Re: [PATCH v2 04/14] aio: make aio_context_acquire()/aio_context_release() a no-op

[Kevin Wolf]

Am 19.12.2023 um 22:23 hat Stefan Hajnoczi geschrieben:
> The following hack makes the test pass but there are larger safety
> issues that I'll need to look at on Wednesday:

I see, you're taking the same approach as in the SCSI layer: Don't make
things thread-safe, but just always access them from the same thread.

In theory this should be okay, but I'm almost sure that at least
nbd_drained_poll() must then run in the same AioContext, too.

> diff --git a/nbd/server.c b/nbd/server.c
> index 895cf0a752..cf4b7d5c6d 100644
> --- a/nbd/server.c
> +++ b/nbd/server.c
> @@ -1617,7 +1617,7 @@ static void nbd_drained_begin(void *opaque)
>      }
>  }
> 
> -static void nbd_drained_end(void *opaque)
> +static void nbd_resume_clients(void *opaque)
>  {
>      NBDExport *exp = opaque;
>      NBDClient *client;
> @@ -1628,6 +1628,15 @@ static void nbd_drained_end(void *opaque)
>      }
>  }
> 
> +static void nbd_drained_end(void *opaque)
> +{
> +    NBDExport *exp = opaque;
> +
> +    /* TODO how to make sure exp doesn't go away? */

blk_exp_ref()?

> +    /* TODO what if AioContext changes before this runs? */
> +    aio_bh_schedule_oneshot(nbd_export_aio_context(exp),
> nbd_resume_clients, exp);

We could increase client->nb_requests if we change it to be accessed
atomically. Then nbd_drained_poll() will make any AioContext change wait
for the BH.

Or maybe aio_wait_bh_oneshot() would already solve both problems?

> +}
> +

Kevin