Patchwork [RFC,qom-cpu] linux-user: Avoid conditional cpu_reset()

login
register
mail settings
Submitter Andreas Färber
Date July 10, 2013, 4:30 p.m.
Message ID <1373473838-1163-1-git-send-email-afaerber@suse.de>
Download mbox | patch
Permalink /patch/258161/
State New
Headers show

Comments

Andreas Färber - July 10, 2013, 4:30 p.m.
Some CPUs reset as part of cpu_init(), some others were reset
afterwards, some not at all. While some targets didn't implement a
cpu_[state_]reset() function, QOM cpu_reset() is always available.
There's nothing wrong with resetting twice on startup, so drop
the #ifdef.

Suggested-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Andreas Färber <afaerber@suse.de>
Cc: Eduardo Habkost <ehabkost@redhat.com>
---
 This had been discussed as a possible cleanup for the #ifdef.
 I am uncertain whether we should do this since it hides the TODO item 
 of investigating ppc and sparc CPU reset.

 linux-user/main.c | 2 --
 1 file changed, 2 deletions(-)
Eduardo Habkost - July 10, 2013, 6:43 p.m.
On Wed, Jul 10, 2013 at 06:30:38PM +0200, Andreas Färber wrote:
> Some CPUs reset as part of cpu_init(), some others were reset
> afterwards, some not at all. While some targets didn't implement a
> cpu_[state_]reset() function, QOM cpu_reset() is always available.
> There's nothing wrong with resetting twice on startup, so drop
> the #ifdef.
> 
> Suggested-by: Peter Maydell <peter.maydell@linaro.org>
> Signed-off-by: Andreas Färber <afaerber@suse.de>
> Cc: Eduardo Habkost <ehabkost@redhat.com>
> ---
>  This had been discussed as a possible cleanup for the #ifdef.
>  I am uncertain whether we should do this since it hides the TODO item 
>  of investigating ppc and sparc CPU reset.
> 
>  linux-user/main.c | 2 --
>  1 file changed, 2 deletions(-)
> 
> diff --git a/linux-user/main.c b/linux-user/main.c
> index 7f15d3d..e904d8c 100644
> --- a/linux-user/main.c
> +++ b/linux-user/main.c
> @@ -3637,9 +3637,7 @@ int main(int argc, char **argv, char **envp)
>          fprintf(stderr, "Unable to find CPU definition\n");
>          exit(1);
>      }
> -#if defined(TARGET_SPARC) || defined(TARGET_PPC)
>      cpu_reset(ENV_GET_CPU(env));
> -#endif

Most of the cpu_reset() implementations I have looked at (including
sparc) contain something like:

    memset(env, 0, offsetof(CPUXXXState, breakpoints));

Isn't this clearing userspace registers that are not supposed to be
touched by clone()?
Peter Maydell - July 10, 2013, 6:51 p.m.
On 10 July 2013 19:43, Eduardo Habkost <ehabkost@redhat.com> wrote:
> On Wed, Jul 10, 2013 at 06:30:38PM +0200, Andreas Färber wrote:
>> Some CPUs reset as part of cpu_init(), some others were reset
>> afterwards, some not at all. While some targets didn't implement a
>> cpu_[state_]reset() function, QOM cpu_reset() is always available.
>> There's nothing wrong with resetting twice on startup, so drop
>> the #ifdef.
>>
>> Suggested-by: Peter Maydell <peter.maydell@linaro.org>
>> Signed-off-by: Andreas Färber <afaerber@suse.de>
>> Cc: Eduardo Habkost <ehabkost@redhat.com>
>> ---
>>  This had been discussed as a possible cleanup for the #ifdef.
>>  I am uncertain whether we should do this since it hides the TODO item
>>  of investigating ppc and sparc CPU reset.
>>
>>  linux-user/main.c | 2 --
>>  1 file changed, 2 deletions(-)
>>
>> diff --git a/linux-user/main.c b/linux-user/main.c
>> index 7f15d3d..e904d8c 100644
>> --- a/linux-user/main.c
>> +++ b/linux-user/main.c
>> @@ -3637,9 +3637,7 @@ int main(int argc, char **argv, char **envp)
>>          fprintf(stderr, "Unable to find CPU definition\n");
>>          exit(1);
>>      }
>> -#if defined(TARGET_SPARC) || defined(TARGET_PPC)
>>      cpu_reset(ENV_GET_CPU(env));
>> -#endif
>
> Most of the cpu_reset() implementations I have looked at (including
> sparc) contain something like:
>
>     memset(env, 0, offsetof(CPUXXXState, breakpoints));
>
> Isn't this clearing userspace registers that are not supposed to be
> touched by clone()?

