tracing: Restructure trace_clock_global() to never block
authorSteven Rostedt (VMware) <rostedt@goodmis.org>
Fri, 30 Apr 2021 16:17:58 +0000 (12:17 -0400)
committerSteven Rostedt (VMware) <rostedt@goodmis.org>
Fri, 30 Apr 2021 17:48:07 +0000 (13:48 -0400)
commitaafe104aa9096827a429bc1358f8260ee565b7cc
tree0602ecdabb5d1e4ce9e3fd0583c9c549ec5a494d
parent785e3c0a3a870e72dc530856136ab4c8dd207128
tracing: Restructure trace_clock_global() to never block

It was reported that a fix to the ring buffer recursion detection would
cause a hung machine when performing suspend / resume testing. The
following backtrace was extracted from debugging that case:

Call Trace:
 trace_clock_global+0x91/0xa0
 __rb_reserve_next+0x237/0x460
 ring_buffer_lock_reserve+0x12a/0x3f0
 trace_buffer_lock_reserve+0x10/0x50
 __trace_graph_return+0x1f/0x80
 trace_graph_return+0xb7/0xf0
 ? trace_clock_global+0x91/0xa0
 ftrace_return_to_handler+0x8b/0xf0
 ? pv_hash+0xa0/0xa0
 return_to_handler+0x15/0x30
 ? ftrace_graph_caller+0xa0/0xa0
 ? trace_clock_global+0x91/0xa0
 ? __rb_reserve_next+0x237/0x460
 ? ring_buffer_lock_reserve+0x12a/0x3f0
 ? trace_event_buffer_lock_reserve+0x3c/0x120
 ? trace_event_buffer_reserve+0x6b/0xc0
 ? trace_event_raw_event_device_pm_callback_start+0x125/0x2d0
 ? dpm_run_callback+0x3b/0xc0
 ? pm_ops_is_empty+0x50/0x50
 ? platform_get_irq_byname_optional+0x90/0x90
 ? trace_device_pm_callback_start+0x82/0xd0
 ? dpm_run_callback+0x49/0xc0

With the following RIP:

RIP: 0010:native_queued_spin_lock_slowpath+0x69/0x200

Since the fix to the recursion detection would allow a single recursion to
happen while tracing, this lead to the trace_clock_global() taking a spin
lock and then trying to take it again:

ring_buffer_lock_reserve() {
  trace_clock_global() {
    arch_spin_lock() {
      queued_spin_lock_slowpath() {
        /* lock taken */
        (something else gets traced by function graph tracer)
          ring_buffer_lock_reserve() {
            trace_clock_global() {
              arch_spin_lock() {
                queued_spin_lock_slowpath() {
                /* DEAD LOCK! */

Tracing should *never* block, as it can lead to strange lockups like the
above.

Restructure the trace_clock_global() code to instead of simply taking a
lock to update the recorded "prev_time" simply use it, as two events
happening on two different CPUs that calls this at the same time, really
doesn't matter which one goes first. Use a trylock to grab the lock for
updating the prev_time, and if it fails, simply try again the next time.
If it failed to be taken, that means something else is already updating
it.

Link: https://lkml.kernel.org/r/20210430121758.650b6e8a@gandalf.local.home
Cc: stable@vger.kernel.org
Tested-by: Konstantin Kharlamov <hi-angel@yandex.ru>
Tested-by: Todd Brandt <todd.e.brandt@linux.intel.com>
Fixes: b02414c8f045 ("ring-buffer: Fix recursion protection transitions between interrupt context") # started showing the problem
Fixes: 14131f2f98ac3 ("tracing: implement trace_clock_*() APIs") # where the bug happened
Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=212761
Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
kernel/trace/trace_clock.c