Comment 0 for bug 1774336

Daniel Axtens (daxtens) wrote :

== SRU Justification ==

Oops during heavy NFS + FSCache use:

[81738.886634] FS-Cache:
[81738.888281] FS-Cache: Assertion failed
[81738.889461] FS-Cache: 6 == 5 is false
[81738.890625] ------------[ cut here ]------------
[81738.891706] kernel BUG at /build/linux-hVVhWi/linux-4.4.0/fs/fscache/operation.c:494!

6 == 5 represents an operation being DEAD when it was not expected to be.

There is a race in fscache and cachefiles.

One thread is in cachefiles_read_waiter:
 1) object->work_lock is taken.
 2) the operation is added to the to_do list.
 3) the work lock is dropped.
 4) fscache_enqueue_retrieval is called, which takes a reference.

Another thread is in cachefiles_read_copier:
 1) object->work_lock is taken
 2) an item is popped off the to_do list.
 3) object->work_lock is dropped.
 4) some processing is done on the item, and fscache_put_retrieval() is called, dropping a reference.

Now if the this process in cachefiles_read_copier takes place *between* steps 3 and 4 in cachefiles_read_waiter, a reference will be dropped before it is taken, which leads to the objects reference count hitting zero, which leads to lifecycle events for the object happening too soon, leading to the assertion failure later on.

(This is simplified and clarified from the original upstream analysis for this patch at and from a similar patch with a different approach to fixing the bug at

Move fscache_enqueue_retrieval under the lock in cachefiles_read_waiter. This means that the object cannot be popped off the to_do list until it is in a fully consistent state with the reference taken.

A user has run ~100 hours of NFS stress tests and not seen this bug recur.

[Regression Potential]
 - Limited to fscache/cachefiles.
 - The change makes things more conservative (doing more under lock) so that's reassuring.
 - There may be performance impacts but none have been observed so far.