[PATCH] arm64/mm: Do not write ASID generation to ttbr0

[james.morse@arm.com (James Morse)]

Hi Julien,

On 05/12/17 17:38, Julien Thierry wrote:
> When writing the user ASID to ttbr0, 16bit get copied to ttbr0, potentially
> including part of the ASID generation in the ttbr0.ASID. If the kernel is
> using less than 16bits ASIDs and the other ttbr0 bits aren't RES0, two
> tasks using the same mm context might end up running with different
> ttbr0.ASID values.
> This would be triggered by one of the threads being scheduled before a
> roll-over and the second one scheduled after a roll-over.
> 
> Pad the generation out of the 16bits of the mm id that are written to
> ttbr0. Thus, what the hardware sees is what the kernel considers ASID.

Is this to fix a system with a mix of 8/16 asid bits? Or someone being clever
and reducing asid_bits to stress rollover?

(I think we should do something like this so we can test rollover.)


> diff --git a/arch/arm64/mm/context.c b/arch/arm64/mm/context.c
> index 6f40170..a7c72d4 100644
> --- a/arch/arm64/mm/context.c
> +++ b/arch/arm64/mm/context.c
> @@ -180,6 +190,13 @@ static u64 new_context(struct mm_struct *mm, unsigned int cpu)
>  	/* We're out of ASIDs, so increment the global generation count */
>  	generation = atomic64_add_return_relaxed(ASID_FIRST_VERSION,
>  						 &asid_generation);
> +
> +	/*
> +	 * It is unlikely the generation will ever overflow, but if this
> +	 * happens, let it be known strange things can occur.
> +	 */
> +	WARN_ON(generation == ASID_FIRST_VERSION);

Won't generation wrap to zero, not '1<<16'?

asid_generation-is-never-zero is one of the nasty things in this code.

In check_and_switch_context() we switch to the 'fast path' where the current
asid is usable if: the generation matches and the active_asid isn't 0 (which
indicates a rollover has occurred).

from mm/context.c::check_and_switch_context():
> 	if (!((asid ^ atomic64_read(&asid_generation)) >> asid_bits)
>	    && atomic64_xchg_relaxed(&per_cpu(active_asids, cpu), asid))
>		goto switch_mm_fastpath;


If asid_generation is ever zero, (even briefly) multiple new tasks with
different pages-tables will pass the generation test, and run with asid=0.

Short of xchg(ASID_MAX_VERSION, ASID_FIRST_VERSION), every time, just in case, I
don't think this is something we can handle.

But, this thing is layers on layers of subtle behaviour, so I may have missed
something...


Instead, do you think we can duplicate just the asid bits (asid & ASID_MASK)
into a new 16bit field in mm_context_t, which we then use instead of the ASID()
and mmid macros? (We only set a new asid in one place as returned by new_context()).

This would let context.c's asid_bits be arbitrary, as the hardware only uses the
masked value it exposes in the new field.

This doesn't solve the mixed 8/16 bit case. I think we should force those
systems to use 8bit asids via cpufeature to preserve as much of the generation
counter as possible.


Thanks,

James