tree 800e4f23f569ef09ebcb865841943972a27f19a6
parent b35225cf1d6867438b3b67469b43482bf783d8bf
author Steven Rostedt (Red Hat) <rostedt@goodmis.org> 1389889305 -0500
committer Steven Rostedt <rostedt@goodmis.org> 1389979571 -0500

preempt: Show preempt disable stack on schedule bugs

There's been a few times where a function was called that did not have a
matching preempt_enable() for the preempt_disable().

For example, if a function deep in a call chain was called, and did a
preempt_disable() but due to an early exit of the routine, did not do a
matching preempt_enable(), it may not warn about this until later.

bar() {
  preempt_disable();
  [...]
  if (ret < 0)
     return;
  [...]
}

foo() {
  [...]
  bar();
  [...]
}

foo();
schedule();

Getting the "scheduling while atomic" warning detects the bug, but it
does not let you know where the bug happened. It may take several trials
to figure out where the missing preempt_enable() was. Some times those
functions are very deep and that makes it even more difficult to find.

The original real-time patch had a feature, written by Ingo Molnar, to trace
all the locations that disabled preemption for each process. On a bug, it
would report where the preemption was disabled.

I started with his code and modified it to be a bit simpler, as he never
meant that code to be mainline. First, as preempt disabling is really more
of a CPU state and not a process state, I made the stacks just per cpu
arrays. This eliminates any changes needed by sched.h or the task_struct.

The only times the preempt trace is needed is when scheduling while atomic
or might sleep bugs are found. This information is only printed in those
cases (not on all bug reports).

I renamed the config option from PREEMPT_TRACE to PREEMPT_STACK to prevent
confusion from the tracing facility PREEMPT_TRACER that traces the length
that preemption is disabled.

Using a struct to save the ip and parent ip together also keeps the code
a bit cleaner than using two different arrays.

Using a module that creates a kernel thread that calls a function
bad_preempt_disable() that just calls preempt_disable() with no matching
preempt_enable(), and then calls schedule, the output can be seen as the
following:

BUG: scheduling while atomic: task1/3889/0x00000002
Modules linked in: sched_bug(O) [..]
CPU: 7 PID: 3889 Comm: task1 Tainted: G           O 3.13.0-rc8-test+ #28
Hardware name: Hewlett-Packard HP Compaq Pro 6300 SFF/339A, BIOS K01 v02.05 05/07/2012
 0000000000000007 ffff8800c87a1dd8 ffffffff81627676 ffff88011ebcf5d0
 ffff8800ca0eaf00 ffff8800c87a1df8 ffffffff81623825 0000000000000002
 ffff88011ebd3600 ffff8800c87a1e78 ffffffff8162b0f5 ffff8800c87a1e28
Call Trace:
 [<ffffffff81627676>] dump_stack+0x4f/0x7c
 [<ffffffff81623825>] __schedule_bug+0x59/0x6e
 [<ffffffff8162b0f5>] __schedule+0x6b5/0x7e0
 [<ffffffff816329f1>] ? preempt_count_sub+0xb1/0x100
 [<ffffffffa06d9020>] ? bad_preempt_disable+0x20/0x20 [sched_bug]
 [<ffffffff8162b339>] schedule+0x29/0x70
 [<ffffffffa06d9038>] dead_thread+0x18/0x78 [sched_bug]
 [<ffffffff81073559>] kthread+0xc9/0xe0
 [<ffffffff81620000>] ? pci_add_new_bus+0x150/0x3b0
 [<ffffffff81073490>] ? flush_kthread_worker+0xb0/0xb0
 [<ffffffff8163676c>] ret_from_fork+0x7c/0xb0
 [<ffffffff81073490>] ? flush_kthread_worker+0xb0/0xb0
---------------------------
| preempt count: 00000002 ]
| 2-level deep critical section nesting:
----------------------------------------
.. [<ffffffffa06d9013>] .... bad_preempt_disable+0x13/0x20 [sched_bug]
.....[<ffffffffa06d9033>] ..   ( <= dead_thread+0x13/0x78 [sched_bug])
.. [<ffffffff8162aa89>] .... __schedule+0x49/0x7e0
.....[<ffffffff8162b339>] ..   ( <= schedule+0x29/0x70)

Cc: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>