Illumos 5008 - lock contention (rrw_exit) while running a read only load

5008 lock contention (rrw_exit) while running a read only load
Reviewed by: Matthew Ahrens <[email protected]>
Reviewed by: George Wilson <[email protected]>
Reviewed by: Alex Reece <[email protected]>
Reviewed by: Christopher Siden <[email protected]>
Reviewed by: Richard Yao <[email protected]>
Reviewed by: Saso Kiselkov <[email protected]>
Approved by: Garrett D'Amore <[email protected]>

Porting notes:

This patch ported perfectly cleanly to ZoL.  During testing 100% cached
small-block reads, extreme contention was noticed on rrl->rr_lock from
rrw_exit() due to the frequent entering and leaving ZPL.  Illumos picked
up this patch from FreeBSD and it also helps under Linux.

On a 1-minute 4K cached read test with 10 fio processes pinned to a single
socket on a 4-socket (10 thread per socket) NUMA system, contentions on
rrl->rr_lock were reduced from 508799 to 43085.

Ported-by: Tim Chase <[email protected]>
Signed-off-by: Brian Behlendorf <[email protected]>
Closes #3555

6 files changed, 126 insertions, 13 deletions

diff --git a/include/sys/rrwlock.h b/include/sys/rrwlock.h
index d2bdff495..7a328fd68 100644
--- a/include/sys/rrwlock.h
+++ b/include/sys/rrwlock.h
@@ -83,6 +83,31 @@ void rrw_tsd_destroy(void *arg);
 #define	RRW_LOCK_HELD(x) \
 	(rrw_held(x, RW_WRITER) || rrw_held(x, RW_READER))
 
+/*
+ * A reader-mostly lock implementation, tuning above reader-writer locks
+ * for hightly parallel read acquisitions, pessimizing write acquisitions.
+ *
+ * This should be a prime number.  See comment in rrwlock.c near
+ * RRM_TD_LOCK() for details.
+ */
+#define	RRM_NUM_LOCKS		17
+typedef struct rrmlock {
+	rrwlock_t	locks[RRM_NUM_LOCKS];
+} rrmlock_t;
+
+void rrm_init(rrmlock_t *rrl, boolean_t track_all);
+void rrm_destroy(rrmlock_t *rrl);
+void rrm_enter(rrmlock_t *rrl, krw_t rw, void *tag);
+void rrm_enter_read(rrmlock_t *rrl, void *tag);
+void rrm_enter_write(rrmlock_t *rrl);
+void rrm_exit(rrmlock_t *rrl, void *tag);
+boolean_t rrm_held(rrmlock_t *rrl, krw_t rw);
+
+#define	RRM_READ_HELD(x)	rrm_held(x, RW_READER)
+#define	RRM_WRITE_HELD(x)	rrm_held(x, RW_WRITER)
+#define	RRM_LOCK_HELD(x) \
+	(rrm_held(x, RW_WRITER) || rrm_held(x, RW_READER))
+
 #ifdef	__cplusplus
 }
 #endif
diff --git a/include/sys/zfs_vfsops.h b/include/sys/zfs_vfsops.h
index c4db2a911..4073d82f8 100644
--- a/include/sys/zfs_vfsops.h
+++ b/include/sys/zfs_vfsops.h
@@ -67,7 +67,7 @@ typedef struct zfs_sb {
 	boolean_t	z_atime;	/* enable atimes mount option */
 	boolean_t	z_relatime;	/* enable relatime mount option */
 	boolean_t	z_unmounted;	/* unmounted */
-	rrwlock_t	z_teardown_lock;
+	rrmlock_t	z_teardown_lock;
 	krwlock_t	z_teardown_inactive_lock;
 	list_t		z_all_znodes;	/* all znodes in the fs */
 	uint64_t	z_nr_znodes;	/* number of znodes in the fs */
diff --git a/include/sys/zfs_znode.h b/include/sys/zfs_znode.h
index 79ca4f7e9..7e73cf991 100644
--- a/include/sys/zfs_znode.h
+++ b/include/sys/zfs_znode.h
@@ -250,7 +250,7 @@ typedef struct znode {
 /* Called on entry to each ZFS vnode and vfs operation  */
 #define	ZFS_ENTER(zsb) \
 	{ \
-		rrw_enter_read(&(zsb)->z_teardown_lock, FTAG); \
+		rrm_enter_read(&(zsb)->z_teardown_lock, FTAG); \
 		if ((zsb)->z_unmounted) { \
 			ZFS_EXIT(zsb); \
 			return (EIO); \
@@ -260,7 +260,7 @@ typedef struct znode {
 /* Must be called before exiting the vop */
 #define	ZFS_EXIT(zsb) \
 	{ \
-		rrw_exit(&(zsb)->z_teardown_lock, FTAG); \
+		rrm_exit(&(zsb)->z_teardown_lock, FTAG); \
 	}
 
