* Frederic Weisbecker wrote:
> On Tue, Oct 08, 2013 at 10:05:21PM +0200, Peter Zijlstra wrote:
> > On Tue, Oct 08, 2013 at 09:55:59PM +0200, Frederic Weisbecker wrote:
> > > 1) make breakpoints independant from perf. The drawback is that we must
> > > then
> > > add seperate hooks on context sw
Some more comments.
> - your patches might break apps/ABI
Can you please explain that a bit more. We have a lot of CONFIG options
that disable syscalls, /sys, lots of stuff. Whoever uses them needs
to know what they are doing. I thought it was pretty
much consensus that Linux is supposed to be
> So I test-built a config close to your config with both tracing and perf
> on and off (note, I had OPROFILE and KVM in a module), and got the
> following kernel sizes:
Yes I mistakenly included KVM (I think that was the difference)
Without KVM it's ~272k text, 96k BSS data delta.
Still big, b
> You'd think that, but for whatever reason, ftrace/perf oopses still happen.
Hiding bugs seems like a poor use of the CONFIG option. It would
be better to figure out a way to catch them earlier. Perhaps
trinity needs to run more often? any chance of a fengguang style nightly
service for mainline
On Tue, Oct 08, 2013 at 10:05:21PM +0200, Peter Zijlstra wrote:
> On Tue, Oct 08, 2013 at 09:55:59PM +0200, Frederic Weisbecker wrote:
> > 1) make breakpoints independant from perf. The drawback is that we must then
> > add seperate hooks on context switch for ptrace breakpoints. OTOH we get
> > ri
On Tue, Oct 08, 2013 at 10:05:21PM +0200, Peter Zijlstra wrote:
> On Tue, Oct 08, 2013 at 09:55:59PM +0200, Frederic Weisbecker wrote:
> > 1) make breakpoints independant from perf. The drawback is that we must then
> > add seperate hooks on context switch for ptrace breakpoints. OTOH we get
> > ri
On Tue, Oct 08, 2013 at 09:36:13PM +0200, Ingo Molnar wrote:
>
> Peter is working on it - but it's slow. Could you try to disable
> sys_perf_open, ptrace and the NMI watchdog? No perf functionality should
> be used in that case. If you disable CONFIG_FTRACE then no ftrace
> functionality should
On Tue, Oct 08, 2013 at 09:55:59PM +0200, Frederic Weisbecker wrote:
> 1) make breakpoints independant from perf. The drawback is that we must then
> add seperate hooks on context switch for ptrace breakpoints. OTOH we get
> rid of the perf -> breakpoint -> perf circular dependency, which is the ve
On Sat, Oct 05, 2013 at 07:05:52PM +0200, Andi Kleen wrote:
> On Sat, Oct 05, 2013 at 09:08:06AM +0200, Ingo Molnar wrote:
> >
> > * Andi Kleen wrote:
> >
> > > From: Andi Kleen
> > >
> > > As suggested by Ingo.
> >
> > No, you haven't read my suggestion carefully enough so NAK.
>
> Ok I tru
* Steven Rostedt wrote:
> On Tue, Oct 08, 2013 at 08:59:38AM +0200, Ingo Molnar wrote:
> >
> > Secondly and more importantly, visualized with relative sizes, in a
> > feature matrix, makes it clearer what's going on with vmlinux .text:
> >
> > perf-OFF perf-ON
* Dave Jones wrote:
> On Tue, Oct 08, 2013 at 10:22:23AM -0600, David Ahern wrote:
> > On 10/8/13 9:51 AM, Dave Jones wrote:
> > > On Tue, Oct 08, 2013 at 08:59:38AM +0200, Ingo Molnar wrote:
> > >
> > > > You might want to concentrate your efforts from fighting perf
> > > > functionali
On Tue, Oct 08, 2013 at 10:22:23AM -0600, David Ahern wrote:
> On 10/8/13 9:51 AM, Dave Jones wrote:
> > On Tue, Oct 08, 2013 at 08:59:38AM +0200, Ingo Molnar wrote:
> >
> > > You might want to concentrate your efforts from fighting perf
> > > functionality towards decreasing per tracepoin
On 10/8/13 9:51 AM, Dave Jones wrote:
On Tue, Oct 08, 2013 at 08:59:38AM +0200, Ingo Molnar wrote:
> You might want to concentrate your efforts from fighting perf
> functionality towards decreasing per tracepoint overhead instead,
> without hurting kernel functionality and maintainability.
On Tue, Oct 08, 2013 at 08:59:38AM +0200, Ingo Molnar wrote:
> You might want to concentrate your efforts from fighting perf
> functionality towards decreasing per tracepoint overhead instead,
> without hurting kernel functionality and maintainability.
Making it easier to disable perf entir
On Tue, 8 Oct 2013, Ingo Molnar wrote:
> The base cost of perf is +151k - not small but a far cry from your claimed
> +700k ...
Things might get that bad if we enable full event lists in the kernel for
all of x86, as Power is starting to do.
The patch to add all Power7 events cfe0d8ba14a1d9824
On Tue, Oct 08, 2013 at 08:59:38AM +0200, Ingo Molnar wrote:
>
> Secondly and more importantly, visualized with relative sizes, in a
> feature matrix, makes it clearer what's going on with vmlinux .text:
>
> perf-OFF perf-ON
>
> ftrace-OFF
* Ingo Molnar wrote:
> Andi Kleen wrote:
>
> >textdata bss dec hex filename
> > 136926401922416 1478656 17093712104d450 obj/vmlinux
> > 129800921787544 1470464 16238100 f7c614 obj-noperf/vmlinux
[...]
> > Ok I trust you will do a better solution
* Andi Kleen wrote:
> From: Andi Kleen
>
> As suggested by Ingo.
>
> Make HW_BREAKPOINTS a config option. HW_BREAKPOINTS depends
> on perf. This allows disabling PERF_EVENTS for systems that
> don't need it (e.g. anything not used for development)
>
> Disabling PERF_EVENTS saves over 700k of
On Sat, Oct 05, 2013 at 09:08:06AM +0200, Ingo Molnar wrote:
>
> * Andi Kleen wrote:
>
> > From: Andi Kleen
> >
> > As suggested by Ingo.
>
> No, you haven't read my suggestion carefully enough so NAK.
Ok I trust you will do a better solution then to save the 700+k text.
Ball is in your cour
* Andi Kleen wrote:
> From: Andi Kleen
>
> As suggested by Ingo.
No, you haven't read my suggestion carefully enough so NAK.
Thanks,
Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo i
On Fri, Oct 04, 2013 at 02:39:48PM -0700, Andi Kleen wrote:
> From: Andi Kleen
>
> As suggested by Ingo.
>
> Make HW_BREAKPOINTS a config option. HW_BREAKPOINTS depends
> on perf. This allows disabling PERF_EVENTS for systems that
> don't need it (e.g. anything not used for development)
>
> Dis
From: Andi Kleen
As suggested by Ingo.
Make HW_BREAKPOINTS a config option. HW_BREAKPOINTS depends
on perf. This allows disabling PERF_EVENTS for systems that
don't need it (e.g. anything not used for development)
Disabling PERF_EVENTS saves over 700k of kernel text (~5% of my config)
and signi
22 matches
Mail list logo
