[XFS] On machines with more than 8 cpus, when running parallel I/O
threads, the incore superblock lock becomes the limiting factor for
buffered write throughput. Make the contended fields in the incore
superblock use per-cpu counters so that there is no global lock to limit
scalability.
SGI-PV: 946630
SGI-Modid: xfs-linux-melb:xfs-kern:25106a
Signed-off-by: David Chinner <dgc@sgi.com>
Signed-off-by: Nathan Scott <nathans@sgi.com>
diff --git a/fs/xfs/xfs_mount.c b/fs/xfs/xfs_mount.c
index 62188ea..9b43b7b 100644
--- a/fs/xfs/xfs_mount.c
+++ b/fs/xfs/xfs_mount.c
@@ -51,11 +51,31 @@
STATIC void xfs_uuid_unmount(xfs_mount_t *mp);
STATIC void xfs_unmountfs_wait(xfs_mount_t *);
+
+#ifdef HAVE_PERCPU_SB
+STATIC void xfs_icsb_destroy_counters(xfs_mount_t *);
+STATIC void xfs_icsb_balance_counter(xfs_mount_t *, xfs_sb_field_t, int);
+STATIC void xfs_icsb_sync_counters(xfs_mount_t *);
+STATIC int xfs_icsb_modify_counters(xfs_mount_t *, xfs_sb_field_t,
+ int, int);
+STATIC int xfs_icsb_modify_counters_locked(xfs_mount_t *, xfs_sb_field_t,
+ int, int);
+
+#else
+
+#define xfs_icsb_destroy_counters(mp) do { } while (0)
+#define xfs_icsb_balance_counter(mp, a, b) do { } while (0)
+#define xfs_icsb_sync_counters(mp) do { } while (0)
+#define xfs_icsb_modify_counters(mp, a, b, c) do { } while (0)
+#define xfs_icsb_modify_counters_locked(mp, a, b, c) do { } while (0)
+
+#endif
+
static const struct {
- short offset;
- short type; /* 0 = integer
- * 1 = binary / string (no translation)
- */
+ short offset;
+ short type; /* 0 = integer
+ * 1 = binary / string (no translation)
+ */
} xfs_sb_info[] = {
{ offsetof(xfs_sb_t, sb_magicnum), 0 },
{ offsetof(xfs_sb_t, sb_blocksize), 0 },
@@ -113,7 +133,11 @@
{
xfs_mount_t *mp;
- mp = kmem_zalloc(sizeof(*mp), KM_SLEEP);
+ mp = kmem_zalloc(sizeof(xfs_mount_t), KM_SLEEP);
+
+ if (xfs_icsb_init_counters(mp)) {
+ mp->m_flags |= XFS_MOUNT_NO_PERCPU_SB;
+ }
AIL_LOCKINIT(&mp->m_ail_lock, "xfs_ail");
spinlock_init(&mp->m_sb_lock, "xfs_sb");
@@ -136,8 +160,8 @@
*/
void
xfs_mount_free(
- xfs_mount_t *mp,
- int remove_bhv)
+ xfs_mount_t *mp,
+ int remove_bhv)
{
if (mp->m_ihash)
xfs_ihash_free(mp);
@@ -177,6 +201,7 @@
VFS_REMOVEBHV(vfsp, &mp->m_bhv);
}
+ xfs_icsb_destroy_counters(mp);
kmem_free(mp, sizeof(xfs_mount_t));
}
@@ -527,6 +552,10 @@
ASSERT(XFS_BUF_VALUSEMA(bp) <= 0);
}
+ xfs_icsb_balance_counter(mp, XFS_SBS_ICOUNT, 0);
+ xfs_icsb_balance_counter(mp, XFS_SBS_IFREE, 0);
+ xfs_icsb_balance_counter(mp, XFS_SBS_FDBLOCKS, 0);
+
mp->m_sb_bp = bp;
xfs_buf_relse(bp);
ASSERT(XFS_BUF_VALUSEMA(bp) > 0);
@@ -1154,6 +1183,9 @@
sbp = xfs_getsb(mp, 0);
if (!(XFS_MTOVFS(mp)->vfs_flag & VFS_RDONLY ||
XFS_FORCED_SHUTDOWN(mp))) {
+
+ xfs_icsb_sync_counters(mp);
+
/*
* mark shared-readonly if desired
*/
@@ -1227,7 +1259,6 @@
xfs_trans_log_buf(tp, bp, first, last);
}
-
/*
* xfs_mod_incore_sb_unlocked() is a utility routine common used to apply
* a delta to a specified field in the in-core superblock. Simply
@@ -1237,7 +1268,7 @@
*
* The SB_LOCK must be held when this routine is called.
