Race condition between iget_locked() and evict_inodes()

[Anton Altaparmakov]

Hi Al,

I think there is a race condition between iget_locked() and evict_inodes().

evict_inodes() checks i_count and if zero proceeds to take i_lock then set I_FREEING and eventually disposes of the inode.

But a concurrent iget_locked() takes i_lock and then increments i_count.

Thus if the events happen in this order:

evict_inodes()				iget_locked() in find_inode_fast()
atomic_read(&inode->i_count) -> 0	take i_lock
wait on i_lock				__iget(inode); -> i_count now 1
set I_FREEING				drop i_lock
evict()					return inode to caller

The inode is now gone due to the evict() call whilst it is happily being used with an elevated i_count by the iget_locked() calling process.  It seems to me like evict_inodes() should be checking i_count inside i_lock either as the only check or it should at least re-check it.

Please tell me what I am missing here.  I assume there must be something providing exclusion and I am just too blind to see it but I thought it worth bringing to your attention in case it really is simply broken.

Best regards,

	Anton
-- 
Anton Altaparmakov <anton at tuxera.com> (replace at with @)
Lead in File System Development, Tuxera Inc., http://www.tuxera.com/
Linux NTFS maintainer