vfio: selftests: Align BAR mmaps for efficient IOMMU mapping

Update vfio_pci_bar_map() to align BAR mmaps for efficient huge page
mappings. The manual mmap alignment can be removed once mmap(!MAP_FIXED)
on vfio device fds improves to automatically return well-aligned
addresses.

Also add MADV_HUGEPAGE, which encourages the kernel to use huge pages
(e.g. when /sys/kernel/mm/transparent_hugepage/enabled is set to "madvise").

Drop MAP_FILE from mmap(). It is an ignored compatibility flag.

Signed-off-by: Alex Mastro <amastro@fb.com>
Reviewed-by: David Matlack <dmatlack@google.com>
Tested-by: David Matlack <dmatlack@google.com>
Link: https://lore.kernel.org/r/20260114-map-mmio-test-v3-2-44e036d95e64@fb.com
Signed-off-by: Alex Williamson <alex@shazbot.org>

3 files changed, 57 insertions, 1 deletions

diff --git a/tools/testing/selftests/vfio/lib/include/libvfio.h b/tools/testing/selftests/vfio/lib/include/libvfio.h
index 279ddcd70194..1b6da54cc2cb 100644
--- a/tools/testing/selftests/vfio/lib/include/libvfio.h
+++ b/tools/testing/selftests/vfio/lib/include/libvfio.h
@@ -23,4 +23,13 @@
 const char *vfio_selftests_get_bdf(int *argc, char *argv[]);
 char **vfio_selftests_get_bdfs(int *argc, char *argv[], int *nr_bdfs);
 
+/*
+ * Reserve virtual address space of size at an address satisfying
+ * (vaddr % align) == offset.
+ *
+ * Returns the reserved vaddr. The caller is responsible for unmapping
+ * the returned region.
+ */
+void *mmap_reserve(size_t size, size_t align, size_t offset);
+
 #endif /* SELFTESTS_VFIO_LIB_INCLUDE_LIBVFIO_H */
diff --git a/tools/testing/selftests/vfio/lib/libvfio.c b/tools/testing/selftests/vfio/lib/libvfio.c
index a23a3cc5be69..3a3d1ed635c1 100644
--- a/tools/testing/selftests/vfio/lib/libvfio.c
+++ b/tools/testing/selftests/vfio/lib/libvfio.c
@@ -2,6 +2,9 @@
 
 #include <stdio.h>
 #include <stdlib.h>
+#include <sys/mman.h>
+
+#include <linux/align.h>
 
 #include "../../../kselftest.h"
 #include <libvfio.h>
@@ -76,3 +79,25 @@ const char *vfio_selftests_get_bdf(int *argc, char *argv[])
 
 	return vfio_selftests_get_bdfs(argc, argv, &nr_bdfs)[0];
 }
+
+void *mmap_reserve(size_t size, size_t align, size_t offset)
+{
+	void *map_base, *map_align;
+	size_t delta;
+
+	VFIO_ASSERT_GT(align, offset);
+	delta = align - offset;
+
+	map_base = mmap(NULL, size + align, PROT_NONE,
+			MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+	VFIO_ASSERT_NE(map_base, MAP_FAILED);
+
+	map_align = (void *)(ALIGN((uintptr_t)map_base + delta, align) - delta);
+
+	if (map_align > map_base)
+		VFIO_ASSERT_EQ(munmap(map_base, map_align - map_base), 0);
+
+	VFIO_ASSERT_EQ(munmap(map_align + size, map_base + align - map_align), 0);
+
+	return map_align;
+}
diff --git a/tools/testing/selftests/vfio/lib/vfio_pci_device.c b/tools/testing/selftests/vfio/lib/vfio_pci_device.c
index fac4c0ecadef..4e5871f1ebc3 100644
--- a/tools/testing/selftests/vfio/lib/vfio_pci_device.c
+++ b/tools/testing/selftests/vfio/lib/vfio_pci_device.c
@@ -11,10 +11,14 @@
 #include <sys/ioctl.h>
 #include <sys/mman.h>
 
+#include <linux/align.h>
 #include <linux/iommufd.h>
+#include <linux/kernel.h>
 #include <linux/limits.h>
+#include <linux/log2.h>
 #include <linux/mman.h>
 #include <linux/overflow.h>
+#include <linux/sizes.h>
 #include <linux/types.h>
 #include <linux/vfio.h>
 
@@ -123,20 +127,38 @@ static void vfio_pci_region_get(struct vfio_pci_device *device, int index,
 static void vfio_pci_bar_map(struct vfio_pci_device *device, int index)
 {
 	struct vfio_pci_bar *bar = &device->bars[index];
+	size_t align, size;
 	int prot = 0;
+	void *vaddr;
 
 	VFIO_ASSERT_LT(index, PCI_STD_NUM_BARS);
 	VFIO_ASSERT_NULL(bar->vaddr);
 	VFIO_ASSERT_TRUE(bar->info.flags & VFIO_REGION_INFO_FLAG_MMAP);
+	VFIO_ASSERT_TRUE(is_power_of_2(bar->info.size));
 
 	if (bar->info.flags & VFIO_REGION_INFO_FLAG_READ)
 		prot |= PROT_READ;
 	if (bar->info.flags & VFIO_REGION_INFO_FLAG_WRITE)
 		prot |= PROT_WRITE;
 
-	bar->vaddr = mmap(NULL, bar->info.size, prot, MAP_FILE | MAP_SHARED,
+	size = bar->info.size;
+
+	/*
+	 * Align BAR mmaps to improve page fault granularity during potential
+	 * subsequent IOMMU mapping of these BAR vaddr. 1G for x86 is the
+	 * largest hugepage size across any architecture, so no benefit from
+	 * larger alignment. BARs smaller than 1G will be aligned by their
+	 * power-of-two size, guaranteeing sufficient alignment for smaller
+	 * hugepages, if present.
+	 */
+	align = min_t(size_t, size, SZ_1G);
+
+	vaddr = mmap_reserve(size, align, 0);
+	bar->vaddr = mmap(vaddr, size, prot, MAP_SHARED | MAP_FIXED,
 			  device->fd, bar->info.offset);
 	VFIO_ASSERT_NE(bar->vaddr, MAP_FAILED);
+
+	madvise(bar->vaddr, size, MADV_HUGEPAGE);
 }
 
 static void vfio_pci_bar_unmap(struct vfio_pci_device *device, int index)