 /* Verifies the znode is valid */
diff --git a/module/zfs/rrwlock.c b/module/zfs/rrwlock.c
index 29a22534e..51394c01c 100644
--- a/module/zfs/rrwlock.c
+++ b/module/zfs/rrwlock.c
@@ -305,3 +305,91 @@ rrw_tsd_destroy(void *arg)
 		    (void *)curthread, (void *)rn->rn_rrl);
 	}
 }
+
+/*
+ * A reader-mostly lock implementation, tuning above reader-writer locks
+ * for hightly parallel read acquisitions, while pessimizing writes.
+ *
+ * The idea is to split single busy lock into array of locks, so that
+ * each reader can lock only one of them for read, depending on result
+ * of simple hash function.  That proportionally reduces lock congestion.
+ * Writer same time has to sequentially aquire write on all the locks.
+ * That makes write aquisition proportionally slower, but in places where
+ * it is used (filesystem unmount) performance is not critical.
+ *
+ * All the functions below are direct wrappers around functions above.
+ */
+void
+rrm_init(rrmlock_t *rrl, boolean_t track_all)
+{
+	int i;
+
+	for (i = 0; i < RRM_NUM_LOCKS; i++)
+		rrw_init(&rrl->locks[i], track_all);
+}
+
+void
+rrm_destroy(rrmlock_t *rrl)
+{
+	int i;
+
+	for (i = 0; i < RRM_NUM_LOCKS; i++)
+		rrw_destroy(&rrl->locks[i]);
+}
+
+void
+rrm_enter(rrmlock_t *rrl, krw_t rw, void *tag)
+{
+	if (rw == RW_READER)
+		rrm_enter_read(rrl, tag);
+	else
+		rrm_enter_write(rrl);
+}
+
+/*
+ * This maps the current thread to a specific lock.  Note that the lock
+ * must be released by the same thread that acquired it.  We do this
+ * mapping by taking the thread pointer mod a prime number.  We examine
+ * only the low 32 bits of the thread pointer, because 32-bit division
+ * is faster than 64-bit division, and the high 32 bits have little
+ * entropy anyway.
+ */
+#define	RRM_TD_LOCK()	(((uint32_t)(uintptr_t)(curthread)) % RRM_NUM_LOCKS)
+
+void
+rrm_enter_read(rrmlock_t *rrl, void *tag)
+{
+	rrw_enter_read(&rrl->locks[RRM_TD_LOCK()], tag);
+}
+
+void
+rrm_enter_write(rrmlock_t *rrl)
+{
+	int i;
+
+	for (i = 0; i < RRM_NUM_LOCKS; i++)
+		rrw_enter_write(&rrl->locks[i]);
+}
+
+void
+rrm_exit(rrmlock_t *rrl, void *tag)
+{
+	int i;
+
+	if (rrl->locks[0].rr_writer == curthread) {
+		for (i = 0; i < RRM_NUM_LOCKS; i++)
+			rrw_exit(&rrl->locks[i], tag);
+	} else {
+		rrw_exit(&rrl->locks[RRM_TD_LOCK()], tag);
+	}
+}
+
+boolean_t
+rrm_held(rrmlock_t *rrl, krw_t rw)
+{
+	if (rw == RW_WRITER) {
+		return (rrw_held(&rrl->locks[0], rw));
+	} else {
+		return (rrw_held(&rrl->locks[RRM_TD_LOCK()], rw));
+	}
+}
diff --git a/module/zfs/zfs_ioctl.c b/module/zfs/zfs_ioctl.c
index c44927036..d997616ae 100644
--- a/module/zfs/zfs_ioctl.c
+++ b/module/zfs/zfs_ioctl.c
@@ -1451,7 +1451,7 @@ zfs_sb_hold(const char *name, void *tag, zfs_sb_t **zsbp, boolean_t writer)
 	if (get_zfs_sb(name, zsbp) != 0)
 		error = zfs_sb_create(name, zsbp);
 	if (error == 0) {
-		rrw_enter(&(*zsbp)->z_teardown_lock, (writer) ? RW_WRITER :
+		rrm_enter(&(*zsbp)->z_teardown_lock, (writer) ? RW_WRITER :
 		    RW_READER, tag);
 		if ((*zsbp)->z_unmounted) {
 			/*
@@ -1459,7 +1459,7 @@ zfs_sb_hold(const char *name, void *tag, zfs_sb_t **zsbp, boolean_t writer)
 			 * thread should be just about to disassociate the
 			 * objset from the zsb.
 			 */
-			rrw_exit(&(*zsbp)->z_teardown_lock, tag);
+			rrm_exit(&(*zsbp)->z_teardown_lock, tag);
 			return (SET_ERROR(EBUSY));
 		}
 	}
@@ -1469,7 +1469,7 @@ zfs_sb_hold(const char *name, void *tag, zfs_sb_t **zsbp, boolean_t writer)
 static void
 zfs_sb_rele(zfs_sb_t *zsb, void *tag)
 {
-	rrw_exit(&zsb->z_teardown_lock, tag);
+	rrm_exit(&zsb->z_teardown_lock, tag);
 