You're thinking about the other cpu_reset() -- this one is in
main.c and happens only for the main thread, immediately
after we've created that thread's CPU, and before we set up
its registers for initial program start. The cpu_reset()
that sometimes happens in the clone() path is in syscall.c
(and is addressed by patch
http://patchwork.ozlabs.org/patch/257232/ )

-- PMM
Eduardo Habkost - July 10, 2013, 7:02 p.m.
On Wed, Jul 10, 2013 at 07:51:32PM +0100, Peter Maydell wrote:
> On 10 July 2013 19:43, Eduardo Habkost <ehabkost@redhat.com> wrote:
> > On Wed, Jul 10, 2013 at 06:30:38PM +0200, Andreas Färber wrote:
> >> Some CPUs reset as part of cpu_init(), some others were reset
> >> afterwards, some not at all. While some targets didn't implement a
> >> cpu_[state_]reset() function, QOM cpu_reset() is always available.
> >> There's nothing wrong with resetting twice on startup, so drop
> >> the #ifdef.
> >>
> >> Suggested-by: Peter Maydell <peter.maydell@linaro.org>
> >> Signed-off-by: Andreas Färber <afaerber@suse.de>
> >> Cc: Eduardo Habkost <ehabkost@redhat.com>
> >> ---
> >>  This had been discussed as a possible cleanup for the #ifdef.
> >>  I am uncertain whether we should do this since it hides the TODO item
> >>  of investigating ppc and sparc CPU reset.
> >>
> >>  linux-user/main.c | 2 --
> >>  1 file changed, 2 deletions(-)
> >>
> >> diff --git a/linux-user/main.c b/linux-user/main.c
> >> index 7f15d3d..e904d8c 100644
> >> --- a/linux-user/main.c
> >> +++ b/linux-user/main.c
> >> @@ -3637,9 +3637,7 @@ int main(int argc, char **argv, char **envp)
> >>          fprintf(stderr, "Unable to find CPU definition\n");
> >>          exit(1);
> >>      }
> >> -#if defined(TARGET_SPARC) || defined(TARGET_PPC)
> >>      cpu_reset(ENV_GET_CPU(env));
> >> -#endif
> >
> > Most of the cpu_reset() implementations I have looked at (including
> > sparc) contain something like:
> >
> >     memset(env, 0, offsetof(CPUXXXState, breakpoints));
> >
> > Isn't this clearing userspace registers that are not supposed to be
> > touched by clone()?
> 
> You're thinking about the other cpu_reset() -- this one is in
> main.c and happens only for the main thread, immediately
> after we've created that thread's CPU, and before we set up
> its registers for initial program start. The cpu_reset()
> that sometimes happens in the clone() path is in syscall.c
> (and is addressed by patch
> http://patchwork.ozlabs.org/patch/257232/ )

Oh, nevermind. We discussed do_fork() for so long that I didn't even
notice we had _another_ cpu_reset() #ifdef on main.c.

In this case the patch looks simple and safe, to me.

Patch

diff --git a/linux-user/main.c b/linux-user/main.c
index 7f15d3d..e904d8c 100644
--- a/linux-user/main.c
+++ b/linux-user/main.c
@@ -3637,9 +3637,7 @@  int main(int argc, char **argv, char **envp)
         fprintf(stderr, "Unable to find CPU definition\n");
         exit(1);
     }
-#if defined(TARGET_SPARC) || defined(TARGET_PPC)
     cpu_reset(ENV_GET_CPU(env));
-#endif
 
     thread_cpu = ENV_GET_CPU(env);