len) \

fs/xfs/xfs_log_recover.c +2 -3

index b181b5f57a19..724a76d87564 100644
@@ -196,7 +196,7 @@ xlog_bread_noalign(
 	bp->b_io_length = nbblks;
 	bp->b_error = 0;
 
-	error = xfs_buf_submit_wait(bp);
+	error = xfs_buf_submit(bp);
 	if (error && !XFS_FORCED_SHUTDOWN(log->l_mp))
 		xfs_buf_ioerror_alert(bp, __func__);
 	return error;

@@ -5707,7 +5707,7 @@ xlog_do_recover(
 	bp->b_flags |= XBF_READ;
 	bp->b_ops = &xfs_sb_buf_ops;
 
-	error = xfs_buf_submit_wait(bp);
+	error = xfs_buf_submit(bp);
 	if (error) {
 		if (!XFS_FORCED_SHUTDOWN(mp)) {
 			xfs_buf_ioerror_alert(bp, __func__);
--

To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

Quoted 15 lines · Brian Foster

Quoted 7 lines · Christoph Hellwig

Quoted 6 lines · Brian Foster

Not totally sure I follow... do you essentially mean to rename xfs_buf_submit_wait() -> xfs_buf_submit() and make the iowait conditional on !XBF_ASYNC and absence of some new "sync_nowait" parameter to the function? Could you clarify how you envision the updated xfs_buf_submit() function signature to look?

Quoted 6 lines

Quoted 3 lines

If I'm following correctly, that seems fairly reasonable at first thought. This is a separate patch though (refactoring the interface vs. refactoring the implementation to fix a bug).

Well. I'd suggest something like the patch below for patch 1:

Quoted 6 lines

Ok, codewise I don't have much of a preference, but I don't think it's worth redoing the regression testing and lowmem testing and whatnot just to change how the guts are refactored here. What's the endgame? I came up with the following on top of patch 2. Compile tested only, and I can refold the _common() helper back into the caller and invert the nowait logic or whatnot..

Whatever you all decide, would you mind adding a comment to xfs_buf.c describing how delwri buffers are supposed to work, particularly with regard to what are the buffer states specific to delwri lists, and how the wait list that gets passed around works? I need to check my assumptions about how this mechanism functions. :)

(AKA I looked at both versions, thought 'this delwri queue list stuff looks rather clever...' and then promptly got distracted by mkfs-config and other stuff. :/)

--D

Quoted 197 lines

Brian

--- 8< ---

fs/xfs/xfs_buf.c +23 -49

index 311ca301b7fd..89d8cedf2828 100644
@@ -757,11 +757,7 @@ _xfs_buf_read(
 	bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD);
 	bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
 
-	if (flags & XBF_ASYNC) {
-		xfs_buf_submit(bp);
-		return 0;
-	}
-	return xfs_buf_submit_wait(bp);
+	return xfs_buf_submit(bp);
 }
 
 xfs_buf_t *

@@ -846,7 +842,7 @@ xfs_buf_read_uncached(
 	bp->b_flags |= XBF_READ;
 	bp->b_ops = ops;
 
-	xfs_buf_submit_wait(bp);
+	xfs_buf_submit(bp);
 	if (bp->b_error) {
 		int	error = bp->b_error;
 		xfs_buf_relse(bp);

@@ -1249,7 +1245,7 @@ xfs_bwrite(
 	bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q |
 			 XBF_WRITE_FAIL | XBF_DONE);
 
-	error = xfs_buf_submit_wait(bp);
+	error = xfs_buf_submit(bp);
 	if (error) {
 		xfs_force_shutdown(bp->b_target->bt_mount,
 				   SHUTDOWN_META_IO_ERROR);

@@ -1459,7 +1455,7 @@ _xfs_buf_ioapply(
  * itself.
  */
 static int
-__xfs_buf_submit(
+__xfs_buf_submit_common(
 	struct xfs_buf	*bp)
 {
 	trace_xfs_buf_submit(bp, _RET_IP_);

@@ -1505,32 +1501,6 @@ __xfs_buf_submit(
 	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);
-}
-
 /*
  * Wait for I/O completion of a sync buffer and return the I/O error code.
  */

@@ -1549,30 +1519,33 @@ xfs_buf_iowait(
  * Synchronous buffer IO submission path, read or write.
  */
 int
-xfs_buf_submit_wait(
-	struct xfs_buf	*bp)
+__xfs_buf_submit(
+	struct xfs_buf	*bp,
+	bool		sync_nowait)
 {
 	int		error;
 
-	ASSERT(!(bp->b_flags & XBF_ASYNC));
-
 	/*
-	 * For synchronous IO, the IO does not inherit the submitters reference
-	 * count, nor the buffer lock. Hence we cannot release the reference we
-	 * are about to take until we've waited for all IO completion to occur,
-	 * including any xfs_buf_ioend_async() work that may be pending.
+	 * Grab a reference so the buffer does not go away underneath us. For
+	 * async buffers, I/O completion drops the callers reference, which
+	 * could occur before submission returns.
 	 */
 	xfs_buf_hold(bp);
 
