Skip to contentSkip to conversation

fs: optimize inotify/fsnotify code for unwatched files

PATCH In review Dave Hansen 2 participants 4 messages Jun 30, 2015 – Jul 1, 2015

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/3 fs: optimize inotify/fsnotify code for unwatched files

From: Dave Hansen <[email protected]>

I have a _tiny_ microbenchmark that sits in a loop and writes single bytes to a file. Writing one byte to a tmpfs file is around 2x slower than reading one byte from a file, which is a _bit_ more than I expecte. This is a dumb benchmark, but I think it's hard to deny that write() is a hot path and we should avoid unnecessary overhead there.

I did a 'perf record' of 30-second samples of read and write. The top item in a diffprofile is srcu_read_lock() from fsnotify(). There are active inotify fd's from systemd, but nothing is actually listening to the file or its part of the filesystem.

I *think* we can avoid taking the srcu_read_lock() for the common case where there are no actual marks on the file. This means that there will both be nothing to notify for *and* implies that there is no need for clearing the ignore mask.

This patch gave a 13.1% speedup in writes/second on my test, which is an improvement from the 10.8% that I saw with the last version.

Cc Andrew Morton <[email protected]>
Cc Jan Kara <[email protected]>
Cc Al Viro <[email protected]>
Cc Eric Paris <[email protected]>
Cc John McCutchan <[email protected]>
Cc Robert Love <[email protected]>
Cc Andi Kleen <[email protected]>

---

 b/fs/notify/fsnotify.c |   10 ++++++++++
 1 file changed, 10 insertions(+)
fs/notify/fsnotify.c renamed from fs/notify/fsnotify.c~optimize-fsnotify +10 -0
diff -puN fs/notify/fsnotify.c~optimize-fsnotify fs/notify/fsnotify.c
@@ -205,6 +205,16 @@ int fsnotify(struct inode *to_tell, __u3
 		mnt = NULL;
 
 	/*
+	 * Optimization: srcu_read_lock() has a memory barrier which can
+	 * be expensive.  It protects walking the *_fsnotify_marks lists.
+	 * However, if we do not walk the lists, we do not have to do
+	 * SRCU because we have no references to any objects and do not
+	 * need SRCU to keep them "alive".
+	 */
+	if (hlist_empty(&to_tell->i_fsnotify_marks) &&
+	    (!mnt || hlist_empty(&mnt->mnt_fsnotify_marks)))
+		return 0;
+	/*
 	 * if this is a modify event we may need to clear the ignored masks
 	 * otherwise return if neither the inode nor the vfsmount care about
 	 * this type of event.

_

#1 lore raw

PATCH 2/3 fs: use RCU for free_super() vs. __sb_start_write()

From: Dave Hansen <[email protected]>

Currently, __sb_start_write() and freeze_super() can race with each other. __sb_start_write() uses a smp_mb() to ensure that freeze_super() can see its write to sb->s_writers.counter and that it can see freeze_super()'s update to sb->s_writers.frozen. This all seems to work fine.

But, this smp_mb() makes __sb_start_write() the single hottest function in the kernel if I sit in a loop and do tiny write()s to tmpfs over and over. This is on a very small 2-core system, so it will only get worse on larger systems.

This _seems_ like an ideal case for RCU. __sb_start_write() is the RCU read-side and is in a very fast, performance-sensitive path. freeze_super() is the RCU writer and is in an extremely rare non-performance-sensitive path.

Instead of doing and smp_wmb() in __sb_start_write(), we do rcu_read_lock(). This ensures that a CPU doing freeze_super() can not proceed past its synchronize_rcu() until the grace period has ended and the 's_writers.frozen = SB_FREEZE_WRITE' is visible to __sb_start_write().

This patch increases the number of writes/second that I can do by 5.6% over the last patch. The combined total increase is 19.7%.

---

 b/fs/super.c |   63 ++++++++++++++++++++++++++++++++++++-----------------------
 1 file changed, 39 insertions(+), 24 deletions(-)
fs/super.c renamed from fs/super.c~rcu-__sb_start_write +39 -24
diff -puN fs/super.c~rcu-__sb_start_write fs/super.c
@@ -1190,27 +1190,21 @@ static void acquire_freeze_lock(struct s
  */
 int __sb_start_write(struct super_block *sb, int level, bool wait)
 {
-retry:
-	if (unlikely(sb->s_writers.frozen >= level)) {
+	rcu_read_lock();
+	while (unlikely(sb->s_writers.frozen >= level)) {
+		rcu_read_unlock();
 		if (!wait)
 			return 0;
 		wait_event(sb->s_writers.wait_unfrozen,
 			   sb->s_writers.frozen < level);
+		rcu_read_lock();
 	}
 
