Re: [PATCH v4 4/4] sparc64: Add support for ADI (Application Data Integrity)

[Khalid Aziz]

On 01/16/2017 09:39 PM, David Miller wrote:
> From: Khalid Aziz <khalid.aziz@xxxxxxxxxx>
> Date: Wed, 11 Jan 2017 09:12:54 -0700
>
> > diff --git a/arch/sparc/kernel/mdesc.c b/arch/sparc/kernel/mdesc.c
> > index 8a6982d..68b03bf 100644
> > --- a/arch/sparc/kernel/mdesc.c
> > +++ b/arch/sparc/kernel/mdesc.c
> > @@ -20,6 +20,7 @@
> >  #include <asm/uaccess.h>
> >  #include <asm/oplib.h>
> >  #include <asm/smp.h>
> > +#include <asm/adi.h>
> >
> >  /* Unlike the OBP device tree, the machine description is a full-on
> >   * DAG.  An arbitrary number of ARCs are possible from one
> > @@ -1104,5 +1105,8 @@ void __init sun4v_mdesc_init(void)
> >
> >  	cur_mdesc = hp;
> >
> > +#ifdef CONFIG_SPARC64
>
> mdesc.c is only built on sparc64, this ifdef is superfluous.

Good point. I will fix it.

> > +/* Update the state of MCDPER register in current task's mm context before
> > + * dup so the dup'd task will inherit flags in this register correctly.
> > + * Current task may have updated flags since it started running.
> > + */
> > +int arch_dup_task_struct(struct task_struct *dst, struct task_struct *src)
> > +{
> > +	if (adi_capable() && src->mm) {
> > +		register unsigned long tmp_mcdper;
> > +
> > +		__asm__ __volatile__(
> > +			".word 0x83438000\n\t"	/* rd %mcdper, %g1 */
> > +			"mov %%g1, %0\n\t"
> > +			: "=r" (tmp_mcdper)
> > +			:
> > +			: "g1");
> > +		src->mm->context.mcdper = tmp_mcdper;
>
> I don't like the idea of duplicating 'mm' state using the task struct
> copy.  Why do not the MM handling interfaces handle this properly?
>
> Maybe it means you've abstracted the ADI register handling in the
> wrong place.  Maybe it's a thread property which is "pushed" from
> the MM context.

I see what you are saying. This code updates mm->context.mcdper for the 
source thread with the current state of MCDPER since MCDPER can be 
changed by a userspace process any time. When userspace changes MCDPER, 
it is not saved into mm->context.mcdper until a context switch happens. 
This means during the timeslice for a thread, its mm->context.mcdper may 
not reflect the current value of MCDPER. Updating it ensures dup_mm() 
will copy the real current value of MCDPER into the newly forked thread. 
arch_dup_mmap() looks like a more appropriate place to do this. Do you 
agree?

Thanks,
Khalid