On Mon, 2009-04-20 at 15:10 +0200, Ralf Baechle wrote:
> On Mon, Apr 20, 2009 at 10:01:55AM +0200, Ralf Baechle wrote:
>
> > > I have divided ftrace to several commits in the above git tree, hope you
> > > can check it, thx :-)
> > >
> > > in addition to the static/dynamic/graph function tracer & system call
> > > tracer implementation, a mips specific ring_buffer_time_stamp
> > > (kernel/trace/ring_buffer.c) is also implemented to get 1us precision
> > > time, this is very important to make ftrace available in mips,
> > > otherwise, we can only get 1ms precision time for the original
> > > ring_buffer_time_stamp is based on sched_clock(jiffies based).
> > >
> > > perhaps we can implement a more precise sched_clock directly, just as
> > > x86 does(native_sched_clock, tsc based), but in mips, there is only a
> > > 32bit timer count which will quickly overflow, so it will need an extra
> > > overflow protection, which may influence the other parts of the kernel.
> >
> > My git clone is still running to I'm commenting only on the patches you
> > posted earlier. #ifdef-MIPS'ing things into the generic kernel code
> > definately won't be an acceptable way to get µs resolution.
>
just fix it via moving mips specific ring_buffer_time_stamp to
arch/mips/kernel/ftrace.c via defining the original one as
__attribute__(weak).
in arch/mips/kernel/ftrace.c:
u64 native_ring_buffer_time_stamp(...) {
...
}
u64 ring_buffer_time_stamp(...)
__attribute__((alias("native_ring_buffer_time_stamp")));
in kernel/trace/ring_buffer.c:
-u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
+ u64 __attribute__((weak)) ring_buffer_time_stamp(struct ring_buffer
*buffer, int cpu)
*** not push to the git tree yet.
> In changeset e67f78d663a84ef93aa12c3c8c1adf3033c4f9a1 you introduce
> <asm/rwsem.h> but because RWSEM_GENERIC_SPINLOCK is always set that file
> won't ever be included.
>
Yes, currently, I'm trying to fix it.
> Ralf
