Re: [PATCH v6 2/6] mm: swap: free_swap_and_cache_nr() as batched free_swap_and_cache()

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

[...]
> 
> [...]
> 
>> +
>> +/**
>> + * swap_pte_batch - detect a PTE batch for a set of contiguous swap entries
>> + * @start_ptep: Page table pointer for the first entry.
>> + * @max_nr: The maximum number of table entries to consider.
>> + * @entry: Swap entry recovered from the first table entry.
>> + *
>> + * Detect a batch of contiguous swap entries: consecutive (non-present) PTEs
>> + * containing swap entries all with consecutive offsets and targeting the same
>> + * swap type.
>> + *
> 
> Likely you should document that any swp pte bits are ignored? ()

Now that I understand what swp pte bits are, I think the simplest thing is to
just make this function always consider the pte bits by using pte_same() as you
suggest below? I don't think there is ever a case for ignoring the swp pte bits?
And then I don't need to do anything special for uffd-wp either (below you
suggested not doing batching when the VMA has uffd enabled).

Any concerns?

> 
>> + * max_nr must be at least one and must be limited by the caller so scanning
>> + * cannot exceed a single page table.
>> + *
>> + * Return: the number of table entries in the batch.
>> + */
>> +static inline int swap_pte_batch(pte_t *start_ptep, int max_nr,
>> +                 swp_entry_t entry)
>> +{
>> +    const pte_t *end_ptep = start_ptep + max_nr;
>> +    unsigned long expected_offset = swp_offset(entry) + 1;
>> +    unsigned int expected_type = swp_type(entry);
>> +    pte_t *ptep = start_ptep + 1;
>> +
>> +    VM_WARN_ON(max_nr < 1);
>> +    VM_WARN_ON(non_swap_entry(entry));
>> +
>> +    while (ptep < end_ptep) {
>> +        pte_t pte = ptep_get(ptep);
>> +
>> +        if (pte_none(pte) || pte_present(pte))
>> +            break;
>> +
>> +        entry = pte_to_swp_entry(pte);
>> +
>> +        if (non_swap_entry(entry) ||
>> +            swp_type(entry) != expected_type ||
>> +            swp_offset(entry) != expected_offset)
>> +            break;
>> +
>> +        expected_offset++;
>> +        ptep++;
>> +    }
>> +
>> +    return ptep - start_ptep;
>> +}
> 
> Looks very clean :)
> 
> I was wondering whether we could similarly construct the expected swp PTE and
> only check pte_same.
> 
> expected_pte = __swp_entry_to_pte(__swp_entry(expected_type, expected_offset));

So planning to do this.

> 
> ... or have a variant to increase only the swp offset for an existing pte. But
> non-trivial due to the arch-dependent format.

not this - I agree this will be difficult due to per-arch changes. I'd rather
just do the generic version and leave the compiler to do the best it can to
simplify and optimize.

> 
> But then, we'd fail on mismatch of other swp pte bits.
> 
> 
> On swapin, when reusing this function (likely!), we'll might to make sure that
> the PTE bits match as well.
> 
> See below regarding uffd-wp.
> 
> 
>>   #endif /* CONFIG_MMU */
>>     void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio,
>> diff --git a/mm/madvise.c b/mm/madvise.c
>> index 1f77a51baaac..070bedb4996e 100644
>> --- a/mm/madvise.c
>> +++ b/mm/madvise.c
>> @@ -628,6 +628,7 @@ static int madvise_free_pte_range(pmd_t *pmd, unsigned
>> long addr,
>>       struct folio *folio;
>>       int nr_swap = 0;
>>       unsigned long next;
>> +    int nr, max_nr;
>>         next = pmd_addr_end(addr, end);
>>       if (pmd_trans_huge(*pmd))
>> @@ -640,7 +641,8 @@ static int madvise_free_pte_range(pmd_t *pmd, unsigned
>> long addr,
>>           return 0;
>>       flush_tlb_batched_pending(mm);
>>       arch_enter_lazy_mmu_mode();
>> -    for (; addr != end; pte++, addr += PAGE_SIZE) {
>> +    for (; addr != end; pte += nr, addr += PAGE_SIZE * nr) {
>> +        nr = 1;
>>           ptent = ptep_get(pte);
>>             if (pte_none(ptent))
>> @@ -655,9 +657,11 @@ static int madvise_free_pte_range(pmd_t *pmd, unsigned
>> long addr,
>>                 entry = pte_to_swp_entry(ptent);
>>               if (!non_swap_entry(entry)) {
>> -                nr_swap--;
>> -                free_swap_and_cache(entry);
>> -                pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
>> +                max_nr = (end - addr) / PAGE_SIZE;
>> +                nr = swap_pte_batch(pte, max_nr, entry);
>> +                nr_swap -= nr;
>> +                free_swap_and_cache_nr(entry, nr);
>> +                clear_not_present_full_ptes(mm, addr, pte, nr, tlb->fullmm);
>>               } else if (is_hwpoison_entry(entry) ||
>>                      is_poisoned_swp_entry(entry)) {
>>                   pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
>> diff --git a/mm/memory.c b/mm/memory.c
>> index 7dc6c3d9fa83..ef2968894718 100644
>> --- a/mm/memory.c
>> +++ b/mm/memory.c
>> @@ -1637,12 +1637,13 @@ static unsigned long zap_pte_range(struct mmu_gather
>> *tlb,
>>                   folio_remove_rmap_pte(folio, page, vma);
>>               folio_put(folio);
>>           } else if (!non_swap_entry(entry)) {
>> -            /* Genuine swap entry, hence a private anon page */
>> +            max_nr = (end - addr) / PAGE_SIZE;
>> +            nr = swap_pte_batch(pte, max_nr, entry);
>> +            /* Genuine swap entries, hence a private anon pages */
>>               if (!should_zap_cows(details))
>>                   continue;
>> -            rss[MM_SWAPENTS]--;
>> -            if (unlikely(!free_swap_and_cache(entry)))
>> -                print_bad_pte(vma, addr, ptent, NULL);
>> +            rss[MM_SWAPENTS] -= nr;
>> +            free_swap_and_cache_nr(entry, nr);
>>           } else if (is_migration_entry(entry)) {
>>               folio = pfn_swap_entry_folio(entry);
>>               if (!should_zap_folio(details, folio))
>> @@ -1665,8 +1666,8 @@ static unsigned long zap_pte_range(struct mmu_gather *tlb,
>>               pr_alert("unrecognized swap entry 0x%lx\n", entry.val);
>>               WARN_ON_ONCE(1);
>>           }
>> -        pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
>> -        zap_install_uffd_wp_if_needed(vma, addr, pte, 1, details, ptent);
>> +        clear_not_present_full_ptes(mm, addr, pte, nr, tlb->fullmm);
> 
> For zap_install_uffd_wp_if_needed(), the uffd-wp bit has to match.
> 
> zap_install_uffd_wp_if_needed() will use the uffd-wp information in
> ptent->pteval to make a decision whether to place PTE_MARKER_UFFD_WP markers.
> 
> On mixture, you either lose some or place too many markers.
> 
> A simple workaround would be to disable any such batching if the VMA does have
> uffd-wp enabled.

Rather than this, I'll just consider all the swp pte bits when batching.

> 
>> +        zap_install_uffd_wp_if_needed(vma, addr, pte, nr, details, ptent);
>>       } while (pte += nr, addr += PAGE_SIZE * nr, addr != end);

[...]