vlib: PCI rework to support VFIO

Also fixes old ixge driver, so it works with recent physmem changes and
vfio.

Change-Id: Id4be74b34daed47cd281a77eec43d6692340d882
Signed-off-by: Damjan Marion <damarion@cisco.com>

7 files changed, 835 insertions, 432 deletions

diff --git a/src/plugins/dpdk/device/init.c b/src/plugins/dpdk/device/init.c
index 55904f018da..20272853035 100755
--- a/src/plugins/dpdk/device/init.c
+++ b/src/plugins/dpdk/device/init.c
@@ -677,24 +677,37 @@ dpdk_lib_init (dpdk_main_t * dm)
 static void
 dpdk_bind_devices_to_uio (dpdk_config_main_t * conf)
 {
-  vlib_pci_main_t *pm = &pci_main;
   clib_error_t *error;
-  vlib_pci_device_t *d;
   u8 *pci_addr = 0;
   int num_whitelisted = vec_len (conf->dev_confs);
+  vlib_pci_device_info_t *d = 0;
+  vlib_pci_addr_t *addr = 0, *addrs;
 
+  addrs = vlib_pci_get_all_dev_addrs ();
   /* *INDENT-OFF* */
-  pool_foreach (d, pm->pci_devs, ({
+  vec_foreach (addr, addrs)
+    {
     dpdk_device_config_t * devconf = 0;
     vec_reset_length (pci_addr);
-    pci_addr = format (pci_addr, "%U%c", format_vlib_pci_addr, &d->bus_address, 0);
+    pci_addr = format (pci_addr, "%U%c", format_vlib_pci_addr, addr, 0);
+    if (d)
+    {
+      vlib_pci_free_device_info (d);
+      d = 0;
+      }
+    d = vlib_pci_get_device_info (addr, &error);
+    if (error)
+    {
+      clib_error_report (error);
+      continue;
+    }
 
     if (d->device_class != PCI_CLASS_NETWORK_ETHERNET && d->device_class != PCI_CLASS_PROCESSOR_CO)
       continue;
 
     if (num_whitelisted)
       {
-	uword * p = hash_get (conf->device_config_index_by_pci_addr, d->bus_address.as_u32);
+	uword * p = hash_get (conf->device_config_index_by_pci_addr, addr->as_u32);
 
 	if (!p)
 	  continue;
@@ -736,23 +749,24 @@ dpdk_bind_devices_to_uio (dpdk_config_main_t * conf)
         continue;
       }
 
-    error = vlib_pci_bind_to_uio (d, (char *) conf->uio_driver_name);
+    error = vlib_pci_bind_to_uio (addr, (char *) conf->uio_driver_name);
 
     if (error)
       {
 	if (devconf == 0)
 	  {
 	    pool_get (conf->dev_confs, devconf);
-	    hash_set (conf->device_config_index_by_pci_addr, d->bus_address.as_u32,
+	    hash_set (conf->device_config_index_by_pci_addr, addr->as_u32,
 		      devconf - conf->dev_confs);
-	    devconf->pci_addr.as_u32 = d->bus_address.as_u32;
+	    devconf->pci_addr.as_u32 = addr->as_u32;
 	  }
 	devconf->is_blacklisted = 1;
 	clib_error_report (error);
       }
-  }));
+  }
   /* *INDENT-ON* */
   vec_free (pci_addr);
+  vlib_pci_free_device_info (d);
 }
 
 static clib_error_t *
diff --git a/src/plugins/ixge/ixge.c b/src/plugins/ixge/ixge.c
index bf9424255b8..62f484e6cca 100644
--- a/src/plugins/ixge/ixge.c
+++ b/src/plugins/ixge/ixge.c
@@ -1395,9 +1395,8 @@ ixge_rx_queue_no_wrap (ixge_main_t * xm,
 	vec_resize (xm->rx_buffers_to_add, n_to_alloc);
 
 	_vec_len (xm->rx_buffers_to_add) = l;
-	n_allocated = vlib_buffer_alloc_from_free_list
-	  (vm, xm->rx_buffers_to_add + l, n_to_alloc,
-	   xm->vlib_buffer_free_list_index);
+	n_allocated =
+	  vlib_buffer_alloc (vm, xm->rx_buffers_to_add + l, n_to_alloc);
 	_vec_len (xm->rx_buffers_to_add) += n_allocated;
 
 	/* Handle transient allocation failure */
@@ -1462,6 +1461,7 @@ ixge_rx_queue_no_wrap (ixge_main_t * xm,
 	  to_rx += 2;
 	  to_add -= 2;
 
+#if 0
 	  ASSERT (VLIB_BUFFER_KNOWN_ALLOCATED ==
 		  vlib_buffer_is_known (vm, bi0));
 	  ASSERT (VLIB_BUFFER_KNOWN_ALLOCATED ==
@@ -1470,6 +1470,7 @@ ixge_rx_queue_no_wrap (ixge_main_t * xm,
 		  vlib_buffer_is_known (vm, fi0));
 	  ASSERT (VLIB_BUFFER_KNOWN_ALLOCATED ==
 		  vlib_buffer_is_known (vm, fi1));
+#endif
 
 	  b0 = vlib_get_buffer (vm, bi0);
 	  b1 = vlib_get_buffer (vm, bi1);
@@ -1678,10 +1679,12 @@ ixge_rx_queue_no_wrap (ixge_main_t * xm,
 	  to_rx += 1;
 	  to_add -= 1;
 
+#if 0
 	  ASSERT (VLIB_BUFFER_KNOWN_ALLOCATED ==
 		  vlib_buffer_is_known (vm, bi0));
 	  ASSERT (VLIB_BUFFER_KNOWN_ALLOCATED ==
 		  vlib_buffer_is_known (vm, fi0));
+#endif
 
 	  b0 = vlib_get_buffer (vm, bi0);
 
@@ -2253,8 +2256,9 @@ format_ixge_device_name (u8 * s, va_list * args)
   u32 i = va_arg (*args, u32);
   ixge_main_t *xm = &ixge_main;
   ixge_device_t *xd = vec_elt_at_index (xm->devices, i);
-  return format (s, "TenGigabitEthernet%U",
-		 format_vlib_pci_handle, &xd->pci_device.bus_address);
+  vlib_pci_addr_t *addr = vlib_pci_get_addr (xd->pci_dev_handle);
+  return format (s, "TenGigabitEthernet%x/%x/%x/%x",
+		 addr->domain, addr->bus, addr->slot, addr->function);
 }
 
 #define IXGE_COUNTER_IS_64_BIT (1 << 0)
@@ -2355,8 +2359,12 @@ format_ixge_device (u8 * s, va_list * args)
 
   {
 
-    s = format (s, "\n%UPCIe %U", format_white_space, indent + 2,
-		format_vlib_pci_link_speed, &xd->pci_device);
+    vlib_pci_addr_t *addr = vlib_pci_get_addr (xd->pci_dev_handle);
+    vlib_pci_device_info_t *d = vlib_pci_get_device_info (addr, 0);
+
+    if (d)
+      s = format (s, "\n%UPCIe %U", format_white_space, indent + 2,
+		  format_vlib_pci_link_speed, d);
   }
 
   s = format (s, "\n%U", format_white_space, indent + 2);
@@ -2477,13 +2485,6 @@ ixge_dma_init (ixge_device_t * xd, vlib_rx_or_tx_t rt, u32 queue_index)
   if (!xm->n_bytes_in_rx_buffer)
     xm->n_bytes_in_rx_buffer = IXGE_N_BYTES_IN_RX_BUFFER;
   xm->n_bytes_in_rx_buffer = round_pow2 (xm->n_bytes_in_rx_buffer, 1024);
-  if (!xm->vlib_buffer_free_list_index)
-    {
-      xm->vlib_buffer_free_list_index =
-	vlib_buffer_get_or_create_free_list (vm, xm->n_bytes_in_rx_buffer,
-					     "ixge rx");
-      ASSERT (xm->vlib_buffer_free_list_index != 0);
-    }
 
   if (!xm->n_descriptors[rt])
     xm->n_descriptors[rt] = 4 * VLIB_FRAME_SIZE;
@@ -2509,28 +2510,25 @@ ixge_dma_init (ixge_device_t * xd, vlib_rx_or_tx_t rt, u32 queue_index)
     {
       u32 n_alloc, i;
 
-      n_alloc = vlib_buffer_alloc_from_free_list
-	(vm, dq->descriptor_buffer_indices,
-	 vec_len (dq->descriptor_buffer_indices),
-	 xm->vlib_buffer_free_list_index);
+      n_alloc = vlib_buffer_alloc (vm, dq->descriptor_buffer_indices,
+				   vec_len (dq->descriptor_buffer_indices));
       ASSERT (n_alloc == vec_len (dq->descriptor_buffer_indices));
       for (i = 0; i < n_alloc; i++)
 	{
-	  vlib_buffer_t *b =
-	    vlib_get_buffer (vm, dq->descriptor_buffer_indices[i]);
 	  dq->descriptors[i].rx_to_hw.tail_address =
-	    vlib_physmem_virtual_to_physical (vm, xm->physmem_region,
-					      b->data);
+	    vlib_get_buffer_data_physical_address (vm,
+						   dq->descriptor_buffer_indices
+						   [i]);
 	}
     }
   else
     {
       u32 i;
 
