dpdk: move DPDK vfio hack to dpdk plugin

Change-Id: I806cbf8c6c49643fe6c317bcceab93c1b9d441ab
Signed-off-by: Damjan Marion <damarion@cisco.com>

4 files changed, 70 insertions, 50 deletions

diff --git a/src/plugins/dpdk/buffer.c b/src/plugins/dpdk/buffer.c
index f6379a1da4a..47d66484b3d 100644
--- a/src/plugins/dpdk/buffer.c
+++ b/src/plugins/dpdk/buffer.c
@@ -43,6 +43,10 @@
  * Allocate/free network buffers.
  */
 
+#include <unistd.h>
+#include <linux/vfio.h>
+#include <sys/ioctl.h>
+
 #include <rte_config.h>
 
 #include <rte_common.h>
@@ -71,11 +75,13 @@
 #include <rte_version.h>
 
 #include <vlib/vlib.h>
+#include <vlib/unix/unix.h>
+#include <vlib/pci/pci.h>
+#include <vlib/linux/vfio.h>
 #include <vnet/vnet.h>
 #include <dpdk/device/dpdk.h>
 #include <dpdk/device/dpdk_priv.h>
 
-
 STATIC_ASSERT (VLIB_BUFFER_PRE_DATA_SIZE == RTE_PKTMBUF_HEADROOM,
 	       "VLIB_BUFFER_PRE_DATA_SIZE must be equal to RTE_PKTMBUF_HEADROOM");
 
@@ -91,6 +97,7 @@ typedef struct
 
 typedef struct
 {
+  int vfio_container_fd;
   dpdk_buffer_per_thread_data *ptd;
 } dpdk_buffer_main_t;
 
@@ -459,11 +466,38 @@ dpdk_packet_template_init (vlib_main_t * vm,
   vlib_worker_thread_barrier_release (vm);
 }
 
+static clib_error_t *
+scan_vfio_fd (void *arg, u8 * path_name, u8 * file_name)
+{
+  dpdk_buffer_main_t *dbm = &dpdk_buffer_main;
+  linux_vfio_main_t *lvm = &vfio_main;
+  const char fn[] = "/dev/vfio/vfio";
+  char buff[sizeof (fn)] = { 0 };
+  int fd;
+  u8 *path = format (0, "%v%c", path_name, 0);
+
+  if (readlink ((char *) path, buff, sizeof (fn)) + 1 != sizeof (fn))
+    goto done;
+
+  if (strncmp (fn, buff, sizeof (fn)))
+    goto done;
+
+  fd = atoi ((char *) file_name);
+  if (fd != lvm->container_fd)
+    dbm->vfio_container_fd = fd;
+
+done:
+  vec_free (path);
+  return 0;
+}
+
 clib_error_t *
 dpdk_pool_create (vlib_main_t * vm, u8 * pool_name, u32 elt_size,
 		  u32 num_elts, u32 pool_priv_size, u16 cache_size, u8 numa,
-		  struct rte_mempool **_mp, vlib_physmem_region_index_t * pri)
+		  struct rte_mempool ** _mp,
+		  vlib_physmem_region_index_t * pri)
 {
+  dpdk_buffer_main_t *dbm = &dpdk_buffer_main;
   struct rte_mempool *mp;
   vlib_physmem_region_t *pr;
   clib_error_t *error = 0;
@@ -501,6 +535,33 @@ dpdk_pool_create (vlib_main_t * vm, u8 * pool_name, u32 elt_size,
 
   _mp[0] = mp;
 
+  /* DPDK currently doesn't provide API to map DMA memory for empty mempool
+     so we are using this hack, will be nice to have at least API to get
+     VFIO container FD */
+  if (dbm->vfio_container_fd == -1)
+    foreach_directory_file ("/proc/self/fd", scan_vfio_fd, 0, 0);
+
+  if (dbm->vfio_container_fd != -1)
+    {
+      struct vfio_iommu_type1_dma_map dm = { 0 };
+      int i, rv = 0;
+      dm.argsz = sizeof (struct vfio_iommu_type1_dma_map);
+      dm.flags = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE;
+
+      /* *INDENT-OFF* */
+      vec_foreach_index (i, pr->page_table)
+	{
+	  dm.vaddr = pointer_to_uword (pr->mem) + (i << pr->log2_page_size);
+	  dm.size = 1 << pr->log2_page_size;
+	  dm.iova = pr->page_table[i];
+	  if ((rv = ioctl (dbm->vfio_container_fd, VFIO_IOMMU_MAP_DMA, &dm)))
+	    break;
+	}
+      /* *INDENT-ON* */
+      if (rv != 0 && rv != EINVAL)
+	clib_unix_warning ("ioctl(VFIO_IOMMU_MAP_DMA) pool '%s'", pool_name);
+    }
+
   return 0;
 }
 
@@ -665,8 +726,12 @@ dpdk_buffer_init (vlib_main_t * vm)
 {
   dpdk_buffer_main_t *dbm = &dpdk_buffer_main;
   vlib_thread_main_t *tm = vlib_get_thread_main ();
+
   vec_validate_aligned (dbm->ptd, tm->n_vlib_mains - 1,
 			CLIB_CACHE_LINE_BYTES);
+
+  dbm->vfio_container_fd = -1;
+
   return 0;
 }
 
diff --git a/src/vlib/linux/physmem.c b/src/vlib/linux/physmem.c
index cf2961b4cf9..1a449921857 100644
--- a/src/vlib/linux/physmem.c
+++ b/src/vlib/linux/physmem.c
@@ -174,16 +174,8 @@ unix_physmem_region_alloc (vlib_main_t * vm, char *name, u32 size,
 	}
     }
 
