Re: [PATCH net-next] xen-netback: improve guest-receive-side flow control

[annie li <annie.li@oracle.com>]

On 2013/11/29 0:27, Wei Liu wrote:
> On Thu, Nov 28, 2013 at 04:05:52PM +0000, Paul Durrant wrote:
>>> -----Original Message-----
>>> From: Wei Liu [mailto:wei.liu2@citrix.com]
>>> Sent: 28 November 2013 15:51
>>> To: Paul Durrant
>>> Cc: xen-devel@lists.xen.org; Wei Liu; Ian Campbell; David Vrabel
>>> Subject: Re: [PATCH net-next] xen-netback: improve guest-receive-side flow
>>> control
>>>
>>> On Thu, Nov 28, 2013 at 01:11:12PM +0000, Paul Durrant wrote:
>>>> The flow control code relies on a double pass of tke skb, firstly to count
>>>> the number of ring slots needed to supply sufficient grant references, and
>>>> another to actually consume those references and build the grant copy
>>>> operations. It transpires that, in some scenarios, the initial count and the
>>>> number of references consumed differs and, after this happens a number
>>> of
>>>> times, flow control is completely broken.

Can you describe it in more details? What kind of scenarios does this 
issue happen? I assume current xenvif_count_skb_slots() returns accurate 
value now.

>>> The removal of xenvif_count_skb_slots() is OK because you're now
>>> estimating minimum slots, but the actual number of slots consumed needs
>>> to be acurate. Any reason you don't want to use xenvif_gop_skb()?
>>>
>> I can't. xenvif_gop_skb() pulls requests off the ring and if there aren't enough there it will crash'n'burn as far as I can tell. So, I need an upper bound on the number of slots before I call it.
>>
> I get what you mean. We can at least use some of the snippet in it
> without pulling requests. See above.
>

I prefer accurate value since it would help to save more slots. But I am 
not sure what the flow control issue is, does using similar mechanism of 
xenvif_count_skb_slots here would cause this issue too?

Thanks
Annie