Re: [PATCH 5/5] tracing: reset iterator in t_start()

[Wang Liming <liming.wang@windriver.com>]

Li Zefan wrote:
>> Another version:
>> Since we have saved current (struct tracer *) in m->private in .next, in
>> .start, we don't need to call .next to find the one that should be
>> printed in 2nd or nth time.
>>
> 
> I don't like this for 2 reasons.
> 
> 1. It's strange that @pos is not incremented in next().
Yes, it's strang, but we know that @pos sometimes is not necessary, such in this 
position.

> 
> 2. 
>    t_stop()
>      mutex_unlock()
>                           unregister_tracer(t)
>    t_start()
>      mutex_lock()
>        t = m->private
>        ...
>        t = t-next.
> 
>   We access t->next though @t was unregistered. This is not
>   good, though it does no harm here.
OK, it's a realy race problem if we call unregister_tracer.
btw: who realy calls this function? :)

Liming Wang
> 
>> diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
>> index cae34c6..02cdccc 100644
>> --- a/kernel/trace/trace.c
>> +++ b/kernel/trace/trace.c
>> @@ -2055,8 +2055,6 @@ t_next(struct seq_file *m, void *v, loff_t *pos)
>>  {
>>         struct tracer *t = m->private;
>>
>> -       (*pos)++;
>> -
>>         if (t)
>>                 t = t->next;
>>
>> @@ -2068,11 +2066,8 @@ t_next(struct seq_file *m, void *v, loff_t *pos)
>>  static void *t_start(struct seq_file *m, loff_t *pos)
>>  {
>>         struct tracer *t = m->private;
>> -       loff_t l = 0;
>>
>>         mutex_lock(&trace_types_lock);
>> -       for (; t && l < *pos; t = t_next(m, t, &l))
>> -               ;
>>
>>         return t;
>>  }
>>
>>
>