On 1/17/18 4:21 AM, Thomas Gleixner wrote:
On Thu, 4 Jan 2018, Yang Shi wrote:
There are nested loops on debug objects free path, sometimes it may take
over hundred thousands of loops, then cause soft lockup with !CONFIG_PREEMPT
occasionally, like below:
Please trim back traces. The whole module info and whatever is completely
irrelevant.
Yes, sure.
@@ -768,6 +771,10 @@ static void __debug_check_no_obj_freed(const void
*address, unsigned long size)
debug_objects_maxchain = cnt;
max_loops += cnt;
+
+ if (max_loops > 10000 && ((max_loops % 10000) == 0)
+ && suppress_lockup != 0)
+ touch_softlockup_watchdog();
This is voodoo programming.
There are two things which can be done here:
1) The collected objects can be put on a global free list and work
scheduled to free them piecewise.
I don't get your point here. objects free has already been done in a
work. free_object() -> schedule_work()
Do you mean free all of them out of the for loop in a batch? Then don't
call free_object() in the for loop?
2) We can do a cond_resched() if not in atomic context and interrupts are
enabled.
I did try this before I went with touching softlockup watchdog approach.
The problem is in_atomic() can't tell if it is in atomic context on
non-preempt kernel. For preempt kernel, it is easy.
Thanks,
Yang
I rather prefer to make that stuff robust than having crystal ball
constants and magic debugfs knobs.
Thanks,
tglx