xen-netback: count number required slots for an skb more carefully
authorDavid Vrabel <david.vrabel@citrix.com>
Wed, 11 Sep 2013 13:52:48 +0000 (14:52 +0100)
committerDavid S. Miller <davem@davemloft.net>
Fri, 13 Sep 2013 03:22:13 +0000 (23:22 -0400)
When a VM is providing an iSCSI target and the LUN is used by the
backend domain, the generated skbs for direct I/O writes to the disk
have large, multi-page skb->data but no frags.

With some lengths and starting offsets, xen_netbk_count_skb_slots()
would be one short because the simple calculation of
DIV_ROUND_UP(skb_headlen(), PAGE_SIZE) was not accounting for the
decisions made by start_new_rx_buffer() which does not guarantee
responses are fully packed.

For example, a skb with length < 2 pages but which spans 3 pages would
be counted as requiring 2 slots but would actually use 3 slots.

skb->data:

    |        1111|222222222222|3333        |

Fully packed, this would need 2 slots:

    |111122222222|22223333    |

But because the 2nd page wholy fits into a slot it is not split across
slots and goes into a slot of its own:

    |1111        |222222222222|3333        |

Miscounting the number of slots means netback may push more responses
than the number of available requests.  This will cause the frontend
to get very confused and report "Too many frags/slots".  The frontend
never recovers and will eventually BUG.

Fix this by counting the number of required slots more carefully.  In
xen_netbk_count_skb_slots(), more closely follow the algorithm used by
xen_netbk_gop_skb() by introducing xen_netbk_count_frag_slots() which
is the dry-run equivalent of netbk_gop_frag_copy().

Signed-off-by: David Vrabel <david.vrabel@citrix.com>
Acked-by: Ian Campbell <ian.campbell@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/xen-netback/netback.c

index 956130c..f3e591c 100644 (file)
@@ -212,6 +212,49 @@ static bool start_new_rx_buffer(int offset, unsigned long size, int head)
        return false;
 }
 
+struct xenvif_count_slot_state {
+       unsigned long copy_off;
+       bool head;
+};
+
+unsigned int xenvif_count_frag_slots(struct xenvif *vif,
+                                    unsigned long offset, unsigned long size,
+                                    struct xenvif_count_slot_state *state)
+{
+       unsigned count = 0;
+
+       offset &= ~PAGE_MASK;
+
+       while (size > 0) {
+               unsigned long bytes;
+
+               bytes = PAGE_SIZE - offset;
+
+               if (bytes > size)
+                       bytes = size;
+
+               if (start_new_rx_buffer(state->copy_off, bytes, state->head)) {
+                       count++;
+                       state->copy_off = 0;
+               }
+
+               if (state->copy_off + bytes > MAX_BUFFER_OFFSET)
+                       bytes = MAX_BUFFER_OFFSET - state->copy_off;
+
+               state->copy_off += bytes;
+
+               offset += bytes;
+               size -= bytes;
+
+               if (offset == PAGE_SIZE)
+                       offset = 0;
+
+               state->head = false;
+       }
+
+       return count;
+}
+
 /*
  * Figure out how many ring slots we're going to need to send @skb to
  * the guest. This function is essentially a dry run of
@@ -219,48 +262,39 @@ static bool start_new_rx_buffer(int offset, unsigned long size, int head)
  */
 unsigned int xenvif_count_skb_slots(struct xenvif *vif, struct sk_buff *skb)
 {
+       struct xenvif_count_slot_state state;
        unsigned int count;
-       int i, copy_off;
+       unsigned char *data;
+       unsigned i;
 
-       count = DIV_ROUND_UP(skb_headlen(skb), PAGE_SIZE);
+       state.head = true;
+       state.copy_off = 0;
 
-       copy_off = skb_headlen(skb) % PAGE_SIZE;
+       /* Slot for the first (partial) page of data. */
+       count = 1;
 
+       /* Need a slot for the GSO prefix for GSO extra data? */
        if (skb_shinfo(skb)->gso_size)
                count++;
 
-       for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
-               unsigned long size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
-               unsigned long offset = skb_shinfo(skb)->frags[i].page_offset;
-               unsigned long bytes;
-
-               offset &= ~PAGE_MASK;
-
-               while (size > 0) {
-                       BUG_ON(offset >= PAGE_SIZE);
-                       BUG_ON(copy_off > MAX_BUFFER_OFFSET);
-
-                       bytes = PAGE_SIZE - offset;
-
-                       if (bytes > size)
-                               bytes = size;
+       data = skb->data;
+       while (data < skb_tail_pointer(skb)) {
+               unsigned long offset = offset_in_page(data);
+               unsigned long size = PAGE_SIZE - offset;
 
-                       if (start_new_rx_buffer(copy_off, bytes, 0)) {
-                               count++;
-                               copy_off = 0;
-                       }
+               if (data + size > skb_tail_pointer(skb))
+                       size = skb_tail_pointer(skb) - data;
 
-                       if (copy_off + bytes > MAX_BUFFER_OFFSET)
-                               bytes = MAX_BUFFER_OFFSET - copy_off;
+               count += xenvif_count_frag_slots(vif, offset, size, &state);
 
-                       copy_off += bytes;
+               data += size;
+       }
 
-                       offset += bytes;
-                       size -= bytes;
+       for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
+               unsigned long size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
+               unsigned long offset = skb_shinfo(skb)->frags[i].page_offset;
 
-                       if (offset == PAGE_SIZE)
-                               offset = 0;
-               }
+               count += xenvif_count_frag_slots(vif, offset, size, &state);
        }
        return count;
 }