Re: [PATCH (V9) 2/2] xen: Add V4V implementation

[David Vrabel <david.vrabel@citrix.com>]

On 28/05/13 20:43, Ross Philipson wrote:
> Setup of v4v domains a domain gets created and cleanup
> when a domain die. Wire up the v4v hypercall.
[...]
> +static void
> +v4v_pending_notify(struct domain *caller_d, struct hlist_head *to_notify)
> +{
> +    struct hlist_node *node, *next;
> +    struct v4v_pending_ent *pending_ent;
> +
> +    ASSERT(rw_is_locked(&v4v_lock));
> +
> +    hlist_for_each_entry_safe(pending_ent, node, next, to_notify, node)
> +    {
> +        hlist_del(&pending_ent->node);
> +        v4v_signal_domid(pending_ent->id);
> +        xfree(pending_ent);
> +    }
> +
> +}
> +
> +static void
> +v4v_pending_find(struct domain *d, struct v4v_ring_info *ring_info,
> +                 uint32_t payload_space, struct hlist_head *to_notify)
> +{
> +    struct hlist_node *node, *next;
> +    struct v4v_pending_ent *ent;
> +
> +    ASSERT(rw_is_locked(&d->v4v->lock));
> +
> +    spin_lock(&ring_info->lock);
> +    hlist_for_each_entry_safe(ent, node, next, &ring_info->pending, node)
> +    {
> +        if ( payload_space >= ent->len )
> +        {
> +            hlist_del(&ent->node);
> +            hlist_add_head(&ent->node, to_notify);
> +        }
> +    }
> +    spin_unlock(&ring_info->lock);
> +}

This "space in destination ring" notification does not seem like it
would scale very well for: a) many senders into the same ring; or b) a
single sender to many (full) rings.

In (a), all domains are notified, even though only one will win the race
and put a message on the ring.  In (b) the guest must retry all pending
transmits as it doesn't know which destination ring now has space.

David