-      dq->tx.head_index_write_back =
-	vlib_physmem_alloc (vm,
-			    vm->buffer_main->buffer_pools[0].physmem_region,
-			    &error, CLIB_CACHE_LINE_BYTES);
+      dq->tx.head_index_write_back = vlib_physmem_alloc (vm,
+							 xm->physmem_region,
+							 &error,
+							 CLIB_CACHE_LINE_BYTES);
 
       for (i = 0; i < dq->n_descriptors; i++)
 	dq->descriptors[i].tx = xm->tx_descriptor_template;
@@ -2543,9 +2541,7 @@ ixge_dma_init (ixge_device_t * xd, vlib_rx_or_tx_t rt, u32 queue_index)
     u64 a;
 
     a =
-      vlib_physmem_virtual_to_physical (vm,
-					vm->buffer_main->
-					buffer_pools[0].physmem_region,
+      vlib_physmem_virtual_to_physical (vm, xm->physmem_region,
 					dq->descriptors);
     dr->descriptor_address[0] = a & 0xFFFFFFFF;
     dr->descriptor_address[1] = a >> (u64) 32;
@@ -2571,11 +2567,8 @@ ixge_dma_init (ixge_device_t * xd, vlib_rx_or_tx_t rt, u32 queue_index)
 	/* Make sure its initialized before hardware can get to it. */
 	dq->tx.head_index_write_back[0] = dq->head_index;
 
-	a =
-	  vlib_physmem_virtual_to_physical (vm,
-					    vm->buffer_main->
-					    buffer_pools[0].physmem_region,
-					    dq->tx.head_index_write_back);
+	a = vlib_physmem_virtual_to_physical (vm, xm->physmem_region,
+					      dq->tx.head_index_write_back);
 	dr->tx.head_index_write_back_address[0] = /* enable bit */ 1 | a;
 	dr->tx.head_index_write_back_address[1] = (u64) a >> (u64) 32;
       }
@@ -2838,10 +2831,11 @@ VLIB_INIT_FUNCTION (ixge_init);
 
 
 static void
-ixge_pci_intr_handler (vlib_pci_device_t * dev)
+ixge_pci_intr_handler (vlib_pci_dev_handle_t h)
 {
   ixge_main_t *xm = &ixge_main;
   vlib_main_t *vm = xm->vlib_main;
+  uword private_data = vlib_pci_get_private_data (h);
 
   vlib_node_set_interrupt_pending (vm, ixge_input_node.index);
 
@@ -2849,26 +2843,32 @@ ixge_pci_intr_handler (vlib_pci_device_t * dev)
   {
     vlib_node_runtime_t *rt =
       vlib_node_get_runtime (vm, ixge_input_node.index);
-    rt->runtime_data[0] |= 1 << dev->private_data;
+    rt->runtime_data[0] |= 1 << private_data;
   }
 }
 
 static clib_error_t *
-ixge_pci_init (vlib_main_t * vm, vlib_pci_device_t * dev)
+ixge_pci_init (vlib_main_t * vm, vlib_pci_dev_handle_t h)
 {
   ixge_main_t *xm = &ixge_main;
-  clib_error_t *error;
+  clib_error_t *error = 0;
   void *r;
   ixge_device_t *xd;
+  vlib_pci_addr_t *addr = vlib_pci_get_addr (h);
+  vlib_pci_device_info_t *d = vlib_pci_get_device_info (addr, 0);
 
   /* Allocate physmem region for DMA buffers */
-  error = vlib_physmem_region_alloc (vm, "ixge decriptors", 2 << 20, 0,
-				     VLIB_PHYSMEM_F_INIT_MHEAP,
-				     &xm->physmem_region);
+  if (xm->physmem_region_allocated == 0)
+    {
+      error = vlib_physmem_region_alloc (vm, "ixge decriptors", 2 << 20, 0,
+					 VLIB_PHYSMEM_F_INIT_MHEAP,
+					 &xm->physmem_region);
+      xm->physmem_region_allocated = 1;
+    }
   if (error)
     return error;
 
-  error = vlib_pci_map_resource (dev, 0, &r);
+  error = vlib_pci_map_resource (h, 0, &r);
   if (error)
     return error;
 
@@ -2879,13 +2879,14 @@ ixge_pci_init (vlib_main_t * vm, vlib_pci_device_t * dev)
       ixge_input_node.function = ixge_input_multiarch_select ();
     }
 
-  xd->pci_device = dev[0];
-  xd->device_id = xd->pci_device.config0.header.device_id;
+  xd->pci_dev_handle = h;
+  xd->device_id = d->device_id;
   xd->regs = r;
   xd->device_index = xd - xm->devices;
-  xd->pci_function = dev->bus_address.function;
+  xd->pci_function = addr->function;
   xd->per_interface_next_index = ~0;
 
+  vlib_pci_set_private_data (h, xd->device_index);
 
   /* Chip found so enable node. */
   {
@@ -2894,7 +2895,7 @@ ixge_pci_init (vlib_main_t * vm, vlib_pci_device_t * dev)
 			  ? VLIB_NODE_STATE_POLLING
 			  : VLIB_NODE_STATE_INTERRUPT));
 
-    dev->private_data = xd->device_index;
+    //dev->private_data = xd->device_index;
   }
 
   if (vec_len (xm->devices) == 1)
@@ -2903,12 +2904,12 @@ ixge_pci_init (vlib_main_t * vm, vlib_pci_device_t * dev)
       xm->process_node_index = ixge_process_node.index;
     }
 
-  error = vlib_pci_bus_master_enable (dev);
+  error = vlib_pci_bus_master_enable (h);
 
   if (error)
     return error;
 
-  return vlib_pci_intr_enable (dev);
+  return vlib_pci_intr_enable (h);
 }
 
 /* *INDENT-OFF* */
diff --git a/src/plugins/ixge/ixge.h b/src/plugins/ixge/ixge.h
index a87af1ab04b..c766397525b 100644
--- a/src/plugins/ixge/ixge.h
+++ b/src/plugins/ixge/ixge.h
@@ -1208,7 +1208,7 @@ typedef struct
   u32 per_interface_next_index;
 
   /* PCI bus info. */
-  vlib_pci_device_t pci_device;
+  vlib_pci_dev_handle_t pci_dev_handle;
 
   /* From PCI config space header. */
   ixge_pci_device_id_t device_id;
@@ -1253,8 +1253,6 @@ typedef struct
 
   u32 n_descriptors_per_cache_line;
 
-  u32 vlib_buffer_free_list_index;
-
   u32 process_node_index;
 
   /* Template and mask for initializing/validating TX descriptors. */
@@ -1268,6 +1266,7 @@ typedef struct
   f64 time_last_stats_update;
 
   vlib_physmem_region_index_t physmem_region;
+  int physmem_region_allocated;
 } ixge_main_t;
 
 extern ixge_main_t ixge_main;
diff --git a/src/vlib/buffer_funcs.h b/src/vlib/buffer_funcs.h
index 08f5b3e9010..1ea3c0829b8 100644
--- a/src/vlib/buffer_funcs.h
+++ b/src/vlib/buffer_funcs.h
@@ -162,14 +162,12 @@ vlib_buffer_contents (vlib_main_t * vm, u32 buffer_index, u8 * contents)
 always_inline u64
 vlib_get_buffer_data_physical_address (vlib_main_t * vm, u32 buffer_index)
 {
-  vlib_physmem_region_index_t pri;
+  vlib_buffer_main_t *bm = vm->buffer_main;
   vlib_buffer_t *b = vlib_get_buffer (vm, buffer_index);
-  pri = vm->buffer_main->buffer_pools[b->buffer_pool_index].physmem_region;
-  return vlib_physmem_offset_to_physical (vm, pri,
-					  (((uword) buffer_index) <<
-					   CLIB_LOG2_CACHE_LINE_BYTES) +
-					  STRUCT_OFFSET_OF (vlib_buffer_t,
-							    data));
+  vlib_buffer_pool_t *pool = vec_elt_at_index (bm->buffer_pools,
+					       b->buffer_pool_index);
+
+  return vlib_physmem_virtual_to_physical (vm, pool->physmem_region, b->data);
 }
 
 /** \brief Prefetch buffer metadata by buffer index
diff --git a/src/vlib/linux/pci.c b/src/vlib/linux/pci.c
index 790f168a355..24e2bb594a3 100644
--- a/src/vlib/linux/pci.c
+++ b/src/vlib/linux/pci.c
@@ -51,11 +51,16 @@
 #include <net/if.h>
 #include <linux/ethtool.h>
 #include <linux/sockios.h>
+#include <linux/vfio.h>
+#include <sys/eventfd.h>
+
+static const char *sysfs_pci_dev_path = "/sys/bus/pci/devices";
+static const char *sysfs_pci_drv_path = "/sys/bus/pci/drivers";
 
 typedef struct
 {
-  /* /sys/bus/pci/devices/... directory name for this device. */
-  u8 *dev_dir_name;
+  vlib_pci_dev_handle_t handle;
+  vlib_pci_addr_t addr;
 
