"Erich E. Hoover" ehoover@mymail.mines.edu writes:
Well, maybe I screwed something up - but when I tried to access the descriptor like that it caused segfaults in some of the tests. What I found was that neither create_async_queue or create_async increment the refcount on the file descriptor, but create_async does increment the refcount on the async queue so that the async queue will live throughout the life of the async. So, what appeared to happen is this:
- A file descriptor is created.
- An async operation is queued on the file descriptor.
- The file descriptor is closed (queuing the async to wake up and
contact user space). 4) The file descriptor is destroyed because its refcount reaches zero. 5) The async operation returns from user space with completion information for the operation. 6) Accessing the completion from the no-longer existing file descriptor causes a crash.
I tried having the async and/or the async queue increment the refcount for the file descriptor, but that caused a host of other problems (mainly file descriptors never closing). So, I'm pretty sure that the completion information needs to be held separately in the async queue
- but I can go back and double check that this is the case if you'd
like.
It may have to be moved to the queue upon shutdown when the file descriptor is released, but I don't see any reason to maintain two copies of the same information.