[PATCH 5.15.y 1/4] vduse: Use fixed 4KB bounce pages for non-4KB page size

[Sasha Levin <sashal@kernel.org>]

From: Sheng Zhao <sheng.zhao@bytedance.com>

[ Upstream commit 3fc3068e7247c94dec08e93fea422a1bb649bfe5 ]

The allocation granularity of bounce pages is PAGE_SIZE. This may cause
even small IO requests to occupy an entire bounce page exclusively. The
kind of memory waste will be more significant when PAGE_SIZE is larger
than 4KB (e.g. arm64 with 64KB pages).

So, optimize it by using fixed 4KB bounce maps and iova allocation
granularity. A single IO request occupies at least a 4KB bounce page
instead of the entire memory page of PAGE_SIZE.

Signed-off-by: Sheng Zhao <sheng.zhao@bytedance.com>
Message-Id: <20250925113516.60305-1-sheng.zhao@bytedance.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Stable-dep-of: 9c1523803445 ("VDUSE: avoid leaking information to userspace")
Signed-off-by: Sasha Levin <sashal@kernel.org>
---
 drivers/vdpa/vdpa_user/iova_domain.c | 64 +++++++++++++++++++---------
 drivers/vdpa/vdpa_user/iova_domain.h |  5 +++
 2 files changed, 50 insertions(+), 19 deletions(-)

diff --git a/drivers/vdpa/vdpa_user/iova_domain.c b/drivers/vdpa/vdpa_user/iova_domain.c
index 0678c251419736..35363ca0b49bd4 100644
--- a/drivers/vdpa/vdpa_user/iova_domain.c
+++ b/drivers/vdpa/vdpa_user/iova_domain.c
@@ -103,19 +103,38 @@ void vduse_domain_clear_map(struct vduse_iova_domain *domain,
 static int vduse_domain_map_bounce_page(struct vduse_iova_domain *domain,
 					 u64 iova, u64 size, u64 paddr)
 {
-	struct vduse_bounce_map *map;
+	struct vduse_bounce_map *map, *head_map;
+	struct page *tmp_page;
 	u64 last = iova + size - 1;
 
 	while (iova <= last) {
-		map = &domain->bounce_maps[iova >> PAGE_SHIFT];
+		/*
+		 * When PAGE_SIZE is larger than 4KB, multiple adjacent bounce_maps will
+		 * point to the same memory page of PAGE_SIZE. Since bounce_maps originate
+		 * from IO requests, we may not be able to guarantee that the orig_phys
+		 * values of all IO requests within the same 64KB memory page are contiguous.
+		 * Therefore, we need to store them separately.
+		 *
+		 * Bounce pages are allocated on demand. As a result, it may occur that
+		 * multiple bounce pages corresponding to the same 64KB memory page attempt
+		 * to allocate memory simultaneously, so we use cmpxchg to handle this
+		 * concurrency.
+		 */
+		map = &domain->bounce_maps[iova >> BOUNCE_MAP_SHIFT];
 		if (!map->bounce_page) {
-			map->bounce_page = alloc_page(GFP_ATOMIC);
-			if (!map->bounce_page)
-				return -ENOMEM;
+			head_map = &domain->bounce_maps[(iova & PAGE_MASK) >> BOUNCE_MAP_SHIFT];
+			if (!head_map->bounce_page) {
+				tmp_page = alloc_page(GFP_ATOMIC);
+				if (!tmp_page)
+					return -ENOMEM;
+				if (cmpxchg(&head_map->bounce_page, NULL, tmp_page))
+					__free_page(tmp_page);
+			}
+			map->bounce_page = head_map->bounce_page;
 		}
 		map->orig_phys = paddr;
-		paddr += PAGE_SIZE;
-		iova += PAGE_SIZE;
+		paddr += BOUNCE_MAP_SIZE;
+		iova += BOUNCE_MAP_SIZE;
 	}
 	return 0;
 }
@@ -127,12 +146,17 @@ static void vduse_domain_unmap_bounce_page(struct vduse_iova_domain *domain,
 	u64 last = iova + size - 1;
 
 	while (iova <= last) {
-		map = &domain->bounce_maps[iova >> PAGE_SHIFT];
+		map = &domain->bounce_maps[iova >> BOUNCE_MAP_SHIFT];
 		map->orig_phys = INVALID_PHYS_ADDR;
-		iova += PAGE_SIZE;
+		iova += BOUNCE_MAP_SIZE;
 	}
 }
 
