xfs: Improve scalability of busy extent tracking

When we free a metadata extent, we record it in the per-AG busy
extent array so that it is not re-used before the freeing
transaction hits the disk. This array is fixed size, so when it
overflows we make further allocation transactions synchronous
because we cannot track more freed extents until those transactions
hit the disk and are completed. Under heavy mixed allocation and
freeing workloads with large log buffers, we can overflow this array
quite easily.

Further, the array is sparsely populated, which means that inserts
need to search for a free slot, and array searches often have to
search many more slots that are actually used to check all the
busy extents. Quite inefficient, really.

To enable this aspect of extent freeing to scale better, we need
a structure that can grow dynamically. While in other areas of
XFS we have used radix trees, the extents being freed are at random
locations on disk so are better suited to being indexed by an rbtree.

So, use a per-AG rbtree indexed by block number to track busy
extents.  This incures a memory allocation when marking an extent
busy, but should not occur too often in low memory situations. This
should scale to an arbitrary number of extents so should not be a
limitation for features such as in-memory aggregation of
transactions.

However, there are still situations where we can't avoid allocating
busy extents (such as allocation from the AGFL). To minimise the
overhead of such occurences, we need to avoid doing a synchronous
log force while holding the AGF locked to ensure that the previous
transactions are safely on disk before we use the extent. We can do
this by marking the transaction doing the allocation as synchronous
rather issuing a log force.

Because of the locking involved and the ordering of transactions,
the synchronous transaction provides the same guarantees as a
synchronous log force because it ensures that all the prior
transactions are already on disk when the synchronous transaction
hits the disk. i.e. it preserves the free->allocate order of the
extent correctly in recovery.

By doing this, we avoid holding the AGF locked while log writes are
in progress, hence reducing the length of time the lock is held and
therefore we increase the rate at which we can allocate and free
from the allocation group, thereby increasing overall throughput.

The only problem with this approach is that when a metadata buffer is
marked stale (e.g. a directory block is removed), then buffer remains
pinned and locked until the log goes to disk. The issue here is that
if that stale buffer is reallocated in a subsequent transaction, the
attempt to lock that buffer in the transaction will hang waiting
the log to go to disk to unlock and unpin the buffer. Hence if
someone tries to lock a pinned, stale, locked buffer we need to
push on the log to get it unlocked ASAP. Effectively we are trading
off a guaranteed log force for a much less common trigger for log
force to occur.

Ideally we should not reallocate busy extents. That is a much more
complex fix to the problem as it involves direct intervention in the
allocation btree searches in many places. This is left to a future
set of modifications.

Finally, now that we track busy extents in allocated memory, we
don't need the descriptors in the transaction structure to point to
them. We can replace the complex busy chunk infrastructure with a
simple linked list of busy extents. This allows us to remove a large
chunk of code, making the overall change a net reduction in code
size.

Signed-off-by: Dave Chinner <david@fromorbit.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Alex Elder <aelder@sgi.com>
diff --git a/fs/xfs/xfs_alloc.c b/fs/xfs/xfs_alloc.c
index 94cddbf..a7fbe8a 100644
--- a/fs/xfs/xfs_alloc.c
+++ b/fs/xfs/xfs_alloc.c
@@ -46,11 +46,9 @@
 #define	XFSA_FIXUP_BNO_OK	1
 #define	XFSA_FIXUP_CNT_OK	2
 
-STATIC void
-xfs_alloc_search_busy(xfs_trans_t *tp,
-		    xfs_agnumber_t agno,
-		    xfs_agblock_t bno,
-		    xfs_extlen_t len);
+static int
+xfs_alloc_busy_search(struct xfs_mount *mp, xfs_agnumber_t agno,
+		    xfs_agblock_t bno, xfs_extlen_t len);
 