-#if 0
-  if ((vpm->flags & VLIB_PHYSMEM_MAIN_F_HAVE_IOMMU) ||
-      (vpm->flags & VLIB_PHYSMEM_MAIN_F_HAVE_PAGEMAP) == 0)
-    for (i = 0; i < pr->n_pages; i++)
-      vec_add1 (pr->page_table, pointer_to_uword (pr->mem) +
-		i * (1 << pr->log2_page_size));
-  else
-#endif
-    pr->page_table = clib_mem_vm_get_paddr (pr->mem, pr->log2_page_size,
-					    pr->n_pages);
+  pr->page_table = clib_mem_vm_get_paddr (pr->mem, pr->log2_page_size,
+					  pr->n_pages);
 
   linux_vfio_dma_map_regions (vm);
 
@@ -224,7 +216,6 @@ clib_error_t *
 unix_physmem_init (vlib_main_t * vm)
 {
   vlib_physmem_main_t *vpm = &physmem_main;
-  linux_vfio_main_t *lvm = &vfio_main;
   clib_error_t *error = 0;
   u64 *pt = 0;
 
@@ -241,9 +232,6 @@ unix_physmem_init (vlib_main_t * vm)
   if ((error = linux_vfio_init (vm)))
     return error;
 
-  if (lvm->flags & LINUX_VFIO_F_HAVE_IOMMU)
-    vpm->flags |= VLIB_PHYSMEM_MAIN_F_HAVE_IOMMU;
-
   vm->os_physmem_alloc_aligned = unix_physmem_alloc_aligned;
   vm->os_physmem_free = unix_physmem_free;
   vm->os_physmem_region_alloc = unix_physmem_region_alloc;
diff --git a/src/vlib/linux/vfio.c b/src/vlib/linux/vfio.c
index e39bf01fdc0..621dfb2ce01 100644
--- a/src/vlib/linux/vfio.c
+++ b/src/vlib/linux/vfio.c
@@ -73,7 +73,7 @@ map_regions (vlib_main_t * vm, int fd)
 	  int rv;
 	  dm.vaddr = pointer_to_uword (pr->mem) + (i << pr->log2_page_size);
 	  dm.size = 1 << pr->log2_page_size;
-	  dm.iova = pr->page_table[i];
+	  dm.iova = dm.vaddr;
 	  if ((rv = ioctl (fd, VFIO_IOMMU_MAP_DMA, &dm)))
 	    return rv;
         }
@@ -82,30 +82,6 @@ map_regions (vlib_main_t * vm, int fd)
   return 0;
 }
 
-static clib_error_t *
-scan_vfio_fd (void *arg, u8 * path_name, u8 * file_name)
-{
-  linux_vfio_main_t *lvm = &vfio_main;
-  const char fn[] = "/dev/vfio/vfio";
-  char buff[sizeof (fn)] = { 0 };
-  int fd;
-  u8 *path = format (0, "%v%c", path_name, 0);
-
-  if (readlink ((char *) path, buff, sizeof (fn)) + 1 != sizeof (fn))
-    goto done;
-
-  if (strncmp (fn, buff, sizeof (fn)))
-    goto done;
-
-  fd = atoi ((char *) file_name);
-  if (fd != lvm->container_fd)
-    lvm->ext_container_fd = atoi ((char *) file_name);
-
-done:
-  vec_free (path);
-  return 0;
-}
-
 void
 linux_vfio_dma_map_regions (vlib_main_t * vm)
 {
@@ -113,12 +89,6 @@ linux_vfio_dma_map_regions (vlib_main_t * vm)
 
   if (lvm->container_fd != -1)
     map_regions (vm, lvm->container_fd);
-
-  if (lvm->ext_container_fd == -1)
-    foreach_directory_file ("/proc/self/fd", scan_vfio_fd, 0, 0);
-
-  if (lvm->ext_container_fd != -1)
-    map_regions (vm, lvm->ext_container_fd);
 }
 
 static linux_pci_vfio_iommu_group_t *
@@ -252,8 +222,6 @@ linux_vfio_init (vlib_main_t * vm)
   linux_vfio_main_t *lvm = &vfio_main;
   int fd;
 
-  lvm->ext_container_fd = -1;
-
   fd = open ("/dev/vfio/vfio", O_RDWR);
 
   /* check if iommu is available */
diff --git a/src/vlib/linux/vfio.h b/src/vlib/linux/vfio.h
index 8e0758cf2e4..e23cee19931 100644
--- a/src/vlib/linux/vfio.h
+++ b/src/vlib/linux/vfio.h
@@ -28,7 +28,6 @@ typedef struct
   u32 flags;
 #define LINUX_VFIO_F_HAVE_IOMMU		(1 << 0)
   int container_fd;
-  int ext_container_fd;		/* container fd used by external library, i.e DPDK */
 
   /* VFIO */
   int iommu_mode;