[PATCH v2 0/2] sched/tracing: sched_switch prev_state reported as TASK_RUNNING when it's not

[Valentin Schneider <valentin.schneider@arm.com>]

Hi folks,

Problem
=======

Abhijeet pointed out that the following sequence of trace events can be
observed:

  cat-1676  [001] d..3 103.010411: sched_waking: comm=grep pid=1677 prio=120 target_cpu=004
  grep-1677 [004] d..2 103.010440: sched_switch: prev_comm=grep prev_pid=1677 prev_prio=120 prev_state=R 0x0 ==> next_comm=swapper/4 next_pid=0 next_prio=120
  <idle>-0  [004] dNh3 103.0100459: sched_wakeup: comm=grep pid=1677 prio=120 target_cpu=004

IOW, not-yet-woken task gets presented as runnable and switched out in
favor of the idle task... Dietmar and I had a look, turns out this sequence
can happen: 

		      p->__state = TASK_INTERRUPTIBLE;
		      __schedule()
			deactivate_task(p);
  ttwu()
    READ !p->on_rq
    p->__state=TASK_WAKING
			trace_sched_switch()
			  __trace_sched_switch_state()
			    task_state_index()
			      return 0;

TASK_WAKING isn't in TASK_REPORT, hence why task_state_index(p) returns 0.
This can happen as of commit c6e7bd7afaeb ("sched/core: Optimize ttwu()
spinning on p->on_cpu") which punted the TASK_WAKING write to the other
side of

  smp_cond_load_acquire(&p->on_cpu, !VAL);

in ttwu().

Uwe reported on #linux-rt what I think is a similar issue with
TASK_RTLOCK_WAIT on PREEMPT_RT; again that state isn't in TASK_REPORT so
you get prev_state=0 despite the task blocking on a lock.

Both of those are very confusing for tooling or even human observers.

Patches
=======

For the TASK_WAKING case, pushing the prev_state read in __schedule() down
to __trace_sched_switch_state() seems sufficient. That's patch 1.

For TASK_RTLOCK_WAIT, it's a bit trickier. One could use ->saved_state as
prev_state, but
a) that is also susceptible to racing (see ttwu() / ttwu_state_match())
b) some lock substitutions (e.g. rt_spin_lock()) leave ->saved_state as
   TASK_RUNNING.

Patch 2 adds TASK_RTLOCK_WAIT to TASK_REPORT instead.

Testing
=======

Briefly tested on an Arm Juno via ftrace+hackbench against
o tip/sched/core@82762d2af31a
o v5.16-rt15-rebase w/ CONFIG_PREEMPT_RT


I also had a look at the __schedule() disassembly as suggested by Peter; on x86
prev_state gets pushed to the stack and popped prior to the trace event static
call, the rest seems mostly unchanged (GCC 9.3).

Revisions
=========

RFC v1 -> v2
++++++++++++

o Collected tags (Steven, Sebastian)
o Patched missing trace_switch event clients (Steven)

Cheers,
Valentin

Valentin Schneider (2):
  sched/tracing: Don't re-read p->state when emitting sched_switch event
  sched/tracing: Add TASK_RTLOCK_WAIT to TASK_REPORT

 fs/proc/array.c                   |  3 ++-
 include/linux/sched.h             | 28 +++++++++++++++++-----------
 include/trace/events/sched.h      | 12 ++++++++----
 kernel/sched/core.c               |  4 ++--
 kernel/trace/fgraph.c             |  4 +++-
 kernel/trace/ftrace.c             |  4 +++-
 kernel/trace/trace_events.c       |  8 ++++++--
 kernel/trace/trace_osnoise.c      |  4 +++-
 kernel/trace/trace_sched_switch.c |  1 +
 kernel/trace/trace_sched_wakeup.c |  1 +
 10 files changed, 46 insertions(+), 23 deletions(-)

-- 
2.25.1