 /*
  * Prototypes for per-ag allocation routines
@@ -540,9 +538,16 @@
 				be32_to_cpu(agf->agf_length));
 			xfs_alloc_log_agf(args->tp, args->agbp,
 						XFS_AGF_FREEBLKS);
-			/* search the busylist for these blocks */
-			xfs_alloc_search_busy(args->tp, args->agno,
-					args->agbno, args->len);
+			/*
+			 * Search the busylist for these blocks and mark the
+			 * transaction as synchronous if blocks are found. This
+			 * avoids the need to block due to a synchronous log
+			 * force to ensure correct ordering as the synchronous
+			 * transaction will guarantee that for us.
+			 */
+			if (xfs_alloc_busy_search(args->mp, args->agno,
+						args->agbno, args->len))
+				xfs_trans_set_sync(args->tp);
 		}
 		if (!args->isfl)
 			xfs_trans_mod_sb(args->tp,
@@ -1693,7 +1698,7 @@
 	 * when the iclog commits to disk.  If a busy block is allocated,
 	 * the iclog is pushed up to the LSN that freed the block.
 	 */
-	xfs_alloc_mark_busy(tp, agno, bno, len);
+	xfs_alloc_busy_insert(tp, agno, bno, len);
 	return 0;
 
  error0:
@@ -1989,14 +1994,20 @@
 	*bnop = bno;
 
 	/*
-	 * As blocks are freed, they are added to the per-ag busy list
-	 * and remain there until the freeing transaction is committed to
-	 * disk.  Now that we have allocated blocks, this list must be
-	 * searched to see if a block is being reused.  If one is, then
-	 * the freeing transaction must be pushed to disk NOW by forcing
-	 * to disk all iclogs up that transaction's LSN.
+	 * As blocks are freed, they are added to the per-ag busy list and
+	 * remain there until the freeing transaction is committed to disk.
+	 * Now that we have allocated blocks, this list must be searched to see
+	 * if a block is being reused.  If one is, then the freeing transaction
+	 * must be pushed to disk before this transaction.
+	 *
+	 * We do this by setting the current transaction to a sync transaction
+	 * which guarantees that the freeing transaction is on disk before this
+	 * transaction. This is done instead of a synchronous log force here so
+	 * that we don't sit and wait with the AGF locked in the transaction
+	 * during the log force.
 	 */
-	xfs_alloc_search_busy(tp, be32_to_cpu(agf->agf_seqno), bno, 1);
+	if (xfs_alloc_busy_search(mp, be32_to_cpu(agf->agf_seqno), bno, 1))
+		xfs_trans_set_sync(tp);
 	return 0;
 }
 