*/
-STATIC int
+int
xfs_mod_incore_sb_unlocked(xfs_mount_t *mp, xfs_sb_field_t field,
int delta, int rsvd)
{
@@ -1406,9 +1437,26 @@
unsigned long s;
int status;
- s = XFS_SB_LOCK(mp);
- status = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
- XFS_SB_UNLOCK(mp, s);
+ /* check for per-cpu counters */
+ switch (field) {
+#ifdef HAVE_PERCPU_SB
+ case XFS_SBS_ICOUNT:
+ case XFS_SBS_IFREE:
+ case XFS_SBS_FDBLOCKS:
+ if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
+ status = xfs_icsb_modify_counters(mp, field,
+ delta, rsvd);
+ break;
+ }
+ /* FALLTHROUGH */
+#endif
+ default:
+ s = XFS_SB_LOCK(mp);
+ status = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
+ XFS_SB_UNLOCK(mp, s);
+ break;
+ }
+
return status;
}
@@ -1445,8 +1493,26 @@
* from the loop so we'll fall into the undo loop
* below.
*/
- status = xfs_mod_incore_sb_unlocked(mp, msbp->msb_field,
- msbp->msb_delta, rsvd);
+ switch (msbp->msb_field) {
+#ifdef HAVE_PERCPU_SB
+ case XFS_SBS_ICOUNT:
+ case XFS_SBS_IFREE:
+ case XFS_SBS_FDBLOCKS:
+ if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
+ status = xfs_icsb_modify_counters_locked(mp,
+ msbp->msb_field,
+ msbp->msb_delta, rsvd);
+ break;
+ }
+ /* FALLTHROUGH */
+#endif
+ default:
+ status = xfs_mod_incore_sb_unlocked(mp,
+ msbp->msb_field,
+ msbp->msb_delta, rsvd);
+ break;
+ }
+
if (status != 0) {
break;
}
@@ -1463,8 +1529,28 @@
if (status != 0) {
msbp--;
while (msbp >= msb) {
- status = xfs_mod_incore_sb_unlocked(mp,
- msbp->msb_field, -(msbp->msb_delta), rsvd);
+ switch (msbp->msb_field) {
+#ifdef HAVE_PERCPU_SB
+ case XFS_SBS_ICOUNT:
+ case XFS_SBS_IFREE:
+ case XFS_SBS_FDBLOCKS:
+ if (!(mp->m_flags & XFS_MOUNT_NO_PERCPU_SB)) {
+ status =
+ xfs_icsb_modify_counters_locked(mp,
+ msbp->msb_field,
+ -(msbp->msb_delta),
+ rsvd);
+ break;
+ }
+ /* FALLTHROUGH */
+#endif
+ default:
+ status = xfs_mod_incore_sb_unlocked(mp,
+ msbp->msb_field,
+ -(msbp->msb_delta),
+ rsvd);
+ break;
+ }
ASSERT(status == 0);
msbp--;
}
@@ -1577,3 +1663,445 @@
xfs_mod_sb(tp, fields);
xfs_trans_commit(tp, 0, NULL);
}
+
+
+#ifdef HAVE_PERCPU_SB
+/*
+ * Per-cpu incore superblock counters
+ *
+ * Simple concept, difficult implementation
+ *
+ * Basically, replace the incore superblock counters with a distributed per cpu
+ * counter for contended fields (e.g. free block count).
+ *
+ * Difficulties arise in that the incore sb is used for ENOSPC checking, and
+ * hence needs to be accurately read when we are running low on space. Hence
+ * there is a method to enable and disable the per-cpu counters based on how
+ * much "stuff" is available in them.
+ *
+ * Basically, a counter is enabled if there is enough free resource to justify
+ * running a per-cpu fast-path. If the per-cpu counter runs out (i.e. a local
+ * ENOSPC), then we disable the counters to synchronise all callers and
+ * re-distribute the available resources.
+ *
+ * If, once we redistributed the available resources, we still get a failure,
+ * we disable the per-cpu counter and go through the slow path.
+ *
+ * The slow path is the current xfs_mod_incore_sb() function. This means that
+ * when we disable a per-cpu counter, we need to drain it's resources back to
+ * the global superblock. We do this after disabling the counter to prevent
+ * more threads from queueing up on the counter.
+ *
+ * Essentially, this means that we still need a lock in the fast path to enable
+ * synchronisation between the global counters and the per-cpu counters. This
+ * is not a problem because the lock will be local to a CPU almost all the time
+ * and have little contention except when we get to ENOSPC conditions.
