Managed to map/pin DMA memory

6 files changed, 185 insertions, 18 deletions

diff --git a/.gitignore b/.gitignore
index f45ab32..8d716c0 100644
--- a/.gitignore
+++ b/.gitignore
@@ -3,7 +3,9 @@
 *.cmd
 Module.symvers
 modules.order
-test_dma
 vfio-netem.mod.c
 .tmp_versions/
 tags
+ether_test
+map_unmap
+
diff --git a/README b/README
index a62cce1..ff0b8ac 100644
--- a/README
+++ b/README
@@ -20,3 +20,10 @@ sudo ./pcimem /sys/devices/pci0000:00/0000:00:1c.3/0000:04:00.0/resource2 0 w
 sudo ./pcimem /sys/devices/pci0000:00/0000:00:1c.3/0000:04:00.0/resource4 16368 w
 
 run ./test_dma
+
+# Enable kernel debug
+echo file drivers/iommu/iommu.c line 1533-1600 +p > /sys/kernel/debug/dynamic_debug/control
+or
+echo file drivers/iommu/iommu.c func iommu_map +p > /sys/kernel/debug/dynamic_debug/control
+or
+echo file drivers/iommu/iommu.c  +p > /sys/kernel/debug/dynamic_debug/control
diff --git a/build.sh b/build.sh
index 2053d28..54ef7e7 100755
--- a/build.sh
+++ b/build.sh
@@ -2,7 +2,9 @@
 what=$1
 if [ -n "$1" -a "$1" = 'clean' ]; then
 	make -C /lib/modules/`uname -r`/build M=`pwd` modules clean
+	rm map_unmap ether_test -f
 else
 	make -C /lib/modules/`uname -r`/build M=`pwd` modules
-	gcc test.c -o test_dma -Werror -Wall -Wunused -O2
+	gcc ether-test.c -o ether_test -Werror -Wall -Wunused -O2
+	gcc vfio-iommu-map-unmap.c -o map_unmap -Werror -Wall -Wunused -O2
 fi
diff --git a/test.c b/ether-test.c
index 764b8d3..002aaf6 100644
--- a/test.c
+++ b/ether-test.c
@@ -13,6 +13,7 @@
 #include "reg_api.h"
 
 const char glob_uuid[] = "d0b24768-9c51-11e7-8de5-c7b0c2769e69";
+/* iommu alignment */
 #define TEST_SIZE 1024 * 1024
 #define EOR (1 << 30)
 #define OWN (1 << 31)
diff --git a/vfio-iommu-map-unmap.c b/vfio-iommu-map-unmap.c
new file mode 100644
index 0000000..a5d5c46
--- /dev/null
+++ b/vfio-iommu-map-unmap.c
@@ -0,0 +1,161 @@
+#include <errno.h>
+#include <libgen.h>
+#include <fcntl.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <unistd.h>
+#include <sys/ioctl.h>
+#include <sys/mman.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+
+#include <linux/ioctl.h>
+#include <linux/vfio.h>
+
+#define MAP_SIZE (1UL * 1024 * 1024 * 1024)
+#define MAP_CHUNK (4 * 1024)
+#define REALLOC_INTERVAL 30
+
+void usage(char *name)
+{
+	printf("usage: %s <group id> <allocs> \n", name);
+}
+
+int main(int argc, char **argv)
+{
+	int ret, container, group;
+	char path[64];
+	unsigned long i, count, max;
+	void **maps;
+	struct vfio_group_status group_status = {
+		.argsz = sizeof(group_status)
+	};
+	struct vfio_iommu_type1_dma_map dma_map = {
+		.argsz = sizeof(dma_map)
+	};
+	struct vfio_iommu_type1_dma_unmap dma_unmap = {
+		.argsz = sizeof(dma_unmap)
+	};
+
+	if (argc != 3) {
+		usage(argv[0]);
+		return -1;
+	}
+
+	container = open("/dev/vfio/vfio", O_RDWR);
+	if (container < 0) {
+		printf("Failed to open /dev/vfio/vfio, %d (%s)\n",
+		       container, strerror(errno));
+		return container;
+	}
+
+	snprintf(path, sizeof(path), "/dev/vfio/%d", atoi(argv[1]));
+	group = open(path, O_RDWR);
+	if (group < 0) {
+		printf("Failed to open %s, %d (%s)\n",
+		       path, group, strerror(errno));
+		return group;
+	}
+	max = atoi(argv[2]);
+
+	ret = ioctl(group, VFIO_GROUP_GET_STATUS, &group_status);
+	if (ret) {
+		printf("ioctl(VFIO_GROUP_GET_STATUS) failed\n");
+		return ret;
+	}
+
+	if (!(group_status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
+		printf("Group not viable, are all devices attached to vfio?\n");
+		return -1;
+	}
+
+	ret = ioctl(group, VFIO_GROUP_SET_CONTAINER, &container);
+	if (ret) {
+		printf("Failed to set group container\n");
+		return ret;
+	}
+
+	ret = ioctl(container, VFIO_SET_IOMMU, VFIO_TYPE1_IOMMU);
+	if (ret) {
+		printf("Failed to set IOMMU\n");
+		return ret;
+	}
+
+	/* Test code */
+	dma_map.flags = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE;
+	dma_map.size = MAP_CHUNK;
+	dma_unmap.size = MAP_SIZE;
+	dma_unmap.iova = 0;
+
+	/* Track our mmaps for re-use */
+	maps = malloc(sizeof(void *) * (MAP_SIZE/dma_map.size));
+	if (!maps) {
+		printf("Failed to allocate map (%s)\n", strerror(errno));
+		return -1;
+	}
+
+	memset(maps, 0, sizeof(void *) * (MAP_SIZE/dma_map.size));
+
+	for (count = 0; count < max; count++) {
+
+		/* Every REALLOC_INTERVAL, dump our mappings to give THP something to collapse */
+		if (count % REALLOC_INTERVAL == 0) {
+			for (i = 0; i < MAP_SIZE/dma_map.size; i++) {
+				if (maps[i]) {
+					munmap(maps[i], dma_map.size);
+					maps[i] = NULL;
+				}
+			}
+			if (count) {
+				printf("\t%ld\n", count);
+				//return 0;
+			}
+			printf("|");
+			fflush(stdout);
+		}
+
+		/* Map MAP_CHUNK at a time, each chunk is pinned on map, so THP can't do anything until unmap */
+		for (i = dma_map.iova = 0; i < MAP_SIZE/dma_map.size; i++, dma_map.iova += dma_map.size) {
+			if (!maps[i]) {
+				maps[i] = mmap(NULL, dma_map.size,
+						PROT_READ | PROT_WRITE,
+						MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+				if (maps[i] == MAP_FAILED) {
+					printf("Failed to mmap memory (%s)\n", strerror(errno));
+					return -1;
+				}
+			}
+
+			ret = madvise(maps[i], dma_map.size, MADV_HUGEPAGE);
+			if (ret) {
+				printf("Madvise failed (%s)\n", strerror(errno));
+			}
+
+			dma_map.vaddr = (unsigned long)maps[i];
+
+			ret = ioctl(container, VFIO_IOMMU_MAP_DMA, &dma_map);
+			if (ret) {
+				printf("Failed to map memory (%s)\n",
+					strerror(errno));
+				return ret;
+			}
+		}
+
+		printf("+");
+		fflush(stdout);
+
+		/* Unmap everything at once */
+		ret = ioctl(container, VFIO_IOMMU_UNMAP_DMA, &dma_unmap);
+		if (ret) {
+			printf("Failed to unmap memory (%s)\n", strerror(errno));
+			return ret;
+		}
+
+		printf("-");
+		fflush(stdout);
+	}
+	printf("\n");
+
+	return 0;
+}
diff --git a/vfio-netem.c b/vfio-netem.c
index 2083007..97ff951 100644
--- a/vfio-netem.c
+++ b/vfio-netem.c
@@ -138,19 +138,16 @@ static const struct attribute_group *vfio_attr_groups[] = {
 
