On Mon, 2012-05-07 at 13:09 -0700, Suresh Siddha wrote:
On Mon, 2012-05-07 at 12:15 -0700, Linus Torvalds wrote:
> > Ugh, I hate this one.
> > 
> > We're making the exit path more expensive for almost no gain. The FPU
> > state is dead in 99.9% of all cases.
> > 
> > Why isn't this a core-dump-only case?
> > 
> 
> Today we do this unlazy_fpu() (which is what prepare_to_copy does)
> already as part of the first schedule() in the exit path. I am just
> making it explicit by calling prepare_to_copy() before exit_mm() which
> synchronizes all the threads before the first thread goes with the core
> dump.
> 
> I don't think I am adding any more cost to the existing path, unless I
> am missing something.

My quick test confirmed my understanding. For example, mainline is 
doing fpu_save_init() some 1770 times when a tsk is in the TASK_DEAD state,
during boot of a two socket system.

So based on your suggestion, made the first patch as core-dump only case and
added the third patch which clears the fpu state during exit.

thanks.

Suresh Siddha (3):
  coredump: flush the fpu exit state for proper multi-threaded core
    dump
  x86, xsave: remove thread_has_fpu() bug check in
    __sanitize_i387_state()
  x86, fpu: clear the fpu state during thread exit

 arch/x86/kernel/process.c |   19 +++++++++++++------
 arch/x86/kernel/xsave.c   |    2 --
 kernel/exit.c             |    5 +++++
 3 files changed, 18 insertions(+), 8 deletions(-)

-- 
1.7.6.5