On Thu, Jan 29, 2026 at 9:35 AM Shani Peretz <[email protected]> wrote:
>
> This commit fixes a use-after-free that causes the application to crash
> on shutdown (detected by ASAN).
>
> The vhost library uses a background event dispatch thread that monitors
> fds with epoll. It runs in an infinite loop, waiting for I/O events
> and calling callbacks when they occur.
>
> During cleanup, a race condition existed:
>
>   Main Thread:                    Event Dispatch Thread:
>   1. Remove fds from fdset        while (1) {
>   2. Close file descriptors           epoll_wait() [gets interrupted]
>   3. rte_eal_cleanup()                [continues loop]
>   4. Unmap hugepages                  Accesses fdset...   CRASH
>                                   }
>
> There was no explicit cleanup of the fdset structure.
> The fdset structure is allocated with rte_zmalloc() and the memory would
> only be reclaimed at application shutdown when rte_eal_cleanup() is called,
> which invokes rte_eal_memory_detach() to unmap all the hugepage memory.
> Meanwhile, the event dispatch thread could still be running and accessing
> the fdset.
>
> The code had a `destroy` flag that the event dispatch thread checked,
> but it was never set during cleanup, and the code never waited for
> the thread to actually exit before freeing memory.
>
> To fix this, the commit implements fdset_destroy() that sets the destroy
> flag with mutex protection, waits for thread termination, and cleans up
> all resources including the fdset memory.
>
> Update socket.c to call fdset_destroy() when the last vhost-user socket
> is unregistered.
>
> Fixes: 0e38b42bf61c ("vhost: manage FD with epoll")
> Cc: [email protected]
>
> Signed-off-by: Shani Peretz <[email protected]>
>
> -----------------
> v3:
> removed vduse implementation from this fix
>
> ---
>  lib/vhost/fd_man.c | 45 ++++++++++++++++++++++++++++++++++++++++++++-
>  lib/vhost/fd_man.h |  1 +
>  lib/vhost/socket.c |  7 +++++++
>  3 files changed, 52 insertions(+), 1 deletion(-)
>
Reviewed-by: Maxime Coquelin <[email protected]>

Thanks,
Maxime

Reply via email to