 	if (zsb->z_sb) {
 		deactivate_super(zsb->z_sb);
diff --git a/module/zfs/zfs_vfsops.c b/module/zfs/zfs_vfsops.c
index ae1bc324b..a7005a2a1 100644
--- a/module/zfs/zfs_vfsops.c
+++ b/module/zfs/zfs_vfsops.c
@@ -771,7 +771,7 @@ zfs_sb_create(const char *osname, zfs_sb_t **zsbp)
 	mutex_init(&zsb->z_lock, NULL, MUTEX_DEFAULT, NULL);
 	list_create(&zsb->z_all_znodes, sizeof (znode_t),
 	    offsetof(znode_t, z_link_node));
-	rrw_init(&zsb->z_teardown_lock, B_FALSE);
+	rrm_init(&zsb->z_teardown_lock, B_FALSE);
 	rw_init(&zsb->z_teardown_inactive_lock, NULL, RW_DEFAULT, NULL);
 	rw_init(&zsb->z_fuid_lock, NULL, RW_DEFAULT, NULL);
 
@@ -890,7 +890,7 @@ zfs_sb_free(zfs_sb_t *zsb)
 	mutex_destroy(&zsb->z_znodes_lock);
 	mutex_destroy(&zsb->z_lock);
 	list_destroy(&zsb->z_all_znodes);
-	rrw_destroy(&zsb->z_teardown_lock);
+	rrm_destroy(&zsb->z_teardown_lock);
 	rw_destroy(&zsb->z_teardown_inactive_lock);
 	rw_destroy(&zsb->z_fuid_lock);
 	for (i = 0; i != ZFS_OBJ_MTX_SZ; i++)
@@ -1221,7 +1221,7 @@ zfs_sb_teardown(zfs_sb_t *zsb, boolean_t unmounting)
 		}
 	}
 
-	rrw_enter(&zsb->z_teardown_lock, RW_WRITER, FTAG);
+	rrm_enter(&zsb->z_teardown_lock, RW_WRITER, FTAG);
 
 	if (!unmounting) {
 		/*
@@ -1252,7 +1252,7 @@ zfs_sb_teardown(zfs_sb_t *zsb, boolean_t unmounting)
 	 */
 	if (!unmounting && (zsb->z_unmounted || zsb->z_os == NULL)) {
 		rw_exit(&zsb->z_teardown_inactive_lock);
-		rrw_exit(&zsb->z_teardown_lock, FTAG);
+		rrm_exit(&zsb->z_teardown_lock, FTAG);
 		return (SET_ERROR(EIO));
 	}
 
@@ -1280,7 +1280,7 @@ zfs_sb_teardown(zfs_sb_t *zsb, boolean_t unmounting)
 	 */
 	if (unmounting) {
 		zsb->z_unmounted = B_TRUE;
-		rrw_exit(&zsb->z_teardown_lock, FTAG);
+		rrm_exit(&zsb->z_teardown_lock, FTAG);
 		rw_exit(&zsb->z_teardown_inactive_lock);
 	}
 
@@ -1599,7 +1599,7 @@ zfs_resume_fs(zfs_sb_t *zsb, const char *osname)
 	znode_t *zp;
 	uint64_t sa_obj = 0;
 
-	ASSERT(RRW_WRITE_HELD(&zsb->z_teardown_lock));
+	ASSERT(RRM_WRITE_HELD(&zsb->z_teardown_lock));
 	ASSERT(RW_WRITE_HELD(&zsb->z_teardown_inactive_lock));
 
 	/*
@@ -1663,7 +1663,7 @@ zfs_resume_fs(zfs_sb_t *zsb, const char *osname)
 bail:
 	/* release the VFS ops */
 	rw_exit(&zsb->z_teardown_inactive_lock);
-	rrw_exit(&zsb->z_teardown_lock, FTAG);
+	rrm_exit(&zsb->z_teardown_lock, FTAG);
 
 	if (err) {
 		/*