   /* Resource file descriptors. */
   int *resource_fds;
@@ -69,40 +74,260 @@ typedef struct
   /* Minor device for uio device. */
   u32 uio_minor;
 
+  /* vfio */
+  int vfio_device_fd;
+
   /* Index given by clib_file_add. */
   u32 clib_file_index;
 
+  /* Interrupt handler */
+  void (*interrupt_handler) (vlib_pci_dev_handle_t h);
+
+  /* private data */
+  uword private_data;
+
 } linux_pci_device_t;
 
+typedef struct
+{
+  int group;
+  int fd;
+  int refcnt;
+} linux_pci_vfio_iommu_group_t;
+
 /* Pool of PCI devices. */
 typedef struct
 {
   vlib_main_t *vlib_main;
   linux_pci_device_t *linux_pci_devices;
+
+  /* VFIO */
+  int vfio_container_fd;
+  int vfio_iommu_mode;
+
+  /* pool of IOMMU groups */
+  linux_pci_vfio_iommu_group_t *iommu_groups;
+
+  /* iommu group pool index by group id  hash */
+  uword *iommu_pool_index_by_group;
+
 } linux_pci_main_t;
 
 extern linux_pci_main_t linux_pci_main;
 
+linux_pci_device_t *
+linux_pci_get_device (vlib_pci_dev_handle_t h)
+{
+  linux_pci_main_t *lpm = &linux_pci_main;
+  return pool_elt_at_index (lpm->linux_pci_devices, h);
+}
+
+uword
+vlib_pci_get_private_data (vlib_pci_dev_handle_t h)
+{
+  linux_pci_device_t *d = linux_pci_get_device (h);
+  return d->private_data;
+}
+
+void
+vlib_pci_set_private_data (vlib_pci_dev_handle_t h, uword private_data)
+{
+  linux_pci_device_t *d = linux_pci_get_device (h);
+  d->private_data = private_data;
+}
+
+vlib_pci_addr_t *
+vlib_pci_get_addr (vlib_pci_dev_handle_t h)
+{
+  linux_pci_device_t *lpm = linux_pci_get_device (h);
+  return &lpm->addr;
+}
+
 /* Call to allocate/initialize the pci subsystem.
    This is not an init function so that users can explicitly enable
    pci only when it's needed. */
 clib_error_t *pci_bus_init (vlib_main_t * vm);
 
-clib_error_t *vlib_pci_bind_to_uio (vlib_pci_device_t * d,
-				    char *uio_driver_name);
-
 linux_pci_main_t linux_pci_main;
 
+vlib_pci_device_info_t *
+vlib_pci_get_device_info (vlib_pci_addr_t * addr, clib_error_t ** error)
+{
+  linux_pci_main_t *lpm = &linux_pci_main;
+  clib_error_t *err;
+  vlib_pci_device_info_t *di;
+  u8 *f = 0;
+  u32 tmp;
+  int fd;
+
+  di = clib_mem_alloc (sizeof (vlib_pci_device_info_t));
+  memset (di, 0, sizeof (vlib_pci_device_info_t));
+  di->addr.as_u32 = addr->as_u32;
+
+  u8 *dev_dir_name = format (0, "%s/%U", sysfs_pci_dev_path,
+			     format_vlib_pci_addr, addr);
+
+  f = format (0, "%v/config%c", dev_dir_name, 0);
+  fd = open ((char *) f, O_RDWR);
+
+  /* Try read-only access if write fails. */
+  if (fd < 0)
+    fd = open ((char *) f, O_RDONLY);
+
+  if (fd < 0)
+    {
+      err = clib_error_return_unix (0, "open `%s'", f);
+      goto error;
+    }
+
+  /* You can only read more that 64 bytes of config space as root; so we try to
+     read the full space but fall back to just the first 64 bytes. */
+  if (read (fd, &di->config_data, sizeof (di->config_data)) !=
+      sizeof (di->config_data)
+      && read (fd, &di->config0,
+	       sizeof (di->config0)) != sizeof (di->config0))
+    {
+      err = clib_error_return_unix (0, "read `%s'", f);
+      close (fd);
+      goto error;
+    }
+
+  {
+    static pci_config_header_t all_ones;
+    if (all_ones.vendor_id == 0)
+      memset (&all_ones, ~0, sizeof (all_ones));
+
+    if (!memcmp (&di->config0.header, &all_ones, sizeof (all_ones)))
+      {
+	err = clib_error_return (0, "invalid PCI config for `%s'", f);
+	close (fd);
+	goto error;
+      }
+  }
+
+  if (di->config0.header.header_type == 0)
+    pci_config_type0_little_to_host (&di->config0);
+  else
+    pci_config_type1_little_to_host (&di->config1);
+
+  di->numa_node = -1;
+  vec_reset_length (f);
+  f = format (f, "%v/numa_node%c", dev_dir_name, 0);
+  err = clib_sysfs_read ((char *) f, "%u", &di->numa_node);
+  if (err)
+    {
+      di->numa_node = -1;
+      clib_error_free (err);
+    }
+
+  vec_reset_length (f);
+  f = format (f, "%v/class%c", dev_dir_name, 0);
+  err = clib_sysfs_read ((char *) f, "0x%x", &tmp);
+  if (err)
+    goto error;
+  di->device_class = tmp >> 8;
+
+  vec_reset_length (f);
+  f = format (f, "%v/vendor%c", dev_dir_name, 0);
+  err = clib_sysfs_read ((char *) f, "0x%x", &tmp);
+  if (err)
+    goto error;
+  di->vendor_id = tmp;
+
+  vec_reset_length (f);
+  f = format (f, "%v/device%c", dev_dir_name, 0);
+  err = clib_sysfs_read ((char *) f, "0x%x", &tmp);
+  if (err)
+    goto error;
+  di->device_id = tmp;
+
+  vec_reset_length (f);
+  f = format (f, "%v/driver%c", dev_dir_name, 0);
+  di->driver_name = clib_sysfs_link_to_name ((char *) f);
+
+  di->iommu_group = -1;
+  if (lpm->vfio_container_fd != -1)
+    {
+      u8 *tmpstr;
+      vec_reset_length (f);
+      f = format (f, "%v/iommu_group%c", dev_dir_name, 0);
+      tmpstr = clib_sysfs_link_to_name ((char *) f);
+      if (tmpstr)
+	{
+	  di->iommu_group = atoi ((char *) tmpstr);
+	  vec_free (tmpstr);
+	}
+    }
+
+  vec_reset_length (f);
+  f = format (f, "%v/vpd%c", dev_dir_name, 0);
+  fd = open ((char *) f, O_RDONLY);
+  if (fd >= 0)
+    {
+      while (1)
+	{
+	  u8 tag[3];
+	  u8 *data = 0;
+	  int len;
+
+	  if (read (fd, &tag, 3) != 3)
+	    break;
+
+	  if (tag[0] != 0x82 && tag[0] != 0x90 && tag[0] != 0x91)
+	    break;
+
+	  len = (tag[2] << 8) | tag[1];
+	  vec_validate (data, len);
+
+	  if (read (fd, data, len) != len)
+	    {
+	      vec_free (data);
+	      break;
+	    }
+	  if (tag[0] == 0x82)
+	    di->product_name = data;
+	  else if (tag[0] == 0x90)
+	    di->vpd_r = data;
+	  else if (tag[0] == 0x91)
+	    di->vpd_w = data;
+
+	  data = 0;
+	}
+      close (fd);
+    }
+
+  goto done;
+
+error:
+  vlib_pci_free_device_info (di);
+  di = 0;
+
+done:
+  vec_free (f);
+  vec_free (dev_dir_name);
+  if (error)
+    *error = err;
+  else
+    clib_error_free (err);
+  return di;
+}
+
 clib_error_t *
-vlib_pci_bind_to_uio (vlib_pci_device_t * d, char *uio_driver_name)
+vlib_pci_bind_to_uio (vlib_pci_addr_t * addr, char *uio_drv_name)
 {
   clib_error_t *error = 0;
   u8 *s = 0, *driver_name = 0;
   DIR *dir = 0;
   struct dirent *e;
+  vlib_pci_device_info_t *di;
   int fd, clear_driver_override = 0;
-  u8 *dev_dir_name = format (0, "/sys/bus/pci/devices/%U",
-			     format_vlib_pci_addr, &d->bus_address);
+  u8 *dev_dir_name = format (0, "%s/%U", sysfs_pci_dev_path,
+			     format_vlib_pci_addr, addr);
+
+  di = vlib_pci_get_device_info (addr, &error);
+
+  if (error)
+    return error;
 
   s = format (s, "%v/driver%c", dev_dir_name, 0);
   driver_name = clib_sysfs_link_to_name ((char *) s);
@@ -117,13 +342,13 @@ vlib_pci_bind_to_uio (vlib_pci_device_t * d, char *uio_driver_name)
   /* walk trough all linux interfaces and if interface belonging to
      this device is founf check if interface is admin up  */
   dir = opendir ("/sys/class/net");
-  s = format (s, "%U%c", format_vlib_pci_addr, &d->bus_address, 0);
+  s = format (s, "%U%c", format_vlib_pci_addr, addr, 0);
 
   if (!dir)
     {
       error = clib_error_return (0, "Skipping PCI device %U: failed to "
 				 "read /sys/class/net",
-				 format_vlib_pci_addr, &d->bus_address);
+				 format_vlib_pci_addr, addr);
       goto done;
     }
 
