[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

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




On 2013-7-12 16:49, Wei Liu wrote:
On Thu, Jul 11, 2013 at 09:52:52PM +0800, annie li wrote:
On 2013-7-11 14:01, annie li wrote:
On 2013-7-11 13:14, Matt Wilson wrote:
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@xxxxxxxxxx>

[ 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@xxxxxxxxxx>
Reviewed-by: Matt Wilson <msw@xxxxxxxxxx>
[ msw: minor code cleanups, rewrote commit message, adjusted code
   to count RX slots instead of meta structures ]
Signed-off-by: Matt Wilson <msw@xxxxxxxxxx>
Cc: Annie Li <annie.li@xxxxxxxxxx>
Cc: Wei Liu <wei.liu2@xxxxxxxxxx>
Cc: Ian Campbell <Ian.Campbell@xxxxxxxxxx>
Cc: netdev@xxxxxxxxxxxxxxx
Cc: xen-devel@xxxxxxxxxxxxxxxxxxxx
---
  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)) {
head is always 1 here when processing the header data, so it is
impossible to start a new buffer for header data with larger
header size, and get_next_rx_buffer is never called. I assume this
would not work well for skb with larger header data size. Have you
run network test with this patch?
Sorry, I forgot the offset == MAX_BUFFER_OFFSET case and
misunderstand your patch, please ignore my last comments. Your patch
keeps the original DIV_ROUND_UP and changes the mechanism in
netbk_gop_frag_copy to make slots same with
xen_netbk_count_skb_slots. All Roads Lead to Rome!:-)

Actually I'm now in favor of Matt's approach as it a) fix the bug, b)
make efficient use of the ring.

Annie, Ian, what's your opinion?

Me too. Looks his approach not only fix this bug, but also saves the ring usage for header data, and this is good. Combining my patch with his is not a good idea, so I think we can use his patch.

Thanks
Annie


Wei.

Thanks
Annie


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.