@@ -2201,7 +2212,7 @@
 			be32_to_cpu(agf->agf_levels[XFS_BTNUM_CNTi]);
 		spin_lock_init(&pag->pagb_lock);
 		pag->pagb_count = 0;
-		memset(pag->pagb_list, 0, sizeof(pag->pagb_list));
+		pag->pagb_tree = RB_ROOT;
 		pag->pagf_init = 1;
 	}
 #ifdef DEBUG
@@ -2479,127 +2490,263 @@
  * list is reused, the transaction that freed it must be forced to disk
  * before continuing to use the block.
  *
- * xfs_alloc_mark_busy - add to the per-ag busy list
- * xfs_alloc_clear_busy - remove an item from the per-ag busy list
+ * xfs_alloc_busy_insert - add to the per-ag busy list
+ * xfs_alloc_busy_clear - remove an item from the per-ag busy list
+ * xfs_alloc_busy_search - search for a busy extent
+ */
+
+/*
+ * Insert a new extent into the busy tree.
+ *
+ * The busy extent tree is indexed by the start block of the busy extent.
+ * there can be multiple overlapping ranges in the busy extent tree but only
+ * ever one entry at a given start block. The reason for this is that
+ * multi-block extents can be freed, then smaller chunks of that extent
+ * allocated and freed again before the first transaction commit is on disk.
+ * If the exact same start block is freed a second time, we have to wait for
+ * that busy extent to pass out of the tree before the new extent is inserted.
+ * There are two main cases we have to handle here.
+ *
+ * The first case is a transaction that triggers a "free - allocate - free"
+ * cycle. This can occur during btree manipulations as a btree block is freed
+ * to the freelist, then allocated from the free list, then freed again. In
+ * this case, the second extxpnet free is what triggers the duplicate and as
+ * such the transaction IDs should match. Because the extent was allocated in
+ * this transaction, the transaction must be marked as synchronous. This is
+ * true for all cases where the free/alloc/free occurs in the one transaction,
+ * hence the addition of the ASSERT(tp->t_flags & XFS_TRANS_SYNC) to this case.
+ * This serves to catch violations of the second case quite effectively.
+ *
+ * The second case is where the free/alloc/free occur in different
+ * transactions. In this case, the thread freeing the extent the second time
+ * can't mark the extent busy immediately because it is already tracked in a
+ * transaction that may be committing.  When the log commit for the existing
+ * busy extent completes, the busy extent will be removed from the tree. If we
+ * allow the second busy insert to continue using that busy extent structure,
+ * it can be freed before this transaction is safely in the log.  Hence our
+ * only option in this case is to force the log to remove the existing busy
+ * extent from the list before we insert the new one with the current
+ * transaction ID.
+ *
+ * The problem we are trying to avoid in the free-alloc-free in separate
+ * transactions is most easily described with a timeline:
+ *
+ *      Thread 1	Thread 2	Thread 3	xfslogd
+ *	xact alloc
+ *	free X
+ *	mark busy
+ *	commit xact
+ *	free xact
+ *			xact alloc
+ *			alloc X
+ *			busy search
+ *			mark xact sync
+ *			commit xact
+ *			free xact
+ *			force log
+ *			checkpoint starts
+ *			....
+ *					xact alloc
+ *					free X
+ *					mark busy
+ *					finds match
+ *					*** KABOOM! ***
+ *					....
+ *							log IO completes
+ *							unbusy X
+ *			checkpoint completes
+ *
+ * By issuing a log force in thread 3 @ "KABOOM", the thread will block until
+ * the checkpoint completes, and the busy extent it matched will have been
+ * removed from the tree when it is woken. Hence it can then continue safely.
+ *
+ * However, to ensure this matching process is robust, we need to use the
+ * transaction ID for identifying transaction, as delayed logging results in
+ * the busy extent and transaction lifecycles being different. i.e. the busy
+ * extent is active for a lot longer than the transaction.  Hence the
+ * transaction structure can be freed and reallocated, then mark the same
+ * extent busy again in the new transaction. In this case the new transaction
+ * will have a different tid but can have the same address, and hence we need
+ * to check against the tid.
+ *
+ * Future: for delayed logging, we could avoid the log force if the extent was
+ * first freed in the current checkpoint sequence. This, however, requires the
+ * ability to pin the current checkpoint in memory until this transaction
+ * commits to ensure that both the original free and the current one combine
+ * logically into the one checkpoint. If the checkpoint sequences are
+ * different, however, we still need to wait on a log force.
  */
 void
-xfs_alloc_mark_busy(xfs_trans_t *tp,
-		    xfs_agnumber_t agno,
-		    xfs_agblock_t bno,
-		    xfs_extlen_t len)
+xfs_alloc_busy_insert(
+	struct xfs_trans	*tp,
+	xfs_agnumber_t		agno,
+	xfs_agblock_t		bno,
+	xfs_extlen_t		len)
 {
-	xfs_perag_busy_t	*bsy;
+	struct xfs_busy_extent	*new;
+	struct xfs_busy_extent	*busyp;
 	struct xfs_perag	*pag;
-	int			n;
+	struct rb_node		**rbp;
+	struct rb_node		*parent;
+	int			match;
 
-	pag = xfs_perag_get(tp->t_mountp, agno);
+
+	new = kmem_zalloc(sizeof(struct xfs_busy_extent), KM_MAYFAIL);
+	if (!new) {
+		/*
+		 * No Memory!  Since it is now not possible to track the free
+		 * block, make this a synchronous transaction to insure that
+		 * the block is not reused before this transaction commits.
+		 */
+		trace_xfs_alloc_busy(tp, agno, bno, len, 1);
+		xfs_trans_set_sync(tp);
+		return;
+	}
+
+	new->agno = agno;
+	new->bno = bno;
+	new->length = len;
+	new->tid = xfs_log_get_trans_ident(tp);
+
+	INIT_LIST_HEAD(&new->list);
+
+	/* trace before insert to be able to see failed inserts */
+	trace_xfs_alloc_busy(tp, agno, bno, len, 0);
+
+	pag = xfs_perag_get(tp->t_mountp, new->agno);
+restart:
 	spin_lock(&pag->pagb_lock);
+	rbp = &pag->pagb_tree.rb_node;
+	parent = NULL;
+	busyp = NULL;
+	match = 0;
+	while (*rbp && match >= 0) {
+		parent = *rbp;
+		busyp = rb_entry(parent, struct xfs_busy_extent, rb_node);
 
