xfs: implement lazy removal for the dquot freelist

Do not remove dquots from the freelist when we grab a reference to them in
xfs_qm_dqlookup, but leave them on the freelist util scanning notices that
they have a reference.  This speeds up the lookup fastpath, and greatly
simplifies the lock ordering constraints.  Note that the same scheme is
used by the VFS inode and dentry caches.

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

4 files changed, 30 insertions, 65 deletions

diff --git a/fs/xfs/xfs_dquot.c b/fs/xfs/xfs_dquot.c
index 3f94f24..35d2b8a 100644
--- a/fs/xfs/xfs_dquot.c
+++ b/fs/xfs/xfs_dquot.c
@@ -722,58 +722,25 @@ xfs_qm_dqlookup(
 		 * dqlock to look at the id field of the dquot, since the
 		 * id can't be modified without the hashlock anyway.
 		 */
-		if (be32_to_cpu(dqp->q_core.d_id) == id && dqp->q_mount == mp) {
-			trace_xfs_dqlookup_found(dqp);
+		if (be32_to_cpu(dqp->q_core.d_id) != id || dqp->q_mount != mp)
+			continue;
 
-			/*
-			 * All in core dquots must be on the dqlist of mp
-			 */
-			ASSERT(!list_empty(&dqp->q_mplist));
-
-			xfs_dqlock(dqp);
-			if (dqp->q_nrefs == 0) {
-				ASSERT(!list_empty(&dqp->q_freelist));
-				if (!mutex_trylock(&xfs_Gqm->qm_dqfrlist_lock)) {
-					trace_xfs_dqlookup_want(dqp);
-
-					/*
-					 * We may have raced with dqreclaim_one()
-					 * (and lost). So, flag that we don't
-					 * want the dquot to be reclaimed.
-					 */
-					dqp->dq_flags |= XFS_DQ_WANT;
-					xfs_dqunlock(dqp);
-					mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
-					xfs_dqlock(dqp);
-					dqp->dq_flags &= ~(XFS_DQ_WANT);
-				}
-
-				if (dqp->q_nrefs == 0) {
-					/* take it off the freelist */
-					trace_xfs_dqlookup_freelist(dqp);
-					list_del_init(&dqp->q_freelist);
-					xfs_Gqm->qm_dqfrlist_cnt--;
-				}
-				XFS_DQHOLD(dqp);
-				mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
-			} else {
-				XFS_DQHOLD(dqp);
-			}
+		trace_xfs_dqlookup_found(dqp);
 
-			/*
-			 * move the dquot to the front of the hashchain
-			 */
-			ASSERT(mutex_is_locked(&qh->qh_lock));
-			list_move(&dqp->q_hashlist, &qh->qh_list);
-			trace_xfs_dqlookup_done(dqp);
-			*O_dqpp = dqp;
-			return 0;
-		}
+		xfs_dqlock(dqp);
+		XFS_DQHOLD(dqp);
+
+		/*
+		 * move the dquot to the front of the hashchain
+		 */
+		list_move(&dqp->q_hashlist, &qh->qh_list);
+		trace_xfs_dqlookup_done(dqp);
+		*O_dqpp = dqp;
+		return 0;
 	}
 
 	*O_dqpp = NULL;
-	ASSERT(mutex_is_locked(&qh->qh_lock));
-	return (1);
+	return 1;
 }
 
 /*
@@ -1033,8 +1000,10 @@ xfs_qm_dqput(
 		if (--dqp->q_nrefs == 0) {
 			trace_xfs_dqput_free(dqp);
 
-			list_add_tail(&dqp->q_freelist, &xfs_Gqm->qm_dqfrlist);
-			xfs_Gqm->qm_dqfrlist_cnt++;
+			if (list_empty(&dqp->q_freelist)) {
+				list_add_tail(&dqp->q_freelist, &xfs_Gqm->qm_dqfrlist);
+				xfs_Gqm->qm_dqfrlist_cnt++;
+			}
 
