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

Re: [Xen-devel] xen-netfront possibly rides the rocket too often

Wednesday, May 14, 2014, 10:06:41 PM, you wrote:

> On 14/05/14 20:49, Zoltan Kiss wrote:
>> On 13/05/14 19:21, Stefan Bader wrote:
>>> Since I am not deeply familiar with the networking code, I wonder
>>> about two things:
>>> - is there something that should limit the skb data length from all frags
>>>    to stay below the 64K which the definition of MAX_SKB_FRAGS hints?
>> I think netfront should be able to handle 64K packets at most.
>>> - is multiple frags having offsets expected?
>> Yes, since compound pages a frag could over the 4K page boundary. The
>> problem is, that in the netback/front protocol the assumption is that
>> every slot is a single page, because grant operations could be done only
>> on a 4K page. And every slot ends up as a frag (expect maybe the first,
>> it can happen it is grant copied straight to the linear buffer),
>> therefore the frontend cannot send an skb which occupies more than
>> MAX_SKB_FRAGS individual 4k page.
>> The problem is known for a while, the solution is not, unfortunately.

> I think the worst case scenario is when every frag and the linear buffer 
> contains 2 bytes, which are overlapping a page boundary (that's 
> (17+1)*2=36 so far), plus 15 of them have a 4k page in the middle of 
> them, so, a 1+4096+1 byte buffer can span over 3 page. That's 51 
> individual pages.
> With the previous grant copy implementation there would be the option to 
> modify backend and coalesce everything into a well formed skb. That 
> would be a minor change there. But with grant mapping it's harder.
> Slots of compound pages could be mapped to adjacent pages to Dom0, maybe 
> somehow you can present them as compound pages in Dom0 as well. But in 
> MFN space they wouldn't be contiguous, you need SWIOTLB or use IOMMU to 
> hide that from the devices. Plus, what happens when you can't find 
> adjacent pending slots?
> I think we would be better off at the moment with trying to compact 
> these skbs a bit. Usually they overflow the limit by one or two, which 
> means we should reallocate one or two frag, or the linear buffer to 
> decrease the number of 4K pages used.

How does virtio-net handle this, the would probably have ran into the same 
problems ?


> Zoli

Xen-devel mailing list



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