xfs: on-stack delayed write buffer lists

Queue delwri buffers on a local on-stack list instead of a per-buftarg one,
and write back the buffers per-process instead of by waking up xfsbufd.

This is now easily doable given that we have very few places left that write
delwri buffers:

 - log recovery:
	Only done at mount time, and already forcing out the buffers
	synchronously using xfs_flush_buftarg

 - quotacheck:
	Same story.

 - dquot reclaim:
	Writes out dirty dquots on the LRU under memory pressure.  We might
	want to look into doing more of this via xfsaild, but it's already
	more optimal than the synchronous inode reclaim that writes each
	buffer synchronously.

 - xfsaild:
	This is the main beneficiary of the change.  By keeping a local list
	of buffers to write we reduce latency of writing out buffers, and
	more importably we can remove all the delwri list promotions which
	were hitting the buffer cache hard under sustained metadata loads.

The implementation is very straight forward - xfs_buf_delwri_queue now gets
a new list_head pointer that it adds the delwri buffers to, and all callers
need to eventually submit the list using xfs_buf_delwi_submit or
xfs_buf_delwi_submit_nowait.  Buffers that already are on a delwri list are
skipped in xfs_buf_delwri_queue, assuming they already are on another delwri
list.  The biggest change to pass down the buffer list was done to the AIL
pushing. Now that we operate on buffers the trylock, push and pushbuf log
item methods are merged into a single push routine, which tries to lock the
item, and if possible add the buffer that needs writeback to the buffer list.
This leads to much simpler code than the previous split but requires the
individual IOP_PUSH instances to unlock and reacquire the AIL around calls
to blocking routines.

Given that xfsailds now also handle writing out buffers, the conditions for
log forcing and the sleep times needed some small changes.  The most
important one is that we consider an AIL busy as long we still have buffers
to push, and the other one is that we do increment the pushed LSN for
buffers that are under flushing at this moment, but still count them towards
the stuck items for restart purposes.  Without this we could hammer on stuck
items without ever forcing the log and not make progress under heavy random
delete workloads on fast flash storage devices.

[ Dave Chinner:
	- rebase on previous patches.
	- improved comments for XBF_DELWRI_Q handling
	- fix XBF_ASYNC handling in queue submission (test 106 failure)
	- rename delwri submit function buffer list parameters for clarity
	- xfs_efd_item_push() should return XFS_ITEM_PINNED ]

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Mark Tinguely <tinguely@sgi.com>
Signed-off-by: Ben Myers <bpm@sgi.com>

1 files changed, 131 insertions, 210 deletions

diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c
index 6819b5163e33..b82fc5c67fed 100644
--- a/fs/xfs/xfs_buf.c
+++ b/fs/xfs/xfs_buf.c
@@ -42,7 +42,6 @@
 #include "xfs_trace.h"
 
 static kmem_zone_t *xfs_buf_zone;
-STATIC int xfsbufd(void *);
 
 static struct workqueue_struct *xfslogd_workqueue;
 
@@ -144,8 +143,17 @@ void
 xfs_buf_stale(
 	struct xfs_buf	*bp)
 {
+	ASSERT(xfs_buf_islocked(bp));
+
 	bp->b_flags |= XBF_STALE;
-	xfs_buf_delwri_dequeue(bp);
+
+	/*
+	 * Clear the delwri status so that a delwri queue walker will not
+	 * flush this buffer to disk now that it is stale. The delwri queue has
+	 * a reference to the buffer, so this is safe to do.
+	 */
+	bp->b_flags &= ~_XBF_DELWRI_Q;
+
 	atomic_set(&(bp)->b_lru_ref, 0);
 	if (!list_empty(&bp->b_lru)) {
 		struct xfs_buftarg *btp = bp->b_target;
@@ -592,10 +600,10 @@ _xfs_buf_read(
 {
 	int			status;
 
-	ASSERT(!(flags & (XBF_DELWRI|XBF_WRITE)));
+	ASSERT(!(flags & XBF_WRITE));
 	ASSERT(bp->b_bn != XFS_BUF_DADDR_NULL);
 
