lkml.org 
[lkml]   [2017]   [Nov]   [30]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH RT] crypto: limit more FPU-enabled sections
On 2017-11-30 10:29:48 [-0500], Steven Rostedt wrote:
> > +static void camellia_fpu_end_rt(struct crypt_priv *ctx)
> > +{
> > +#if CONFIG_PREEMPT_RT_FULL
> > + bool fpu_enabled = ctx->fpu_enabled;
> > +
> > + if (!fpu_enabled)
> > + return;
> > + camellia_fpu_end(fpu_enabled);
> > + ctx->fpu_enabled = false;
> > +#endif
> > +}
> > +
> > +static void camellia_fpu_sched_rt(struct crypt_priv *ctx)
> > +{
> > +#if CONFIG_PREEMPT_RT_FULL
> > + bool fpu_enabled = ctx->fpu_enabled;
> > +
> > + if (!fpu_enabled || !tif_need_resched_now())
> > + return;
> > + camellia_fpu_end(fpu_enabled);
>
> I haven't looked deeply, but why does this call the camellia_fpu_end()
> but other *_fpu_sched_rt() do not call the equivalent?

There is lrw_encrypt() which as "end" function after the end of the full
operation. So encrypt_callback() might be invoked multiple times but
only the first invocation does the fpu-enable part.

#1 We need to disable-FTP after one invocation of encrypt_callback()
because blkcipher_walk_done() might alloc/free/map memory.

#2 That camellia_fpu_sched_rt() is only RT specific in order to enable
preempt if need-resched is set for a RT task. For !RT we continue.

> > + kernel_fpu_end();
> > + /* schedule due to preemptible */
> > + kernel_fpu_begin();
> > +#endif
> > +}
> > +
>
> These are duplicate functions. Shouldn't they go into a header file?
>
> Also, they are very similar:

That whole thing is kind of ugly, yes. That *_fpu_end_rt() use a struct
crypt_priv but this is defined in each cipher file.

I would keep it for now until we decide if we keep fixing those things
are disable for RT because it is not worth it. As I said, this is only
x86 specific and we have something similar on ARM with NEON.
But I guess the decision here will be postponed until someone posts
numbers with SW and with this workaround and cyclictest with a cycle of
1 - 10 ms.

Sebastian

\
 
 \ /
  Last update: 2017-11-30 16:42    [W:2.262 / U:0.032 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site