> @@ -437,18 +437,31 @@ static inline void __flush_tlb_all(void)
>  /*
>   * flush one page in the kernel mapping
>   */
> -static inline void __flush_tlb_one(unsigned long addr)
> +static inline void __flush_tlb_one_kernel(unsigned long addr)
>  {
>       count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ONE);
> -     __flush_tlb_single(addr);
> +
> +     /*
> +      * If PTI is off, then __flush_tlb_one_user() is just INVLPG or its
> +      * paravirt equivalent.  Even with PCID, this is sufficient: we only
> +      * use PCID if we also use global PTEs for the kernel mapping, and
> +      * INVLPG flushes global translations across all address spaces.

This looks good.

> +      * If PTI is on, then the kernel is mapped with non-global PTEs, and
> +      * __flush_tlb_one_user() will flush the given address for the current
> +      * kernel address space and for its usermode counterpart, but it goes
> +      * not flush it for other address spaces.
> +      */
> +     __flush_tlb_one_user(addr);

s/goes/does/

It also goes off and flushes the address out of the user asid.  That
_seems_ a bit goofy, but it is needed for addresses that might be mapped
into the user asid, so it's definitely safe.  Might be worth calling out.

Maybe add a (if one exists) or something like:

... kernel address space and for its usermode counterpart (if one exists).

>       if (!static_cpu_has(X86_FEATURE_PTI))
>               return;
>  
>       /*
> -      * __flush_tlb_single() will have cleared the TLB entry for this ASID,
> -      * but since kernel space is replicated across all, we must also
> -      * invalidate all others.
> +      * See above.  We need to propagate the flush to all other address
> +      * spaces.  In principle, we only need to propagate it to kernelmode
> +      * address spaces, but the extra bookkeeping we would need is not
> +      * worth it.
>        */
>       invalidate_other_asid();
>  }

That comment is true, except if we were invalidating a user-mapped
address.  Right?

We've just been pretending so far for the purposes of TLB invalidation
that all kernel addresses are potentially user-mapped.

The name change looks really good to me, though.  Thanks for doing this.

Reply via email to