-	bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_DELWRI | XBF_READ_AHEAD);
+	bp->b_flags &= ~(XBF_WRITE | XBF_ASYNC | XBF_READ_AHEAD);
 	bp->b_flags |= flags & (XBF_READ | XBF_ASYNC | XBF_READ_AHEAD);
 
 	status = xfs_buf_iorequest(bp);
@@ -855,7 +863,7 @@ xfs_buf_rele(
 			spin_unlock(&pag->pag_buf_lock);
 		} else {
 			xfs_buf_lru_del(bp);
-			ASSERT(!(bp->b_flags & (XBF_DELWRI|_XBF_DELWRI_Q)));
+			ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
 			rb_erase(&bp->b_rbnode, &pag->pag_buf_tree);
 			spin_unlock(&pag->pag_buf_lock);
 			xfs_perag_put(pag);
@@ -915,13 +923,6 @@ xfs_buf_lock(
 	trace_xfs_buf_lock_done(bp, _RET_IP_);
 }
 
-/*
- *	Releases the lock on the buffer object.
- *	If the buffer is marked delwri but is not queued, do so before we
- *	unlock the buffer as we need to set flags correctly.  We also need to
- *	take a reference for the delwri queue because the unlocker is going to
- *	drop their's and they don't know we just queued it.
- */
 void
 xfs_buf_unlock(
 	struct xfs_buf		*bp)
@@ -1019,10 +1020,11 @@ xfs_bwrite(
 {
 	int			error;
 
+	ASSERT(xfs_buf_islocked(bp));
+
 	bp->b_flags |= XBF_WRITE;
-	bp->b_flags &= ~(XBF_ASYNC | XBF_READ);
+	bp->b_flags &= ~(XBF_ASYNC | XBF_READ | _XBF_DELWRI_Q);
 
-	xfs_buf_delwri_dequeue(bp);
 	xfs_bdstrat_cb(bp);
 
 	error = xfs_buf_iowait(bp);
@@ -1254,7 +1256,7 @@ xfs_buf_iorequest(
 {
 	trace_xfs_buf_iorequest(bp, _RET_IP_);
 
-	ASSERT(!(bp->b_flags & XBF_DELWRI));
+	ASSERT(!(bp->b_flags & _XBF_DELWRI_Q));
 
 	if (bp->b_flags & XBF_WRITE)
 		xfs_buf_wait_unpin(bp);
@@ -1435,11 +1437,9 @@ xfs_free_buftarg(
 {
 	unregister_shrinker(&btp->bt_shrinker);
 
-	xfs_flush_buftarg(btp, 1);
 	if (mp->m_flags & XFS_MOUNT_BARRIER)
 		xfs_blkdev_issue_flush(btp);
 
-	kthread_stop(btp->bt_task);
 	kmem_free(btp);
 }
 
@@ -1491,20 +1491,6 @@ xfs_setsize_buftarg(
 	return xfs_setsize_buftarg_flags(btp, blocksize, sectorsize, 1);
 }
 
-STATIC int
-xfs_alloc_delwri_queue(
-	xfs_buftarg_t		*btp,
-	const char		*fsname)
-{
-	INIT_LIST_HEAD(&btp->bt_delwri_queue);
-	spin_lock_init(&btp->bt_delwri_lock);
-	btp->bt_flags = 0;
-	btp->bt_task = kthread_run(xfsbufd, btp, "xfsbufd/%s", fsname);
-	if (IS_ERR(btp->bt_task))
-		return PTR_ERR(btp->bt_task);
-	return 0;
-}
-
 xfs_buftarg_t *
 xfs_alloc_buftarg(
 	struct xfs_mount	*mp,
@@ -1527,8 +1513,6 @@ xfs_alloc_buftarg(
 	spin_lock_init(&btp->bt_lru_lock);
 	if (xfs_setsize_buftarg_early(btp, bdev))
 		goto error;
-	if (xfs_alloc_delwri_queue(btp, fsname))
-		goto error;
 	btp->bt_shrinker.shrink = xfs_buftarg_shrink;
 	btp->bt_shrinker.seeks = DEFAULT_SEEKS;
 	register_shrinker(&btp->bt_shrinker);
@@ -1539,125 +1523,52 @@ error:
 	return NULL;
 }
 
