eventfd: don't take the spinlock in eventfd_poll

The spinlock in eventfd_poll is trying to protect the count of events so
it can decide if it should return POLLIN, POLLERR, or POLLOUT.  But,
because of the way we drop the lock after calling poll_wait, and drop it
again before returning, we have the same pile of races with the lock as we
do with a single read of ctx->count().

This replaces the lock with a read barrier and single read.

eventfd_write does a single bump of ctx->count, so this should not add new
races with adding events.  eventfd_read is similar, it will do a single
decrement with the lock held, and so we're making the race with concurrent
readers slightly larger.

This spinlock is the top CPU user in kernel code during one of our
workloads.  Removing it gives us a ~2% boost.

Signed-off-by: Chris Mason <clm@fb.com>
Cc: Davide Libenzi <davidel@xmailserver.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>

1 files changed, 6 insertions, 5 deletions

diff --git a/fs/eventfd.c b/fs/eventfd.c
index 4b0a226024fa..439e6f0177f3 100644
--- a/fs/eventfd.c
+++ b/fs/eventfd.c
@@ -119,17 +119,18 @@ static unsigned int eventfd_poll(struct file *file, poll_table *wait)
 	struct eventfd_ctx *ctx = file->private_data;
 	unsigned int events = 0;
 	unsigned long flags;
+	unsigned int count;
 
 	poll_wait(file, &ctx->wqh, wait);
+	smp_rmb();
+	count = ctx->count;
 
-	spin_lock_irqsave(&ctx->wqh.lock, flags);
-	if (ctx->count > 0)
+	if (count > 0)
 		events |= POLLIN;
-	if (ctx->count == ULLONG_MAX)
+	if (count == ULLONG_MAX)
 		events |= POLLERR;
-	if (ULLONG_MAX - 1 > ctx->count)
+	if (ULLONG_MAX - 1 > count)
 		events |= POLLOUT;
-	spin_unlock_irqrestore(&ctx->wqh.lock, flags);
 
 	return events;
 }