Skip to contentSkip to conversation

vfs: Fix missed wakeup in I_NEW handling

PATCH RESEND In review Jan Kara 2 participants 2 messages Mar 7, 2012 – Mar 8, 2012

Subsystems (2)
lore
Share & tools
lore

Jan Kara

author maintainer
#0 lore raw

PATCH RESEND 1/1 vfs: Fix missed wakeup in I_NEW handling

Commit 250df6ed removed wake_up_inode() (in particular a memory barrier before wake_up_bit()) on the basis that i_state transitions are protected by i_lock. That would be fine if all the readers of i_state were using i_lock as well. But wait_on_inode() doesn't use i_lock and thus the following can happen due to reordering:

CPU 1 CPU 2
unlock_new_inode()
spin_lock(&inode->i_lock);

  wake_up_bit(&inode->i_state, __I_NEW);
					wait_on_inode()
					  wait_on_bit(&inode->i_state, __I_NEW);

inode->i_state &= ~I_NEW;
^^^ this store was reordered
spin_unlock(&inode->i_lock);

And waiter on CPU2 sleeps forever (or for a really long time).

We fix the issue by using i_lock in wait_on_inode() in the spirit of commit 250df6ed.

---
 fs/inode.c                |   22 ++++++++++++++++++++++
 include/linux/writeback.h |    5 -----
 2 files changed, 22 insertions(+), 5 deletions(-)

Al, could you please merge this patch? It fixes a real deadlock...

fs/inode.c +22 -0
index fb10d86..e768f9e 100644
@@ -954,6 +954,28 @@ EXPORT_SYMBOL(lockdep_annotate_inode_mutex_key);
 #endif
 
 /**
+ * wait_on_inode - wait while inode is in I_NEW state
+ * @inode:	inode to wait for
+ *
+ * This function waits until inode is fully initialized and exits new state
+ */
+static void wait_on_inode(struct inode *inode)
+{
+	DEFINE_WAIT_BIT(wq, &inode->i_state, __I_NEW);
+	wait_queue_head_t *wqh;
+
+	might_sleep();
+	wqh = bit_waitqueue(&inode->i_state, __I_NEW);
+	spin_lock(&inode->i_lock);
+	while (inode->i_state & I_NEW) {
+		spin_unlock(&inode->i_lock);
+		__wait_on_bit(wqh, &wq, inode_wait, TASK_UNINTERRUPTIBLE);
+		spin_lock(&inode->i_lock);
+	}
+	spin_unlock(&inode->i_lock);
+}
+
+/**
  * unlock_new_inode - clear the I_NEW state and wake up any waiters
  * @inode:	new inode to unlock
  *
include/linux/writeback.h +0 -5
index 995b8bf..e2dbc70 100644
@@ -96,11 +96,6 @@ long wb_do_writeback(struct bdi_writeback *wb, int force_wait);
 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason);
 
 /* writeback.h requires fs.h; it, too, is not included from here. */
-static inline void wait_on_inode(struct inode *inode)
-{
-	might_sleep();
-	wait_on_bit(&inode->i_state, __I_NEW, inode_wait, TASK_UNINTERRUPTIBLE);
-}
 static inline void inode_sync_wait(struct inode *inode)
 {
 	might_sleep();
Signature · 1 line
1.7.1
maintainer
#1 lore raw
Quoted 12 lines · Jan Kara

Commit 250df6ed removed wake_up_inode() (in particular a memory barrier before wake_up_bit()) on the basis that i_state transitions are protected by i_lock. That would be fine if all the readers of i_state were using i_lock as well. But wait_on_inode() doesn't use i_lock and thus the following can happen due to reordering:

   CPU 1				CPU 2
unlock_new_inode()
  spin_lock(&inode->i_lock);
  wake_up_bit(&inode->i_state, __I_NEW);
					wait_on_inode()
					  wait_on_bit(&inode->i_state, __I_NEW);
Quoted 8 lines

inode->i_state &= ~I_NEW;
^^^ this store was reordered
spin_unlock(&inode->i_lock);

And waiter on CPU2 sleeps forever (or for a really long time).

We fix the issue by using i_lock in wait_on_inode() in the spirit of commit 250df6ed.

Applied (along with a bunch of other stuff; will push to Linus tomorrow)