lkml.org 
[lkml]   [2024]   [May]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH 5/5] selftests/bpf: a simple benchmark tool for /proc/<pid>/maps APIs
On Sat, May 4, 2024 at 2:57 PM Andrii Nakryiko
<andrii.nakryiko@gmail.com> wrote:
>
> On Sat, May 4, 2024 at 8:29 AM Greg KH <gregkh@linuxfoundation.org> wrote:
> >
> > On Fri, May 03, 2024 at 05:30:06PM -0700, Andrii Nakryiko wrote:
> > > Implement a simple tool/benchmark for comparing address "resolution"
> > > logic based on textual /proc/<pid>/maps interface and new binary
> > > ioctl-based PROCFS_PROCMAP_QUERY command.
> >
> > Of course an artificial benchmark of "read a whole file" vs. "a tiny
> > ioctl" is going to be different, but step back and show how this is
> > going to be used in the real world overall. Pounding on this file is
> > not a normal operation, right?
> >
>
> It's not artificial at all. It's *exactly* what, say, blazesym library
> is doing (see [0], it's Rust and part of the overall library API, I
> think C code in this patch is way easier to follow for someone not
> familiar with implementation of blazesym, but both implementations are
> doing exactly the same sequence of steps). You can do it even less
> efficiently by parsing the whole file, building an in-memory lookup
> table, then looking up addresses one by one. But that's even slower
> and more memory-hungry. So I didn't even bother implementing that, it
> would put /proc/<pid>/maps at even more disadvantage.
>
> Other applications that deal with stack traces (including perf) would
> be doing one of those two approaches, depending on circumstances and
> level of sophistication of code (and sensitivity to performance).

The code in perf doing this is here:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/perf/util/synthetic-events.c#n440
The code is using the api/io.h code:
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/tools/lib/api/io.h
Using perf to profile perf it was observed time was spent allocating
buffers and locale related activities when using stdio, so io is a
lighter weight alternative, albeit with more verbose code than fscanf.
You could add this as an alternate /proc/<pid>/maps reader, we have a
similar benchmark in `perf bench internals synthesize`.

Thanks,
Ian

> [0] https://github.com/libbpf/blazesym/blob/ee9b48a80c0b4499118a1e8e5d901cddb2b33ab1/src/normalize/user.rs#L193
>
> > thanks,
> >
> > greg k-h
>

\
 
 \ /
  Last update: 2024-05-27 18:16    [W:0.068 / U:1.168 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site