-
 /*
- *	Delayed write buffer handling
+ * Add a buffer to the delayed write list.
+ *
+ * This queues a buffer for writeout if it hasn't already been.  Note that
+ * neither this routine nor the buffer list submission functions perform
+ * any internal synchronization.  It is expected that the lists are thread-local
+ * to the callers.
+ *
+ * Returns true if we queued up the buffer, or false if it already had
+ * been on the buffer list.
  */
-void
+bool
 xfs_buf_delwri_queue(
-	xfs_buf_t		*bp)
+	struct xfs_buf		*bp,
+	struct list_head	*list)
 {
-	struct xfs_buftarg	*btp = bp->b_target;
-
-	trace_xfs_buf_delwri_queue(bp, _RET_IP_);
-
+	ASSERT(xfs_buf_islocked(bp));
 	ASSERT(!(bp->b_flags & XBF_READ));
 
-	spin_lock(&btp->bt_delwri_lock);
-	if (!list_empty(&bp->b_list)) {
-		/* if already in the queue, move it to the tail */
-		ASSERT(bp->b_flags & _XBF_DELWRI_Q);
-		list_move_tail(&bp->b_list, &btp->bt_delwri_queue);
-	} else {
-		/* start xfsbufd as it is about to have something to do */
-		if (list_empty(&btp->bt_delwri_queue))
-			wake_up_process(bp->b_target->bt_task);
-
-		atomic_inc(&bp->b_hold);
-		bp->b_flags |= XBF_DELWRI | _XBF_DELWRI_Q | XBF_ASYNC;
-		list_add_tail(&bp->b_list, &btp->bt_delwri_queue);
-	}
-	bp->b_queuetime = jiffies;
-	spin_unlock(&btp->bt_delwri_lock);
-}
-
-void
-xfs_buf_delwri_dequeue(
-	xfs_buf_t		*bp)
-{
-	int			dequeued = 0;
-
-	spin_lock(&bp->b_target->bt_delwri_lock);
-	if ((bp->b_flags & XBF_DELWRI) && !list_empty(&bp->b_list)) {
-		ASSERT(bp->b_flags & _XBF_DELWRI_Q);
-		list_del_init(&bp->b_list);
-		dequeued = 1;
+	/*
+	 * If the buffer is already marked delwri it already is queued up
+	 * by someone else for imediate writeout.  Just ignore it in that
+	 * case.
+	 */
+	if (bp->b_flags & _XBF_DELWRI_Q) {
+		trace_xfs_buf_delwri_queued(bp, _RET_IP_);
+		return false;
 	}
-	bp->b_flags &= ~(XBF_DELWRI|_XBF_DELWRI_Q);
-	spin_unlock(&bp->b_target->bt_delwri_lock);
-
-	if (dequeued)
-		xfs_buf_rele(bp);
-
-	trace_xfs_buf_delwri_dequeue(bp, _RET_IP_);
-}
 
-/*
- * If a delwri buffer needs to be pushed before it has aged out, then promote
- * it to the head of the delwri queue so that it will be flushed on the next
- * xfsbufd run. We do this by resetting the queuetime of the buffer to be older
- * than the age currently needed to flush the buffer. Hence the next time the
- * xfsbufd sees it is guaranteed to be considered old enough to flush.
- */
-void
-xfs_buf_delwri_promote(
-	struct xfs_buf	*bp)
-{
-	struct xfs_buftarg *btp = bp->b_target;
-	long		age = xfs_buf_age_centisecs * msecs_to_jiffies(10) + 1;
-
-	ASSERT(bp->b_flags & XBF_DELWRI);
-	ASSERT(bp->b_flags & _XBF_DELWRI_Q);
+	trace_xfs_buf_delwri_queue(bp, _RET_IP_);
 