 			/*
 			 * If we just added a udquot to the freelist, then
diff --git a/fs/xfs/xfs_qm.c b/fs/xfs/xfs_qm.c
index be1df68..6a0c4f0 100644
--- a/fs/xfs/xfs_qm.c
+++ b/fs/xfs/xfs_qm.c
@@ -517,13 +517,12 @@ xfs_qm_dqpurge_int(
 	 * get them off mplist and hashlist, but leave them on freelist.
 	 */
 	list_for_each_entry_safe(dqp, n, &q->qi_dqlist, q_mplist) {
-		/*
-		 * It's OK to look at the type without taking dqlock here.
-		 * We're holding the mplist lock here, and that's needed for
-		 * a dqreclaim.
-		 */
-		if ((dqp->dq_flags & dqtype) == 0)
+		xfs_dqlock(dqp);
+		if ((dqp->dq_flags & dqtype) == 0) {
+			xfs_dqunlock(dqp);
 			continue;
+		}
+		xfs_dqunlock(dqp);
 
 		if (!mutex_trylock(&dqp->q_hash->qh_lock)) {
 			nrecl = q->qi_dqreclaims;
@@ -1692,14 +1691,15 @@ again:
 		xfs_dqlock(dqp);
 
 		/*
-		 * We are racing with dqlookup here. Naturally we don't
-		 * want to reclaim a dquot that lookup wants. We release the
-		 * freelist lock and start over, so that lookup will grab
-		 * both the dquot and the freelistlock.
+		 * This dquot has already been grabbed by dqlookup.
+		 * Remove it from the freelist and try again.
 		 */
-		if (dqp->dq_flags & XFS_DQ_WANT) {
+		if (dqp->q_nrefs) {
 			trace_xfs_dqreclaim_want(dqp);
 			XQM_STATS_INC(xqmstats.xs_qm_dqwants);
+
+			list_del_init(&dqp->q_freelist);
+			xfs_Gqm->qm_dqfrlist_cnt--;
 			restarts++;
 			startagain = 1;
 			goto dqunlock;
diff --git a/fs/xfs/xfs_quota.h b/fs/xfs/xfs_quota.h
index cbafdee..487653d 100644
--- a/fs/xfs/xfs_quota.h
+++ b/fs/xfs/xfs_quota.h
@@ -87,7 +87,6 @@ typedef struct xfs_dqblk {
 #define XFS_DQ_PROJ		0x0002		/* project quota */
 #define XFS_DQ_GROUP		0x0004		/* a group quota */
 #define XFS_DQ_DIRTY		0x0008		/* dquot is dirty */
-#define XFS_DQ_WANT		0x0010		/* for lookup/reclaim race */
 
 #define XFS_DQ_ALLTYPES		(XFS_DQ_USER|XFS_DQ_PROJ|XFS_DQ_GROUP)
 
@@ -95,8 +94,7 @@ typedef struct xfs_dqblk {
 	{ XFS_DQ_USER,		"USER" }, \
 	{ XFS_DQ_PROJ,		"PROJ" }, \
 	{ XFS_DQ_GROUP,		"GROUP" }, \
-	{ XFS_DQ_DIRTY,		"DIRTY" }, \
-	{ XFS_DQ_WANT,		"WANT" }
+	{ XFS_DQ_DIRTY,		"DIRTY" }
 
 /*
  * In the worst case, when both user and group quotas are on,
diff --git a/fs/xfs/xfs_trace.h b/fs/xfs/xfs_trace.h
index 4940357..a9d5b1e 100644
--- a/fs/xfs/xfs_trace.h
+++ b/fs/xfs/xfs_trace.h
@@ -743,8 +743,6 @@ DEFINE_DQUOT_EVENT(xfs_dqtobp_read);
 DEFINE_DQUOT_EVENT(xfs_dqread);
 DEFINE_DQUOT_EVENT(xfs_dqread_fail);
 DEFINE_DQUOT_EVENT(xfs_dqlookup_found);
-DEFINE_DQUOT_EVENT(xfs_dqlookup_want);
-DEFINE_DQUOT_EVENT(xfs_dqlookup_freelist);
 DEFINE_DQUOT_EVENT(xfs_dqlookup_done);
 DEFINE_DQUOT_EVENT(xfs_dqget_hit);
 DEFINE_DQUOT_EVENT(xfs_dqget_miss);