[PATCH RFC] xen-netback: remove guest RX path dependence on MAX_SKB_FRAGS

[PATCH RFC] xen-netback: remove guest RX path dependence on MAX_SKB_FRAGS

[Wei Liu]

This dependence is undesirable and logically incorrect.

It's undesirable because Xen network protocol should not depend on a
OS-specific constant.

It's incorrect because the ring slots required doesn't correspond to the
number of frags a SKB has (consider compound page frags).

This patch removes this dependence by correctly counting the ring slots
required.

Signed-off-by: Wei Liu <wei.liu2@citrix.com>
---
 drivers/net/xen-netback/interface.c |   12 +++++++---
 drivers/net/xen-netback/netback.c   |   41 +++++++++++++----------------------
 2 files changed, 24 insertions(+), 29 deletions(-)

diff --git a/drivers/net/xen-netback/interface.c b/drivers/net/xen-netback/interface.c
index 087d2db..5c10e87 100644
--- a/drivers/net/xen-netback/interface.c
+++ b/drivers/net/xen-netback/interface.c
@@ -96,18 +96,24 @@ static irqreturn_t xenvif_interrupt(int irq, void *dev_id)
 static int xenvif_start_xmit(struct sk_buff *skb, struct net_device *dev)
 {
 	struct xenvif *vif = netdev_priv(dev);
+	unsigned int ring_slots_required;
 
 	BUG_ON(skb->dev != dev);
 
 	if (vif->netbk == NULL)
 		goto drop;
 
+	ring_slots_required = xen_netbk_count_skb_slots(vif, skb);
+
+	/* Reserve ring slots for the worst-case number of fragments. */
+	vif->rx_req_cons_peek += ring_slots_required;
+
 	/* Drop the packet if the target domain has no receive buffers. */
-	if (!xenvif_rx_schedulable(vif))
+	if (!xenvif_rx_schedulable(vif)) {
+		vif->rx_req_cons_peek -= ring_slots_required;
 		goto drop;
+	}
 
-	/* Reserve ring slots for the worst-case number of fragments. */
-	vif->rx_req_cons_peek += xen_netbk_count_skb_slots(vif, skb);
 	xenvif_get(vif);
 
 	if (vif->can_queue && xen_netbk_must_stop_queue(vif))
diff --git a/drivers/net/xen-netback/netback.c b/drivers/net/xen-netback/netback.c
index 64828de..1b2b71b 100644
--- a/drivers/net/xen-netback/netback.c
+++ b/drivers/net/xen-netback/netback.c
@@ -289,24 +289,12 @@ static void xen_netbk_kick_thread(struct xen_netbk *netbk)
 	wake_up(&netbk->wq);
 }
 
-static int max_required_rx_slots(struct xenvif *vif)
-{
-	int max = DIV_ROUND_UP(vif->dev->mtu, PAGE_SIZE);
-
-	/* XXX FIXME: RX path dependent on MAX_SKB_FRAGS */
-	if (vif->can_sg || vif->gso || vif->gso_prefix)
-		max += MAX_SKB_FRAGS + 1; /* extra_info + frags */
-
-	return max;
-}
-
 int xen_netbk_rx_ring_full(struct xenvif *vif)
 {
-	RING_IDX peek   = vif->rx_req_cons_peek;
-	RING_IDX needed = max_required_rx_slots(vif);
+	RING_IDX peek = vif->rx_req_cons_peek;
 
-	return ((vif->rx.sring->req_prod - peek) < needed) ||
-	       ((vif->rx.rsp_prod_pvt + XEN_NETIF_RX_RING_SIZE - peek) < needed);
+	return ((vif->rx.sring->req_prod < peek) ||
+		(vif->rx.rsp_prod_pvt + XEN_NETIF_RX_RING_SIZE < peek));
 }
 
 int xen_netbk_must_stop_queue(struct xenvif *vif)
@@ -314,8 +302,7 @@ int xen_netbk_must_stop_queue(struct xenvif *vif)
 	if (!xen_netbk_rx_ring_full(vif))
 		return 0;
 
-	vif->rx.sring->req_event = vif->rx_req_cons_peek +
-		max_required_rx_slots(vif);
+	vif->rx.sring->req_event = vif->rx_req_cons_peek + 1;
 	mb(); /* request notification /then/ check the queue */
 
 	return xen_netbk_rx_ring_full(vif);
@@ -675,7 +662,6 @@ static void xen_netbk_rx_action(struct xen_netbk *netbk)
 	struct sk_buff *skb;
 	LIST_HEAD(notify);
 	int ret;
-	int nr_frags;
 	int count;
 	unsigned long offset;
 	struct skb_cb_overlay *sco;
@@ -690,20 +676,23 @@ static void xen_netbk_rx_action(struct xen_netbk *netbk)
 	count = 0;
 
 	while ((skb = skb_dequeue(&netbk->rx_queue)) != NULL) {
+		unsigned int ring_slots_required;
 		vif = netdev_priv(skb->dev);
-		nr_frags = skb_shinfo(skb)->nr_frags;
 
 		sco = (struct skb_cb_overlay *)skb->cb;
-		sco->meta_slots_used = netbk_gop_skb(skb, &npo);
-
-		count += nr_frags + 1;
 
