Messages in this thread | | | Date | Sat, 4 Oct 2008 10:40:02 +0200 | From | Ingo Molnar <> | Subject | Re: [PATCH 0/3] ring-buffer: less locking and only disable preemption |
| |
* Steven Rostedt <rostedt@goodmis.org> wrote:
> Ingo, > > These patches need to be put through the ringer. Could you add them to > your ring-buffer branch, so we can test them out before putting them > into your master branch.
hey, in fact your latest iteration already tested out so well on a wide range of boxes that i've merged it all into tip/tracing/core already.
I'll reuse tip/tracing/ring-buffer for these latest 3 patches (merge it up to tip/tracing/core and add these three patches) but it's a delta, i.e. the whole ring-buffer approach is ready for prime time i think.
Hm, do we do deallocation of the buffers already when we switch tracers?
> The following patches bring the ring buffer closer to a lockless > solution. They move the locking only to the actual moving the > tail/write pointer from one page to the next. Interrupts are now > enabled during most of the writes.
very nice direction!
> A lot of the locking protection is still within the ftrace > infrastructure. The last patch takes some of that away. > > The function tracer cannot be reentrant just due to the nature that it > traces everything, and can cause recursion issues.
Correct, and that's by far the yuckiest aspect of it. And there's another aspect: NMIs. We've still got the tip/tracing/nmisafe angle with these commits:
d979781: ftrace: mark lapic_wd_event() notrace c2c27ba: ftrace: ignore functions that cannot be kprobe-ed 431e946: ftrace: do not trace NMI contexts 1eda930: x86, tracing, nmisafe: fix threadinfo_ -> TI_ rename fallout 84c2ca9: sched: sched_clock() improvement: use in_nmi() 0d84b78: x86 NMI-safe INT3 and Page Fault a04464b: x86_64 page fault NMI-safe b335389: Change avr32 active count bit a581cbd: Change Alpha active count bit eca0999: Stringify support commas
but i'm not yet fully convinced about the NMI angle, the practical cross section to random lowlevel x86 code is wider than any sched_clock() impact for example. We might be best off avoiding it: force-disable the NMI watchdog when we trace?
Ingo
| |