xfs: parallelize inode inactivation

Split the inode inactivation work into per-AG work items so that we can
take advantage of parallelization.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>

6 files changed, 104 insertions, 16 deletions

diff --git a/fs/xfs/scrub/common.c b/fs/xfs/scrub/common.c
index abe88fa756aa..299567271e01 100644
--- a/fs/xfs/scrub/common.c
+++ b/fs/xfs/scrub/common.c
@@ -910,6 +910,7 @@ xchk_stop_reaping(
 {
 	sc->flags |= XCHK_REAPING_DISABLED;
 	xfs_stop_block_reaping(sc->mp);
+	xfs_inactive_cancel_work(sc->mp);
 }
 
 /* Restart background reaping of resources. */
@@ -917,6 +918,7 @@ void
 xchk_start_reaping(
 	struct xfs_scrub	*sc)
 {
+	xfs_inactive_schedule_work(sc->mp, 0);
 	xfs_start_block_reaping(sc->mp);
 	sc->flags &= ~XCHK_REAPING_DISABLED;
 }
diff --git a/fs/xfs/xfs_icache.c b/fs/xfs/xfs_icache.c
index 296cf2f2c417..4cce17eb3d6c 100644
--- a/fs/xfs/xfs_icache.c
+++ b/fs/xfs/xfs_icache.c
@@ -1981,12 +1981,12 @@ xfs_start_block_reaping(
 /* Queue a new inode inactivation pass if there are reclaimable inodes. */
 static void
 xfs_inactive_work_queue(
-	struct xfs_mount        *mp)
+	struct xfs_perag	*pag)
 {
 	rcu_read_lock();
-	if (radix_tree_tagged(&mp->m_perag_tree, XFS_ICI_RECLAIM_TAG))
-		queue_delayed_work(mp->m_inactive_workqueue,
-				&mp->m_inactive_work,
+	if (pag->pag_ici_inactive)
+		queue_delayed_work(pag->pag_mount->m_inactive_workqueue,
+				&pag->pag_inactive_work,
 				msecs_to_jiffies(xfs_syncd_centisecs / 6 * 10));
 	rcu_read_unlock();
 }
@@ -2012,7 +2012,7 @@ xfs_perag_set_inactive_tag(
 	 * take a while, so we allow the deferral of an already-scheduled
 	 * inactivation on the grounds that we prefer batching.
 	 */
-	xfs_inactive_work_queue(mp);
+	xfs_inactive_work_queue(pag);
 
 	trace_xfs_perag_set_reclaim(mp, pag->pag_agno, -1, _RET_IP_);
 }
@@ -2121,6 +2121,28 @@ xfs_inactive_inode(
 }
 
 /*
+ * Inactivate the inodes in an AG. Even if the filesystem is corrupted, we
+ * still need to clear the INACTIVE iflag so that we can move on to reclaiming
+ * the inode.
+ */
+static int
+xfs_inactive_inodes_pag(
+	struct xfs_perag	*pag)
+{
+	DEFINE_INACTIVE_CTX(inctx, NULL);
+	int			error;
+
+	error = xfs_ici_walk_ag(pag, 0, xfs_inactive_grab, xfs_inactive_inode,
+			NULL, &inctx, XFS_ICI_RECLAIM_TAG);
+
+	/* If we inactivated any inodes at all, we need to kick reclaim. */
+	if (inctx.kick_reclaim)
+		xfs_reclaim_work_queue(pag->pag_mount);
+
+	return error;
+}
+
+/*
  * Walk the AGs and reclaim the inodes in them. Even if the filesystem is
  * corrupted, we still need to clear the INACTIVE iflag so that we can move
  * on to reclaiming the inode.
@@ -2148,8 +2170,9 @@ void
 xfs_inactive_worker(
 	struct work_struct	*work)
 {
-	struct xfs_mount	*mp = container_of(to_delayed_work(work),
-					struct xfs_mount, m_inactive_work);
+	struct xfs_perag	*pag = container_of(to_delayed_work(work),
+					struct xfs_perag, pag_inactive_work);
+	struct xfs_mount	*mp = pag->pag_mount;
 	int			error;
 