+ *
+ * Basically, this lock becomes a barrier that enables us to lock out the fast
+ * path while we do things like enabling and disabling counters and
+ * synchronising the counters.
+ *
+ * Locking rules:
+ *
+ * 1. XFS_SB_LOCK() before picking up per-cpu locks
+ * 2. per-cpu locks always picked up via for_each_online_cpu() order
+ * 3. accurate counter sync requires XFS_SB_LOCK + per cpu locks
+ * 4. modifying per-cpu counters requires holding per-cpu lock
+ * 5. modifying global counters requires holding XFS_SB_LOCK
+ * 6. enabling or disabling a counter requires holding the XFS_SB_LOCK
+ * and _none_ of the per-cpu locks.
+ *
+ * Disabled counters are only ever re-enabled by a balance operation
+ * that results in more free resources per CPU than a given threshold.
+ * To ensure counters don't remain disabled, they are rebalanced when
+ * the global resource goes above a higher threshold (i.e. some hysteresis
+ * is present to prevent thrashing).
+ *
+ * Note: hotplug CPUs not yet supported
+ */
+int
+xfs_icsb_init_counters(
+ xfs_mount_t *mp)
+{
+ xfs_icsb_cnts_t *cntp;
+ int i;
+
+ mp->m_sb_cnts = alloc_percpu(xfs_icsb_cnts_t);
+ if (mp->m_sb_cnts == NULL)
+ return -ENOMEM;
+
+ for_each_online_cpu(i) {
+ cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
+ spin_lock_init(&cntp->icsb_lock);
+ }
+ /*
+ * start with all counters disabled so that the
+ * initial balance kicks us off correctly
+ */
+ mp->m_icsb_counters = -1;
+ return 0;
+}
+
+STATIC void
+xfs_icsb_destroy_counters(
+ xfs_mount_t *mp)
+{
+ if (mp->m_sb_cnts)
+ free_percpu(mp->m_sb_cnts);
+}
+
+
+STATIC inline void
+xfs_icsb_lock_all_counters(
+ xfs_mount_t *mp)
+{
+ xfs_icsb_cnts_t *cntp;
+ int i;
+
+ for_each_online_cpu(i) {
+ cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
+ spin_lock(&cntp->icsb_lock);
+ }
+}
+
+STATIC inline void
+xfs_icsb_unlock_all_counters(
+ xfs_mount_t *mp)
+{
+ xfs_icsb_cnts_t *cntp;
+ int i;
+
+ for_each_online_cpu(i) {
+ cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
+ spin_unlock(&cntp->icsb_lock);
+ }
+}
+
+STATIC void
+xfs_icsb_count(
+ xfs_mount_t *mp,
+ xfs_icsb_cnts_t *cnt,
+ int flags)
+{
+ xfs_icsb_cnts_t *cntp;
+ int i;
+
+ memset(cnt, 0, sizeof(xfs_icsb_cnts_t));
+
+ if (!(flags & XFS_ICSB_LAZY_COUNT))
+ xfs_icsb_lock_all_counters(mp);
+
+ for_each_online_cpu(i) {
+ cntp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, i);
+ cnt->icsb_icount += cntp->icsb_icount;
+ cnt->icsb_ifree += cntp->icsb_ifree;
+ cnt->icsb_fdblocks += cntp->icsb_fdblocks;
+ }
+
+ if (!(flags & XFS_ICSB_LAZY_COUNT))
+ xfs_icsb_unlock_all_counters(mp);
+}
+
+STATIC int
+xfs_icsb_counter_disabled(
+ xfs_mount_t *mp,
+ xfs_sb_field_t field)
+{
+ ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
