Merge git://git.kernel.org/pub/scm/linux/kernel/git/mason/btrfs-unstable

* git://git.kernel.org/pub/scm/linux/kernel/git/mason/btrfs-unstable: (37 commits)
  Btrfs: Make sure dir is non-null before doing S_ISGID checks
  Btrfs: Fix memory leak in cache_drop_leaf_ref
  Btrfs: don't return congestion in write_cache_pages as often
  Btrfs: Only prep for btree deletion balances when nodes are mostly empty
  Btrfs: fix btrfs_unlock_up_safe to walk the entire path
  Btrfs: change btrfs_del_leaf to drop locks earlier
  Btrfs: Change btrfs_truncate_inode_items to stop when it hits the inode
  Btrfs: Don't try to compress pages past i_size
  Btrfs: join the transaction in __btrfs_setxattr
  Btrfs: Handle SGID bit when creating inodes
  Btrfs: Make btrfs_drop_snapshot work in larger and more efficient chunks
  Btrfs: Change btree locking to use explicit blocking points
  Btrfs: hash_lock is no longer needed
  Btrfs: disable leak debugging checks in extent_io.c
  Btrfs: sort references by byte number during btrfs_inc_ref
  Btrfs: async threads should try harder to find work
  Btrfs: selinux support
  Btrfs: make btrfs acls selectable
  Btrfs: Catch missed bios in the async bio submission thread
  Btrfs: fix readdir on 32 bit machines
  ...

1 files changed, 190 insertions, 18 deletions

diff --git a/fs/btrfs/locking.c b/fs/btrfs/locking.c
index 39bae7761db..68fd9ccf180 100644
--- a/fs/btrfs/locking.c
+++ b/fs/btrfs/locking.c
@@ -26,45 +26,215 @@
 #include "locking.h"
 
 /*
- * locks the per buffer mutex in an extent buffer.  This uses adaptive locks
- * and the spin is not tuned very extensively.  The spinning does make a big
- * difference in almost every workload, but spinning for the right amount of
- * time needs some help.
- *
- * In general, we want to spin as long as the lock holder is doing btree
- * searches, and we should give up if they are in more expensive code.
+ * btrfs_header_level() isn't free, so don't call it when lockdep isn't
+ * on
  */
+#ifdef CONFIG_DEBUG_LOCK_ALLOC
+static inline void spin_nested(struct extent_buffer *eb)
+{
+	spin_lock_nested(&eb->lock, BTRFS_MAX_LEVEL - btrfs_header_level(eb));
+}
+#else
+static inline void spin_nested(struct extent_buffer *eb)
+{
+	spin_lock(&eb->lock);
+}
+#endif
 
-int btrfs_tree_lock(struct extent_buffer *eb)
+/*
+ * Setting a lock to blocking will drop the spinlock and set the
+ * flag that forces other procs who want the lock to wait.  After
+ * this you can safely schedule with the lock held.
+ */
+void btrfs_set_lock_blocking(struct extent_buffer *eb)
 {
-	int i;
+	if (!test_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags)) {
+		set_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags);
+		spin_unlock(&eb->lock);
+	}
+	/* exit with the spin lock released and the bit set */
+}
 
