* Peter Xu (pet...@redhat.com) wrote:
> We have a logic in await_return_path_close_on_source() that we will explicitly
> shutdown the socket when migration encounters errors.  However it could be 
> racy
> because from_dst_file could have been reset right after checking it but before
> passing it to qemu_file_shutdown() by the rp_thread.
> 
> Fix it by shutdown() on the src file instead.  Since they must be a pair of
> qemu files, shutdown on either of them will work the same.
> 
> Since at it, drop the check for from_dst_file directly, which makes the
> behavior even more predictable.

So while the existing code maybe racy, I'm not sure that this change
keeps the semantics; the channel may well have dup()'d the fd's for the
two directions, and I'm not convinced that a shutdown() on one will
necessarily impact the other; and if the shutdown doesn't happen the
rp_thread might not exit, and we might block on the koin.

Why don't we solve this a different way - how about we move the:
    ms->rp_state.from_dst_file = NULL;
    qemu_fclose(rp);

out of the source_return_path_thread and put it in
await_return_path_close_on_source, immediately after the join?
Then we *know* that the the rp thread isn't messing with it.

Dave

> Reported-by: Dr. David Alan Gilbert <dgilb...@redhat.com>
> Signed-off-by: Peter Xu <pet...@redhat.com>
> ---
>  migration/migration.c | 9 ++++++---
>  1 file changed, 6 insertions(+), 3 deletions(-)
> 
> diff --git a/migration/migration.c b/migration/migration.c
> index 21b94f75a3..4f48cde796 100644
> --- a/migration/migration.c
> +++ b/migration/migration.c
> @@ -2882,12 +2882,15 @@ static int 
> await_return_path_close_on_source(MigrationState *ms)
>       * rp_thread will exit, however if there's an error we need to cause
>       * it to exit.
>       */
> -    if (qemu_file_get_error(ms->to_dst_file) && ms->rp_state.from_dst_file) {
> +    if (qemu_file_get_error(ms->to_dst_file)) {
>          /*
>           * shutdown(2), if we have it, will cause it to unblock if it's stuck
> -         * waiting for the destination.
> +         * waiting for the destination.  We do shutdown on to_dst_file should
> +         * also shutdown the from_dst_file as they're in a pair. We 
> explicilty
> +         * don't operate on from_dst_file because it's potentially racy
> +         * (rp_thread could have reset it in parallel).
>           */
> -        qemu_file_shutdown(ms->rp_state.from_dst_file);
> +        qemu_file_shutdown(ms->to_dst_file);
>          mark_source_rp_bad(ms);
>      }
>      trace_await_return_path_close_on_source_joining();
> -- 
> 2.31.1
> 
-- 
Dr. David Alan Gilbert / dgilb...@redhat.com / Manchester, UK


Reply via email to