 	/*
@@ -2164,12 +2187,33 @@ xfs_inactive_worker(
 	if (!sb_start_write_trylock(mp->m_super))
 		return;
 
-	error = xfs_inactive_inodes(mp, NULL);
+	error = xfs_inactive_inodes_pag(pag);
 	if (error && error != -EAGAIN)
 		xfs_err(mp, "inode inactivation failed, error %d", error);
 
 	sb_end_write(mp->m_super);
-	xfs_inactive_work_queue(mp);
+	xfs_inactive_work_queue(pag);
+}
+
+/* Wait for all background inactivation work to finish. */
+static void
+xfs_inactive_flush(
+	struct xfs_mount	*mp)
+{
+	struct xfs_perag	*pag;
+	xfs_agnumber_t		agno = 0;
+
+	while ((pag = xfs_perag_get_tag(mp, agno, XFS_ICI_RECLAIM_TAG))) {
+		bool		flush;
+
+		agno = pag->pag_agno + 1;
+		spin_lock(&pag->pag_ici_lock);
+		flush = pag->pag_ici_inactive > 0;
+		spin_unlock(&pag->pag_ici_lock);
+		if (flush)
+			flush_delayed_work(&pag->pag_inactive_work);
+		xfs_perag_put(pag);
+	}
 }
 
 /* Flush all inode inactivation work that might be queued. */
@@ -2177,8 +2221,8 @@ void
 xfs_inactive_force(
 	struct xfs_mount	*mp)
 {
-	queue_delayed_work(mp->m_inactive_workqueue, &mp->m_inactive_work, 0);
-	flush_delayed_work(&mp->m_inactive_work);
+	xfs_inactive_schedule_work(mp, 0);
+	xfs_inactive_flush(mp);
 }
 
 /*
@@ -2190,7 +2234,42 @@ void
 xfs_inactive_shutdown(
 	struct xfs_mount	*mp)
 {
-	cancel_delayed_work_sync(&mp->m_inactive_work);
-	flush_workqueue(mp->m_inactive_workqueue);
+	xfs_inactive_cancel_work(mp);
 	xfs_inactive_inodes(mp, NULL);
 }
+
+/* Cancel all queued inactivation work. */
+void
+xfs_inactive_cancel_work(
+	struct xfs_mount	*mp)
+{
+	struct xfs_perag	*pag;
+	xfs_agnumber_t		agno = 0;
+
+	while ((pag = xfs_perag_get_tag(mp, agno, XFS_ICI_RECLAIM_TAG))) {
+		agno = pag->pag_agno + 1;
+		cancel_delayed_work_sync(&pag->pag_inactive_work);
+		xfs_perag_put(pag);
+	}
+	flush_workqueue(mp->m_inactive_workqueue);
+}
+
+/* Reschedule background inactivation work. */
+void
+xfs_inactive_schedule_work(
+	struct xfs_mount	*mp,
+	unsigned long		delay)
+{
+	struct xfs_perag	*pag;
+	xfs_agnumber_t		agno = 0;
+
+	while ((pag = xfs_perag_get_tag(mp, agno, XFS_ICI_RECLAIM_TAG))) {
+		agno = pag->pag_agno + 1;
+		spin_lock(&pag->pag_ici_lock);
+		if (pag->pag_ici_inactive)
+			queue_delayed_work(mp->m_inactive_workqueue,
+					&pag->pag_inactive_work, delay);
+		spin_unlock(&pag->pag_ici_lock);
+		xfs_perag_put(pag);
+	}
+}
diff --git a/fs/xfs/xfs_icache.h b/fs/xfs/xfs_icache.h
index eaddfba46f3e..f0a18238bc86 100644
--- a/fs/xfs/xfs_icache.h
+++ b/fs/xfs/xfs_icache.h
@@ -86,5 +86,7 @@ void xfs_inactive_worker(struct work_struct *work);
 int xfs_inactive_inodes(struct xfs_mount *mp, struct xfs_eofblocks *eofb);
 void xfs_inactive_force(struct xfs_mount *mp);
 void xfs_inactive_shutdown(struct xfs_mount *mp);
+void xfs_inactive_cancel_work(struct xfs_mount *mp);
+void xfs_inactive_schedule_work(struct xfs_mount *mp, unsigned long delay);
 
