Re: [PATCH v4 2/4] block/nbd.c: Add yank feature

[Stefan Hajnoczi <stefanha@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 1348 bytes --]

On Mon, May 25, 2020 at 05:44:26PM +0200, Lukas Straub wrote:
> @@ -1395,6 +1407,15 @@ static int nbd_client_reopen_prepare(BDRVReopenState *state,
>      return 0;
>  }
> 
> +static void nbd_yank(void *opaque)
> +{
> +    BlockDriverState *bs = opaque;
> +    BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
> +
> +    qio_channel_shutdown(QIO_CHANNEL(s->sioc), QIO_CHANNEL_SHUTDOWN_BOTH, NULL);

qio_channel_shutdown() is not guaranteed to be thread-safe. Please
document new assumptions that are being introduced.

Today we can more or less get away with it (although TLS sockets are a
little iffy) because it boils down the a shutdown(2) system call. I
think it would be okay to update the qio_channel_shutdown() and
.io_shutdown() documentation to clarify that this is thread-safe.

> +    atomic_set(&s->state, NBD_CLIENT_QUIT);

docs/devel/atomics.rst says:

  No barriers are implied by ``atomic_read`` and ``atomic_set`` in either Linux
  or QEMU.

Other threads might not see the latest value of s->state because this is
a weakly ordered memory access.

I haven't audited the NBD code in detail, but if you want the other
threads to always see NBD_CLIENT_QUIT then s->state should be set before
calling qio_channel_shutdown() using a stronger atomics API like
atomic_load_acquire()/atomic_store_release().

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]