 static int vfio_net_create(struct kobject *kobj, struct mdev_device *mdev)
 {
-	struct device *dev = mdev_parent_dev(mdev);
-	//struct iommu_domain *domain;
-
-	dev_info(dev, "MDEV-NETEM: Create on parent\n");
 
-	//domain = iommu_get_domain_for_dev(dev);
-	//if (domain)
-		//iommu_detach_device(domain, dev);
-	if (dev && dev->driver && dev->driver->name)
-		printk("AA %s\n", dev->driver->name);
+	struct device *dev = mdev_parent_dev(mdev);
+	struct iommu_domain *domain;
 
-	if (dev->parent && dev->parent->driver && dev->parent->driver->name)
-		printk("BB %s\n", dev->parent->driver->name);
+	dev_info(dev, "MDEV-NETEM: Create mdev for %s\n", dev_name(dev));
+	domain = iommu_get_domain_for_dev(dev->parent->parent->parent);
+	if (domain) {
+		printk("AA Detach domain\n");
+		iommu_detach_device(domain, dev->parent->parent->parent);
+	}
 
 	/* Inherit parent's iommu_ops */
 	if (dev && dev->parent && dev->parent->bus &&
@@ -158,7 +155,8 @@ static int vfio_net_create(struct kobject *kobj, struct mdev_device *mdev)
 		mdev->dev.bus->iommu_ops = dev->parent->bus->iommu_ops;
 		//XXX dont do that
 		//mdev->dev.iommu_group = dev->parent->iommu_group;
-		dev_info(dev, "MDEV-NETEM: Set mdev iommu_ops\n");
+		dev_info(dev, "MDEV-NETEM: Inherit %s iommu_ops\n",
+			 dev_name(dev->parent));
 	}
 
 	return 0;
@@ -338,9 +336,6 @@ int vfio_vring_init(struct device *dev)
 	}
 
 	err = mdev_register_device(dev, &vfio_net_ops);
-	if (!err)
-		dev_info(dev, "MDEV-NETEM: Register for driver %s",
-			 pdev->driver->name);
 
 	return err;
 
@@ -351,7 +346,6 @@ out_inv_dev:
 
 void vfio_vring_destroy(struct device *dev)
 {
-	dev_info(dev, "MDEV-NETEM: Unregistering\n");
 	mdev_unregister_device(dev);
 }