Re: [PATCH v1 11/14] arm64/mm: Wire up PTE_CONT for user mappings

[Ryan Roberts <ryan.roberts@arm.com>]

On 30/06/2023 02:54, John Hubbard wrote:
> On 6/22/23 07:42, Ryan Roberts wrote:
>> With the ptep API sufficiently refactored, we can now introduce a new
>> "contpte" API layer, which transparently manages the PTE_CONT bit for
>> user mappings. Whenever it detects a set of PTEs that meet the
>> requirements for a contiguous range, the PTEs are re-painted with the
>> PTE_CONT bit.
>>
>> This initial change provides a baseline that can be optimized in future
>> commits. That said, fold/unfold operations (which imply tlb
>> invalidation) are avoided where possible with a few tricks for
>> access/dirty bit management.
>>
>> Write-enable and write-protect modifications are likely non-optimal and
>> likely incure a regression in fork() performance. This will be addressed
>> separately.
>>
>> Signed-off-by: Ryan Roberts <ryan.roberts@arm.com>
>> ---
> 
> Hi Ryan!
> 
> While trying out the full series from your gitlab features/granule_perf/all
> branch, I found it necessary to EXPORT a symbol in order to build this.

Thanks for the bug report!

> Please see below:
> 
> ...
>> +
>> +pte_t contpte_ptep_get(pte_t *ptep, pte_t orig_pte)
>> +{
>> +    /*
>> +     * Gather access/dirty bits, which may be populated in any of the ptes
>> +     * of the contig range. We are guarranteed to be holding the PTL, so any
>> +     * contiguous range cannot be unfolded or otherwise modified under our
>> +     * feet.
>> +     */
>> +
>> +    pte_t pte;
>> +    int i;
>> +
>> +    ptep = contpte_align_down(ptep);
>> +
>> +    for (i = 0; i < CONT_PTES; i++, ptep++) {
>> +        pte = __ptep_get(ptep);
>> +
>> +        /*
>> +         * Deal with the partial contpte_ptep_get_and_clear_full() case,
>> +         * where some of the ptes in the range may be cleared but others
>> +         * are still to do. See contpte_ptep_get_and_clear_full().
>> +         */
>> +        if (pte_val(pte) == 0)
>> +            continue;
>> +
>> +        if (pte_dirty(pte))
>> +            orig_pte = pte_mkdirty(orig_pte);
>> +
>> +        if (pte_young(pte))
>> +            orig_pte = pte_mkyoung(orig_pte);
>> +    }
>> +
>> +    return orig_pte;
>> +}
> 
> Here we need something like this, in order to get it to build in all
> possible configurations:
> 
> EXPORT_SYMBOL_GPL(contpte_ptep_get);
> 
> (and a corresponding "#include linux/export.h" at the top of the file).
> 
> Because, the static inline functions invoke this routine, above.


A quick grep through the drivers directory shows:

ptep_get() is used by:
  - drivers/gpu/drm/i915/gem/selftests/i915_gem_mman.c
  - drivers/misc/sgi-gru/grufault.c
  - drivers/vfio/vfio_iommu_type1.c
  - drivers/xen/privcmd.c

ptep_set_at() is used by:
  - drivers/gpu/drm/i915/i915_mm.c
  - drivers/xen/xlate_mmu.c

None of the other symbols are called, but I guess it is possible that out of
tree modules are calling others.

So on the basis that these symbols were previously pure inline, I propose to
export all the contpte_* symbols using EXPORT_SYMBOL() so that anything that was
previously calling them successfully continue to do so. Will include in v2.

Thanks,
Ryan


> 
> thanks,