Re: [PATCH v4 2/3] gup: introduce unpin_user_folio_dirty_locked()

[David Hildenbrand <david@redhat.com>]

On 17.06.25 11:21, lizhe.67@bytedance.com wrote:
> On Tue, 17 Jun 2025 09:43:56 +0200, david@redhat.com wrote:
>   
>> On 17.06.25 06:18, lizhe.67@bytedance.com wrote:
>>> From: Li Zhe <lizhe.67@bytedance.com>
>>>
>>> When vfio_unpin_pages_remote() is called with a range of addresses that
>>> includes large folios, the function currently performs individual
>>> put_pfn() operations for each page. This can lead to significant
>>> performance overheads, especially when dealing with large ranges of pages.
>>>
>>> This patch optimize this process by batching the put_pfn() operations.
>>>
>>> The performance test results, based on v6.15, for completing the 16G VFIO
>>> IOMMU DMA unmapping, obtained through unit test[1] with slight
>>> modifications[2], are as follows.
>>>
>>> Base(v6.15):
>>> ./vfio-pci-mem-dma-map 0000:03:00.0 16
>>> ------- AVERAGE (MADV_HUGEPAGE) --------
>>> VFIO MAP DMA in 0.047 s (338.6 GB/s)
>>> VFIO UNMAP DMA in 0.138 s (116.2 GB/s)
>>> ------- AVERAGE (MAP_POPULATE) --------
>>> VFIO MAP DMA in 0.280 s (57.2 GB/s)
>>> VFIO UNMAP DMA in 0.312 s (51.3 GB/s)
>>> ------- AVERAGE (HUGETLBFS) --------
>>> VFIO MAP DMA in 0.052 s (308.3 GB/s)
>>> VFIO UNMAP DMA in 0.139 s (115.1 GB/s)
>>>
>>> Map[3] + This patchset:
>>> ------- AVERAGE (MADV_HUGEPAGE) --------
>>> VFIO MAP DMA in 0.028 s (563.9 GB/s)
>>> VFIO UNMAP DMA in 0.049 s (325.1 GB/s)
>>> ------- AVERAGE (MAP_POPULATE) --------
>>> VFIO MAP DMA in 0.294 s (54.4 GB/s)
>>> VFIO UNMAP DMA in 0.296 s (54.1 GB/s)
>>> ------- AVERAGE (HUGETLBFS) --------
>>> VFIO MAP DMA in 0.033 s (485.1 GB/s)
>>> VFIO UNMAP DMA in 0.049 s (324.4 GB/s)
>>>
>>> For large folio, we achieve an approximate 64% performance improvement
>>> in the VFIO UNMAP DMA item. For small folios, the performance test
>>> results appear to show no significant changes.
>>>
>>> [1]: https://github.com/awilliam/tests/blob/vfio-pci-mem-dma-map/vfio-pci-mem-dma-map.c
>>> [2]: https://lore.kernel.org/all/20250610031013.98556-1-lizhe.67@bytedance.com/
>>> [3]: https://lore.kernel.org/all/20250529064947.38433-1-lizhe.67@bytedance.com/
>>>
>>> Signed-off-by: Li Zhe <lizhe.67@bytedance.com>
>>> ---
>>>    drivers/vfio/vfio_iommu_type1.c | 35 +++++++++++++++++++++++++++++----
>>>    1 file changed, 31 insertions(+), 4 deletions(-)
>>>
>>> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
>>> index e952bf8bdfab..159ba80082a8 100644
>>> --- a/drivers/vfio/vfio_iommu_type1.c
>>> +++ b/drivers/vfio/vfio_iommu_type1.c
>>> @@ -806,11 +806,38 @@ static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
>>>    				    bool do_accounting)
>>>    {
>>>    	long unlocked = 0, locked = vpfn_pages(dma, iova, npage);
>>> -	long i;
>>>    
>>> -	for (i = 0; i < npage; i++)
>>> -		if (put_pfn(pfn++, dma->prot))
>>> -			unlocked++;
>>> +	while (npage) {
>>> +		long nr_pages = 1;
>>> +
>>> +		if (!is_invalid_reserved_pfn(pfn)) {
>>> +			struct page *page = pfn_to_page(pfn);
>>> +			struct folio *folio = page_folio(page);
>>> +			long folio_pages_num = folio_nr_pages(folio);
>>> +
>>> +			/*
>>> +			 * For a folio, it represents a physically
>>> +			 * contiguous set of bytes, and all of its pages
>>> +			 * share the same invalid/reserved state.
>>> +			 *
>>> +			 * Here, our PFNs are contiguous. Therefore, if we
>>> +			 * detect that the current PFN belongs to a large
>>> +			 * folio, we can batch the operations for the next
>>> +			 * nr_pages PFNs.
>>> +			 */
>>> +			if (folio_pages_num > 1)
>>> +				nr_pages = min_t(long, npage,
>>> +					folio_pages_num -
>>> +					folio_page_idx(folio, page));
>>> +
>>
>> (I know I can be a pain :) )
> 
> No, not at all! I really appreciate you taking the time to review my
> patch.
> 
>> But the long comment indicates that this is confusing.
>>
>>
>> That is essentially the logic in gup_folio_range_next().
>>
>> What about factoring that out into a helper like
>>
>> /*
>>    * TODO, returned number includes the provided current page.
>>    */
>> unsigned long folio_remaining_pages(struct folio *folio,
>> 	struct pages *pages, unsigned long max_pages)
>> {
>> 	if (!folio_test_large(folio))
>> 		return 1;
>> 	return min_t(unsigned long, max_pages,
>> 		     folio_nr_pages(folio) - folio_page_idx(folio, page));
>> }
>>
>>
>> Then here you would do
>>
>> if (!is_invalid_reserved_pfn(pfn)) {
>> 	struct page *page = pfn_to_page(pfn);
>> 	struct folio *folio = page_folio(page);
>>
>> 	/* We can batch-process pages belonging to the same folio. */
>> 	nr_pages = folio_remaining_pages(folio, page, npage);
>>
>> 	unpin_user_folio_dirty_locked(folio, nr_pages,
>> 				      dma->prot & IOMMU_WRITE);
>> 	unlocked += nr_pages;
>> }
> 
> Yes, this indeed makes the code much more comprehensible. Do you think
> the implementation of the patch as follows look viable to you? I have
> added some brief comments on top of your work to explain why we can
> batch-process pages belonging to the same folio. This was suggested by
> Alex[1].
> 
> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
> index e952bf8bdfab..d7653f4c10d5 100644
> --- a/drivers/vfio/vfio_iommu_type1.c
> +++ b/drivers/vfio/vfio_iommu_type1.c
> @@ -801,16 +801,43 @@ static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
>          return pinned;
>   }
>   
> +/* Returned number includes the provided current page. */
> +static inline unsigned long folio_remaining_pages(struct folio *folio,
> +               struct page *page, unsigned long max_pages)
> +{
> +       if (!folio_test_large(folio))
> +               return 1;
> +       return min_t(unsigned long, max_pages,
> +                    folio_nr_pages(folio) - folio_page_idx(folio, page));
> +}

Note that I think that should go somewhere into mm.h, and also get used 
by GUP. So factoring it out from GUP and then using it here.

-- 
Cheers,

David / dhildenb