@@ -173,8 +398,7 @@ vlib_pci_bind_to_uio (vlib_pci_device_t * d, char *uio_driver_name)
 	{
 	  error = clib_error_return (0, "Skipping PCI device %U as host "
 				     "interface %s is up",
-				     format_vlib_pci_addr, &d->bus_address,
-				     e->d_name);
+				     format_vlib_pci_addr, addr, e->d_name);
 	  close (fd);
 	  goto done;
 	}
@@ -184,26 +408,26 @@ vlib_pci_bind_to_uio (vlib_pci_device_t * d, char *uio_driver_name)
   vec_reset_length (s);
 
   s = format (s, "%v/driver/unbind%c", dev_dir_name, 0);
-  clib_sysfs_write ((char *) s, "%U", format_vlib_pci_addr, &d->bus_address);
+  clib_sysfs_write ((char *) s, "%U", format_vlib_pci_addr, addr);
   vec_reset_length (s);
 
   s = format (s, "%v/driver_override%c", dev_dir_name, 0);
   if (access ((char *) s, F_OK) == 0)
     {
-      clib_sysfs_write ((char *) s, "%s", uio_driver_name);
+      clib_sysfs_write ((char *) s, "%s", uio_drv_name);
       clear_driver_override = 1;
     }
   else
     {
       vec_reset_length (s);
-      s = format (s, "/sys/bus/pci/drivers/%s/new_id%c", uio_driver_name, 0);
-      clib_sysfs_write ((char *) s, "0x%04x 0x%04x", d->vendor_id,
-			d->device_id);
+      s = format (s, "%s/%s/new_id%c", sysfs_pci_drv_path, uio_drv_name, 0);
+      clib_sysfs_write ((char *) s, "0x%04x 0x%04x", di->vendor_id,
+			di->device_id);
     }
   vec_reset_length (s);
 
-  s = format (s, "/sys/bus/pci/drivers/%s/bind%c", uio_driver_name, 0);
-  clib_sysfs_write ((char *) s, "%U", format_vlib_pci_addr, &d->bus_address);
+  s = format (s, "%s/%s/bind%c", sysfs_pci_drv_path, uio_drv_name, 0);
+  clib_sysfs_write ((char *) s, "%U", format_vlib_pci_addr, addr);
   vec_reset_length (s);
 
   if (clear_driver_override)
@@ -240,24 +464,42 @@ scan_uio_dir (void *arg, u8 * path_name, u8 * file_name)
 static clib_error_t *
 linux_pci_uio_read_ready (clib_file_t * uf)
 {
-  vlib_pci_main_t *pm = &pci_main;
-  vlib_pci_device_t *d;
   int __attribute__ ((unused)) rv;
+  vlib_pci_dev_handle_t h = uf->private_data;
+  linux_pci_device_t *p = linux_pci_get_device (h);
 
   u32 icount;
   rv = read (uf->file_descriptor, &icount, 4);
 
-  d = pool_elt_at_index (pm->pci_devs, uf->private_data);
-
-  if (d->interrupt_handler)
-    d->interrupt_handler (d);
+  if (p->interrupt_handler)
+    p->interrupt_handler (h);
 
-  vlib_pci_intr_enable (d);
+  vlib_pci_intr_enable (h);
 
   return /* no error */ 0;
 }
 
 static clib_error_t *
+linux_pci_vfio_unmask_intx (linux_pci_device_t * d)
+{
+  clib_error_t *err = 0;
+  struct vfio_irq_set i = {
+    .argsz = sizeof (struct vfio_irq_set),
+    .start = 0,
+    .count = 1,
+    .index = VFIO_PCI_INTX_IRQ_INDEX,
+    .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
+  };
+
+  if (ioctl (d->vfio_device_fd, VFIO_DEVICE_SET_IRQS, &i) < 0)
+    {
+      err = clib_error_return_unix (0, "ioctl(VFIO_DEVICE_SET_IRQS) '%U'",
+				    format_vlib_pci_addr, &d->addr);
+    }
+  return err;
+}
+
+static clib_error_t *
 linux_pci_uio_error_ready (clib_file_t * uf)
 {
   u32 error_index = (u32) uf->private_data;
@@ -265,74 +507,293 @@ linux_pci_uio_error_ready (clib_file_t * uf)
   return clib_error_return (0, "pci device %d: error", error_index);
 }
 
-static void
-add_device (vlib_pci_device_t * dev, linux_pci_device_t * pdev)
+static clib_error_t *
+linux_pci_vfio_read_ready (clib_file_t * uf)
+{
+  int __attribute__ ((unused)) rv;
+  vlib_pci_dev_handle_t h = uf->private_data;
+  linux_pci_device_t *p = linux_pci_get_device (h);
+
+  u64 icount;
+  rv = read (uf->file_descriptor, &icount, sizeof (icount));
+
+  if (p->interrupt_handler)
+    p->interrupt_handler (h);
+
+  linux_pci_vfio_unmask_intx (p);
+
+  return /* no error */ 0;
+}
+
+static clib_error_t *
+linux_pci_vfio_error_ready (clib_file_t * uf)
+{
+  u32 error_index = (u32) uf->private_data;
+
+  return clib_error_return (0, "pci device %d: error", error_index);
+}
+
+static clib_error_t *
+add_device_uio (vlib_main_t * vm, linux_pci_device_t * p,
+		vlib_pci_device_info_t * di, pci_device_registration_t * r)
+{
+  clib_error_t *err = 0;
+  clib_file_t t = { 0 };
+  u8 *s = 0;
+
+  p->addr.as_u32 = di->addr.as_u32;
+  p->uio_fd = -1;
+
+  s = format (s, "%s/%U/config%c", sysfs_pci_dev_path,
+	      format_vlib_pci_addr, &di->addr, 0);
+
+  p->config_fd = open ((char *) s, O_RDWR);
+  vec_reset_length (s);
+
+  if (p->config_fd == -1)
+    {
+      err = clib_error_return_unix (0, "open '%s'", s);
+      goto error;
+    }
+
+  s = format (0, "%s/%U/uio", sysfs_pci_dev_path,
+	      format_vlib_pci_addr, &di->addr);
+  foreach_directory_file ((char *) s, scan_uio_dir, p,	/* scan_dirs */
+			  1);
+  vec_reset_length (s);
+
+  s = format (s, "/dev/uio%d%c", p->uio_minor, 0);
+  p->uio_fd = open ((char *) s, O_RDWR);
+  if (p->uio_fd < 0)
+    {
+      err = clib_error_return_unix (0, "open '%s'", s);
+      goto error;
+    }
+
+  t.read_function = linux_pci_uio_read_ready;
+  t.file_descriptor = p->uio_fd;
+  t.error_function = linux_pci_uio_error_ready;
+  t.private_data = p->handle;
+
+  p->clib_file_index = clib_file_add (&file_main, &t);
+  p->interrupt_handler = r->interrupt_handler;
+  err = r->init_function (vm, p->handle);
+
+error:
+  free (s);
+  if (err)
+    {
+      close (p->config_fd);
+      close (p->uio_fd);
+    }
+  return err;
+}
+
+static linux_pci_vfio_iommu_group_t *
+get_vfio_iommu_group (int group)
 {
-  vlib_pci_main_t *pm = &pci_main;
   linux_pci_main_t *lpm = &linux_pci_main;
-  linux_pci_device_t *l;
+  uword *p;
 
-  pool_get (lpm->linux_pci_devices, l);
-  l[0] = pdev[0];
+  p = hash_get (lpm->iommu_pool_index_by_group, group);
 
-  l->dev_dir_name = vec_dup (l->dev_dir_name);
+  return p ? pool_elt_at_index (lpm->iommu_groups, p[0]) : 0;
+}
 
-  dev->os_handle = l - lpm->linux_pci_devices;
+static clib_error_t *
+open_vfio_iommu_group (int group)
+{
+  linux_pci_main_t *lpm = &linux_pci_main;
+  linux_pci_vfio_iommu_group_t *g;
+  clib_error_t *err = 0;
+  struct vfio_group_status group_status;
+  u8 *s = 0;
+  int fd;
 
