Re: [Xen-devel] [PATCH RFC] xen-netback: calculate the number of slots required for large MTU vifs

[Matt Wilson <msw@amazon.com>]

On Wed, Jul 10, 2013 at 08:37:03PM +0100, Wei Liu wrote:
> On Wed, Jul 10, 2013 at 09:13:33AM +0100, Wei Liu wrote:
> > On Tue, Jul 09, 2013 at 10:40:59PM +0000, Matt Wilson wrote:
> > > From: Xi Xiong <xixiong@amazon.com>
> > > 
> > > [ note: I've just cherry picked this onto net-next, and only compile
> > >   tested. This a RFC only. -msw ]
> > > 
> > 
> > Should probably rebase it on net.git because it is a bug fix. Let's
> > worry about that later...

*nod*

> > > Currently the number of RX slots required to transmit a SKB to
> > > xen-netfront can be miscalculated when an interface uses a MTU larger
> > > than PAGE_SIZE. If the slot calculation is wrong, xen-netback can
> > > pause the queue indefinitely or reuse slots. The former manifests as a
> > > loss of connectivity to the guest (which can be restored by lowering
> > > the MTU set on the interface). The latter manifests with "Bad grant
> > > reference" messages from Xen such as:
> > > 
> > > (XEN) grant_table.c:1797:d0 Bad grant reference 264241157
> > > 
> > > and kernel messages within the guest such as:
> > > 
> > > [  180.419567] net eth0: Invalid extra type: 112
> > > [  180.868620] net eth0: rx->offset: 0, size: 4294967295
> > > [  180.868629] net eth0: rx->offset: 0, size: 4294967295
> > > 
> > > BUG_ON() assertions can also be hit if RX slots are exhausted while
> > > handling a SKB.
> > > 
> > > This patch changes xen_netbk_rx_action() to count the number of RX
> > > slots actually consumed by netbk_gop_skb() instead of using nr_frags + 1.
> > > This prevents under-counting the number of RX slots consumed when a
> > > SKB has a large linear buffer.
> > > 
> > > Additionally, we now store the estimated number of RX slots required
> > > to handle a SKB in the cb overlay. This value is used to determine if
> > > the next SKB in the queue can be processed.
> > > 
> > > Finally, the logic in start_new_rx_buffer() can cause RX slots to be
> > > wasted when setting up copy grant table operations for SKBs with large
> > > linear buffers. For example, a SKB with skb_headlen() equal to 8157
> > > bytes that starts 64 bytes 64 bytes from the start of the page will
> > 
> > Duplicated "64 bytes". And this change looks like an improvement not a
> > bug fix. Probably submit a separate patch for this?

Argh, I knew it was in there somewhere (since you pointed it out in
Dubin :-).

Maybe it could be a separate patch. I think the description is also a
bit confusing. I'll work on rewording it.

