Messages in this thread | | | Date | Thu, 15 Jan 2015 20:34:55 +0100 | From | Oleg Nesterov <> | Subject | Re: [RFC PATCH 06/11] x86,fpu: lazily skip fpu restore with eager fpu mode, too |
| |
On 01/14, Rik van Riel wrote: > > On 01/14/2015 01:36 PM, Oleg Nesterov wrote: > > >> @@ -466,6 +462,10 @@ static inline void switch_fpu_finish(void) > >> > >> __thread_fpu_begin(tsk); > >> > >> + /* The FPU registers already have this task's FPU state. */ + > >> if (fpu_lazy_restore(tsk, raw_smp_processor_id())) + return; + > > > > Now that this is called before return to user-mode, I am not sure > > this is correct. Note that __kernel_fpu_begin() doesn't clear > > fpu_owner_task if use_eager_fpu(). > > However, __kernel_fpu_begin() does call __thread_clear_has_fpu(), > which clears the per-cpu fpu_owner variable, which is also > evaluated by fpu_lazy_restore(), so I think this is actually > correct.
Sure, but only if __thread_has_fpu().
But please ignore. My comment was confusing, sorry. What I actually tried to say is that this patch is another reason why (I think) we should start with kernel_fpu_begin/end.
If nothing else:
1. interrupted_kernel_fpu_idle() should not fail if use_eager_fpu() && !__thread_has_fpu(), otherwise your changes will introduce the performance regression.
And in fact I think that it should only fail if kernel_fpu_begin() is already in progress.
2. And in this case this_cpu_write(fpu_owner_task, NULL) can't depend on use_eager_fpu().
And in fact I think it should not depend in any case, this only adds more confusion.
Please look at the initial cleanups I sent.
Oleg.
| |