-  {
-    u8 *uio_dir = format (0, "%s/uio", l->dev_dir_name);
-    foreach_directory_file ((char *) uio_dir, scan_uio_dir, l,	/* scan_dirs */
-			    1);
-    vec_free (uio_dir);
-  }
+  g = get_vfio_iommu_group (group);
+  if (g)
+    {
+      g->refcnt++;
+      return 0;
+    }
+  s = format (s, "/dev/vfio/%u%c", group, 0);
+  fd = open ((char *) s, O_RDWR);
+  if (fd < 0)
+    clib_error_return_unix (0, "open '%s'", s);
 
-  {
-    char *uio_name = (char *) format (0, "/dev/uio%d%c", l->uio_minor, 0);
-    l->uio_fd = open (uio_name, O_RDWR);
-    if (l->uio_fd < 0)
-      clib_unix_error ("open `%s'", uio_name);
-    vec_free (uio_name);
-  }
+  group_status.argsz = sizeof (group_status);
+  if (ioctl (fd, VFIO_GROUP_GET_STATUS, &group_status) < 0)
+    {
+      err = clib_error_return_unix (0, "ioctl(VFIO_GROUP_GET_STATUS) '%s'",
+				    s);
+      goto error;
+    }
 
-  {
-    clib_file_t template = { 0 };
+  if (!(group_status.flags & VFIO_GROUP_FLAGS_VIABLE))
+    {
+      err = clib_error_return (0, "iommu group %d is not viable (not all "
+			       "devices in this group bound to vfio-pci)",
+			       group);
+      goto error;
+    }
+
+  if (ioctl (fd, VFIO_GROUP_SET_CONTAINER, &lpm->vfio_container_fd) < 0)
+    {
+      err = clib_error_return_unix (0, "ioctl(VFIO_GROUP_SET_CONTAINER) '%s'",
+				    s);
+      goto error;
+    }
 
-    template.read_function = linux_pci_uio_read_ready;
-    template.file_descriptor = l->uio_fd;
-    template.error_function = linux_pci_uio_error_ready;
-    template.private_data = dev - pm->pci_devs;
+  if (lpm->vfio_iommu_mode == 0)
+    {
+      if (ioctl (lpm->vfio_container_fd, VFIO_SET_IOMMU, VFIO_TYPE1_IOMMU) <
+	  0)
+	{
+	  err = clib_error_return_unix (0, "ioctl(VFIO_SET_IOMMU) "
+					"'/dev/vfio/vfio'");
+	  goto error;
+	}
+      lpm->vfio_iommu_mode = VFIO_TYPE1_IOMMU;
+    }
 
-    l->clib_file_index = clib_file_add (&file_main, &template);
-  }
+
+  pool_get (lpm->iommu_groups, g);
+  g->fd = fd;
+  g->refcnt = 1;
+  hash_set (lpm->iommu_pool_index_by_group, group, g - lpm->iommu_groups);
+  vec_free (s);
+  return 0;
+error:
+  close (fd);
+  return err;
 }
 
-static void
-linux_pci_device_free (linux_pci_device_t * l)
+static clib_error_t *
+add_device_vfio (vlib_main_t * vm, linux_pci_device_t * p,
+		 vlib_pci_device_info_t * di, pci_device_registration_t * r)
 {
-  int i;
-  for (i = 0; i < vec_len (l->resource_fds); i++)
-    if (l->resource_fds[i] > 0)
-      close (l->resource_fds[i]);
-  if (l->config_fd > 0)
-    close (l->config_fd);
-  if (l->uio_fd > 0)
-    close (l->uio_fd);
-  vec_free (l->resource_fds);
-  vec_free (l->dev_dir_name);
+  linux_pci_vfio_iommu_group_t *g;
+  struct vfio_device_info device_info;
+  struct vfio_irq_info irq_info = { 0 };
+  clib_error_t *err = 0;
+  u8 *s = 0;
+  int dfd;
+
+  p->addr.as_u32 = di->addr.as_u32;
+
+  if (di->driver_name == 0 ||
+      (strcmp ("vfio-pci", (char *) di->driver_name) != 0))
+    return clib_error_return (0, "Device '%U' (iommu group %d) not bound to "
+			      "vfio-pci", format_vlib_pci_addr, &di->addr,
+			      di->iommu_group);
+
+  if ((err = open_vfio_iommu_group (di->iommu_group)))
+    return err;
+
+  g = get_vfio_iommu_group (di->iommu_group);
+
+  s = format (s, "%U%c", format_vlib_pci_addr, &di->addr, 0);
+  if ((dfd = ioctl (g->fd, VFIO_GROUP_GET_DEVICE_FD, (char *) s)) < 0)
+    {
+      err = clib_error_return_unix (0, "ioctl(VFIO_GROUP_GET_DEVICE_FD) '%U'",
+				    format_vlib_pci_addr, &di->addr);
+      goto error;
+    }
+  vec_reset_length (s);
+  p->vfio_device_fd = dfd;
+
+  device_info.argsz = sizeof (device_info);
+  if (ioctl (dfd, VFIO_DEVICE_GET_INFO, &device_info) < 0)
+    {
+      err = clib_error_return_unix (0, "ioctl(VFIO_DEVICE_GET_INFO) '%U'",
+				    format_vlib_pci_addr, &di->addr);
+      goto error;
+    }
+
+  irq_info.argsz = sizeof (struct vfio_irq_info);
+  irq_info.index = VFIO_PCI_INTX_IRQ_INDEX;
+  if (ioctl (dfd, VFIO_DEVICE_GET_IRQ_INFO, &irq_info) < 0)
+    {
+      err = clib_error_return_unix (0, "ioctl(VFIO_DEVICE_GET_IRQ_INFO) "
+				    "'%U'", format_vlib_pci_addr, &di->addr);
+      goto error;
+    }
+
+  /* reset if device supports it */
+  if (device_info.flags & VFIO_DEVICE_FLAGS_RESET)
+    if (ioctl (dfd, VFIO_DEVICE_RESET) < 0)
+      {
+	err = clib_error_return_unix (0, "ioctl(VFIO_DEVICE_RESET) '%U'",
+				      format_vlib_pci_addr, &di->addr);
+	goto error;
+      }
+
+  if ((irq_info.flags & VFIO_IRQ_INFO_EVENTFD) && irq_info.count == 1)
+    {
+      u8 buf[sizeof (struct vfio_irq_set) + sizeof (int)] = { 0 };
+      struct vfio_irq_set *irq_set = (struct vfio_irq_set *) buf;
+      clib_file_t t = { 0 };
+      int efd = eventfd (0, EFD_NONBLOCK);
+
+      irq_set->argsz = sizeof (buf);
+      irq_set->count = 1;
+      irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
+      irq_set->flags =
+	VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
+      clib_memcpy (&irq_set->data, &efd, sizeof (int));
+
+      if (ioctl (dfd, VFIO_DEVICE_SET_IRQS, irq_set) < 0)
+	{
+	  err = clib_error_return_unix (0, "ioctl(VFIO_DEVICE_SET_IRQS) '%U'",
+					format_vlib_pci_addr, &di->addr);
+	  goto error;
+	}
+
+      t.read_function = linux_pci_vfio_read_ready;
+      t.file_descriptor = efd;
+      t.error_function = linux_pci_vfio_error_ready;
+      t.private_data = p->handle;
+      p->clib_file_index = clib_file_add (&file_main, &t);
+
+      /* unmask the interrupt */
+      linux_pci_vfio_unmask_intx (p);
+    }
+
+  p->interrupt_handler = r->interrupt_handler;
+
+  s = format (s, "%s/%U/config%c", sysfs_pci_dev_path,
+	      format_vlib_pci_addr, &di->addr, 0);
+
+  p->config_fd = open ((char *) s, O_RDWR);
+  vec_reset_length (s);
+
+  if (p->config_fd == -1)
+    {
+      err = clib_error_return_unix (0, "open '%s'", s);
+      goto error;
+    }
+
+  err = r->init_function (vm, p->handle);
+
+error:
+  vec_free (s);
+  if (err)
+    {
+      close (dfd);
+      close (p->config_fd);
+    }
+  return err;
 }
 
 /* Configuration space read/write. */
 clib_error_t *
-vlib_pci_read_write_config (vlib_pci_device_t * dev,
+vlib_pci_read_write_config (vlib_pci_dev_handle_t h,
 			    vlib_read_or_write_t read_or_write,
 			    uword address, void *data, u32 n_bytes)
 {
-  linux_pci_main_t *lpm = &linux_pci_main;
-  linux_pci_device_t *p;
+  linux_pci_device_t *p = linux_pci_get_device (h);
   int n;
 
-  p = pool_elt_at_index (lpm->linux_pci_devices, dev->os_handle);
-
   if (read_or_write == VLIB_READ)
     n = pread (p->config_fd, data, n_bytes, address);
   else
@@ -347,11 +808,10 @@ vlib_pci_read_write_config (vlib_pci_device_t * dev,
 }
 
 static clib_error_t *
-os_map_pci_resource_internal (uword os_handle,
-			      u32 resource, u8 * addr, void **result)
+vlib_pci_map_resource_int (vlib_pci_dev_handle_t h,
+			   u32 resource, u8 * addr, void **result)
 {
-  linux_pci_main_t *pm = &linux_pci_main;
-  linux_pci_device_t *p;
+  linux_pci_device_t *p = linux_pci_get_device (h);
   struct stat stat_buf;
   u8 *file_name;
   int fd;
@@ -359,9 +819,10 @@ os_map_pci_resource_internal (uword os_handle,
   int flags = MAP_SHARED;
 
   error = 0;
-  p = pool_elt_at_index (pm->linux_pci_devices, os_handle);
 
