On 12/22/2016 5:10 PM, Jason Gunthorpe wrote:
> There is no guarantee how a process will be exiting, specifically there
> is no guarantee that there is only 1 thread. This means what atexit
> handlers can safely do is very limited.
>
> In particular freeing resources that would otherwise be freed by the OS
> is absolutely wrong in a multi-threaded environment.
>
> - neigh.c: This deletes the netlink socket
> - ocrdma_main.c: These delete locks and free devices
> - cma.c: This frees pds and other objects
> - amp.c: We don't need to wait for a thread to exit if we are
> exiting.
>
> This fixes a crash on-exit observed with at least rping where a
> thread calls exit() while the main thread is still running.
> This creates a race where rdma_destroy_id() runs concurrently
> with ucma_cleanup() and one of them will randomly crash with a jump
> to NULL.
>
> This will create some leaked memory complaints from valgrind
> mem checkers. These need to be solved with sensible resource
> tracking to free the objects once all possible users are gone
> and not relying on atexit.
>
> Signed-off-by: Jason Gunthorpe <jgunthorpe@xxxxxxxxxxxxxxxxxxxx>
This all looks sane, thanks, applied.
--
Doug Ledford <dledford@xxxxxxxxxx>
GPG Key ID: B826A3330E572FDD
Key fingerprint = AE6B 1BDA 122B 23B4 265B 1274 B826 A333 0E57 2FDD
Attachment:
signature.asc
Description: OpenPGP digital signature