 #ifdef CONFIG_LOCKDEP
 	acquire_freeze_lock(sb, level, !wait, _RET_IP_);
 #endif
 	percpu_counter_inc(&sb->s_writers.counter[level-1]);
-	/*
-	 * Make sure counter is updated before we check for frozen.
-	 * freeze_super() first sets frozen and then checks the counter.
-	 */
-	smp_mb();
-	if (unlikely(sb->s_writers.frozen >= level)) {
-		__sb_end_write(sb, level);
-		goto retry;
-	}
+	rcu_read_unlock();
 	return 1;
 }
 EXPORT_SYMBOL(__sb_start_write);
@@ -1254,6 +1248,29 @@ static void sb_wait_write(struct super_b
 	} while (writers);
 }
 
+static void __thaw_super(struct super_block *sb)
+{
+	sb->s_writers.frozen = SB_UNFROZEN;
+	/*
+	 * RCU protects us against races where we are taking
+	 * s_writers.frozen in to a less permissive state.  When
+	 * that happens, __sb_start_write() might not yet have
+	 * seen our write and might still increment
+	 * s_writers.counter.
+	 *
+	 * Here, however, we are transitioning to a _more_
+	 * permissive state.  The filesystem is frozen and no
+	 * writes to s_writers.counter are being permitted.
+	 *
+	 * A smp_wmb() is sufficient here because we just need
+	 * to ensure that new calls __sb_start_write() are
+	 * allowed, not that _concurrent_ calls have finished.
+	 */
+	smp_wmb();
+	wake_up(&sb->s_writers.wait_unfrozen);
+	deactivate_locked_super(sb);
+}
+
 /**
  * freeze_super - lock the filesystem and force it into a consistent state
  * @sb: the super to lock
@@ -1312,7 +1329,13 @@ int freeze_super(struct super_block *sb)
 
 	/* From now on, no new normal writers can start */
 	sb->s_writers.frozen = SB_FREEZE_WRITE;
-	smp_wmb();
+	/*
+	 * After we synchronize_rcu(), we have ensured that everyone
+	 * who reads sb->s_writers.frozen under rcu_read_lock() can
+	 * now see our update.  This pretty much means that
+	 * __sb_start_write() will not allow any new writers.
+	 */
+	synchronize_rcu();
 
 	/* Release s_umount to preserve sb_start_write -> s_umount ordering */
 	up_write(&sb->s_umount);
@@ -1322,7 +1345,7 @@ int freeze_super(struct super_block *sb)
 	/* Now we go and block page faults... */
 	down_write(&sb->s_umount);
 	sb->s_writers.frozen = SB_FREEZE_PAGEFAULT;
-	smp_wmb();
+	synchronize_rcu();
 
 	sb_wait_write(sb, SB_FREEZE_PAGEFAULT);
 
@@ -1331,7 +1354,7 @@ int freeze_super(struct super_block *sb)
 
 	/* Now wait for internal filesystem counter */
 	sb->s_writers.frozen = SB_FREEZE_FS;
-	smp_wmb();
+	synchronize_rcu();
 	sb_wait_write(sb, SB_FREEZE_FS);
 
 	if (sb->s_op->freeze_fs) {
@@ -1339,11 +1362,7 @@ int freeze_super(struct super_block *sb)
 		if (ret) {
 			printk(KERN_ERR
 				"VFS:Filesystem freeze failed\n");
-			sb->s_writers.frozen = SB_UNFROZEN;
-			smp_wmb();
-			wake_up(&sb->s_writers.wait_unfrozen);
-			deactivate_locked_super(sb);
-			return ret;
+			__thaw_super(sb);
 		}
 	}
 	/*
@@ -1386,11 +1405,7 @@ int thaw_super(struct super_block *sb)
 	}
 
 out:
-	sb->s_writers.frozen = SB_UNFROZEN;
-	smp_wmb();
-	wake_up(&sb->s_writers.wait_unfrozen);
-	deactivate_locked_super(sb);
-
+	__thaw_super(sb);
 	return 0;
 }
 EXPORT_SYMBOL(thaw_super);

_

#2 lore raw

PATCH 3/3 fs: replace memory barrier in __sb_end_write() with RCU

In reply to Dave Hansen

From: Dave Hansen <[email protected]>

If I sit in a loop and do write()s to small tmpfs files, __sb_end_write() is third-hottest kernel function due to its smp_mb().