-  file_name = format (0, "%v/resource%d%c", p->dev_dir_name, resource, 0);
+  file_name = format (0, "%s/%U/resource%d%c", sysfs_pci_dev_path,
+		      format_vlib_pci_addr, &p->addr, resource, 0);
+
   fd = open ((char *) file_name, O_RDWR);
   if (fd < 0)
     {
@@ -402,240 +863,131 @@ done:
 }
 
 clib_error_t *
-vlib_pci_map_resource (vlib_pci_device_t * dev, u32 resource, void **result)
+vlib_pci_map_resource (vlib_pci_dev_handle_t h, u32 resource, void **result)
 {
-  return (os_map_pci_resource_internal
-	  (dev->os_handle, resource, 0 /* addr */ ,
-	   result));
+  return (vlib_pci_map_resource_int (h, resource, 0 /* addr */ , result));
 }
 
 clib_error_t *
-vlib_pci_map_resource_fixed (vlib_pci_device_t * dev,
+vlib_pci_map_resource_fixed (vlib_pci_dev_handle_t h,
 			     u32 resource, u8 * addr, void **result)
 {
-  return (os_map_pci_resource_internal
-	  (dev->os_handle, resource, addr, result));
+  return (vlib_pci_map_resource_int (h, resource, addr, result));
 }
 
 void
-vlib_pci_free_device (vlib_pci_device_t * dev)
-{
-  linux_pci_main_t *pm = &linux_pci_main;
-  linux_pci_device_t *l;
-
-  l = pool_elt_at_index (pm->linux_pci_devices, dev->os_handle);
-  linux_pci_device_free (l);
-  pool_put (pm->linux_pci_devices, l);
-}
-
-pci_device_registration_t * __attribute__ ((unused))
-pci_device_next_registered (pci_device_registration_t * r)
-{
-  uword i;
-
-  /* Null vendor id marks end of initialized list. */
-  for (i = 0; r->supported_devices[i].vendor_id != 0; i++)
-    ;
-
-  return clib_elf_section_data_next (r, i * sizeof (r->supported_devices[0]));
-}
-
-static clib_error_t *
-init_device_from_registered (vlib_main_t * vm,
-			     vlib_pci_device_t * dev,
-			     linux_pci_device_t * pdev)
+init_device_from_registered (vlib_main_t * vm, vlib_pci_device_info_t * di)
 {
   vlib_pci_main_t *pm = &pci_main;
+  linux_pci_main_t *lpm = &linux_pci_main;
   pci_device_registration_t *r;
   pci_device_id_t *i;
-  clib_error_t *error;
+  clib_error_t *err = 0;
+  linux_pci_device_t *p;
+
+  pool_get (lpm->linux_pci_devices, p);
+  p->handle = p - lpm->linux_pci_devices;
 
   r = pm->pci_device_registrations;
 
   while (r)
     {
       for (i = r->supported_devices; i->vendor_id != 0; i++)
-	if (i->vendor_id == dev->vendor_id && i->device_id == dev->device_id)
+	if (i->vendor_id == di->config0.header.vendor_id &&
+	    i->device_id == di->config0.header.device_id)
 	  {
-	    error = vlib_pci_bind_to_uio (dev, "uio_pci_generic");
-	    if (error)
-	      {
-		clib_error_report (error);
-		continue;
-	      }
-
-	    add_device (dev, pdev);
-	    dev->interrupt_handler = r->interrupt_handler;
-	    return r->init_function (vm, dev);
+	    if (di->iommu_group != -1)
+	      err = add_device_vfio (vm, p, di, r);
+	    else
+	      err = add_device_uio (vm, p, di, r);
+
+	    if (err)
+	      clib_error_report (err);
+	    else
+	      return;
 	  }
       r = r->next_registration;
     }
-  /* No driver, close the PCI config-space FD */
-  close (pdev->config_fd);
-  return 0;
-}
 
-static clib_error_t *
-init_device (vlib_main_t * vm,
-	     vlib_pci_device_t * dev, linux_pci_device_t * pdev)
-{
-  return init_device_from_registered (vm, dev, pdev);
+  /* No driver, close the PCI config-space FD */
+  memset (p, 0, sizeof (linux_pci_device_t));
+  pool_put (lpm->linux_pci_devices, p);
 }
 
 static clib_error_t *
-scan_device (void *arg, u8 * dev_dir_name, u8 * ignored)
+scan_pci_addr (void *arg, u8 * dev_dir_name, u8 * ignored)
 {
-  vlib_main_t *vm = arg;
-  vlib_pci_main_t *pm = &pci_main;
-  int fd;
-  u8 *f;
-  clib_error_t *error = 0;
-  vlib_pci_device_t *dev;
-  linux_pci_device_t pdev = { 0 };
-  u32 tmp;
-
-  f = format (0, "%v/config%c", dev_dir_name, 0);
-  fd = open ((char *) f, O_RDWR);
-
-  /* Try read-only access if write fails. */
-  if (fd < 0)
-    fd = open ((char *) f, O_RDONLY);
-
-  if (fd < 0)
-    {
-      error = clib_error_return_unix (0, "open `%s'", f);
-      goto done;
-    }
-
-  pool_get (pm->pci_devs, dev);
-
-  /* You can only read more that 64 bytes of config space as root; so we try to
-     read the full space but fall back to just the first 64 bytes. */
-  if (read (fd, &dev->config_data, sizeof (dev->config_data)) !=
-      sizeof (dev->config_data)
-      && read (fd, &dev->config0,
-	       sizeof (dev->config0)) != sizeof (dev->config0))
-    {
-      pool_put (pm->pci_devs, dev);
-      error = clib_error_return_unix (0, "read `%s'", f);
-      close (fd);
-      goto done;
-    }
-
-  {
-    static pci_config_header_t all_ones;
-    if (all_ones.vendor_id == 0)
-      memset (&all_ones, ~0, sizeof (all_ones));
-
-    if (!memcmp (&dev->config0.header, &all_ones, sizeof (all_ones)))
-      {
-	pool_put (pm->pci_devs, dev);
-	error = clib_error_return (0, "invalid PCI config for `%s'", f);
-	close (fd);
-	goto done;
-      }
-  }
-
-  if (dev->config0.header.header_type == 0)
-    pci_config_type0_little_to_host (&dev->config0);
-  else
-    pci_config_type1_little_to_host (&dev->config1);
-
-  /* Parse bus, dev, function from directory name. */
-  {
-    unformat_input_t input;
-
-    unformat_init_string (&input, (char *) dev_dir_name,
-			  vec_len (dev_dir_name));
+  vlib_pci_addr_t addr, **addrv = arg;
+  unformat_input_t input;
+  clib_error_t *err = 0;
 
-    if (!unformat (&input, "/sys/bus/pci/devices/%U",
-		   unformat_vlib_pci_addr, &dev->bus_address))
-      abort ();
+  unformat_init_string (&input, (char *) dev_dir_name,
+			vec_len (dev_dir_name));
 
-    unformat_free (&input);
+  if (!unformat (&input, "/sys/bus/pci/devices/%U",
+		 unformat_vlib_pci_addr, &addr))
+    err = clib_error_return (0, "unformat error `%v`", dev_dir_name);
 
-  }
+  unformat_free (&input);
 
+  if (err)
+    return err;
 
-  pdev.config_fd = fd;
-  pdev.dev_dir_name = dev_dir_name;
+  vec_add1 (*addrv, addr);
+  return 0;
+}
 
-  hash_set (pm->pci_dev_index_by_pci_addr, dev->bus_address.as_u32,
-	    dev - pm->pci_devs);
+static int
+pci_addr_cmp (void *v1, void *v2)
+{
+  vlib_pci_addr_t *a1 = v1;
+  vlib_pci_addr_t *a2 = v2;
+
+  if (a1->domain > a2->domain)
+    return 1;
+  if (a1->domain < a2->domain)
+    return -1;
+  if (a1->bus > a2->bus)
+    return 1;
+  if (a1->bus < a2->bus)
+    return -1;
+  if (a1->slot > a2->slot)
+    return 1;
+  if (a1->slot < a2->slot)
+    return -1;
+  if (a1->function > a2->function)
+    return 1;
+  if (a1->function < a2->function)
+    return -1;
+  return 0;
+}
 