+ return test_bit(field, &mp->m_icsb_counters);
+}
+
+STATIC int
+xfs_icsb_disable_counter(
+ xfs_mount_t *mp,
+ xfs_sb_field_t field)
+{
+ xfs_icsb_cnts_t cnt;
+
+ ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
+
+ xfs_icsb_lock_all_counters(mp);
+ if (!test_and_set_bit(field, &mp->m_icsb_counters)) {
+ /* drain back to superblock */
+
+ xfs_icsb_count(mp, &cnt, XFS_ICSB_SB_LOCKED|XFS_ICSB_LAZY_COUNT);
+ switch(field) {
+ case XFS_SBS_ICOUNT:
+ mp->m_sb.sb_icount = cnt.icsb_icount;
+ break;
+ case XFS_SBS_IFREE:
+ mp->m_sb.sb_ifree = cnt.icsb_ifree;
+ break;
+ case XFS_SBS_FDBLOCKS:
+ mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
+ break;
+ default:
+ BUG();
+ }
+ }
+
+ xfs_icsb_unlock_all_counters(mp);
+
+ return 0;
+}
+
+STATIC void
+xfs_icsb_enable_counter(
+ xfs_mount_t *mp,
+ xfs_sb_field_t field,
+ uint64_t count,
+ uint64_t resid)
+{
+ xfs_icsb_cnts_t *cntp;
+ int i;
+
+ ASSERT((field >= XFS_SBS_ICOUNT) && (field <= XFS_SBS_FDBLOCKS));
+
+ xfs_icsb_lock_all_counters(mp);
+ for_each_online_cpu(i) {
+ cntp = per_cpu_ptr(mp->m_sb_cnts, i);
+ switch (field) {
+ case XFS_SBS_ICOUNT:
+ cntp->icsb_icount = count + resid;
+ break;
+ case XFS_SBS_IFREE:
+ cntp->icsb_ifree = count + resid;
+ break;
+ case XFS_SBS_FDBLOCKS:
+ cntp->icsb_fdblocks = count + resid;
+ break;
+ default:
+ BUG();
+ break;
+ }
+ resid = 0;
+ }
+ clear_bit(field, &mp->m_icsb_counters);
+ xfs_icsb_unlock_all_counters(mp);
+}
+
+STATIC void
+xfs_icsb_sync_counters_int(
+ xfs_mount_t *mp,
+ int flags)
+{
+ xfs_icsb_cnts_t cnt;
+ int s;
+
+ /* Pass 1: lock all counters */
+ if ((flags & XFS_ICSB_SB_LOCKED) == 0)
+ s = XFS_SB_LOCK(mp);
+
+ xfs_icsb_count(mp, &cnt, flags);
+
+ /* Step 3: update mp->m_sb fields */
+ if (!xfs_icsb_counter_disabled(mp, XFS_SBS_ICOUNT))
+ mp->m_sb.sb_icount = cnt.icsb_icount;
+ if (!xfs_icsb_counter_disabled(mp, XFS_SBS_IFREE))
+ mp->m_sb.sb_ifree = cnt.icsb_ifree;
+ if (!xfs_icsb_counter_disabled(mp, XFS_SBS_FDBLOCKS))
+ mp->m_sb.sb_fdblocks = cnt.icsb_fdblocks;
+
+ if ((flags & XFS_ICSB_SB_LOCKED) == 0)
+ XFS_SB_UNLOCK(mp, s);
+}
+
+/*
+ * Accurate update of per-cpu counters to incore superblock
+ */
+STATIC void
+xfs_icsb_sync_counters(
+ xfs_mount_t *mp)
+{
+ xfs_icsb_sync_counters_int(mp, 0);
+}
+
+/*
+ * lazy addition used for things like df, background sb syncs, etc
+ */
+void
+xfs_icsb_sync_counters_lazy(
+ xfs_mount_t *mp)
+{
+ xfs_icsb_sync_counters_int(mp, XFS_ICSB_LAZY_COUNT);
+}
+
+/*
+ * Balance and enable/disable counters as necessary.
+ *
+ * Thresholds for re-enabling counters are somewhat magic.
+ * inode counts are chosen to be the same number as single
+ * on disk allocation chunk per CPU, and free blocks is
+ * something far enough zero that we aren't going thrash
+ * when we get near ENOSPC.