> > > consume three RX slots instead of two. This patch changes the "head"
> > > parameter to netbk_gop_frag_copy() to act as a flag. When set,
> > > start_new_rx_buffer() will always place as much data as possible into
> > > each RX slot.
> > > 
> > > Signed-off-by: Xi Xiong <xixiong@amazon.com>
> > > Reviewed-by: Matt Wilson <msw@amazon.com>
> > > [ msw: minor code cleanups, rewrote commit message, adjusted code
> > >   to count RX slots instead of meta structures ]
> > > Signed-off-by: Matt Wilson <msw@amazon.com>
> > > Cc: Annie Li <annie.li@oracle.com>
> > > Cc: Wei Liu <wei.liu2@citrix.com>
> > > Cc: Ian Campbell <Ian.Campbell@citrix.com>
> > > Cc: netdev@vger.kernel.org
> > > Cc: xen-devel@lists.xenproject.org
> > > ---
> > >  drivers/net/xen-netback/netback.c |   51 ++++++++++++++++++++++--------------
> > >  1 files changed, 31 insertions(+), 20 deletions(-)
> > > 
> > > diff --git a/drivers/net/xen-netback/netback.c b/drivers/net/xen-netback/netback.c
> > > index 64828de..82dd207 100644
> > > --- a/drivers/net/xen-netback/netback.c
> > > +++ b/drivers/net/xen-netback/netback.c
> > > @@ -110,6 +110,11 @@ union page_ext {
> > >  	void *mapping;
> > >  };
> > >  
> > > +struct skb_cb_overlay {
> > > +	int meta_slots_used;
> > > +	int peek_slots_count;
> > > +};
> > > +
> > >  struct xen_netbk {
> > >  	wait_queue_head_t wq;
> > >  	struct task_struct *task;
> > > @@ -370,6 +375,7 @@ unsigned int xen_netbk_count_skb_slots(struct xenvif *vif, struct sk_buff *skb)
> > >  {
> > >  	unsigned int count;
> > >  	int i, copy_off;
> > > +	struct skb_cb_overlay *sco;
> > >  
> > >  	count = DIV_ROUND_UP(skb_headlen(skb), PAGE_SIZE);
> > >  
> > > @@ -411,6 +417,9 @@ unsigned int xen_netbk_count_skb_slots(struct xenvif *vif, struct sk_buff *skb)
> > >  				offset = 0;
> > >  		}
> > >  	}
> > > +
> > > +	sco = (struct skb_cb_overlay *) skb->cb;
> > > +	sco->peek_slots_count = count;
> > >  	return count;
> > >  }
> > >  
> > > @@ -443,13 +452,12 @@ static struct netbk_rx_meta *get_next_rx_buffer(struct xenvif *vif,
> > >  }
> > >  
> > >  /*
> > > - * Set up the grant operations for this fragment. If it's a flipping
> > > - * interface, we also set up the unmap request from here.
> > > + * Set up the grant operations for this fragment.
> > >   */
> > >  static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
> > >  				struct netrx_pending_operations *npo,
> > >  				struct page *page, unsigned long size,
> > > -				unsigned long offset, int *head)
> > > +				unsigned long offset, int head, int *first)
> > >  {
> > >  	struct gnttab_copy *copy_gop;
> > >  	struct netbk_rx_meta *meta;
> > > @@ -479,12 +487,12 @@ static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
> > >  		if (bytes > size)
> > >  			bytes = size;
> > >  
> > > -		if (start_new_rx_buffer(npo->copy_off, bytes, *head)) {
> > > +		if (start_new_rx_buffer(npo->copy_off, bytes, head)) {
> > >  			/*
> > >  			 * Netfront requires there to be some data in the head
> > >  			 * buffer.
> > >  			 */
> > > -			BUG_ON(*head);
> > > +			BUG_ON(*first);
> > >  
> > >  			meta = get_next_rx_buffer(vif, npo);
> > >  		}
> > > @@ -529,10 +537,10 @@ static void netbk_gop_frag_copy(struct xenvif *vif, struct sk_buff *skb,
> > >  		}
> > >  
> > >  		/* Leave a gap for the GSO descriptor. */
> > > -		if (*head && skb_shinfo(skb)->gso_size && !vif->gso_prefix)
> > > +		if (*first && skb_shinfo(skb)->gso_size && !vif->gso_prefix)
> > >  			vif->rx.req_cons++;
> > >  
> > > -		*head = 0; /* There must be something in this buffer now. */
> > > +		*first = 0; /* There must be something in this buffer now. */
> > >  
> > >  	}
> > >  }
> > > @@ -558,7 +566,7 @@ static int netbk_gop_skb(struct sk_buff *skb,
> > >  	struct xen_netif_rx_request *req;