-  vec_reset_length (f);
-  f = format (f, "%v/vpd%c", dev_dir_name, 0);
-  fd = open ((char *) f, O_RDONLY);
-  if (fd >= 0)
+vlib_pci_addr_t *
+vlib_pci_get_all_dev_addrs ()
+{
+  vlib_pci_addr_t *addrs = 0;
+  clib_error_t *err;
+  err = foreach_directory_file ((char *) sysfs_pci_dev_path, scan_pci_addr,
+				&addrs, /* scan_dirs */ 0);
+  if (err)
     {
-      while (1)
-	{
-	  u8 tag[3];
-	  u8 *data = 0;
-	  int len;
-
-	  if (read (fd, &tag, 3) != 3)
-	    break;
-
-	  if (tag[0] != 0x82 && tag[0] != 0x90 && tag[0] != 0x91)
-	    break;
-
-	  len = (tag[2] << 8) | tag[1];
-	  vec_validate (data, len);
-
-	  if (read (fd, data, len) != len)
-	    {
-	      vec_free (data);
-	      break;
-	    }
-	  if (tag[0] == 0x82)
-	    dev->product_name = data;
-	  else if (tag[0] == 0x90)
-	    dev->vpd_r = data;
-	  else if (tag[0] == 0x91)
-	    dev->vpd_w = data;
-
-	  data = 0;
-	}
-      close (fd);
+      vec_free (addrs);
+      return 0;
     }
 
-  dev->numa_node = -1;
-  vec_reset_length (f);
-  f = format (f, "%v/numa_node%c", dev_dir_name, 0);
-  clib_sysfs_read ((char *) f, "%u", &dev->numa_node);
-
-  vec_reset_length (f);
-  f = format (f, "%v/class%c", dev_dir_name, 0);
-  clib_sysfs_read ((char *) f, "0x%x", &tmp);
-  dev->device_class = tmp >> 8;
-
-  vec_reset_length (f);
-  f = format (f, "%v/vendor%c", dev_dir_name, 0);
-  clib_sysfs_read ((char *) f, "0x%x", &tmp);
-  dev->vendor_id = tmp;
-
-  vec_reset_length (f);
-  f = format (f, "%v/device%c", dev_dir_name, 0);
-  clib_sysfs_read ((char *) f, "0x%x", &tmp);
-  dev->device_id = tmp;
+  vec_sort_with_function (addrs, pci_addr_cmp);
 
-  error = init_device (vm, dev, &pdev);
-
-  vec_reset_length (f);
-  f = format (f, "%v/driver%c", dev_dir_name, 0);
-  dev->driver_name = clib_sysfs_link_to_name ((char *) f);
-
-done:
-  vec_free (f);
-  return error;
+  return addrs;
 }
 
 clib_error_t *
 linux_pci_init (vlib_main_t * vm)
 {
   vlib_pci_main_t *pm = &pci_main;
+  linux_pci_main_t *lpm = &linux_pci_main;
+  vlib_pci_addr_t *addr = 0, *addrs;
   clib_error_t *error;
+  int fd;
 
   pm->vlib_main = vm;
 
@@ -643,14 +995,36 @@ linux_pci_init (vlib_main_t * vm)
     return error;
 
   ASSERT (sizeof (vlib_pci_addr_t) == sizeof (u32));
-  pm->pci_dev_index_by_pci_addr = hash_create (0, sizeof (uword));
 
-  error = foreach_directory_file ("/sys/bus/pci/devices", scan_device, vm,
-				  /* scan_dirs */ 0);
+  fd = open ("/dev/vfio/vfio", O_RDWR);
 
-  /* Complain and continue. might not be root, etc. */
-  if (error)
-    clib_error_report (error);
+  if ((fd != -1) && (ioctl (fd, VFIO_GET_API_VERSION) != VFIO_API_VERSION))
+    {
+      close (fd);
+      fd = -1;
+    }
+
+  if ((fd != -1) && (ioctl (fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU) == 0))
+    {
+      close (fd);
+      fd = -1;
+    }
+
+  lpm->vfio_container_fd = fd;
+  lpm->iommu_pool_index_by_group = hash_create (0, sizeof (uword));
+
+  addrs = vlib_pci_get_all_dev_addrs ();
+  /* *INDENT-OFF* */
+  vec_foreach (addr, addrs)
+    {
+      vlib_pci_device_info_t *d;
+      if ((d = vlib_pci_get_device_info (addr, 0)))
+	{
+	  init_device_from_registered (vm, d);
+	  vlib_pci_free_device_info (d);
+	}
+    }
+  /* *INDENT-ON* */
 
   return error;
 }
diff --git a/src/vlib/pci/pci.c b/src/vlib/pci/pci.c
index 1f1edab62e6..f50ae2d9edf 100644
--- a/src/vlib/pci/pci.c
+++ b/src/vlib/pci/pci.c
@@ -52,25 +52,24 @@
 
 vlib_pci_main_t pci_main;
 
-vlib_pci_device_t *
-vlib_get_pci_device (vlib_pci_addr_t * addr)
+vlib_pci_device_info_t * __attribute__ ((weak))
+vlib_pci_get_device_info (vlib_pci_addr_t * addr, clib_error_t ** error)
 {
-  vlib_pci_main_t *pm = &pci_main;
-  uword *p;
-  p = hash_get (pm->pci_dev_index_by_pci_addr, addr->as_u32);
-
-  if (p == 0)
-    return 0;
+  if (error)
+    *error = clib_error_return (0, "unsupported");
+  return 0;
+}
 
-  return vec_elt_at_index (pm->pci_devs, p[0]);
+vlib_pci_addr_t * __attribute__ ((weak)) vlib_pci_get_all_dev_addrs ()
+{
+  return 0;
 }
 
 static clib_error_t *
 show_pci_fn (vlib_main_t * vm,
 	     unformat_input_t * input, vlib_cli_command_t * cmd)
 {
-  vlib_pci_main_t *pm = &pci_main;
-  vlib_pci_device_t *d;
+  vlib_pci_addr_t *addr = 0, *addrs;
   int show_all = 0;
   u8 *s = 0;
 
@@ -87,28 +86,36 @@ show_pci_fn (vlib_main_t * vm,
 		   "Address", "Sock", "VID:PID", "Link Speed", "Driver",
 		   "Product Name", "Vital Product Data");
 
+  addrs = vlib_pci_get_all_dev_addrs ();
   /* *INDENT-OFF* */
-  pool_foreach (d, pm->pci_devs, ({
-
-    if (d->device_class != PCI_CLASS_NETWORK_ETHERNET && !show_all)
-      continue;
-
-    vec_reset_length (s);
-
-    if (d->numa_node >= 0)
-      s = format (s, "  %d", d->numa_node);
-
-    vlib_cli_output (vm, "%-13U%-5v%04x:%04x   %-13U%-16s%-32v%U",
-		     format_vlib_pci_addr, &d->bus_address, s,
-		     d->vendor_id, d->device_id,
-		     format_vlib_pci_link_speed, d,
-		     d->driver_name ? (char *) d->driver_name : "",
-		     d->product_name,
-		     format_vlib_pci_vpd, d->vpd_r, 0);
-  }));
-/* *INDENT-ON* */
+  vec_foreach (addr, addrs)
+    {
+      vlib_pci_device_info_t *d;
+      d = vlib_pci_get_device_info (addr, 0);
+
+      if (!d)
+        continue;
+
+      if (d->device_class != PCI_CLASS_NETWORK_ETHERNET && !show_all)
+        continue;
+
+        vec_reset_length (s);
+        if (d->numa_node >= 0)
+	  s = format (s, "  %d", d->numa_node);
+
+        vlib_cli_output (vm, "%-13U%-5v%04x:%04x   %-13U%-16s%-32v%U",
+			 format_vlib_pci_addr, addr, s,
+			 d->vendor_id, d->device_id,
+			 format_vlib_pci_link_speed, d,
+			 d->driver_name ? (char *) d->driver_name : "",
+			 d->product_name,
+			 format_vlib_pci_vpd, d->vpd_r, 0);
+	vlib_pci_free_device_info (d);
+    }
+  /* *INDENT-ON* */
 
   vec_free (s);
+  vec_free (addrs);
   return 0;
 }
 
@@ -138,16 +145,9 @@ format_vlib_pci_addr (u8 * s, va_list * va)
 }
 
 u8 *
-format_vlib_pci_handle (u8 * s, va_list * va)
-{
-  vlib_pci_addr_t *addr = va_arg (*va, vlib_pci_addr_t *);
-  return format (s, "%x/%x/%x", addr->bus, addr->slot, addr->function);
-}
-
-u8 *
 format_vlib_pci_link_speed (u8 * s, va_list * va)
 {
-  vlib_pci_device_t *d = va_arg (*va, vlib_pci_device_t *);
+  vlib_pci_device_info_t *d = va_arg (*va, vlib_pci_device_info_t *);
   pcie_config_regs_t *r =
     pci_config_find_capability (&d->config0, PCI_CAP_ID_PCIE);
   int width;
@@ -250,7 +250,7 @@ VLIB_CLI_COMMAND (show_pci_command, static) = {
 clib_error_t *
 pci_bus_init (vlib_main_t * vm)
 {
-  return 0;
+  return vlib_call_init_function (vm, pci_bus_init);
 }
 