+ */
+#define XFS_ICSB_INO_CNTR_REENABLE 64
+#define XFS_ICSB_FDBLK_CNTR_REENABLE 512
+STATIC void
+xfs_icsb_balance_counter(
+ xfs_mount_t *mp,
+ xfs_sb_field_t field,
+ int flags)
+{
+ uint64_t count, resid = 0;
+ int weight = num_online_cpus();
+ int s;
+
+ if (!(flags & XFS_ICSB_SB_LOCKED))
+ s = XFS_SB_LOCK(mp);
+
+ /* disable counter and sync counter */
+ xfs_icsb_disable_counter(mp, field);
+
+ /* update counters - first CPU gets residual*/
+ switch (field) {
+ case XFS_SBS_ICOUNT:
+ count = mp->m_sb.sb_icount;
+ resid = do_div(count, weight);
+ if (count < XFS_ICSB_INO_CNTR_REENABLE)
+ goto out;
+ break;
+ case XFS_SBS_IFREE:
+ count = mp->m_sb.sb_ifree;
+ resid = do_div(count, weight);
+ if (count < XFS_ICSB_INO_CNTR_REENABLE)
+ goto out;
+ break;
+ case XFS_SBS_FDBLOCKS:
+ count = mp->m_sb.sb_fdblocks;
+ resid = do_div(count, weight);
+ if (count < XFS_ICSB_FDBLK_CNTR_REENABLE)
+ goto out;
+ break;
+ default:
+ BUG();
+ break;
+ }
+
+ xfs_icsb_enable_counter(mp, field, count, resid);
+out:
+ if (!(flags & XFS_ICSB_SB_LOCKED))
+ XFS_SB_UNLOCK(mp, s);
+}
+
+STATIC int
+xfs_icsb_modify_counters_int(
+ xfs_mount_t *mp,
+ xfs_sb_field_t field,
+ int delta,
+ int rsvd,
+ int flags)
+{
+ xfs_icsb_cnts_t *icsbp;
+ long long lcounter; /* long counter for 64 bit fields */
+ int cpu, s, locked = 0;
+ int ret = 0, balance_done = 0;
+
+again:
+ cpu = get_cpu();
+ icsbp = (xfs_icsb_cnts_t *)per_cpu_ptr(mp->m_sb_cnts, cpu),
+ spin_lock(&icsbp->icsb_lock);
+ if (unlikely(xfs_icsb_counter_disabled(mp, field)))
+ goto slow_path;
+
+ switch (field) {
+ case XFS_SBS_ICOUNT:
+ lcounter = icsbp->icsb_icount;
+ lcounter += delta;
+ if (unlikely(lcounter < 0))
+ goto slow_path;
+ icsbp->icsb_icount = lcounter;
+ break;
+
+ case XFS_SBS_IFREE:
+ lcounter = icsbp->icsb_ifree;
+ lcounter += delta;
+ if (unlikely(lcounter < 0))
+ goto slow_path;
+ icsbp->icsb_ifree = lcounter;
+ break;
+
+ case XFS_SBS_FDBLOCKS:
+ BUG_ON((mp->m_resblks - mp->m_resblks_avail) != 0);
+
+ lcounter = icsbp->icsb_fdblocks;
+ lcounter += delta;
+ if (unlikely(lcounter < 0))
+ goto slow_path;
+ icsbp->icsb_fdblocks = lcounter;
+ break;
+ default:
+ BUG();
+ break;
+ }
+ spin_unlock(&icsbp->icsb_lock);
+ put_cpu();
+ if (locked)
+ XFS_SB_UNLOCK(mp, s);
+ return 0;
+
+ /*
+ * The slow path needs to be run with the SBLOCK
+ * held so that we prevent other threads from
+ * attempting to run this path at the same time.
+ * this provides exclusion for the balancing code,
+ * and exclusive fallback if the balance does not
+ * provide enough resources to continue in an unlocked
+ * manner.
+ */
+slow_path:
+ spin_unlock(&icsbp->icsb_lock);
+ put_cpu();
+
+ /* need to hold superblock incase we need
+ * to disable a counter */
+ if (!(flags & XFS_ICSB_SB_LOCKED)) {
+ s = XFS_SB_LOCK(mp);
+ locked = 1;
+ flags |= XFS_ICSB_SB_LOCKED;
+ }
+ if (!balance_done) {
+ xfs_icsb_balance_counter(mp, field, flags);
+ balance_done = 1;
+ goto again;
+ } else {
+ /*
+ * we might not have enough on this local
+ * cpu to allocate for a bulk request.
+ * We need to drain this field from all CPUs
+ * and disable the counter fastpath
+ */
+ xfs_icsb_disable_counter(mp, field);
+ }
+
+ ret = xfs_mod_incore_sb_unlocked(mp, field, delta, rsvd);
+
+ if (locked)
+ XFS_SB_UNLOCK(mp, s);
+ return ret;
+}
+
+STATIC int
+xfs_icsb_modify_counters(
+ xfs_mount_t *mp,
+ xfs_sb_field_t field,
+ int delta,
+ int rsvd)
+{
+ return xfs_icsb_modify_counters_int(mp, field, delta, rsvd, 0);
+}
+
+/*
+ * Called when superblock is already locked
+ */
+STATIC int
+xfs_icsb_modify_counters_locked(
+ xfs_mount_t *mp,
+ xfs_sb_field_t field,
+ int delta,
+ int rsvd)
+{
+ return xfs_icsb_modify_counters_int(mp, field, delta,
+ rsvd, XFS_ICSB_SB_LOCKED);
+}
+#endif