xfs: nest qm_dqfrlist_lock inside the dquot qlock

Allow xfs_qm_dqput to work without trylock loops by nesting the freelist lock
inside the dquot qlock.  In turn that requires trylocks in the reclaim path
instead, but given it's a classic tradeoff between fast and slow path, and
we follow the model of the inode and dentry caches.

Document our new lock order now that it has settled.

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

fs/xfs/xfs_dquot.c

@@ -39,20 +39,19 @@
 #include "xfs_qm.h"
 #include "xfs_trace.h"
 
-
 /*
-   LOCK ORDER
-
-   inode lock		    (ilock)
-   dquot hash-chain lock    (hashlock)
-   xqm dquot freelist lock  (freelistlock
-   mount's dquot list lock  (mplistlock)
-   user dquot lock - lock ordering among dquots is based on the uid or gid
-   group dquot lock - similar to udquots. Between the two dquots, the udquot
-		      has to be locked first.
-   pin lock - the dquot lock must be held to take this lock.
-   flush lock - ditto.
-*/
+ * Lock order:
+ *
+ * ip->i_lock
+ *   qh->qh_lock
+ *     qi->qi_dqlist_lock
+ *       dquot->q_qlock (xfs_dqlock() and friends)
+ *         dquot->q_flush (xfs_dqflock() and friends)
+ *         xfs_Gqm->qm_dqfrlist_lock
+ *
+ * If two dquots need to be locked the order is user before group/project,
+ * otherwise by the lowest id first, see xfs_dqlock2.
+ */
 
 #ifdef DEBUG
 xfs_buftarg_t *xfs_dqerror_target;
@@ -984,69 +983,49 @@ restart:
  */
 void
 xfs_qm_dqput(
-	xfs_dquot_t	*dqp)
+	struct xfs_dquot	*dqp)
 {
-	xfs_dquot_t	*gdqp;
+	struct xfs_dquot	*gdqp;
 
 	ASSERT(dqp->q_nrefs > 0);
 	ASSERT(XFS_DQ_IS_LOCKED(dqp));
 
 	trace_xfs_dqput(dqp);
 
-	if (dqp->q_nrefs != 1) {
-		dqp->q_nrefs--;
+recurse:
+	if (--dqp->q_nrefs > 0) {
 		xfs_dqunlock(dqp);
 		return;
 	}
 
+	trace_xfs_dqput_free(dqp);
+
+	mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
+	if (list_empty(&dqp->q_freelist)) {
+		list_add_tail(&dqp->q_freelist, &xfs_Gqm->qm_dqfrlist);
+		xfs_Gqm->qm_dqfrlist_cnt++;
+	}
+	mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
+
 	/*
-	 * drop the dqlock and acquire the freelist and dqlock
-	 * in the right order; but try to get it out-of-order first
+	 * If we just added a udquot to the freelist, then we want to release
+	 * the gdquot reference that it (probably) has. Otherwise it'll keep
+	 * the gdquot from getting reclaimed.
 	 */
-	if (!mutex_trylock(&xfs_Gqm->qm_dqfrlist_lock)) {
-		trace_xfs_dqput_wait(dqp);
-		xfs_dqunlock(dqp);
-		mutex_lock(&xfs_Gqm->qm_dqfrlist_lock);
-		xfs_dqlock(dqp);
+	gdqp = dqp->q_gdquot;
+	if (gdqp) {
+		xfs_dqlock(gdqp);
+		dqp->q_gdquot = NULL;
 	}
+	xfs_dqunlock(dqp);
 
-	while (1) {
-		gdqp = NULL;
-
-		/* We can't depend on nrefs being == 1 here */
-		if (--dqp->q_nrefs == 0) {
-			trace_xfs_dqput_free(dqp);
-
-			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
-			 * we want to release the gdquot reference that
-			 * it (probably) has. Otherwise it'll keep the
-			 * gdquot from getting reclaimed.
-			 */
-			if ((gdqp = dqp->q_gdquot)) {
-				/*
-				 * Avoid a recursive dqput call
-				 */
-				xfs_dqlock(gdqp);
-				dqp->q_gdquot = NULL;
-			}
-		}
-		xfs_dqunlock(dqp);
-
-		/*
-		 * If we had a group quota inside the user quota as a hint,
-		 * release it now.
-		 */
-		if (! gdqp)
-			break;
+	/*
+	 * If we had a group quota hint, release it now.
+	 */
+	if (gdqp) {
 		dqp = gdqp;
+		goto recurse;
 	}
-	mutex_unlock(&xfs_Gqm->qm_dqfrlist_lock);
 }
 
 /*

fs/xfs/xfs_qm.c

@@ -1668,7 +1668,9 @@ xfs_qm_dqreclaim_one(void)
 restart:
 	list_for_each_entry(dqp, &xfs_Gqm->qm_dqfrlist, q_freelist) {
 		struct xfs_mount *mp = dqp->q_mount;
-		xfs_dqlock(dqp);
+
+		if (!xfs_dqlock_nowait(dqp))
+			continue;
 
 		/*
 		 * This dquot has already been grabbed by dqlookup.