Re: virtio_net and SMP guests

[Dor Laor <dor.laor-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>]

Rusty Russell wrote:
> On Monday 24 December 2007 10:19:19 Dor Laor wrote:
>   
>> Rusty Russell wrote:
>> Looks good to me. The only thing is the naming.. Maybe one can find
>> better name than [dis|en]able_cb since
>> it is more like disable interrupts than disable_cb and enable_cb is more
>> like run_callbacks (and enable interrupts).
>>     
>
> Yes, the nomenclature is a little wishy-washy.  Perhaps we should call 
> them "interrupts" even though (all together) the s390 doesn't have 
> interrupts.
>
>   
>> Actually while looking at it some more, there's might be a performance
>> optimization using your new patch:
>> Up to now, if the tx/rx paths were running out of descriptors they
>> called enable_cb.
>> enable_cb told the host it should trigger an irq the next time it has
>> data for the guest.
>>  From now on, enable_cb will run the callbacks inside shortening latency.
>>
>> btw, I tried to apply this patch on my source tree without luck,  after
>> doing a manual merge, the
>> guest opssed on the BUG_ON. I think it's because my sources are not
>> aligned with yours.
>> Can you please post a mercurial/git repo? Please specify the relatively
>> repository in case you choose mercurial.
>>     
>
> Hmm, I currently publish a patch queue and a subset of that for Linus to pull 
> from.  I could create a git tree from it but it'd be useless to you, since 
> it'd be re-created every time I modify the patch queue...
>
> Rusty.
>   
It's ok with me, anyway I always pull from Avi & you and try push new 
patches back.
btw: I checked your Linux tree and found it a bit old, there was no tx 
coalescing timer, it's probably good
for 2.6.24. Do you plan to post patchset for 25 windows?
Thanks,
Dor

-------------------------------------------------------------------------
This SF.net email is sponsored by: Microsoft
Defy all challenges. Microsoft(R) Visual Studio 2005.
http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/