lkml.org 
[lkml]   [2015]   [Jan]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
Subject[PATCH 0/3] x86, fpu: more eagerfpu cleanups
On 01/15, Oleg Nesterov wrote:
>
> Let me send initial kernel_fpu_begin/end cleanups, I believe they make
> sense anyway and won't conflict with your changes.
>
> This is actually resend, I sent more patches some time ago but they were
> ignored.
>
> Note that (I hope) we can do more changes on top of this series, in
> particular:
>
> - remove all checks from irq_fpu_usable() except in_kernel_fpu
>
> - do not abuse FPU in kernel threads, this makes sense even if
> use_eager_fpu(), and with or without the changes you proposed.

On top of this series.

Initially I was going to make more changes, but then I decided to delay
the cleanups. IMHO this code needs them in any case. math_state_restore()
and its usage doesn't look nice. init_fpu() too, and unlazy_fpu(current)
is simply wrong afaics. Fortunately the only caller of init_fpu(current)
is coredump, so this task can't return to user-mode, still this doesn't
look good. And it should be unified with save_init_fpu(). Which has the
wrong WARN_ON_ONCE(!__thread_has_fpu(tsk)).

And I am not sure that unlazy_fpu() is correct wrt __kernel_fpu_begin(),
but probably this is because I do not know how fpu works. If the nested
__save_init_fpu() is fine, then why (before the previous changes)
__kernel_fpu_begin() does __thread_clear_has_fpu() first?

Rik, to remind, I think that your changes need 1 + 2 at least, to avoid
the performance regression. Perhaps this needs a single patch.

3/3 is not strictly neccessary, but imo it makes sense anyway, even
without your changes. And if we add TIF_LOAD_FPU, it would be nice to
filter out kthreads automatically.


Could someone review this series?

If this makes any sense, I'll try to make the cleanups later.

Oleg.



\
 
 \ /
  Last update: 2015-01-19 20:01    [W:0.347 / U:0.248 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site