-		__skb_queue_tail(&rxq, skb);
+		ring_slots_required = xen_netbk_count_skb_slots(vif, skb);
 
-		/* Filled the batch queue? */
-		/* XXX FIXME: RX path dependent on MAX_SKB_FRAGS */
-		if (count + MAX_SKB_FRAGS >= XEN_NETIF_RX_RING_SIZE)
+		if (count + ring_slots_required >= XEN_NETIF_RX_RING_SIZE) {
+			skb_queue_head(&netbk->rx_queue, skb);
 			break;
+		}
+
+		count += ring_slots_required;
+
+		sco->meta_slots_used = netbk_gop_skb(skb, &npo);
+
+		__skb_queue_tail(&rxq, skb);
 	}
 
 	BUG_ON(npo.meta_prod > ARRAY_SIZE(netbk->meta));
-- 
1.7.10.4

Re: [PATCH RFC] xen-netback: remove guest RX path dependence on MAX_SKB_FRAGS

[Jan Beulich]

>>> On 02.07.13 at 15:40, Wei Liu <wei.liu2@citrix.com> wrote:
> This dependence is undesirable and logically incorrect.
> 
> It's undesirable because Xen network protocol should not depend on a
> OS-specific constant.

But this is not making the protocol dependent upon the constant;
at least in one case the check is merely used as a worst case
estimation (there can't possibly be an skb with more than
MAX_SKB_FRAGS, and hence having as many slots available
implies that at least one more skb can be processed).

>  int xen_netbk_rx_ring_full(struct xenvif *vif)
>  {
> -	RING_IDX peek   = vif->rx_req_cons_peek;
> -	RING_IDX needed = max_required_rx_slots(vif);
> +	RING_IDX peek = vif->rx_req_cons_peek;
>  
> -	return ((vif->rx.sring->req_prod - peek) < needed) ||
> -	       ((vif->rx.rsp_prod_pvt + XEN_NETIF_RX_RING_SIZE - peek) < needed);
> +	return ((vif->rx.sring->req_prod < peek) ||
> +		(vif->rx.rsp_prod_pvt + XEN_NETIF_RX_RING_SIZE < peek));

This transformation is definitely wrong: You need to retain the
code's capability of dealing with the indexes wrapping, i.e.
simple comparisons won't do.

> @@ -690,20 +676,23 @@ static void xen_netbk_rx_action(struct xen_netbk *netbk)
>  	count = 0;
>  
>  	while ((skb = skb_dequeue(&netbk->rx_queue)) != NULL) {
> +		unsigned int ring_slots_required;
>  		vif = netdev_priv(skb->dev);
> -		nr_frags = skb_shinfo(skb)->nr_frags;
>  
>  		sco = (struct skb_cb_overlay *)skb->cb;
> -		sco->meta_slots_used = netbk_gop_skb(skb, &npo);
> -
> -		count += nr_frags + 1;
>  
> -		__skb_queue_tail(&rxq, skb);
> +		ring_slots_required = xen_netbk_count_skb_slots(vif, skb);
>  
> -		/* Filled the batch queue? */
> -		/* XXX FIXME: RX path dependent on MAX_SKB_FRAGS */
> -		if (count + MAX_SKB_FRAGS >= XEN_NETIF_RX_RING_SIZE)
> +		if (count + ring_slots_required >= XEN_NETIF_RX_RING_SIZE) {

Is this really still >= here? The old code, iiuc, used >= as being
equivalent to count + MAX_SKB_FRAGS + 1 > XEN_NETIF_RX_RING_SIZE.

Jan

Re: [PATCH RFC] xen-netback: remove guest RX path dependence on MAX_SKB_FRAGS

[Wei Liu]

On Tue, Jul 02, 2013 at 03:19:12PM +0100, Jan Beulich wrote:
> >>> On 02.07.13 at 15:40, Wei Liu <wei.liu2@citrix.com> wrote:
> > This dependence is undesirable and logically incorrect.
> > 
> > It's undesirable because Xen network protocol should not depend on a
> > OS-specific constant.
> 
> But this is not making the protocol dependent upon the constant;
> at least in one case the check is merely used as a worst case
> estimation (there can't possibly be an skb with more than
> MAX_SKB_FRAGS, and hence having as many slots available
> implies that at least one more skb can be processed).
> 

The "worst case" is a wrong estimation -- consider compound page frags
and RX coalescing, a frag can take up multiple ring slots.

The assumption seems to work at the moment because we haven't hit the
corner cases, we should probably not rely on that.