-	error = __xfs_buf_submit(bp);
-	if (error)
+	error = __xfs_buf_submit_common(bp);
+	if (error) {
+		if (bp->b_flags & XBF_ASYNC)
+			xfs_buf_ioend(bp);
 		goto out;
-	error = xfs_buf_iowait(bp);
+	}
 
+	if (!(bp->b_flags & XBF_ASYNC) && !sync_nowait)
+		error = xfs_buf_iowait(bp);
 out:
 	/*
-	 * all done now, we can release the hold that keeps the buffer
-	 * referenced for the entire IO.
+	 * Release the hold that keeps the buffer referenced for the entire
+	 * I/O. Note that if the buffer is async, it is not safe to reference
+	 * after this release.
 	 */
 	xfs_buf_rele(bp);
 	return error;

@@ -1984,6 +1957,7 @@ xfs_buf_delwri_submit_buffers(
 	LIST_HEAD		(submit_list);
 	int			pinned = 0;
 	struct blk_plug		plug;
+	bool			nowait = false;
 
 	list_sort(NULL, buffer_list, xfs_buf_cmp);
 

@@ -2025,12 +1999,12 @@ xfs_buf_delwri_submit_buffers(
 		if (wait_list) {
 			bp->b_flags &= ~XBF_ASYNC;
 			list_move_tail(&bp->b_list, wait_list);
-			__xfs_buf_submit(bp);
+			nowait = true;
 		} else {
 			bp->b_flags |= XBF_ASYNC;
 			list_del_init(&bp->b_list);
-			xfs_buf_submit(bp);
 		}
+		__xfs_buf_submit(bp, nowait);
 	}
 	blk_finish_plug(&plug);
 

fs/xfs/xfs_buf.h +7 -2

index d24dbd4dac39..9bae5c201003 100644
@@ -298,8 +298,13 @@ extern void __xfs_buf_ioerror(struct xfs_buf *bp, int error,
 		xfs_failaddr_t failaddr);
 #define xfs_buf_ioerror(bp, err) __xfs_buf_ioerror((bp), (err), __this_address)
 extern void xfs_buf_ioerror_alert(struct xfs_buf *, const char *func);
-extern void xfs_buf_submit(struct xfs_buf *bp);
-extern int xfs_buf_submit_wait(struct xfs_buf *bp);
+
+extern int __xfs_buf_submit(struct xfs_buf *bp, bool);
+static inline int xfs_buf_submit(struct xfs_buf *bp)
+{
+	return __xfs_buf_submit(bp, false);
+}
+
 extern void xfs_buf_iomove(xfs_buf_t *, size_t, size_t, void *,
 				xfs_buf_rw_t);
 #define xfs_buf_zero(bp, off, len) \

fs/xfs/xfs_log_recover.c +2 -3

index b181b5f57a19..724a76d87564 100644
@@ -196,7 +196,7 @@ xlog_bread_noalign(
 	bp->b_io_length = nbblks;
 	bp->b_error = 0;
 
-	error = xfs_buf_submit_wait(bp);
+	error = xfs_buf_submit(bp);
 	if (error && !XFS_FORCED_SHUTDOWN(log->l_mp))
 		xfs_buf_ioerror_alert(bp, __func__);
 	return error;

@@ -5707,7 +5707,7 @@ xlog_do_recover(
 	bp->b_flags |= XBF_READ;
 	bp->b_ops = &xfs_sb_buf_ops;
 
-	error = xfs_buf_submit_wait(bp);
+	error = xfs_buf_submit(bp);
 	if (error) {
 		if (!XFS_FORCED_SHUTDOWN(mp)) {
 			xfs_buf_ioerror_alert(bp, __func__);
--

To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

Ok, codewise I don't have much of a preference, but I don't think it's worth redoing the regression testing and lowmem testing and whatnot just to change how the guts are refactored here. What's the endgame?

I came
up with the following on top of patch 2. Compile tested only, and I can refold the _common() helper back into the caller and invert the nowait logic or whatnot..

Quoted 3 lines · Brian Foster

Ok, codewise I don't have much of a preference, but I don't think it's worth redoing the regression testing and lowmem testing and whatnot just to change how the guts are refactored here. What's the endgame?

Avoid the nasty duplication as much as possible. I think your patch already goes a long way towards that, but I'd rather go all the way.

Quoted 4 lines

I came
up with the following on top of patch 2. Compile tested only, and I can refold the _common() helper back into the caller and invert the nowait logic or whatnot..

Mostly looks fine, except that it seems pointless to have __xfs_buf_submit_common around instead of merging it into the only caller.

fs/xfs/xfs_buf.c xfs_buf_iowait( · line 1542

Quoted 1 line

fs/xfs/xfs_buf.c +1 -0

+	if (!(bp->b_flags & XBF_ASYNC) && !sync_nowait)

Also I'd rather pass in a

'bool wait' parameter.

xfs_buf_submit() would set it based on XBF_ASYNC, and the delwri code would just set it to false explicitly.

--
To unsubscribe from this list: send the line "unsubscribe linux-xfs" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html