> > >  	struct netbk_rx_meta *meta;
> > >  	unsigned char *data;
> > > -	int head = 1;
> > > +	int first = 1;
> > >  	int old_meta_prod;
> > >  
> > >  	old_meta_prod = npo->meta_prod;
> > > @@ -594,16 +602,16 @@ static int netbk_gop_skb(struct sk_buff *skb,
> > >  			len = skb_tail_pointer(skb) - data;
> > >  
> > >  		netbk_gop_frag_copy(vif, skb, npo,
> > > -				    virt_to_page(data), len, offset, &head);
> > > +				virt_to_page(data), len, offset, 1, &first);
> > >  		data += len;
> > >  	}
> > >  
> > >  	for (i = 0; i < nr_frags; i++) {
> > >  		netbk_gop_frag_copy(vif, skb, npo,
> > > -				    skb_frag_page(&skb_shinfo(skb)->frags[i]),
> > > -				    skb_frag_size(&skb_shinfo(skb)->frags[i]),
> > > -				    skb_shinfo(skb)->frags[i].page_offset,
> > > -				    &head);
> > > +				skb_frag_page(&skb_shinfo(skb)->frags[i]),
> > > +				skb_frag_size(&skb_shinfo(skb)->frags[i]),
> > > +				skb_shinfo(skb)->frags[i].page_offset,
> > > +				0, &first);
> > >  	}
> > >  
> > >  	return npo->meta_prod - old_meta_prod;
> > > @@ -661,10 +669,6 @@ static void netbk_add_frag_responses(struct xenvif *vif, int status,
> > >  	}
> > >  }
> > >  
> > > -struct skb_cb_overlay {
> > > -	int meta_slots_used;
> > > -};
> > > -
> > >  static void xen_netbk_rx_action(struct xen_netbk *netbk)
> > >  {
> > >  	struct xenvif *vif = NULL, *tmp;
> > > @@ -690,19 +694,26 @@ static void xen_netbk_rx_action(struct xen_netbk *netbk)
> > >  	count = 0;
> > >  
> > >  	while ((skb = skb_dequeue(&netbk->rx_queue)) != NULL) {
> > > +		RING_IDX old_rx_req_cons;
> > > + 
> > >  		vif = netdev_priv(skb->dev);
> > >  		nr_frags = skb_shinfo(skb)->nr_frags;
> > >  
> > > +		old_rx_req_cons = vif->rx.req_cons;
> > >  		sco = (struct skb_cb_overlay *)skb->cb;
> > >  		sco->meta_slots_used = netbk_gop_skb(skb, &npo);
> > >  
> > > -		count += nr_frags + 1;
> > > +		count += vif->rx.req_cons - old_rx_req_cons;
> > >  
> > >  		__skb_queue_tail(&rxq, skb);
> > >  
> > > +		skb = skb_peek(&netbk->rx_queue);
> > > +		if (skb == NULL)
> > > +			break;
> > > +		sco = (struct skb_cb_overlay *) skb->cb;
> > > +
> > >  		/* 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 + sco->peek_slots_count >= XEN_NETIF_RX_RING_SIZE)
> > >  			break;
> > 
> > Using req_cons to count is OK, but since you've already store the number
> > of slots in sco->peek_slots_count before actually queueing the packet,
> > why don't you use that directly?
> > 
> 
> Ah, I must be off my head here. I mixed several patches 1) my RFC patch
> that removes MAX_SKB_FRAGS in RX path, 2) Annie's patch and 3) your
> patch.

Yes, I wanted to get this out since we're talking about multiple
changes in similar areas.

> The reason that your patches uses req_cons instead of pre-calculated
> value is, that value returned by xen_netbk_skb_count_slots is actually
> *wrong* -- that's what Annie tried to fix in her patch.

Yes and no. With this patch, the xen_netbk_count_skb_slots() is now
correct. xen_netbk_count_skb_slots() was under-counting when slots
were inefficiently consumed. With this patch they match.

Since we've had trouble with the counting in the past, I felt like
using the "real" number of slots here was safer and future-proof.

> After fixing xen_netbk_skb_count_slots, we would need the above snippet
> (or the snippet I proposed in my RFC patch) to prevent overruning the
> ring.

Agreed.

> So a proper patch to this issue (not couting RX slots correctly causing
> overrun of the ring) would be the above two aspects combined.

Yes, I think so.

> Comments?

I think that this patch addresses the problem more completely. Annie?

See also the thread from last August:
   http://lists.xen.org/archives/html/xen-devel/2012-12/msg00274.html

--msw