Re: [RFC/PATCH] ftrace: Reduce size of function graph entries

[Namhyung Kim]

On Sat, Jun 25, 2016 at 2:29 AM, Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
> On Sat, 25 Jun 2016 01:15:34 +0900
> Namhyung Kim <namhyung@xxxxxxxxxx> wrote:
>
>> On Fri, Jun 24, 2016 at 12:04:40PM -0400, Steven Rostedt wrote:
>> > On Fri, 24 Jun 2016 15:35:44 +0900
>> > Namhyung Kim <namhyung@xxxxxxxxxx> wrote:
>> >
>> >
>> > > > > diff --git a/include/linux/ftrace.h b/include/linux/ftrace.h
>> > > > > index dea12a6e413b..35c523ba5c59 100644
>> > > > > --- a/include/linux/ftrace.h
>> > > > > +++ b/include/linux/ftrace.h
>> > > > > @@ -751,25 +751,33 @@ extern void ftrace_init(void);
>> > > > >  static inline void ftrace_init(void) { }
>> > > > >  #endif
>> > > > >
>> > > > > +#ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
>> > > > > +# define FTRACE_ALIGNMENT    4
>> > > > > +#else
>> > > > > +# define FTRACE_ALIGNMENT    8
>> > > > > +#endif
>> > > >
>>
>> As far as I can see, the ring buffer has following code in ring_buffer.c:
>>
>>   #define RB_ALIGNMENT                4U
>>   #define RB_MAX_SMALL_DATA   (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
>>   #define RB_EVNT_MIN_SIZE    8U      /* two 32bit words */
>>
>>   #ifndef CONFIG_HAVE_64BIT_ALIGNED_ACCESS
>>   # define RB_FORCE_8BYTE_ALIGNMENT   0
>>   # define RB_ARCH_ALIGNMENT          RB_ALIGNMENT
>>   #else
>>   # define RB_FORCE_8BYTE_ALIGNMENT   1
>>   # define RB_ARCH_ALIGNMENT          8U
>>   #endif
>>
>>   #define RB_ALIGN_DATA               __aligned(RB_ARCH_ALIGNMENT)
>>
>
> Right, what I meant was that we should just define FTRACE_ALIGNMENT
> unconditionally to 4. If CONFIG_HAVE_64BIT_ALIGNED_ACCESS is not set,
> it will add the buffered space regardless.
>
> You already moved "overrun", I don't see anything that would be out of
> alignment if the structure itself is aligned.

In that case if CONFIG_HAVE_64BIT_ALIGNED_ACCESS is set, the ring
buffer is 8-byte aligned but the struct is 4-byte aligned, right? Note
that the function graph tracer saves the data in a local variable (of
the struct) first and copies to the ring buffer later.  Wouldn't it be
a problem?

Thanks,
Namhyung