> >  int xen_netbk_rx_ring_full(struct xenvif *vif)
> >  {
> > -	RING_IDX peek   = vif->rx_req_cons_peek;
> > -	RING_IDX needed = max_required_rx_slots(vif);
> > +	RING_IDX peek = vif->rx_req_cons_peek;
> >  
> > -	return ((vif->rx.sring->req_prod - peek) < needed) ||
> > -	       ((vif->rx.rsp_prod_pvt + XEN_NETIF_RX_RING_SIZE - peek) < needed);
> > +	return ((vif->rx.sring->req_prod < peek) ||
> > +		(vif->rx.rsp_prod_pvt + XEN_NETIF_RX_RING_SIZE < peek));
> 
> This transformation is definitely wrong: You need to retain the
> code's capability of dealing with the indexes wrapping, i.e.
> simple comparisons won't do.
> 

Hmm... so I don't quite understand this wrapping around thing and I have
a question here, what if req_prod < peek so that (req_prod - peek)
results in a very large number. In that case the ring is actually full
(cannot accommodate that packet), however "needed" is only like 20-ish
something, so the first comparision would fail. Are we relying on the
second comparion to tell the ring is full?

The only thing I need to do here is to tell whether rep_prod is ahead of
peek in the first comparison and in the second case there is enough
slots between rsp_prod_pvt and peek, what's the correct line?

> > @@ -690,20 +676,23 @@ static void xen_netbk_rx_action(struct xen_netbk *netbk)
> >  	count = 0;
> >  
> >  	while ((skb = skb_dequeue(&netbk->rx_queue)) != NULL) {
> > +		unsigned int ring_slots_required;
> >  		vif = netdev_priv(skb->dev);
> > -		nr_frags = skb_shinfo(skb)->nr_frags;
> >  
> >  		sco = (struct skb_cb_overlay *)skb->cb;
> > -		sco->meta_slots_used = netbk_gop_skb(skb, &npo);
> > -
> > -		count += nr_frags + 1;
> >  
> > -		__skb_queue_tail(&rxq, skb);
> > +		ring_slots_required = xen_netbk_count_skb_slots(vif, skb);
> >  
> > -		/* Filled the batch queue? */
> > -		/* XXX FIXME: RX path dependent on MAX_SKB_FRAGS */
> > -		if (count + MAX_SKB_FRAGS >= XEN_NETIF_RX_RING_SIZE)
> > +		if (count + ring_slots_required >= XEN_NETIF_RX_RING_SIZE) {
> 
> Is this really still >= here? The old code, iiuc, used >= as being
> equivalent to count + MAX_SKB_FRAGS + 1 > XEN_NETIF_RX_RING_SIZE.
> 

Yes I should use > here.


Wei.

> Jan

Re: [PATCH RFC] xen-netback: remove guest RX path dependence on MAX_SKB_FRAGS

[Jan Beulich]

>>> On 02.07.13 at 17:02, Wei Liu <wei.liu2@citrix.com> wrote:
> On Tue, Jul 02, 2013 at 03:19:12PM +0100, Jan Beulich wrote:
>> >>> On 02.07.13 at 15:40, Wei Liu <wei.liu2@citrix.com> wrote:
>> >  int xen_netbk_rx_ring_full(struct xenvif *vif)
>> >  {
>> > -	RING_IDX peek   = vif->rx_req_cons_peek;
>> > -	RING_IDX needed = max_required_rx_slots(vif);
>> > +	RING_IDX peek = vif->rx_req_cons_peek;
>> >  
>> > -	return ((vif->rx.sring->req_prod - peek) < needed) ||
>> > -	       ((vif->rx.rsp_prod_pvt + XEN_NETIF_RX_RING_SIZE - peek) < needed);
>> > +	return ((vif->rx.sring->req_prod < peek) ||
>> > +		(vif->rx.rsp_prod_pvt + XEN_NETIF_RX_RING_SIZE < peek));
>> 
>> This transformation is definitely wrong: You need to retain the
>> code's capability of dealing with the indexes wrapping, i.e.
>> simple comparisons won't do.
> 
> Hmm... so I don't quite understand this wrapping around thing and I have
> a question here, what if req_prod < peek so that (req_prod - peek)
> results in a very large number. In that case the ring is actually full
> (cannot accommodate that packet), however "needed" is only like 20-ish
> something, so the first comparision would fail. Are we relying on the
> second comparion to tell the ring is full?

Yes, and I'm struggling to understand what the first one actually
checked for. Your new first one in any case doesn't check whether
the ring is full, but whether there's enough slots available on the
ring to consume all the way up to "peek" (plus is having a wrap
around issue just like the second half).

> The only thing I need to do here is to tell whether rep_prod is ahead of
> peek in the first comparison and in the second case there is enough
> slots between rsp_prod_pvt and peek, what's the correct line?

Not sure - in any event, as said above, the first part isn't really
matching the name of the function anymore (so would at least
deserve a comment, to save future readers from struggling to
understand the purpose just like I do with the original one now).

Jan