On Mon, Aug 6, 2018 at 7:05 AM, Masami Hiramatsu <mhira...@kernel.org> wrote: > On Sun, 5 Aug 2018 20:40:49 -0700 > "Joel Fernandes (Google)" <j...@joelfernandes.org> wrote: > >> Recently we tried to make the preemptirqsoff tracer to use irqsoff >> tracepoint probes. However this causes issues as reported by Masami: >> >> [2.271078] Testing tracer preemptirqsoff: .. no entries found ..FAILED! >> [2.381015] WARNING: CPU: 0 PID: 1 at /home/mhiramat/ksrc/linux/kernel/ >> trace/trace.c:1512 run_tracer_selftest+0xf3/0x154 >> >> This is due to the tracepoint code increasing the preempt nesting count >> by calling an additional preempt_disable before calling into the >> preemptoff tracer which messes up the preempt_count() check in >> tracer_hardirqs_off. >> >> To fix this, make the irqsoff tracer probes balance the additional outer >> preempt_disable with a preempt_enable_notrace. > > I've tested it and ensured this fixes the problem. > > Tested-by: Masami Hiramatsu <mhira...@kernel.org>
Thanks! >> The other way to fix this is to just use SRCU for all tracepoints. >> However we can't do that because we can't use NMIs from RCU context. >> >> Fixes: c3bc8fd637a9 ("tracing: Centralize preemptirq tracepoints >> and unify their usage") >> Fixes: e6753f23d961 ("tracepoint: Make rcuidle tracepoint callers use SRCU") >> Reported-by: Masami Hiramatsu <mhira...@kernel.org> >> Signed-off-by: Joel Fernandes (Google) <j...@joelfernandes.org> >> --- >> kernel/trace/trace_irqsoff.c | 26 ++++++++++++++++++++++++++ >> 1 file changed, 26 insertions(+) >> >> diff --git a/kernel/trace/trace_irqsoff.c b/kernel/trace/trace_irqsoff.c >> index 770cd30cda40..ffbf1505d5bc 100644 >> --- a/kernel/trace/trace_irqsoff.c >> +++ b/kernel/trace/trace_irqsoff.c >> @@ -603,14 +603,40 @@ static void irqsoff_tracer_stop(struct trace_array *tr) >> */ >> static void tracer_hardirqs_on(void *none, unsigned long a0, unsigned long >> a1) >> { > > To ensure this function must not be preempted even if we increment preempt > count, I think you should check irq_disabled() whole this process, put below > here. > > if (unlikely(!irq_disabled())) > return; > > Since irq_disabled() will be checked in irq_trace() anyway, so no problem > to return here when !irq_disabled(). IRQs can never be enabled here. The trace hooks are called only after disabling interrupts, or before enabling them. Right? thanks, - Joel