-	if (mutex_trylock(&eb->mutex))
-		return 0;
+/*
+ * clearing the blocking flag will take the spinlock again.
+ * After this you can't safely schedule
+ */
+void btrfs_clear_lock_blocking(struct extent_buffer *eb)
+{
+	if (test_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags)) {
+		spin_nested(eb);
+		clear_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags);
+		smp_mb__after_clear_bit();
+	}
+	/* exit with the spin lock held */
+}
+
+/*
+ * unfortunately, many of the places that currently set a lock to blocking
+ * don't end up blocking for every long, and often they don't block
+ * at all.  For a dbench 50 run, if we don't spin one the blocking bit
+ * at all, the context switch rate can jump up to 400,000/sec or more.
+ *
+ * So, we're still stuck with this crummy spin on the blocking bit,
+ * at least until the most common causes of the short blocks
+ * can be dealt with.
+ */
+static int btrfs_spin_on_block(struct extent_buffer *eb)
+{
+	int i;
 	for (i = 0; i < 512; i++) {
 		cpu_relax();
-		if (mutex_trylock(&eb->mutex))
+		if (!test_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags))
+			return 1;
+		if (need_resched())
+			break;
+	}
+	return 0;
+}
+
+/*
+ * This is somewhat different from trylock.  It will take the
+ * spinlock but if it finds the lock is set to blocking, it will
+ * return without the lock held.
+ *
+ * returns 1 if it was able to take the lock and zero otherwise
+ *
+ * After this call, scheduling is not safe without first calling
+ * btrfs_set_lock_blocking()
+ */
+int btrfs_try_spin_lock(struct extent_buffer *eb)
+{
+	int i;
+
+	spin_nested(eb);
+	if (!test_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags))
+		return 1;
+	spin_unlock(&eb->lock);
+
+	/* spin for a bit on the BLOCKING flag */
+	for (i = 0; i < 2; i++) {
+		if (!btrfs_spin_on_block(eb))
+			break;
+
+		spin_nested(eb);
+		if (!test_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags))
+			return 1;
+		spin_unlock(&eb->lock);
+	}
+	return 0;
+}
+
+/*
+ * the autoremove wake function will return 0 if it tried to wake up
+ * a process that was already awake, which means that process won't
+ * count as an exclusive wakeup.  The waitq code will continue waking
+ * procs until it finds one that was actually sleeping.
+ *
+ * For btrfs, this isn't quite what we want.  We want a single proc
+ * to be notified that the lock is ready for taking.  If that proc
+ * already happen to be awake, great, it will loop around and try for
+ * the lock.
+ *
+ * So, btrfs_wake_function always returns 1, even when the proc that we
+ * tried to wake up was already awake.
+ */
+static int btrfs_wake_function(wait_queue_t *wait, unsigned mode,
+			       int sync, void *key)
+{
+	autoremove_wake_function(wait, mode, sync, key);
+	return 1;
+}
+
+/*
+ * returns with the extent buffer spinlocked.
+ *
+ * This will spin and/or wait as required to take the lock, and then
+ * return with the spinlock held.
+ *
+ * After this call, scheduling is not safe without first calling
+ * btrfs_set_lock_blocking()
+ */
+int btrfs_tree_lock(struct extent_buffer *eb)
+{
+	DEFINE_WAIT(wait);
+	wait.func = btrfs_wake_function;
+
+	while(1) {
+		spin_nested(eb);
+
+		/* nobody is blocking, exit with the spinlock held */
+		if (!test_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags))
 			return 0;
+
+		/*
+		 * we have the spinlock, but the real owner is blocking.
+		 * wait for them
+		 */
+		spin_unlock(&eb->lock);
+
+		/*
+		 * spin for a bit, and if the blocking flag goes away,
+		 * loop around
+		 */
+		if (btrfs_spin_on_block(eb))
+			continue;
+
+		prepare_to_wait_exclusive(&eb->lock_wq, &wait,
+					  TASK_UNINTERRUPTIBLE);
+
+		if (test_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags))
+			schedule();
+
+		finish_wait(&eb->lock_wq, &wait);
 	}
-	cpu_relax();
-	mutex_lock_nested(&eb->mutex, BTRFS_MAX_LEVEL - btrfs_header_level(eb));
 	return 0;
 }
 
+/*
+ * Very quick trylock, this does not spin or schedule.  It returns
+ * 1 with the spinlock held if it was able to take the lock, or it
+ * returns zero if it was unable to take the lock.
+ *
+ * After this call, scheduling is not safe without first calling
+ * btrfs_set_lock_blocking()
+ */
 int btrfs_try_tree_lock(struct extent_buffer *eb)
 {
-	return mutex_trylock(&eb->mutex);
+	if (spin_trylock(&eb->lock)) {
+		if (test_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags)) {
+			/*
+			 * we've got the spinlock, but the real owner is
+			 * blocking.  Drop the spinlock and return failure
+			 */
+			spin_unlock(&eb->lock);
+			return 0;
+		}
+		return 1;
+	}
+	/* someone else has the spinlock giveup */
+	return 0;
 }
 
 int btrfs_tree_unlock(struct extent_buffer *eb)
 {
-	mutex_unlock(&eb->mutex);
+	/*
+	 * if we were a blocking owner, we don't have the spinlock held
+	 * just clear the bit and look for waiters
+	 */
+	if (test_and_clear_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags))
+		smp_mb__after_clear_bit();
+	else
+		spin_unlock(&eb->lock);
+
+	if (waitqueue_active(&eb->lock_wq))
+		wake_up(&eb->lock_wq);
 	return 0;
 }
 
 int btrfs_tree_locked(struct extent_buffer *eb)
 {
-	return mutex_is_locked(&eb->mutex);
+	return test_bit(EXTENT_BUFFER_BLOCKING, &eb->bflags) ||
+			spin_is_locked(&eb->lock);
 }
 
 /*
@@ -75,12 +245,14 @@ int btrfs_path_lock_waiting(struct btrfs_path *path, int level)
 {
 	int i;
 	struct extent_buffer *eb;
+
 	for (i = level; i <= level + 1 && i < BTRFS_MAX_LEVEL; i++) {
 		eb = path->nodes[i];
 		if (!eb)
 			break;
 		smp_mb();
-		if (!list_empty(&eb->mutex.wait_list))
+		if (spin_is_contended(&eb->lock) ||
+		    waitqueue_active(&eb->lock_wq))
 			return 1;
 	}
 	return 0;