 VLIB_INIT_FUNCTION (pci_bus_init);
diff --git a/src/vlib/pci/pci.h b/src/vlib/pci/pci.h
index 2141080936d..e1bc4172e8b 100644
--- a/src/vlib/pci/pci.h
+++ b/src/vlib/pci/pci.h
@@ -43,33 +43,24 @@
 #include <vlib/vlib.h>
 #include <vlib/pci/pci_config.h>
 
+/* *INDENT-OFF* */
 typedef CLIB_PACKED (union
-		     {
-		     struct
-		     {
-u16 domain; u8 bus; u8 slot: 5; u8 function:3;};
-		     u32 as_u32;}) vlib_pci_addr_t;
-
-typedef struct vlib_pci_device
 {
-  /* Operating system handle for this device. */
-  uword os_handle;
-
-  vlib_pci_addr_t bus_address;
-
-  /* First 64 bytes of configuration space. */
-  union
-  {
-    pci_config_type0_regs_t config0;
-    pci_config_type1_regs_t config1;
-    u8 config_data[256];
-  };
-
-  /* Interrupt handler */
-  void (*interrupt_handler) (struct vlib_pci_device * dev);
-
-  /* Driver name */
-  u8 *driver_name;
+  struct
+    {
+      u16 domain;
+      u8 bus;
+      u8 slot: 5;
+      u8 function:3;
+    };
+  u32 as_u32;
+}) vlib_pci_addr_t;
+/* *INDENT-ON* */
+
+typedef struct vlib_pci_device_info
+{
+  /* addr */
+  vlib_pci_addr_t addr;
 
   /* Numa Node */
   int numa_node;
@@ -84,10 +75,42 @@ typedef struct vlib_pci_device
   u8 *vpd_r;
   u8 *vpd_w;
 
-  /* Private data */
-  uword private_data;
+  /* Driver name */
+  u8 *driver_name;
+
+  /* First 64 bytes of configuration space. */
+  union
+  {
+    pci_config_type0_regs_t config0;
+    pci_config_type1_regs_t config1;
+    u8 config_data[256];
+  };
+
+  /* IOMMU Group */
+  int iommu_group;
+
+} vlib_pci_device_info_t;
+
+typedef u32 vlib_pci_dev_handle_t;
 
-} vlib_pci_device_t;
+vlib_pci_device_info_t *vlib_pci_get_device_info (vlib_pci_addr_t * addr,
+						  clib_error_t ** error);
+vlib_pci_addr_t *vlib_pci_get_all_dev_addrs ();
+vlib_pci_addr_t *vlib_pci_get_addr (vlib_pci_dev_handle_t h);
+uword vlib_pci_get_private_data (vlib_pci_dev_handle_t h);
+void vlib_pci_set_private_data (vlib_pci_dev_handle_t h, uword private_data);
+
+static inline void
+vlib_pci_free_device_info (vlib_pci_device_info_t * di)
+{
+  if (!di)
+    return;
+  vec_free (di->product_name);
+  vec_free (di->vpd_r);
+  vec_free (di->vpd_w);
+  vec_free (di->driver_name);
+  clib_mem_free (di);
+}
 
 typedef struct
 {
@@ -97,10 +120,11 @@ typedef struct
 typedef struct _pci_device_registration
 {
   /* Driver init function. */
-  clib_error_t *(*init_function) (vlib_main_t * vm, vlib_pci_device_t * dev);
+  clib_error_t *(*init_function) (vlib_main_t * vm,
+				  vlib_pci_dev_handle_t handle);
 
   /* Interrupt handler */
-  void (*interrupt_handler) (vlib_pci_device_t * dev);
+  void (*interrupt_handler) (vlib_pci_dev_handle_t handle);
 
   /* List of registrations */
   struct _pci_device_registration *next_registration;
@@ -113,9 +137,7 @@ typedef struct _pci_device_registration
 typedef struct
 {
   vlib_main_t *vlib_main;
-  vlib_pci_device_t *pci_devs;
   pci_device_registration_t *pci_device_registrations;
-  uword *pci_dev_index_by_pci_addr;
 } vlib_pci_main_t;
 
 extern vlib_pci_main_t pci_main;
@@ -132,21 +154,21 @@ static void __vlib_add_pci_device_registration_##x (void)       \
 }                                                               \
 __VA_ARGS__ pci_device_registration_t x
 
-clib_error_t *vlib_pci_bind_to_uio (vlib_pci_device_t * d,
+clib_error_t *vlib_pci_bind_to_uio (vlib_pci_addr_t * addr,
 				    char *uio_driver_name);
 
 /* Configuration space read/write. */
-clib_error_t *vlib_pci_read_write_config (vlib_pci_device_t * dev,
+clib_error_t *vlib_pci_read_write_config (vlib_pci_dev_handle_t handle,
 					  vlib_read_or_write_t read_or_write,
 					  uword address,
 					  void *data, u32 n_bytes);
 
 #define _(t)								\
 static inline clib_error_t *						\
-vlib_pci_read_config_##t (vlib_pci_device_t * dev,			\
+vlib_pci_read_config_##t (vlib_pci_dev_handle_t h,			\
 			  uword address, t * data)			\
 {									\
-  return vlib_pci_read_write_config (dev, VLIB_READ,address, data,	\
+  return vlib_pci_read_write_config (h, VLIB_READ,address, data,	\
 				     sizeof (data[0]));			\
 }
 
@@ -158,10 +180,10 @@ _(u8);
 
 #define _(t)								\
 static inline clib_error_t *						\
-vlib_pci_write_config_##t (vlib_pci_device_t * dev, uword address,	\
+vlib_pci_write_config_##t (vlib_pci_dev_handle_t h, uword address,	\
 			   t * data)					\
 {									\
-  return vlib_pci_read_write_config (dev, VLIB_WRITE,			\
+  return vlib_pci_read_write_config (h, VLIB_WRITE,			\
 				   address, data, sizeof (data[0]));	\
 }
 
@@ -172,45 +194,45 @@ _(u8);
 #undef _
 
 static inline clib_error_t *
-vlib_pci_intr_enable (vlib_pci_device_t * dev)
+vlib_pci_intr_enable (vlib_pci_dev_handle_t h)
 {
   u16 command;
   clib_error_t *err;
 
-  err = vlib_pci_read_config_u16 (dev, 4, &command);
+  err = vlib_pci_read_config_u16 (h, 4, &command);
 
   if (err)
     return err;
 
   command &= ~PCI_COMMAND_INTX_DISABLE;
 
-  return vlib_pci_write_config_u16 (dev, 4, &command);
+  return vlib_pci_write_config_u16 (h, 4, &command);
 }
 
 static inline clib_error_t *
-vlib_pci_intr_disable (vlib_pci_device_t * dev)
+vlib_pci_intr_disable (vlib_pci_dev_handle_t h)
 {
   u16 command;
   clib_error_t *err;
 
-  err = vlib_pci_read_config_u16 (dev, 4, &command);
+  err = vlib_pci_read_config_u16 (h, 4, &command);
 
   if (err)
     return err;
 
   command |= PCI_COMMAND_INTX_DISABLE;
 
-  return vlib_pci_write_config_u16 (dev, 4, &command);
+  return vlib_pci_write_config_u16 (h, 4, &command);
 }
 
 static inline clib_error_t *
-vlib_pci_bus_master_enable (vlib_pci_device_t * dev)
+vlib_pci_bus_master_enable (vlib_pci_dev_handle_t h)
 {
   clib_error_t *err;
   u16 command;
 
   /* Set bus master enable (BME) */
-  err = vlib_pci_read_config_u16 (dev, 4, &command);
+  err = vlib_pci_read_config_u16 (h, 4, &command);
 
   if (err)
     return err;
@@ -220,23 +242,18 @@ vlib_pci_bus_master_enable (vlib_pci_device_t * dev)
 
   command |= PCI_COMMAND_BUS_MASTER;
 
-  return vlib_pci_write_config_u16 (dev, 4, &command);
+  return vlib_pci_write_config_u16 (h, 4, &command);
 }
 
-clib_error_t *vlib_pci_map_resource (vlib_pci_device_t * dev, u32 resource,
+clib_error_t *vlib_pci_map_resource (vlib_pci_dev_handle_t h, u32 resource,
 				     void **result);
 
-clib_error_t *vlib_pci_map_resource_fixed (vlib_pci_device_t * dev,
+clib_error_t *vlib_pci_map_resource_fixed (vlib_pci_dev_handle_t h,
 					   u32 resource, u8 * addr,
 					   void **result);
 
-vlib_pci_device_t *vlib_get_pci_device (vlib_pci_addr_t * addr);
-/* Free's device. */
-void vlib_pci_free_device (vlib_pci_device_t * dev);
-
 unformat_function_t unformat_vlib_pci_addr;
 format_function_t format_vlib_pci_addr;
-format_function_t format_vlib_pci_handle;
 format_function_t format_vlib_pci_link_speed;
 format_function_t format_vlib_pci_vpd;