2008-12-07 23:02:40

by Eric Lacombe

[permalink] [raw]
Subject: Re: [x86] do_arch_prctl

Hi,

I'm sorry to insist, but I really want to understand what occurs in this
portion of kernel code. And that's why I resend my previous message with the
hope that someone could enlighten my mind.

Thanks in advance,

Eric

Le lundi 24 novembre 2008 19:22:18 Jeremy Fitzhardinge, vous avez ?crit?:
> Eric Lacombe wrote:
> > Hello,
> >
> > Does the "doit case" (line 822 in ARCH_GET_FS, function do_arch_prctl)
> > exist for performance reasons? Else, why "task->thread.fs" (line 824)
> > does not contain the fs base in the "doit case"?
>
> "doit" gets set when you're operating on yourself. If you're operating
> on another process, then you need to use their task structure values
> rather than the current process's values. If you're doing it to
> yourself, then the task structure may be out of date because its only
> updated on a context switch.

The task_struct is also updated in sys_arch_prctl (ARCH_SET_FS and
ARCH_SET_GS), so not just on a context switch.
How the task structure could be out of date wrt thread.gs and thread.fs?
What could be a typical scenario that could induced gs or fs to be modified and
not thread.gs and thread.fs?

Why we have a difference between ARCH_GET_GS :

> 833 else if (doit) {
> 834 asm("movl %%gs,%0" : "=r" (gsindex));
> 835 if (gsindex)
> 836 rdmsrl(MSR_KERNEL_GS_BASE, base);
> 837 else
> 838 base = task->thread.gs;
> 839 }

and ARCH_GET_FS :

> 821 else if (doit)
> 822 rdmsrl(MSR_FS_BASE, base);

If I follow what you say, why can't we have the same optimization in
ARCH_GET_FS?

thanks,

Eric



2008-12-08 19:10:57

by Jeremy Fitzhardinge

[permalink] [raw]
Subject: Re: [x86] do_arch_prctl

Eric Lacombe wrote:
> I'm sorry to insist, but I really want to understand what occurs in this
> portion of kernel code. And that's why I resend my previous message with the
> hope that someone could enlighten my mind.
>

Well, its quite possible there are no good answers beyond "it needs a
cleanup".
> Thanks in advance,
>
> Eric
>
> Le lundi 24 novembre 2008 19:22:18 Jeremy Fitzhardinge, vous avez ?crit :
>
>> Eric Lacombe wrote:
>>
>>> Hello,
>>>
>>> Does the "doit case" (line 822 in ARCH_GET_FS, function do_arch_prctl)
>>> exist for performance reasons? Else, why "task->thread.fs" (line 824)
>>> does not contain the fs base in the "doit case"?
>>>
>> "doit" gets set when you're operating on yourself. If you're operating
>> on another process, then you need to use their task structure values
>> rather than the current process's values. If you're doing it to
>> yourself, then the task structure may be out of date because its only
>> updated on a context switch.
>>
>
> The task_struct is also updated in sys_arch_prctl (ARCH_SET_FS and
> ARCH_SET_GS), so not just on a context switch.
> How the task structure could be out of date wrt thread.gs and thread.fs?
> What could be a typical scenario that could induced gs or fs to be modified and
> not thread.gs and thread.fs?
>

Not sure. It could just be redundant.

> Why we have a difference between ARCH_GET_GS :
>
>
>> 833 else if (doit) {
>> 834 asm("movl %%gs,%0" : "=r" (gsindex));
>> 835 if (gsindex)
>> 836 rdmsrl(MSR_KERNEL_GS_BASE, base);
>> 837 else
>> 838 base = task->thread.gs;
>> 839 }
>>
>
> and ARCH_GET_FS :
>
>
>> 821 else if (doit)
>> 822 rdmsrl(MSR_FS_BASE, base);
>>
>
> If I follow what you say, why can't we have the same optimization in
> ARCH_GET_FS?
>

I haven't looked into it very closely, but its possible the asymmetry
comes from the fact that there's no swapfs, and so the kernel and
userspace aren't sharing %fs.

J

2008-12-08 20:35:33

by Andi Kleen

[permalink] [raw]
Subject: Re: [x86] do_arch_prctl

Jeremy Fitzhardinge <[email protected]> writes:
>>
>> The task_struct is also updated in sys_arch_prctl (ARCH_SET_FS and
>> ARCH_SET_GS), so not just on a context switch.
>> How the task structure could be out of date wrt thread.gs and thread.fs?
>> What could be a typical scenario that could induced gs or fs to be
>> modified and not thread.gs and thread.fs?
>>
>
> Not sure. It could just be redundant.

The process has just changed fs/gs, but not done a context switch yet.

-Andi

--
[email protected]