 #endif
diff --git a/fs/xfs/xfs_mount.c b/fs/xfs/xfs_mount.c
index 3b8669f6bd24..060e558f44c3 100644
--- a/fs/xfs/xfs_mount.c
+++ b/fs/xfs/xfs_mount.c
@@ -126,6 +126,7 @@ __xfs_free_perag(
 {
 	struct xfs_perag *pag = container_of(head, struct xfs_perag, rcu_head);
 
+	ASSERT(!delayed_work_pending(&pag->pag_inactive_work));
 	ASSERT(atomic_read(&pag->pag_ref) == 0);
 	kmem_free(pag);
 }
@@ -146,6 +147,7 @@ xfs_free_perag(
 		spin_unlock(&mp->m_perag_lock);
 		ASSERT(pag);
 		ASSERT(atomic_read(&pag->pag_ref) == 0);
+		cancel_delayed_work_sync(&pag->pag_inactive_work);
 		xfs_iunlink_destroy(pag);
 		xfs_buf_hash_destroy(pag);
 		mutex_destroy(&pag->pag_ici_reclaim_lock);
@@ -201,6 +203,7 @@ xfs_initialize_perag(
 		pag->pag_mount = mp;
 		spin_lock_init(&pag->pag_ici_lock);
 		mutex_init(&pag->pag_ici_reclaim_lock);
+		INIT_DELAYED_WORK(&pag->pag_inactive_work, xfs_inactive_worker);
 		INIT_RADIX_TREE(&pag->pag_ici_root, GFP_ATOMIC);
 		if (xfs_buf_hash_init(pag))
 			goto out_free_pag;
diff --git a/fs/xfs/xfs_mount.h b/fs/xfs/xfs_mount.h
index 2d675dd37e5b..e5170fbc98e9 100644
--- a/fs/xfs/xfs_mount.h
+++ b/fs/xfs/xfs_mount.h
@@ -177,7 +177,6 @@ typedef struct xfs_mount {
 						     trimming */
 	struct delayed_work	m_cowblocks_work; /* background cow blocks
 						     trimming */
-	struct delayed_work	m_inactive_work; /* background inode inactive */
 	bool			m_update_sb;	/* sb needs update in mount */
 	int64_t			m_low_space[XFS_LOWSP_MAX];
 						/* low free space thresholds */
@@ -416,6 +415,9 @@ typedef struct xfs_perag {
 	/* Blocks reserved for the reverse mapping btree. */
 	struct xfs_ag_resv	pag_rmapbt_resv;
 
+	/* background inode inactivation */
+	struct delayed_work	pag_inactive_work;
+
 	/* reference count */
 	uint8_t			pagf_refcount_level;
 
diff --git a/fs/xfs/xfs_super.c b/fs/xfs/xfs_super.c
index 6ecf4686e8a8..90de69cfc00d 100644
--- a/fs/xfs/xfs_super.c
+++ b/fs/xfs/xfs_super.c
@@ -839,7 +839,8 @@ xfs_init_mount_workqueues(
 		goto out_destroy_eofb;
 
 	mp->m_inactive_workqueue = alloc_workqueue("xfs-inactive/%s",
-			WQ_MEM_RECLAIM|WQ_FREEZABLE, 0, mp->m_fsname);
+			WQ_UNBOUND | WQ_MEM_RECLAIM | WQ_FREEZABLE, 0,
+			mp->m_fsname);
 	if (!mp->m_inactive_workqueue)
 		goto out_destroy_sync;
 
@@ -1641,7 +1642,6 @@ xfs_mount_alloc(
 	INIT_DELAYED_WORK(&mp->m_reclaim_work, xfs_reclaim_worker);
 	INIT_DELAYED_WORK(&mp->m_eofblocks_work, xfs_eofblocks_worker);
 	INIT_DELAYED_WORK(&mp->m_cowblocks_work, xfs_cowblocks_worker);
-	INIT_DELAYED_WORK(&mp->m_inactive_work, xfs_inactive_worker);
 	mp->m_kobj.kobject.kset = xfs_kset;
 	/*
 	 * We don't create the finobt per-ag space reservation until after log