-	/* search pagb_list for an open slot */
-	for (bsy = pag->pagb_list, n = 0;
-	     n < XFS_PAGB_NUM_SLOTS;
-	     bsy++, n++) {
-		if (bsy->busy_tp == NULL) {
+		if (new->bno < busyp->bno) {
+			/* may overlap, but exact start block is lower */
+			rbp = &(*rbp)->rb_left;
+			if (new->bno + new->length > busyp->bno)
+				match = busyp->tid == new->tid ? 1 : -1;
+		} else if (new->bno > busyp->bno) {
+			/* may overlap, but exact start block is higher */
+			rbp = &(*rbp)->rb_right;
+			if (bno < busyp->bno + busyp->length)
+				match = busyp->tid == new->tid ? 1 : -1;
+		} else {
+			match = busyp->tid == new->tid ? 1 : -1;
 			break;
 		}
 	}
-
-	trace_xfs_alloc_busy(tp->t_mountp, agno, bno, len, n);
-
-	if (n < XFS_PAGB_NUM_SLOTS) {
-		bsy = &pag->pagb_list[n];
-		pag->pagb_count++;
-		bsy->busy_start = bno;
-		bsy->busy_length = len;
-		bsy->busy_tp = tp;
-		xfs_trans_add_busy(tp, agno, n);
-	} else {
-		/*
-		 * The busy list is full!  Since it is now not possible to
-		 * track the free block, make this a synchronous transaction
-		 * to insure that the block is not reused before this
-		 * transaction commits.
-		 */
-		xfs_trans_set_sync(tp);
+	if (match < 0) {
+		/* overlap marked busy in different transaction */
+		spin_unlock(&pag->pagb_lock);
+		xfs_log_force(tp->t_mountp, XFS_LOG_SYNC);
+		goto restart;
 	}
+	if (match > 0) {
+		/*
+		 * overlap marked busy in same transaction. Update if exact
+		 * start block match, otherwise combine the busy extents into
+		 * a single range.
+		 */
+		if (busyp->bno == new->bno) {
+			busyp->length = max(busyp->length, new->length);
+			spin_unlock(&pag->pagb_lock);
+			ASSERT(tp->t_flags & XFS_TRANS_SYNC);
+			xfs_perag_put(pag);
+			kmem_free(new);
+			return;
+		}
+		rb_erase(&busyp->rb_node, &pag->pagb_tree);
+		new->length = max(busyp->bno + busyp->length,
+					new->bno + new->length) -
+				min(busyp->bno, new->bno);
+		new->bno = min(busyp->bno, new->bno);
+	} else
+		busyp = NULL;
 
+	rb_link_node(&new->rb_node, parent, rbp);
+	rb_insert_color(&new->rb_node, &pag->pagb_tree);
+
+	list_add(&new->list, &tp->t_busy);
 	spin_unlock(&pag->pagb_lock);
 	xfs_perag_put(pag);
+	kmem_free(busyp);
+}
+
+/*
+ * Search for a busy extent within the range of the extent we are about to
+ * allocate.  You need to be holding the busy extent tree lock when calling
+ * xfs_alloc_busy_search(). This function returns 0 for no overlapping busy
+ * extent, -1 for an overlapping but not exact busy extent, and 1 for an exact
+ * match. This is done so that a non-zero return indicates an overlap that
+ * will require a synchronous transaction, but it can still be
+ * used to distinguish between a partial or exact match.
+ */
+static int
+xfs_alloc_busy_search(
+	struct xfs_mount	*mp,
+	xfs_agnumber_t		agno,
+	xfs_agblock_t		bno,
+	xfs_extlen_t		len)
+{
+	struct xfs_perag	*pag;
+	struct rb_node		*rbp;
+	struct xfs_busy_extent	*busyp;
+	int			match = 0;
+
+	pag = xfs_perag_get(mp, agno);
+	spin_lock(&pag->pagb_lock);
+
+	rbp = pag->pagb_tree.rb_node;
+
+	/* find closest start bno overlap */
+	while (rbp) {
+		busyp = rb_entry(rbp, struct xfs_busy_extent, rb_node);
+		if (bno < busyp->bno) {
+			/* may overlap, but exact start block is lower */
+			if (bno + len > busyp->bno)
+				match = -1;
+			rbp = rbp->rb_left;
+		} else if (bno > busyp->bno) {
+			/* may overlap, but exact start block is higher */
+			if (bno < busyp->bno + busyp->length)
+				match = -1;
+			rbp = rbp->rb_right;
+		} else {
+			/* bno matches busyp, length determines exact match */
+			match = (busyp->length == len) ? 1 : -1;
+			break;
+		}
+	}
+	spin_unlock(&pag->pagb_lock);
+	trace_xfs_alloc_busysearch(mp, agno, bno, len, !!match);
+	xfs_perag_put(pag);
+	return match;
 }
 
