[PATCH 1/5] powerpc/64s/hash: Fix 128TB-512TB virtual address boundary case allocation

[Nicholas Piggin]

On Mon, 6 Nov 2017 16:35:43 +0530
"Aneesh Kumar K.V" <aneesh.kumar at linux.vnet.ibm.com> wrote:

> On 11/06/2017 04:24 PM, Nicholas Piggin wrote:
> > On Mon, 06 Nov 2017 16:08:06 +0530
> > "Aneesh Kumar K.V" <aneesh.kumar at linux.vnet.ibm.com> wrote:
> >   
> >> Nicholas Piggin <npiggin at gmail.com> writes:
> >>  
> >>> When allocating VA space with a hint that crosses 128TB, the SLB addr_limit
> >>> variable is not expanded if addr is not > 128TB, but the slice allocation
> >>> looks at task_size, which is 512TB. This results in slice_check_fit()
> >>> incorrectly succeeding because the slice_count truncates off bit 128 of the
> >>> requested mask, so the comparison to the available mask succeeds.  
> >>
> >>
> >> But then the mask passed to slice_check_fit() is generated using
> >> context.addr_limit as max value. So how did that return succcess? ie,
> >> we get the request mask via
> >>
> >> slice_range_to_mask(addr, len, &mask);
> >>
> >> And the potential/possible mask using
> >>
> >> slice_mask_for_size(mm, psize, &good_mask);
> >>
> >> So how did slice_check_fit() return sucess with
> >>
> >> slice_check_fit(mm, mask, good_mask);  
> > 
> > Because the addr_limit check is used to *limit* the comparison.
> > 
> > The available mask had bit up to 127 set, and the mask had 127 and
> > 128 set. However the 128T addr_limit causes only bits 0-127 to be
> > compared.
> >  
> 
> Should we fix it then via ? I haven't tested this yet. Also this result 
> in us comparing more bits?

I prefer not to rely on that as the fix because we should not be calling
into the slice code with an address beyond addr_limit IMO. There's quite
a few other places that use addr_limit. So I prefer my patch.

You could add this as an extra check, but yes it does result in more bitmap
to test. So if anything I would prefer to go the other way and actually
reduce the scope of *other* bitmap operations that are now using
SLICE_NUM_HIGH by similarly using addr_limit (if there are other
performance critical ones).

We could add some VM_BUG_ON checks to ensure tail bits are zero if
that's a concern.

> 
> modified   arch/powerpc/mm/slice.c
> @@ -169,13 +169,12 @@ static int slice_check_fit(struct mm_struct *mm,
>   			   struct slice_mask mask, struct slice_mask available)
>   {
>   	DECLARE_BITMAP(result, SLICE_NUM_HIGH);
> -	unsigned long slice_count = GET_HIGH_SLICE_INDEX(mm->context.addr_limit);
> 
>   	bitmap_and(result, mask.high_slices,
> -		   available.high_slices, slice_count);
> +		   available.high_slices, SLICE_NUM_HIGH);
> 
>   	return (mask.low_slices & available.low_slices) == mask.low_slices &&
> -		bitmap_equal(result, mask.high_slices, slice_count);
> +		bitmap_equal(result, mask.high_slices, SLICE_NUM_HIGH)
> 
> 
> -aneesh
>