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

[Qi Zheng <zhengqi.arch@bytedance.com>]

On 2025/1/3 16:02, Kevin Brodsky wrote:
> On 03/01/2025 04:48, Qi Zheng wrote:
>> Hi Kevin,
>>
>> On 2025/1/3 00:53, Kevin Brodsky wrote:
>>> On 30/12/2024 10:07, Qi Zheng wrote:
>>>>    static inline void riscv_tlb_remove_ptdesc(struct mmu_gather *tlb,
>>>> void *pt)
>>>>    {
>>>> -    if (riscv_use_sbi_for_rfence())
>>>> +    if (riscv_use_sbi_for_rfence()) {
>>>>            tlb_remove_ptdesc(tlb, pt);
>>>> -    else
>>>> +    } else {
>>>> +        pagetable_dtor(pt);
>>>>            tlb_remove_page_ptdesc(tlb, pt);
>>>
>>> I find the imbalance pretty confusing: pagetable_dtor() is called
>>> explicitly before using tlb_remove_page() but not tlb_remove_ptdesc().
>>> Doesn't that assume that CONFIG_MMU_GATHER_HAVE_TABLE_FREE is selected?
>>> Could we not call pagetable_dtor() from __tlb_batch_free_encoded_pages()
>>> to ensure that the dtor is always called just before freeing, and remove
>>
>> 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().

> 
>>
>> But now I think maybe we can do this in tlb_remove_page_ptdesc(), like
>> this:
>>
>> diff --git a/arch/csky/include/asm/pgalloc.h
>> b/arch/csky/include/asm/pgalloc.h
>> index f1ce5b7b28f22..e45c7e91dcbf9 100644
>> --- a/arch/csky/include/asm/pgalloc.h
>> +++ b/arch/csky/include/asm/pgalloc.h
>> @@ -63,7 +63,6 @@ static inline pgd_t *pgd_alloc(struct mm_struct *mm)
>>
>>   #define __pte_free_tlb(tlb, pte, address)              \
>>   do {                                                   \
>> -       pagetable_dtor(page_ptdesc(pte));               \
>>          tlb_remove_page_ptdesc(tlb, page_ptdesc(pte));  \
>>   } while (0)
>>
>> [...]
>>
>> diff --git a/include/asm-generic/tlb.h b/include/asm-generic/tlb.h
>> index a96d4b440f3da..a59205863f431 100644
>> --- a/include/asm-generic/tlb.h
>> +++ b/include/asm-generic/tlb.h
>> @@ -506,6 +506,7 @@ static inline void tlb_remove_ptdesc(struct
>> mmu_gather *tlb, void *pt)
>>   /* Like tlb_remove_ptdesc, but for page-like page directories. */
>>   static inline void tlb_remove_page_ptdesc(struct mmu_gather *tlb,
>> struct ptdesc *pt)
>>   {
>> +       pagetable_dtor(pt);
>>          tlb_remove_page(tlb, ptdesc_page(pt));
>>   }
> 
> I think this is an interesting idea, it does make arch code easier to
> follow. OTOH it would have been more natural to me to call
> pagetable_dtor() from tlb_remove_page(). I can however see that this
> doesn't work, because tlb_remove_table() is defined as tlb_remove_page()
> if CONFIG_MMU_GATHER_HAVE_TABLE_FREE isn't selected. Which brings me
> back to my earlier question: in that case, aren't we missing a call to
> pagetable_dtor() when using tlb_remove_table() (or tlb_remove_ptdesc())?

Thank you for pointing this out!

Now, there are the following architectures selected 
CONFIG_MMU_GATHER_RCU_TABLE_FREE:

1. arm: select MMU_GATHER_RCU_TABLE_FREE if SMP && ARM_LPAE
2. arm64: select MMU_GATHER_RCU_TABLE_FREE
3. powerpc: select MMU_GATHER_RCU_TABLE_FREE
4. riscv: select MMU_GATHER_RCU_TABLE_FREE if SMP && MMU
5. s390: select MMU_GATHER_RCU_TABLE_FREE
6. sparc: select MMU_GATHER_RCU_TABLE_FREE if SMP
7. x86: select MMU_GATHER_RCU_TABLE_FREE	if PARAVIRT

If CONFIG_MMU_GATHER_TABLE_FREE is selected, an architecture is expected
to provide __tlb_remove_table(). This patch series modifies the
__tlb_remove_table() in arm, arm64, riscv, s390 and x86. Among them,
arm64 and s390 unconditionally select MMU_GATHER_RCU_TABLE_FREE, so we
only need to double-check arm, riscv and x86.

For x86, it was called tlb_remove_page() in the non-PARAVIRT case, and I
added pagetable_dtor() for it explicitly (see patch #11), so this should
be no problem.

For riscv, it will only call tlb_remove_ptdesc() in the case of
SMP && MMU, so this should be no problem.

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.

So a more direct fix might be:

diff --git a/include/asm-generic/tlb.h b/include/asm-generic/tlb.h
index a59205863f431..0a131444a18ca 100644
--- a/include/asm-generic/tlb.h
+++ b/include/asm-generic/tlb.h
@@ -500,6 +500,9 @@ static inline void tlb_remove_page(struct mmu_gather 
*tlb, struct page *page)

  static inline void tlb_remove_ptdesc(struct mmu_gather *tlb, void *pt)
  {
+#ifndef CONFIG_MMU_GATHER_TABLE_FREE
+       pagetable_dtor(pt);
+#endif
         tlb_remove_table(tlb, pt);
  }

Or fix it directly in arm? Like the following:

diff --git a/arch/arm/include/asm/tlb.h b/arch/arm/include/asm/tlb.h
index ea4fbe7b17f6f..cf5d0ca021440 100644
--- a/arch/arm/include/asm/tlb.h
+++ b/arch/arm/include/asm/tlb.h
@@ -43,6 +43,9 @@ __pte_free_tlb(struct mmu_gather *tlb, pgtable_t pte, 
unsigned long addr)
         __tlb_adjust_range(tlb, addr - PAGE_SIZE, 2 * PAGE_SIZE);
  #endif

+#ifndef CONFIG_MMU_GATHER_TABLE_FREE
+       pagetable_dtor(ptdesc);
+#endif
         tlb_remove_ptdesc(tlb, ptdesc);
  }

Thanks,
Qi

> 
> - Kevin