 void
-xfs_alloc_clear_busy(xfs_trans_t *tp,
-		     xfs_agnumber_t agno,
-		     int idx)
+xfs_alloc_busy_clear(
+	struct xfs_mount	*mp,
+	struct xfs_busy_extent	*busyp)
 {
 	struct xfs_perag	*pag;
-	xfs_perag_busy_t	*list;
 
-	ASSERT(idx < XFS_PAGB_NUM_SLOTS);
-	pag = xfs_perag_get(tp->t_mountp, agno);
+	trace_xfs_alloc_unbusy(mp, busyp->agno, busyp->bno,
+						busyp->length);
+
+	ASSERT(xfs_alloc_busy_search(mp, busyp->agno, busyp->bno,
+						busyp->length) == 1);
+
+	list_del_init(&busyp->list);
+
+	pag = xfs_perag_get(mp, busyp->agno);
 	spin_lock(&pag->pagb_lock);
-	list = pag->pagb_list;
-
-	trace_xfs_alloc_unbusy(tp->t_mountp, agno, idx, list[idx].busy_tp == tp);
-
-	if (list[idx].busy_tp == tp) {
-		list[idx].busy_tp = NULL;
-		pag->pagb_count--;
-	}
-
+	rb_erase(&busyp->rb_node, &pag->pagb_tree);
 	spin_unlock(&pag->pagb_lock);
 	xfs_perag_put(pag);
-}
 
-
-/*
- * If we find the extent in the busy list, force the log out to get the
- * extent out of the busy list so the caller can use it straight away.
- */
-STATIC void
-xfs_alloc_search_busy(xfs_trans_t *tp,
-		    xfs_agnumber_t agno,
-		    xfs_agblock_t bno,
-		    xfs_extlen_t len)
-{
-	struct xfs_perag	*pag;
-	xfs_perag_busy_t	*bsy;
-	xfs_agblock_t		uend, bend;
-	xfs_lsn_t		lsn = 0;
-	int			cnt;
-
-	pag = xfs_perag_get(tp->t_mountp, agno);
-	spin_lock(&pag->pagb_lock);
-	cnt = pag->pagb_count;
-
-	/*
-	 * search pagb_list for this slot, skipping open slots. We have to
-	 * search the entire array as there may be multiple overlaps and
-	 * we have to get the most recent LSN for the log force to push out
-	 * all the transactions that span the range.
-	 */
-	uend = bno + len - 1;
-	for (cnt = 0; cnt < pag->pagb_count; cnt++) {
-		bsy = &pag->pagb_list[cnt];
-		if (!bsy->busy_tp)
-			continue;
-
-		bend = bsy->busy_start + bsy->busy_length - 1;
-		if (bno > bend || uend < bsy->busy_start)
-			continue;
-
-		/* (start1,length1) within (start2, length2) */
-		if (XFS_LSN_CMP(bsy->busy_tp->t_commit_lsn, lsn) > 0)
-			lsn = bsy->busy_tp->t_commit_lsn;
-	}
-	spin_unlock(&pag->pagb_lock);
-	xfs_perag_put(pag);
-	trace_xfs_alloc_busysearch(tp->t_mountp, agno, bno, len, lsn);
-
-	/*
-	 * If a block was found, force the log through the LSN of the
-	 * transaction that freed the block
-	 */
-	if (lsn)
-		xfs_log_force_lsn(tp->t_mountp, lsn, XFS_LOG_SYNC);
+	kmem_free(busyp);
 }