lkml.org 
[lkml]   [2010]   [May]   [17]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC] Tracer Ring Buffer splice() vs page cache [was: Re: Perf and ftrace [was Re: PyTimechart]]
* Peter Zijlstra (peterz@infradead.org) wrote:
> On Fri, 2010-05-14 at 14:32 -0400, Mathieu Desnoyers wrote:
>
> > [CCing memory management specialists]
>
> And jet you forgot Jens who wrote it ;-)

oops ! thanks for adding him.

>
> > So I have three questions here:
> >
> > 1 - could we enforce removal of these pages from the page cache by calling
> > "page_cache_release()" before giving these pages back to the ring buffer ?
> >
> > 2 - or maybe is there a page flag we could specify when we allocate them to
> > ask for these pages to never be put in the page cache ? (but they should be
> > still usable as write buffers)
> >
> > 3 - is there something more we need to do to grab a reference on the pages
> > before passing them to splice(), so that when we call page_cache_release()
> > they don't get reclaimed ?
>
> There is no guarantee it is the pagecache they end up in, it could be a
> network packet queue, a pipe, or anything that implements .splice_write.
>
> >From what I understand of splice() is that it assumes it passes
> ownership of the page, you're not supposed to touch them again, non of
> the above three are feasible.

Yup, I've looked more deeply at the splice() code, and I now see why things
don't fall apart in LTTng currently. My implementation seems to be causing
splice() to perform a copy. My ring buffer splice implementation is derived from
kernel/relay.c. I override

pipe_buf_operations release op with:

static void ltt_relay_pipe_buf_release(struct pipe_inode_info *pipe,
struct pipe_buffer *pbuf)
{
}

and

splice_pipe_desc spd_release file op with:

static void ltt_relay_page_release(struct splice_pipe_desc *spd, unsigned int i)
{
}

My understanding is that by keeping 2 references on the pages (the ring buffer +
the pipe), splice safely refuses to move the pages and performs a copy instead.

I'll continue to look into this. One of the things I noticed that that we could
possibly use the "steal()" operation to steal the pages back from the page cache
to repopulate the ring buffer rather than continuously allocating new pages. If
steal() fails for some reasons, then we can fall back on page allocation. I'm
not sure it is safe to assume anything about pages being in the page cache
though. Maybe the safest route is to just allocate new pages for now.

Thoughts ?

Mathieu


--
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com


\
 
 \ /
  Last update: 2010-05-18 00:45    [W:1.049 / U:0.068 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site