Re: [Qemu-devel] [PATCH 2/3] scsi: Introduce scsi_req_cancel_async

[Paolo Bonzini <pbonzini@redhat.com>]

Il 18/09/2014 04:36, Fam Zheng ha scritto:
> +    QTAILQ_FOREACH_SAFE(ref, &req->cancel_deps, next, next) {
> +        SCSIRequest *r = ref->req;
> +        assert(r->cancel_dep_count);
> +        r->cancel_dep_count--;
> +        if (!r->cancel_dep_count && req->bus->info->cancel_dep_complete) {
> +            req->bus->info->cancel_dep_complete(r);
> +        }
> +        QTAILQ_REMOVE(&req->cancel_deps, ref, next);
> +        scsi_req_unref(r);
> +        g_free(ref);
> +    }

I think there is one problem here.

scsi_req_cancel_async can actually be synchronous if you're unlucky
(because bdrv_aio_cancel_async can be synchronous too, for example in
the case of a linux-aio AIOCB).  So you could end up calling
cancel_dep_complete even if the caller intends to cancel more requests.

I think it's better to track the count in virtio-scsi instead.  You can
initialize it similar to bdrv_aio_multiwrite:

    /* Run the aio requests. */
    mcb->num_requests = num_reqs;
    for (i = 0; i < num_reqs; i++) {
        bdrv_co_aio_rw_vector(bs, reqs[i].sector, reqs[i].qiov,
                              reqs[i].nb_sectors, reqs[i].flags,
                              multiwrite_cb, mcb,
                              true);
    }

    return 0;

and decrement the count on every call to the notifier.

This is independent of the choice to make bdrv_aio_cancel_async
semantics stricter.  In the case of bdrv_aio_multiwrite, we know that
bdrv_co_aio_rw_vector is never synchronous, but the code is simply nicer
that way. :)

Paolo