Re: [RFC PATCH v4] Unified trace buffer

[Mathieu Desnoyers <compudj@krystal.dyndns.org>]

* Steven Rostedt (rostedt@goodmis.org) wrote:
> 
> On Thu, 25 Sep 2008, Masami Hiramatsu wrote:
> 
> > Hi Steven,
> > 
> > Steven Rostedt wrote:
> > > This version has been cleaned up a bit. I've been running it as
> > > a back end to ftrace, and it has been handling pretty well.
> > 
> > Thank you for your great work.
> > It seems good to me(especially, encapsulating events :)).
> 
> Thanks!
> 
> > 
> > I have one request of enhancement.
> > 
> > > +static struct ring_buffer_per_cpu *
> > > +ring_buffer_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu)
> > > +{
> > [...]
> > > +	cpu_buffer->pages = kzalloc_node(ALIGN(sizeof(void *) * pages,
> > > +					       cache_line_size()), GFP_KERNEL,
> > > +					 cpu_to_node(cpu));
> > 
> > Here, you are using a slab object for page managing array,
> > the largest object size is 128KB(x86-64), so it can contain
> > 16K pages = 64MB.
> > 
> > As I had improved relayfs, in some rare case(on 64bit arch),
> > we'd like to use larger buffer than 64MB.
> > 
> > http://sourceware.org/ml/systemtap/2008-q2/msg00103.html
> > 
> > So, I think similar hack can be applicable.
> > 
> > Would it be acceptable for the next version?
> 
> I would like to avoid using vmalloc as much as possible, but I do see the 
> limitation here. Here's my compromise.
> 
> Instead of using vmalloc if the page array is greater than one page, 
> how about using vmalloc if the page array is greater than 
> KMALLOC_MAX_SIZE?
> 
> This would let us keep the vmap area free unless we have no choice.
> 
> -- Steve
> 

You could also fallback on a 2-level page array when buffer size is >
64MB. The cost is mainly a supplementary pointer dereference, but one
more should not make sure a big difference overall.

Mathieu

-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68