Re: [RFC] net/core/dst.c : Should'nt dst_run_gc() be more scalable and friendly ?

[Eric Dumazet <dada1@cosmosbay.com>]

On Fri, 17 Aug 2007 16:15:22 +0800
Herbert Xu <herbert@gondor.apana.org.au> wrote:

> On Fri, Aug 17, 2007 at 10:10:30AM +0200, Eric Dumazet wrote:
> >
> > Will a workqueue react the same in case of a DDOS situation, 
> > where softirq could use all CPU cycles to handle incoming
> > packets and feed the GC list, and GC would never
> > have a chance to scan and free some items ?
> 
> Well when that happens the softirqs will be deferred to
> ksoftirqd which should share the CPU fairly with the
> workqueue.

Thats nice :)

I'll code a workqueue based thing in about 10 days after my hollidays,
and perform DOS tests as well.

Thanks for the feedback.

> 
> > About chunk processing, I did it on purpose, to not throw away
> > all CPU cache. Goal is to process entries, but not all of them
> > in a row, especially if we find many yet referenced entries
> > (and thus not candidates to freeing)
> 
> I agree that chunks are desirable for a timer since you'd
> be hogging the CPU otherwise.  However, if you went to a
> workqueue then it's less of a concern and would simplify
> things.  In particular, you won't have to pick a good
> chunk size :)