__sb_end_write() uses the barrier to avoid races with freeze_super() and its calls to sb_wait_write(). But, now that freeze_super() is calling synchronize_rcu() before each sb_wait_write() call, we can use that to our advantage.

The synchronize_rcu() ensures that all __sb_end_write() will see freeze_super()'s updates to s_writers.counter. That, in turn, guarantees that __sb_end_write() will try to wake up any subsequent call by freeze_super() to sb_wait_write().

This improves the number of writes/second I can do by 6.1% on top of the previous patch. The total improvement is 27.1% over a completely unpatched kernel.

---

 b/fs/super.c |   17 +++++++++++++----
 1 file changed, 13 insertions(+), 4 deletions(-)
fs/super.c renamed from fs/super.c~selectively-do-barriers-in-__sb_end_write +13 -4
diff -puN fs/super.c~selectively-do-barriers-in-__sb_end_write fs/super.c
@@ -1146,14 +1146,23 @@ out:
  */
 void __sb_end_write(struct super_block *sb, int level)
 {
+	rcu_read_lock();
 	percpu_counter_dec(&sb->s_writers.counter[level-1]);
 	/*
-	 * Make sure s_writers are updated before we wake up waiters in
-	 * freeze_super().
+	 * We are racing here with freeze_super()'s calls to
+	 * sb_wait_write().  We want to ensure that we call
+	 * wake_up() whenever one of those calls _might_ be
+	 * in sb_wait_write().
+	 *
+	 * Since freeze_super() does a synchronize_rcu() before
+	 * each of its sb_wait_write() calls, it can guarantee
+	 * that it sees our update to s_writers.counter as well
+	 * as that we see its update to s_writers.frozen.
 	 */
-	smp_mb();
-	if (waitqueue_active(&sb->s_writers.wait))
+	if (unlikely(sb->s_writers.frozen >= level))
 		wake_up(&sb->s_writers.wait);
+	rcu_read_unlock();
+
 	rwsem_release(&sb->s_writers.lock_map[level-1], 1, _RET_IP_);
 }
 EXPORT_SYMBOL(__sb_end_write);

_

maintainer Reviewed-by on v1 1/3
#3 lore raw
In reply to Dave Hansen
Quoted 26 lines · Dave Hansen

From: Dave Hansen <[email protected]>

I have a _tiny_ microbenchmark that sits in a loop and writes single bytes to a file. Writing one byte to a tmpfs file is around 2x slower than reading one byte from a file, which is a _bit_ more than I expecte. This is a dumb benchmark, but I think it's hard to deny that write() is a hot path and we should avoid unnecessary overhead there.

I did a 'perf record' of 30-second samples of read and write. The top item in a diffprofile is srcu_read_lock() from fsnotify(). There are active inotify fd's from systemd, but nothing is actually listening to the file or its part of the filesystem.

I *think* we can avoid taking the srcu_read_lock() for the common case where there are no actual marks on the file. This means that there will both be nothing to notify for *and* implies that there is no need for clearing the ignore mask.

This patch gave a 13.1% speedup in writes/second on my test, which is an improvement from the 10.8% that I saw with the last version.

Quoted 9 lines
Cc Andrew Morton <[email protected]>
Cc Jan Kara <[email protected]>
Cc Al Viro <[email protected]>
Cc Eric Paris <[email protected]>
Cc John McCutchan <[email protected]>
Cc Robert Love <[email protected]>
Cc Andi Kleen <[email protected]>

Looks good. You can add:

Honza

Quoted 26 lines

---

 b/fs/notify/fsnotify.c |   10 ++++++++++
 1 file changed, 10 insertions(+)

diff -puN fs/notify/fsnotify.c~optimize-fsnotify fs/notify/fsnotify.c

fs/notify/fsnotify.c renamed from fs/notify/fsnotify.c~optimize-fsnotify +10 -0
@@ -205,6 +205,16 @@ int fsnotify(struct inode *to_tell, __u3
 		mnt = NULL;
 
 	/*
+	 * Optimization: srcu_read_lock() has a memory barrier which can
+	 * be expensive.  It protects walking the *_fsnotify_marks lists.
+	 * However, if we do not walk the lists, we do not have to do
+	 * SRCU because we have no references to any objects and do not
+	 * need SRCU to keep them "alive".
+	 */
+	if (hlist_empty(&to_tell->i_fsnotify_marks) &&
+	    (!mnt || hlist_empty(&mnt->mnt_fsnotify_marks)))
+		return 0;
+	/*
 	 * if this is a modify event we may need to clear the ignored masks
 	 * otherwise return if neither the inode nor the vfsmount care about
 	 * this type of event.

_

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR