lkml.org 
[lkml]   [2007]   [Jan]   [12]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 05/05] Linux Kernel Markers, non optimised architectures
* Nick Piggin (nickpiggin@yahoo.com.au) wrote:
> Mathieu Desnoyers wrote:
> >* Nick Piggin (nickpiggin@yahoo.com.au) wrote:
> >
> >>Mathieu Desnoyers wrote:
> >>
> >>
> >>>+#define MARK(name, format, args...) \
> >>>+ do { \
> >>>+ static marker_probe_func *__mark_call_##name = \
> >>>+ __mark_empty_function; \
> >>>+ volatile static char __marker_enable_##name = 0; \
> >>>+ static const struct __mark_marker_c __mark_c_##name \
> >>>+ __attribute__((section(".markers.c"))) = \
> >>>+ { #name, &__mark_call_##name, format } ; \
> >>>+ static const struct __mark_marker __mark_##name \
> >>>+ __attribute__((section(".markers"))) = \
> >>>+ { &__mark_c_##name, &__marker_enable_##name } ; \
> >>>+ asm volatile ( "" : : "i" (&__mark_##name)); \
> >>>+ __mark_check_format(format, ## args); \
> >>>+ if (unlikely(__marker_enable_##name)) { \
> >>>+ preempt_disable(); \
> >>>+ (*__mark_call_##name)(format, ## args); \
> >>>+ preempt_enable_no_resched(); \
> >>
> >>Why not just preempt_enable() here?
> >>
> >
> >
> >Because the preempt_enable() macro contains preempt_check_resched(), which
> >may call preempt_schedule() which leads us to a call to schedule().
> >Therefore,
> >all those very interesting scheduler functions would cause an infinite
> >recursive scheduler call if we marked schedule() and used preempt_enable()
> >in
> >the marker.
>
> The vast majority of schedule() has preempt turned off, so that shouldn't
> be a problem, if you provide a comment.
>
> >The primary goal for the markers (and the probes that attaches to them) is
> >to
> >have the fewest side-effects possible : any kernel method called from an
> >instrumentation site adds this precise kernel method to the "cannot be
> >instrumented" list, which I want to keep as small possible.
>
> OK, well one problem is that it can cause a resched event to be lost, so
> you might say it has more side-effects without checking resched.
>

I agree : this a side-effect I pointed out in my LTTng presentation last
summer at OLS.

Here is a quick idea of the potentially problematic instrumentation points
(i386 example) :

- with the task_rq_lock held (therefore preemption is disabled, so it's not a
problem)
sched.c wait_task_inactive()
sched.c try_to_wake_up()
sched.c wake_up_new_task()
sched.c sched_migrate_task()

sched.c schedule() after prepare_task_switch call, before context_switch call.
Surrounded by preempt_disable(), preempt_enable_no_resched(), should be ok.

- IRQs : irq_enter()/irq_exit() calls in do_IRQ makes sure that the
preempt_count is incremented. irq_enter() is called with interrupts still
disabled.
kernel/irq/handle.c handle_IRQ_event()

- NMIs : nmi_enter() -> irq_enter() -> add_preempt_count(HARDIRQ_OFFSET) called
with interrupts still disabled.
Therefore, preemption is disabled within trace points in do_nmi.

- traps : GPF, do_trap, do_page_fault, do_debug, spurious_interrupt,
math_emulate.
It is not uncommon for these trap handlers to reenable interrupts very soon.
They do not increment the preemption count.
Therefore, preemption must be expected when these handlers run : we cannot
rely of the fact that hard IRQs would be disabled to prevent the scheduler
from running, as markers becomes a new source of scheduler events.

- local_irq_enable()/local_irq_disable() :
It can call trace_hardirqs_on()/trace_hardirqs_off(). These macros are
sprinkled in _every_ possible context cited above, from trap handlers to
preemptible code.

Other contexts or code location are not a problem (process context, softirq).

If we are sure that we expect calls to preempt_schedule() from each of these
contexts, then it's ok to put preempt_enable(). It is important to note that a
marker would then act as a source of scheduler events in code paths where
disabling interrupts is expected to disable the scheduler.

Mathieu

--
OpenPGP public key: http://krystal.dyndns.org:8080/key/compudj.gpg
Key fingerprint: 8CD5 52C3 8E3C 4140 715F BA06 3F25 A8FE 3BAE 9A68
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2007-01-12 19:23    [W:0.051 / U:0.248 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site