[PATCH v4 10/15] riscv: pgtable: move pagetable_dtor() to __tlb_remove_table()

[Kevin Brodsky]

On 03/01/2025 10:35, Qi Zheng wrote:
> On 2025/1/3 17:13, Qi Zheng wrote:
>> On 2025/1/3 16:02, Kevin Brodsky wrote:
>>> On 03/01/2025 04:48, Qi Zheng wrote:
>>>> [...]
>>>>
>>>> In __tlb_batch_free_encoded_pages(), we can indeed detect PageTable()
>>>> and call pagetable_dtor() to dtor the page table pages.
>>>> But __tlb_batch_free_encoded_pages() is also used to free normal pages
>>>> (not page table pages), so I don't want to add overhead there.
>>>
>>> Interesting, can a tlb batch refer to pages than are not PTPs then?
>>
>> Yes, you can see the caller of __tlb_remove_folio_pages() or
>> tlb_remove_page_size().

I had a brief look but clearly not a good enough one! I hadn't realised
that "table" in tlb_remove_table() means PTP, while "page" in
tlb_remove_page() can mean any page, and it's making more sense now.

[...]

>>
>> For arm, the call to pagetable_dtor() is indeed missed in the
>> non-MMU_GATHER_RCU_TABLE_FREE case. This needs to be fixed. But we
>> can't fix this by adding pagetable_dtor() to tlb_remove_table(),
>> because some architectures call tlb_remove_table() but don't support
>> page table statistics, like sparc.

When I investigated this for my own series, I found that the only case
where ctor/dtor are not called for page-sized page tables is 32-bit
sparc (see table at the end of [1]). However only 64-bit sparc makes use
of tlb_remove_table() (at PTE level, where ctor/dtor are already called).

So really calling pagetable_dtor() from tlb_remove_table() in the
non-MMU_GATHER_TABLE_FREE case seems to be the obvious thing to do.

Once this is done, we should be able to replace all those confusing
calls to tlb_remove_page() on PTPs with tlb_remove_table() and remove
the explicit call to pagetable_dtor(). AIUI this is essentially what
Peter suggested on v3 [2].

[1]
https://lore.kernel.org/linux-mm/20241219164425.2277022-1-kevin.brodsky@arm.com/
[2]
https://lore.kernel.org/linux-mm/20250103111457.GC22934@noisy.programming.kicks-ass.net/

[...]

> Or can we just not let tlb_remove_table() fall back to
> tlb_remove_page()? Like the following:
>
> diff --git a/include/asm-generic/tlb.h b/include/asm-generic/tlb.h
> index a59205863f431..354ffaa4bd120 100644
> --- a/include/asm-generic/tlb.h
> +++ b/include/asm-generic/tlb.h
> @@ -195,8 +195,6 @@
>   *  various ptep_get_and_clear() functions.
>   */
>
> -#ifdef CONFIG_MMU_GATHER_TABLE_FREE
> -
>  struct mmu_table_batch {
>  #ifdef CONFIG_MMU_GATHER_RCU_TABLE_FREE
>         struct rcu_head         rcu;
> @@ -219,16 +217,6 @@ static inline void __tlb_remove_table(void *table)
>
>  extern void tlb_remove_table(struct mmu_gather *tlb, void *table);
>
> -#else /* !CONFIG_MMU_GATHER_HAVE_TABLE_FREE */
> -
> -/*
> - * Without MMU_GATHER_TABLE_FREE the architecture is assumed to have
> page based
> - * page directories and we can use the normal page batching to free
> them.
> - */
> -#define tlb_remove_table(tlb, page) tlb_remove_page((tlb), (page))

We still need a different implementation of tlb_remove_table() in this
case. We could define it inline here:

static inline void tlb_remove_table(struct mmu_gather *tlb, void *table)
{
    struct page *page = table;

    pagetable_dtor(page_ptdesc(page));
    tlb_remove_page(page);
}

- Kevin