On Wed, Oct 28, 2020 at 12:28 PM Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
>
[...]
> > +void *mte_set_mem_tag_range(void *addr, size_t size, u8 tag)
> > +{
> > + void *ptr = addr;
> > +
> > + if ((!system_supports_mte()) || (size == 0))
> > + return addr;
> > +
> > + /* Make sure that size is MTE granule aligned. */
> > + WARN_ON(size & (MTE_GRANULE_SIZE - 1));
> > +
> > + /* Make sure that the address is MTE granule aligned. */
> > + WARN_ON((u64)addr & (MTE_GRANULE_SIZE - 1));
> > +
> > + tag = 0xF0 | tag;
> > + ptr = (void *)__tag_set(ptr, tag);
> > +
> > + mte_assign_mem_tag_range(ptr, size);
>
> This function will be called on production hot paths. I think it makes
> sense to shave off some overheads here.
>
> The additional debug checks may be useful, so maybe we need an
> additional debug mode (debug of MTE/KASAN itself)?
>
> Do we ever call this when !system_supports_mte()? I think we wanted to
> have static_if's higher up the stack. Having additional checks
> scattered across lower-level functions is overhead for every
> malloc/free.
>
> Looking at how this is called from KASAN code.
> KASAN code already ensures addr/size are properly aligned. I think we
> should either remove the duplicate alignment checks, or do them only
> in the additional debugging mode.
> Does KASAN also ensure proper tag value (0xF0 mask)?
>
> KASAN wrapper is inlined in this patch:
> https://linux-review.googlesource.com/c/linux/kernel/git/torvalds/linux/+/3699
> but here we still have 2 non-inlined calls. The
> mte_assign_mem_tag_range is kinda inherent since it's in .S. But then
> I think this wrapper should be inlinable.
>
> Also, can we move mte_assign_mem_tag_range into inline asm in the
> header? This would avoid register spills around the call in
> malloc/free.
>
> The asm code seems to do the rounding of the size up at no additional
> cost (checks remaining size > 0, right?). I think it makes sense to
> document that as the contract and remove the additional round_up(size,
> KASAN_GRANULE_SIZE) in KASAN code.
These are all valid concerns. It would be great to have inline asm
mte_assign_mem_tag_range() implementation. We can also call it
directly from KASAN code without all these additional checks.
Perhaps it makes sense to include this change into the other series
that adds the production mode. And then squash if we decide to put
both changes into a single one.
Vincenzo, could you write a patch that adds inline asm
mte_assign_mem_tag_range() implementation?
