Re: [profile]: [0/23] mmap() support for /proc/profile

[William Lee Irwin III <wli@holomorphy.com>]

On Wed, Jun 23, 2004 at 09:16:46AM +1000, Anton Blanchard wrote:
> Interesting stuff. FYI we did some analysis of the hottest addresses in
> the kernel and profile_lock featured very high up:
> void profile_hook(struct pt_regs * regs)
> {
>         read_lock(&profile_lock);
>         notifier_call_chain(&profile_listeners, 0, regs);
>         read_unlock(&profile_lock);
> }
> Thats 2 atomic operations to the same cacheline per timer interrupt per
> cpu. Considering how rarely timer based profiling is used, perhaps RCU
> or even just a profiling_enabled sysctl flag would help here. Id prefer
> not to compile it out in distro kernels if possible, its a very useful
> feature when required.
> In the mean time, how about this quick fix?

Well, this is a little different. I was more concerned about the
"Heisenberg effect" that the in-kernel copies to fetch profiling data
had upon the data fetched. i.e. instead of idle time and the stuff I
was looking for, I saw copy_to_user() and all kinds of vfs and /proc/
crap instead, which blew what I was looking for completely out of the
top 20. The profiling I did was on UP, which was done in part to
eliminate lock contention as the cause of the phenomena I had observed.

Also, Randy said I should mention that my kerneltop-like thingie is
about 5 times faster (well, uses about 20% of the cpu time) as the
read()-based kerneltop. Which is all well and good, but the reason I
actually needed this was to get rid of the Heisenberg problem.


-- wli