lkml.org 
[lkml]   [2015]   [Aug]   [28]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 11/11] x86, fpu: check CPU-provided sizes against struct declarations
On 08/27/2015 10:25 PM, Ingo Molnar wrote:
> * Dave Hansen <dave@sr71.net> wrote:
>> @@ -447,6 +492,14 @@ static void do_extra_xstate_size_checks(
>> paranoid_xstate_size += xfeature_size(i);
>> }
>> XSTATE_WARN_ON(paranoid_xstate_size != xstate_size);
>> + /*
>> + * Basically, make sure that XSTATE_RESERVE has forced
>> + * xregs_state to be large enough. This is not fatal
>> + * because we reserve a *lot* of extra room in the init
>> + * task struct, but we should at least know we got it
>> + * wrong.
>> + */
>> + XSTATE_WARN_ON(xstate_size > sizeof(struct xregs_state));
>
> So do we need to warn about this? arch_task_struct_size is already dynamic today.

I'm unsure what _actually_ blew up, but I missed adding protection keys
and AVX-512 to XSTATE_RESERVE and the kernel crashed the first time I
did a non-init-state-PKRU XSAVE.

> The only problem would be the init task, which is allocated statically - can we
> fix that?

We could theoretically make it dynamic, but I'm really not sure it's
worth the trouble. I just removed the init_task=INIT_TASK()
initialization to see what would happen and something blew up early
(last I saw on the console was the "early console in setup code").

The current size of the non-XSAVE data in task_struct is ~2k. The xsave
data is 800-something bytes, so say ~1k. Our init_task ends up being
~6k, 3k of which is wasted. On an AVX-512 CPU, that means 1k of waste.

From how early things died, I'm going to go out on a limb and say that
we'll need to bootmem alloc our new dynamic init_task and probably can't
practically wait for the slab to show up. Bootmem can only do full
pages, so our 6k can be trimmed to 4k. On an AVX-512 CPU, the 6k goes
*up* to 8k.

It doesn't look like a fun exercise for 2k of memory savings.




\
 
 \ /
  Last update: 2015-08-28 18:21    [W:0.063 / U:0.860 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site