Re: Re: [BUG] 3.13-rc1 kernel crash when enable all tracepoints

[Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>]

(2013/11/26 17:57), Masami Hiramatsu wrote:
> (2013/11/26 16:20), Jovi Zhangwei wrote:
>> I'm not sure this relate with ftrace:function event, from the
>> crash log, it seems more like caused by lock events.
>>
>> Now I only enable lock events (lockdep compiled in my box), below
>> two commands both can crash system.
> 
> That is not the lockdep, but the lockstat. :)
> 
>>
>> #echo 1 > /sys/kernel/debug/tracing/events/lock/lock_acquire/enable
>> or.
>> #echo 1 > /sys/kernel/debug/tracing/events/lock/lock_release/enable
>>
>> According to the log, it seems like a recursion tracing bug.
>>
>> register lock event -> jump_label_update -> text_poke_bp ->
>> on_each_cpu -> local_apic_timer_interrupt -> ktime_get_update_offsets
>> -> lock_release
> 
> OK, the problem is in the lock_acquire/release events.

Hm, it seems that the bug is caused because lock->name is NULL in
ftrace_raw_event_lock(_acquire). To allocate a buffer on ring_buffer,
ftrace_raw_event_##call tries to measure the length of lock->name
but it causes a memory access violation in ftrace_get_offsets_##call
because lock->name is NULL. :(

I could fix this by checking lock->name in include/trace/events/lock.h

e.g.
        TP_STRUCT__entry(
                __field(unsigned int, flags)
                __string(name, lock->name ?: "NULL")
                __field(void *, lockdep_addr)
        ),

        TP_fast_assign(
                __entry->flags = (trylock ? 1 : 0) | (read ? 2 : 0);
                __assign_str(name, lock->name ?: "NULL");
                __entry->lockdep_addr = lock;
        ),

Thank you,

-- 
Masami HIRAMATSU
IT Management Research Dept. Linux Technology Center
Hitachi, Ltd., Yokohama Research Laboratory
E-mail: masami.hiramatsu.pt@hitachi.com