+static unsigned int offset_in_bounce_page(dma_addr_t addr)
+{
+	return (addr & ~BOUNCE_MAP_MASK);
+}
+
 static void do_bounce(phys_addr_t orig, void *addr, size_t size,
 		      enum dma_data_direction dir)
 {
@@ -163,7 +187,7 @@ static void vduse_domain_bounce(struct vduse_iova_domain *domain,
 				enum dma_data_direction dir)
 {
 	struct vduse_bounce_map *map;
-	unsigned int offset;
+	unsigned int offset, head_offset;
 	void *addr;
 	size_t sz;
 
@@ -171,15 +195,16 @@ static void vduse_domain_bounce(struct vduse_iova_domain *domain,
 		return;
 
 	while (size) {
-		map = &domain->bounce_maps[iova >> PAGE_SHIFT];
-		offset = offset_in_page(iova);
-		sz = min_t(size_t, PAGE_SIZE - offset, size);
+		map = &domain->bounce_maps[iova >> BOUNCE_MAP_SHIFT];
+		head_offset = offset_in_page(iova);
+		offset = offset_in_bounce_page(iova);
+		sz = min_t(size_t, BOUNCE_MAP_SIZE - offset, size);
 
 		if (WARN_ON(!map->bounce_page ||
 			    map->orig_phys == INVALID_PHYS_ADDR))
 			return;
 
-		addr = page_address(map->bounce_page) + offset;
+		addr = page_address(map->bounce_page) + head_offset;
 		do_bounce(map->orig_phys + offset, addr, sz, dir);
 		size -= sz;
 		iova += sz;
@@ -214,7 +239,7 @@ vduse_domain_get_bounce_page(struct vduse_iova_domain *domain, u64 iova)
 	struct page *page = NULL;
 
 	spin_lock(&domain->iotlb_lock);
-	map = &domain->bounce_maps[iova >> PAGE_SHIFT];
+	map = &domain->bounce_maps[iova >> BOUNCE_MAP_SHIFT];
 	if (!map->bounce_page)
 		goto out;
 
@@ -232,7 +257,7 @@ vduse_domain_free_bounce_pages(struct vduse_iova_domain *domain)
 	struct vduse_bounce_map *map;
 	unsigned long pfn, bounce_pfns;
 
-	bounce_pfns = domain->bounce_size >> PAGE_SHIFT;
+	bounce_pfns = domain->bounce_size >> BOUNCE_MAP_SHIFT;
 
 	for (pfn = 0; pfn < bounce_pfns; pfn++) {
 		map = &domain->bounce_maps[pfn];
@@ -242,7 +267,8 @@ vduse_domain_free_bounce_pages(struct vduse_iova_domain *domain)
 		if (!map->bounce_page)
 			continue;
 
-		__free_page(map->bounce_page);
+		if (!((pfn << BOUNCE_MAP_SHIFT) & ~PAGE_MASK))
+			__free_page(map->bounce_page);
 		map->bounce_page = NULL;
 	}
 }
@@ -489,7 +515,7 @@ vduse_domain_create(unsigned long iova_limit, size_t bounce_size)
 	struct vduse_bounce_map *map;
 	unsigned long pfn, bounce_pfns;
 
-	bounce_pfns = PAGE_ALIGN(bounce_size) >> PAGE_SHIFT;
+	bounce_pfns = PAGE_ALIGN(bounce_size) >> BOUNCE_MAP_SHIFT;
 	if (iova_limit <= bounce_size)
 		return NULL;
 
@@ -520,7 +546,7 @@ vduse_domain_create(unsigned long iova_limit, size_t bounce_size)
 	domain->file = file;
 	spin_lock_init(&domain->iotlb_lock);
 	init_iova_domain(&domain->stream_iovad,
-			PAGE_SIZE, IOVA_START_PFN);
+			BOUNCE_MAP_SIZE, IOVA_START_PFN);
 	init_iova_domain(&domain->consistent_iovad,
 			PAGE_SIZE, bounce_pfns);
 
diff --git a/drivers/vdpa/vdpa_user/iova_domain.h b/drivers/vdpa/vdpa_user/iova_domain.h
index 2722d9b8e21aff..b87dac1137d7c7 100644
--- a/drivers/vdpa/vdpa_user/iova_domain.h
+++ b/drivers/vdpa/vdpa_user/iova_domain.h
@@ -19,6 +19,11 @@
 
 #define INVALID_PHYS_ADDR (~(phys_addr_t)0)
 
+#define BOUNCE_MAP_SHIFT	12
+#define BOUNCE_MAP_SIZE	(1 << BOUNCE_MAP_SHIFT)
+#define BOUNCE_MAP_MASK	(~(BOUNCE_MAP_SIZE - 1))
+#define BOUNCE_MAP_ALIGN(addr)	(((addr) + BOUNCE_MAP_SIZE - 1) & ~(BOUNCE_MAP_SIZE - 1))
+
 struct vduse_bounce_map {
 	struct page *bounce_page;
 	u64 orig_phys;
-- 
2.53.0