 	/*
-	 * Check the buffer age before locking the delayed write queue as we
-	 * don't need to promote buffers that are already past the flush age.
+	 * If a buffer gets written out synchronously or marked stale while it
+	 * is on a delwri list we lazily remove it. To do this, the other party
+	 * clears the  _XBF_DELWRI_Q flag but otherwise leaves the buffer alone.
+	 * It remains referenced and on the list.  In a rare corner case it
+	 * might get readded to a delwri list after the synchronous writeout, in
+	 * which case we need just need to re-add the flag here.
 	 */
-	if (bp->b_queuetime < jiffies - age)
-		return;
-	bp->b_queuetime = jiffies - age;
-	spin_lock(&btp->bt_delwri_lock);
-	list_move(&bp->b_list, &btp->bt_delwri_queue);
-	spin_unlock(&btp->bt_delwri_lock);
-}
-
-/*
- * Move as many buffers as specified to the supplied list
- * idicating if we skipped any buffers to prevent deadlocks.
- */
-STATIC int
-xfs_buf_delwri_split(
-	xfs_buftarg_t	*target,
-	struct list_head *list,
-	unsigned long	age)
-{
-	xfs_buf_t	*bp, *n;
-	int		skipped = 0;
-	int		force;
-
-	force = test_and_clear_bit(XBT_FORCE_FLUSH, &target->bt_flags);
-	INIT_LIST_HEAD(list);
-	spin_lock(&target->bt_delwri_lock);
-	list_for_each_entry_safe(bp, n, &target->bt_delwri_queue, b_list) {
-		ASSERT(bp->b_flags & XBF_DELWRI);
-
-		if (!xfs_buf_ispinned(bp) && xfs_buf_trylock(bp)) {
-			if (!force &&
-			    time_before(jiffies, bp->b_queuetime + age)) {
-				xfs_buf_unlock(bp);
-				break;
-			}
-
-			bp->b_flags &= ~(XBF_DELWRI | _XBF_DELWRI_Q);
-			bp->b_flags |= XBF_WRITE;
-			list_move_tail(&bp->b_list, list);
-			trace_xfs_buf_delwri_split(bp, _RET_IP_);
-		} else
-			skipped++;
+	bp->b_flags |= _XBF_DELWRI_Q;
+	if (list_empty(&bp->b_list)) {
+		atomic_inc(&bp->b_hold);
+		list_add_tail(&bp->b_list, list);
 	}
 
-	spin_unlock(&target->bt_delwri_lock);
-	return skipped;
+	return true;
 }
 
 /*
@@ -1683,99 +1594,109 @@ xfs_buf_cmp(
 	return 0;
 }
 
-STATIC int
-xfsbufd(
-	void		*data)
+static int
+__xfs_buf_delwri_submit(
+	struct list_head	*buffer_list,
+	struct list_head	*io_list,
+	bool			wait)
 {
-	xfs_buftarg_t   *target = (xfs_buftarg_t *)data;
-
-	current->flags |= PF_MEMALLOC;
-
-	set_freezable();
+	struct blk_plug		plug;
+	struct xfs_buf		*bp, *n;
+	int			pinned = 0;
+
+	list_for_each_entry_safe(bp, n, buffer_list, b_list) {
+		if (!wait) {
+			if (xfs_buf_ispinned(bp)) {
+				pinned++;
+				continue;
+			}
+			if (!xfs_buf_trylock(bp))
+				continue;
+		} else {
+			xfs_buf_lock(bp);
+		}
 
-	do {
-		long	age = xfs_buf_age_centisecs * msecs_to_jiffies(10);
-		long	tout = xfs_buf_timer_centisecs * msecs_to_jiffies(10);
-		struct list_head tmp;
-		struct blk_plug plug;
+		/*
+		 * Someone else might have written the buffer synchronously or
+		 * marked it stale in the meantime.  In that case only the
+		 * _XBF_DELWRI_Q flag got cleared, and we have to drop the
+		 * reference and remove it from the list here.
+		 */
+		if (!(bp->b_flags & _XBF_DELWRI_Q)) {
+			list_del_init(&bp->b_list);
+			xfs_buf_relse(bp);
+			continue;
+		}
 
-		if (unlikely(freezing(current)))
-			try_to_freeze();
+		list_move_tail(&bp->b_list, io_list);
+		trace_xfs_buf_delwri_split(bp, _RET_IP_);
+	}
 
-		/* sleep for a long time if there is nothing to do. */
-		if (list_empty(&target->bt_delwri_queue))
-			tout = MAX_SCHEDULE_TIMEOUT;
-		schedule_timeout_interruptible(tout);
+	list_sort(NULL, io_list, xfs_buf_cmp);
 
