Re: [Xen-devel] [PATCH] xen/netfront: handle compound page fragments on transmit

[Eric Dumazet <edumazet@xxxxxxxxxx>]

At least TCP skbs are limited to 65536 bytes in tcp_sendmsg()

(around 45 1460-bytes MSS segments)

Thats probably because IPv4  stack only copes with this limit.

This probably could be relaxed for TCP friends, but this kind of skbs should not hit a driver.

On Tue, Nov 20, 2012 at 7:44 AM, Jan Beulich <JBeulich@xxxxxxxx> wrote:

>>> On 20.11.12 at 16:06, Ian Campbell <Ian.Campbell@xxxxxxxxxx> wrote:
> On Tue, 2012-11-20 at 14:32 +0000, Jan Beulich wrote:
>> > @@ -517,15 +540,16 @@ static int xennet_start_xmit(struct sk_buff *skb,
>> > struct net_device *dev)
>> >    grant_ref_t ref;
>> >    unsigned long mfn;
>> >    int notify;
>> > -  int frags = skb_shinfo(skb)->nr_frags;
>> > +  int slots;
>> >    unsigned int offset = offset_in_page(data);
>> >    unsigned int len = skb_headlen(skb);
>> >    unsigned long flags;
>> >
>> > -  frags += DIV_ROUND_UP(offset + len, PAGE_SIZE);
>> > -  if (unlikely(frags > MAX_SKB_FRAGS + 1)) {
>> > -          printk(KERN_ALERT "xennet: skb rides the rocket: %d frags\n",
>> > -                 frags);
>> > +  slots = DIV_ROUND_UP(offset + len, PAGE_SIZE) +
>> > +          xennet_count_skb_frag_slots(skb);
>> > +  if (unlikely(slots > MAX_SKB_FRAGS + 1)) {
>>
>> But still - isn't this wrong now (i.e. can't it now validly exceed the
>> boundary checked for)?
>
> In practice no because of the property that the number of pages backing
> the frags is <= MAX_SKB_FRAGS even if you are using compound pages as
> the frags.

So are you saying that there is something in the system
preventing up to MAX_SKB_FRAGS * SKB_FRAG_PAGE_ORDER
(or NETDEV_FRAG_PAGE_MAX_ORDER) skb-s to be created? I
didn't find any. I do notice that __netdev_alloc_frag() currently
never gets called with a size larger than PAGE_SIZE, but
considering that the function just recently got made capable of
that, I'm sure respective users will show up rather sooner than
later.

Jan

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