fscache: extended "dying" check before emitting EV_INVALIDATE
PATCH 1/1 fscache: extended "dying" check before emitting EV_INVALIDATE
Before emitting an FSCACHE_OBJECT_EV_INVALIDATE event, the function __fscache_invalidate() checks whether the fscache_object is currently "dying". This checks only the current state, not the queued events that will very soon lead to the object's death.
The problem is that fscache_object_state_machine() will BUG("Unsupported event") when there is object->events includes EV_INVALIDATE during "terminal_transit". Even if we would ignore that check, we could run into "Unexpected event in dead state" later because object->work may still be queued.
The solution is to check for EV_RETIRE and EV_RELEASE; if one of those two terminal events is already queued, don't bother to queue EV_INVALIDATE.
--- fs/fscache/cookie.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
fs/fscache/cookie.c +2 -1
@@ -402,7 +402,8 @@ void __fscache_invalidate(struct fscache_cookie *cookie) object =hlist_entry (cookie->backing_objects.first, struct fscache_object, cookie_link); -if (object->state <FSCACHE_OBJECT_DYING ) +if (object->state <FSCACHE_OBJECT_DYING && + (object->events & (FSCACHE_OBJECT_EV_RETIRE |FSCACHE_OBJECT_EV_RELEASE )) ==0 )fscache_raise_event ( object,FSCACHE_OBJECT_EV_INVALIDATE ); }
fs/fscache/cookie.c void __fscache_invalidate(struct fscache_cookie *cookie) · line 406
Quoted 1 line · Max Kellermann
fs/fscache/cookie.c +1
-0+ (object->events & (FSCACHE_OBJECT_EV_RETIRE|FSCACHE_OBJECT_EV_RELEASE)) == 0)
You can't do it like this. EV_RETIRE and EV_RELEASE are bit numbers, so need to be shifted into place.
David
Quoted 4 lines · Max Kellermann
Before emitting an
FSCACHE_OBJECT_EV_INVALIDATEevent, the function__fscache_invalidate()checks whether thefscache_objectis currently "dying". This checks only the current state, not the queued events that will very soon lead to the object's death.
This should not be a problem, I think. EV_RETIRE and EV_RELEASE should only be set by fscache_relinquish_cookie(). If you (the netfs) call that to dispose of a cookie to fscache, you shouldn't thereafter be calling fscache_invalidate() on it. The only case in which __fscache_invalidate() should see a dying object is if the caching backend withdraws it (ie. EV_WITHDRAW is set).
Do you have a backtrace?
David