-		xfs_buf_delwri_split(target, &tmp, age);
-		list_sort(NULL, &tmp, xfs_buf_cmp);
+	blk_start_plug(&plug);
+	list_for_each_entry_safe(bp, n, io_list, b_list) {
+		bp->b_flags &= ~(_XBF_DELWRI_Q | XBF_ASYNC);
+		bp->b_flags |= XBF_WRITE;
 
-		blk_start_plug(&plug);
-		while (!list_empty(&tmp)) {
-			struct xfs_buf *bp;
-			bp = list_first_entry(&tmp, struct xfs_buf, b_list);
+		if (!wait) {
+			bp->b_flags |= XBF_ASYNC;
 			list_del_init(&bp->b_list);
-			xfs_bdstrat_cb(bp);
 		}
-		blk_finish_plug(&plug);
-	} while (!kthread_should_stop());
+		xfs_bdstrat_cb(bp);
+	}
+	blk_finish_plug(&plug);
 
-	return 0;
+	return pinned;
 }
 
 /*
- *	Go through all incore buffers, and release buffers if they belong to
- *	the given device. This is used in filesystem error handling to
- *	preserve the consistency of its metadata.
+ * Write out a buffer list asynchronously.
+ *
+ * This will take the @buffer_list, write all non-locked and non-pinned buffers
+ * out and not wait for I/O completion on any of the buffers.  This interface
+ * is only safely useable for callers that can track I/O completion by higher
+ * level means, e.g. AIL pushing as the @buffer_list is consumed in this
+ * function.
  */
 int
-xfs_flush_buftarg(
-	xfs_buftarg_t	*target,
-	int		wait)
+xfs_buf_delwri_submit_nowait(
+	struct list_head	*buffer_list)
 {
-	xfs_buf_t	*bp;
-	int		pincount = 0;
-	LIST_HEAD(tmp_list);
-	LIST_HEAD(wait_list);
-	struct blk_plug plug;
+	LIST_HEAD		(io_list);
+	return __xfs_buf_delwri_submit(buffer_list, &io_list, false);
+}
 
-	flush_workqueue(xfslogd_workqueue);
+/*
+ * Write out a buffer list synchronously.
+ *
+ * This will take the @buffer_list, write all buffers out and wait for I/O
+ * completion on all of the buffers. @buffer_list is consumed by the function,
+ * so callers must have some other way of tracking buffers if they require such
+ * functionality.
+ */
+int
+xfs_buf_delwri_submit(
+	struct list_head	*buffer_list)
+{
+	LIST_HEAD		(io_list);
+	int			error = 0, error2;
+	struct xfs_buf		*bp;
 
-	set_bit(XBT_FORCE_FLUSH, &target->bt_flags);
-	pincount = xfs_buf_delwri_split(target, &tmp_list, 0);
+	__xfs_buf_delwri_submit(buffer_list, &io_list, true);
 
-	/*
-	 * Dropped the delayed write list lock, now walk the temporary list.
-	 * All I/O is issued async and then if we need to wait for completion
-	 * we do that after issuing all the IO.
-	 */
-	list_sort(NULL, &tmp_list, xfs_buf_cmp);
+	/* Wait for IO to complete. */
+	while (!list_empty(&io_list)) {
+		bp = list_first_entry(&io_list, struct xfs_buf, b_list);
 
-	blk_start_plug(&plug);
-	while (!list_empty(&tmp_list)) {
-		bp = list_first_entry(&tmp_list, struct xfs_buf, b_list);
-		ASSERT(target == bp->b_target);
 		list_del_init(&bp->b_list);
-		if (wait) {
-			bp->b_flags &= ~XBF_ASYNC;
-			list_add(&bp->b_list, &wait_list);
-		}
-		xfs_bdstrat_cb(bp);
-	}
-	blk_finish_plug(&plug);
-
-	if (wait) {
-		/* Wait for IO to complete. */
-		while (!list_empty(&wait_list)) {
-			bp = list_first_entry(&wait_list, struct xfs_buf, b_list);
-
-			list_del_init(&bp->b_list);
-			xfs_buf_iowait(bp);
-			xfs_buf_relse(bp);
-		}
+		error2 = xfs_buf_iowait(bp);
+		xfs_buf_relse(bp);
+		if (!error)
+			error = error2;
 	}
 
-	return pincount;
+	return error;
 }
 
 int __init