fsnotify: next_i is freed during fsnotify_unmount_inodes.
authorJerry Hoemann <jerry.hoemann@hp.com>
Wed, 29 Oct 2014 21:50:22 +0000 (14:50 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 29 Oct 2014 23:33:14 +0000 (16:33 -0700)
During file system stress testing on 3.10 and 3.12 based kernels, the
umount command occasionally hung in fsnotify_unmount_inodes in the
section of code:

                spin_lock(&inode->i_lock);
                if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) {
                        spin_unlock(&inode->i_lock);
                        continue;
                }

As this section of code holds the global inode_sb_list_lock, eventually
the system hangs trying to acquire the lock.

Multiple crash dumps showed:

The inode->i_state == 0x60 and i_count == 0 and i_sb_list would point
back at itself.  As this is not the value of list upon entry to the
function, the kernel never exits the loop.

To help narrow down problem, the call to list_del_init in
inode_sb_list_del was changed to list_del.  This poisons the pointers in
the i_sb_list and causes a kernel to panic if it transverse a freed
inode.

Subsequent stress testing paniced in fsnotify_unmount_inodes at the
bottom of the list_for_each_entry_safe loop showing next_i had become
free.

We believe the root cause of the problem is that next_i is being freed
during the window of time that the list_for_each_entry_safe loop
temporarily releases inode_sb_list_lock to call fsnotify and
fsnotify_inode_delete.

The code in fsnotify_unmount_inodes attempts to prevent the freeing of
inode and next_i by calling __iget.  However, the code doesn't do the
__iget call on next_i

if i_count == 0 or
if i_state & (I_FREEING | I_WILL_FREE)

The patch addresses this issue by advancing next_i in the above two cases
until we either find a next_i which we can __iget or we reach the end of
the list.  This makes the handling of next_i more closely match the
handling of the variable "inode."

The time to reproduce the hang is highly variable (from hours to days.) We
ran the stress test on a 3.10 kernel with the proposed patch for a week
without failure.

During list_for_each_entry_safe, next_i is becoming free causing
the loop to never terminate.  Advance next_i in those cases where
__iget is not done.

Signed-off-by: Jerry Hoemann <jerry.hoemann@hp.com>
Cc: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
Cc: Ken Helias <kenhelias@firemail.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/notify/inode_mark.c

index 9ce062218de9cf2559c02b92a7a201224d1d67fc..e8497144b32342437377748f26a616168455339a 100644 (file)
@@ -288,20 +288,25 @@ void fsnotify_unmount_inodes(struct list_head *list)
                spin_unlock(&inode->i_lock);
 
                /* In case the dropping of a reference would nuke next_i. */
-               if ((&next_i->i_sb_list != list) &&
-                   atomic_read(&next_i->i_count)) {
+               while (&next_i->i_sb_list != list) {
                        spin_lock(&next_i->i_lock);
-                       if (!(next_i->i_state & (I_FREEING | I_WILL_FREE))) {
+                       if (!(next_i->i_state & (I_FREEING | I_WILL_FREE)) &&
+                                               atomic_read(&next_i->i_count)) {
                                __iget(next_i);
                                need_iput = next_i;
+                               spin_unlock(&next_i->i_lock);
+                               break;
                        }
                        spin_unlock(&next_i->i_lock);
+                       next_i = list_entry(next_i->i_sb_list.next,
+                                               struct inode, i_sb_list);
                }
 
                /*
-                * We can safely drop inode_sb_list_lock here because we hold
-                * references on both inode and next_i.  Also no new inodes
-                * will be added since the umount has begun.
+                * We can safely drop inode_sb_list_lock here because either
+                * we actually hold references on both inode and next_i or
+                * end of list.  Also no new inodes will be added since the
+                * umount has begun.
                 */
                spin_unlock(&inode_sb_list_lock);