Re: [PATCH v3] vfio/type1: optimize vfio_pin_pages_remote() for huge folio

[Alex Williamson <alex.williamson@redhat.com>]

On Tue, 20 May 2025 15:00:20 +0800
lizhe.67@bytedance.com wrote:

> From: Li Zhe <lizhe.67@bytedance.com>
> 
> When vfio_pin_pages_remote() is called with a range of addresses that
> includes huge folios, the function currently performs individual
> statistics counting 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 statistics counting
> operations.
> 
> The performance test results for completing the 8G VFIO IOMMU DMA mapping,
> obtained through trace-cmd, are as follows. In this case, the 8G virtual
> address space has been mapped to physical memory using hugetlbfs with
> pagesize=2M.
> 
> Before this patch:
> funcgraph_entry:      # 33813.703 us |  vfio_pin_map_dma();
> 
> After this patch:
> funcgraph_entry:      # 15635.055 us |  vfio_pin_map_dma();

It looks like we're using the same numbers since the initial
implementation, have these results changed?

> 
> Signed-off-by: Li Zhe <lizhe.67@bytedance.com>
> Signed-off-by: Alex Williamson <alex.williamson@redhat.com>

Appreciate the credit, this should probably be Co-developed-by: though.
In general a sign-off is something that needs to be explicitly given.

> ---
> Changelogs:
> 
> v2->v3:
> - Code simplification.
> - Fix some issues in comments.
> 
> v1->v2:
> - Fix some issues in comments and formatting.
> - Consolidate vfio_find_vpfn_range() and vfio_find_vpfn().
> - Move the processing logic for huge folio into the while(true) loop
>   and use a variable with a default value of 1 to indicate the number
>   of consecutive pages.
> 
> v2 patch: https://lore.kernel.org/all/20250519070419.25827-1-lizhe.67@bytedance.com/
> v1 patch: https://lore.kernel.org/all/20250513035730.96387-1-lizhe.67@bytedance.com/
> 
>  drivers/vfio/vfio_iommu_type1.c | 48 +++++++++++++++++++++++++--------
>  1 file changed, 37 insertions(+), 11 deletions(-)
> 
> diff --git a/drivers/vfio/vfio_iommu_type1.c b/drivers/vfio/vfio_iommu_type1.c
> index 0ac56072af9f..48f06ce0e290 100644
> --- a/drivers/vfio/vfio_iommu_type1.c
> +++ b/drivers/vfio/vfio_iommu_type1.c
> @@ -319,15 +319,22 @@ static void vfio_dma_bitmap_free_all(struct vfio_iommu *iommu)
>  /*
>   * Helper Functions for host iova-pfn list
>   */
> -static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
> +
> +/*
> + * Find the first vfio_pfn that overlapping the range
> + * [iova, iova + PAGE_SIZE * npage) in rb tree.
> + */
> +static struct vfio_pfn *vfio_find_vpfn_range(struct vfio_dma *dma,
> +		dma_addr_t iova, unsigned long npage)
>  {
>  	struct vfio_pfn *vpfn;
>  	struct rb_node *node = dma->pfn_list.rb_node;
> +	dma_addr_t end_iova = iova + PAGE_SIZE * npage;
>  
>  	while (node) {
>  		vpfn = rb_entry(node, struct vfio_pfn, node);
>  
> -		if (iova < vpfn->iova)
> +		if (end_iova <= vpfn->iova)
>  			node = node->rb_left;
>  		else if (iova > vpfn->iova)
>  			node = node->rb_right;
> @@ -337,6 +344,11 @@ static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
>  	return NULL;
>  }
>  
> +static inline struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
> +{
> +	return vfio_find_vpfn_range(dma, iova, 1);
> +}
> +
>  static void vfio_link_pfn(struct vfio_dma *dma,
>  			  struct vfio_pfn *new)
>  {
> @@ -681,32 +693,46 @@ static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
>  		 * and rsvd here, and therefore continues to use the batch.
>  		 */
>  		while (true) {
> +			struct folio *folio = page_folio(batch->pages[batch->offset]);
> +			long nr_pages;
> +
>  			if (pfn != *pfn_base + pinned ||
>  			    rsvd != is_invalid_reserved_pfn(pfn))
>  				goto out;
>  
> +			/*
> +			 * Note: The current nr_pages does not achieve the optimal
> +			 * performance in scenarios where folio_nr_pages() exceeds
> +			 * batch->capacity. It is anticipated that future enhancements
> +			 * will address this limitation.
> +			 */
> +			nr_pages = min((long)batch->size, folio_nr_pages(folio) -
> +						folio_page_idx(folio, batch->pages[batch->offset]));

We should use min_t() here, otherwise it looks good to me.

Peter, David, if you wouldn't mind double checking the folio usage
here, I'd appreciate it.  The underlying assumption used here is that
folios always have physically contiguous pages, so we can increment at
the remainder of the folio_nr_pages() rather than iterate each page.
Thanks,

Alex

> +			if (nr_pages > 1 && vfio_find_vpfn_range(dma, iova, nr_pages))
> +				nr_pages = 1;
> +
>  			/*
>  			 * Reserved pages aren't counted against the user,
>  			 * externally pinned pages are already counted against
>  			 * the user.
>  			 */
> -			if (!rsvd && !vfio_find_vpfn(dma, iova)) {
> +			if (!rsvd && (nr_pages > 1 || !vfio_find_vpfn(dma, iova))) {
>  				if (!dma->lock_cap &&
> -				    mm->locked_vm + lock_acct + 1 > limit) {
> +				    mm->locked_vm + lock_acct + nr_pages > limit) {
>  					pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
>  						__func__, limit << PAGE_SHIFT);
>  					ret = -ENOMEM;
>  					goto unpin_out;
>  				}
> -				lock_acct++;
> +				lock_acct += nr_pages;
>  			}
>  
> -			pinned++;
> -			npage--;
> -			vaddr += PAGE_SIZE;
> -			iova += PAGE_SIZE;
> -			batch->offset++;
> -			batch->size--;
> +			pinned += nr_pages;
> +			npage -= nr_pages;
> +			vaddr += PAGE_SIZE * nr_pages;
> +			iova += PAGE_SIZE * nr_pages;
> +			batch->offset += nr_pages;
> +			batch->size -= nr_pages;
>  
>  			if (!batch->size)
>  				break;