From 03889384cee7a198a79447c1ea6aca2c8e54d155 Mon Sep 17 00:00:00 2001 From: Steven Rostedt Date: Fri, 11 Dec 2009 09:48:22 -0500 Subject: [PATCH 1/2] tracing: Add trace_dump_stack() I've been asked a few times about how to find out what is calling some location in the kernel. One way is to use dynamic function tracing and implement the func_stack_trace. But this only finds out who is calling a particular function. It does not tell you who is calling that function and entering a specific if conditional. I have myself implemented a quick version of trace_dump_stack() for this purpose a few times, and just needed it now. This is when I realized that this would be a good tool to have in the kernel like trace_printk(). Using trace_dump_stack() is similar to dump_stack() except that it writes to the trace buffer instead and can be used in critical locations. For example: @@ -5485,8 +5485,12 @@ need_resched_nonpreemptible: if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { if (unlikely(signal_pending_state(prev->state, prev))) prev->state = TASK_RUNNING; - else + else { deactivate_task(rq, prev, 1); + trace_printk("Deactivating task %s:%d\n", + prev->comm, prev->pid); + trace_dump_stack(); + } switch_count = &prev->nvcsw; } Produces: <...>-3249 [001] 296.105269: schedule: Deactivating task ntpd:3249 <...>-3249 [001] 296.105270: => schedule => schedule_hrtimeout_range => poll_schedule_timeout => do_select => core_sys_select => sys_select => system_call_fastpath Signed-off-by: Steven Rostedt --- include/linux/kernel.h | 2 ++ kernel/trace/trace.c | 16 ++++++++++++++++ 2 files changed, 18 insertions(+) diff --git a/include/linux/kernel.h b/include/linux/kernel.h index 3fa4c590cf1..5ad4199fb07 100644 --- a/include/linux/kernel.h +++ b/include/linux/kernel.h @@ -492,6 +492,8 @@ extern int __trace_printk(unsigned long ip, const char *fmt, ...) __attribute__ ((format (printf, 2, 3))); +extern void trace_dump_stack(void); + /* * The double __builtin_constant_p is because gcc will give us an error * if we try to allocate the static variable to fmt if it is not a diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c index 88bd9ae2a9e..f531301b7a3 100644 --- a/kernel/trace/trace.c +++ b/kernel/trace/trace.c @@ -1151,6 +1151,22 @@ void __trace_stack(struct trace_array *tr, unsigned long flags, int skip, __ftrace_trace_stack(tr->buffer, flags, skip, pc); } +/** + * trace_dump_stack - record a stack back trace in the trace buffer + */ +void trace_dump_stack(void) +{ + unsigned long flags; + + if (tracing_disabled || tracing_selftest_running) + return 0; + + local_save_flags(flags); + + /* skipping 3 traces, seems to get us at the caller of this function */ + __ftrace_trace_stack(global_trace.buffer, flags, 3, preempt_count()); +} + void ftrace_trace_userstack(struct ring_buffer *buffer, unsigned long flags, int pc) { From cc51a0fca66658ea710db566ba17e80e3f7d4957 Mon Sep 17 00:00:00 2001 From: Steven Rostedt Date: Fri, 11 Dec 2009 11:54:51 -0500 Subject: [PATCH 2/2] tracing: Add stack trace to irqsoff tracer The irqsoff and friends tracers help in finding causes of latency in the kernel. The also work with the function tracer to show what was happening when interrupts or preemption are disabled. But the function tracer has a bit of an overhead and can cause exagerated readings. Currently, when tracing with /proc/sys/kernel/ftrace_enabled = 0, where the function tracer is disabled, the information that is provided can end up being useless. For example, a 2 and a half millisecond latency only showed: # tracer: preemptirqsoff # # preemptirqsoff latency trace v1.1.5 on 2.6.32 # -------------------------------------------------------------------- # latency: 2463 us, #4/4, CPU#2 | (M:preempt VP:0, KP:0, SP:0 HP:0 #P:4) # ----------------- # | task: -4242 (uid:0 nice:0 policy:0 rt_prio:0) # ----------------- # => started at: _spin_lock_irqsave # => ended at: remove_wait_queue # # # _------=> CPU# # / _-----=> irqs-off # | / _----=> need-resched # || / _---=> hardirq/softirq # ||| / _--=> preempt-depth # |||| /_--=> lock-depth # |||||/ delay # cmd pid |||||| time | caller # \ / |||||| \ | / hackbenc-4242 2d.... 0us!: trace_hardirqs_off <-_spin_lock_irqsave hackbenc-4242 2...1. 2463us+: _spin_unlock_irqrestore <-remove_wait_queue hackbenc-4242 2...1. 2466us : trace_preempt_on <-remove_wait_queue The above lets us know that hackbench with pid 2463 grabbed a spin lock somewhere and enabled preemption at remove_wait_queue. This helps a little but where this actually happened is not informative. This patch adds the stack dump to the end of the irqsoff tracer. This provides the following output: hackbenc-4242 2d.... 0us!: trace_hardirqs_off <-_spin_lock_irqsave hackbenc-4242 2...1. 2463us+: _spin_unlock_irqrestore <-remove_wait_queue hackbenc-4242 2...1. 2466us : trace_preempt_on <-remove_wait_queue hackbenc-4242 2...1. 2467us : => sub_preempt_count => _spin_unlock_irqrestore => remove_wait_queue => free_poll_entry => poll_freewait => do_sys_poll => sys_poll => system_call_fastpath Now we see that the culprit of this latency was the free_poll_entry code. Signed-off-by: Steven Rostedt --- kernel/trace/trace_irqsoff.c | 2 ++ 1 file changed, 2 insertions(+) diff --git a/kernel/trace/trace_irqsoff.c b/kernel/trace/trace_irqsoff.c index 3aa7eaa2114..2974bc7538c 100644 --- a/kernel/trace/trace_irqsoff.c +++ b/kernel/trace/trace_irqsoff.c @@ -151,6 +151,8 @@ check_critical_timing(struct trace_array *tr, goto out_unlock; trace_function(tr, CALLER_ADDR0, parent_ip, flags, pc); + /* Skip 5 functions to get to the irq/preempt enable function */ + __trace_stack(tr, flags, 5, pc); if (data->critical_sequence != max_sequence) goto out_unlock;