FD.io VPP  v18.04-17-g3a0d853
Vector Packet Processing
af_packet.c
Go to the documentation of this file.
1 /*
2  *------------------------------------------------------------------
3  * af_packet.c - linux kernel packet interface
4  *
5  * Copyright (c) 2016 Cisco and/or its affiliates.
6  * Licensed under the Apache License, Version 2.0 (the "License");
7  * you may not use this file except in compliance with the License.
8  * You may obtain a copy of the License at:
9  *
10  * http://www.apache.org/licenses/LICENSE-2.0
11  *
12  * Unless required by applicable law or agreed to in writing, software
13  * distributed under the License is distributed on an "AS IS" BASIS,
14  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15  * See the License for the specific language governing permissions and
16  * limitations under the License.
17  *------------------------------------------------------------------
18  */
19 
20 #include <linux/if_ether.h>
21 #include <linux/if_packet.h>
22 #include <dirent.h>
23 #include <sys/stat.h>
24 #include <sys/types.h>
25 #include <fcntl.h>
26 
27 #include <vppinfra/linux/sysfs.h>
28 #include <vlib/vlib.h>
29 #include <vlib/unix/unix.h>
30 #include <vnet/ip/ip.h>
31 #include <vnet/ethernet/ethernet.h>
32 
34 
36 
37 #define AF_PACKET_DEBUG_SOCKET 0
38 
39 #define AF_PACKET_TX_FRAMES_PER_BLOCK 1024
40 #define AF_PACKET_TX_FRAME_SIZE (2048 * 5)
41 #define AF_PACKET_TX_BLOCK_NR 1
42 #define AF_PACKET_TX_FRAME_NR (AF_PACKET_TX_BLOCK_NR * \
43  AF_PACKET_TX_FRAMES_PER_BLOCK)
44 #define AF_PACKET_TX_BLOCK_SIZE (AF_PACKET_TX_FRAME_SIZE * \
45  AF_PACKET_TX_FRAMES_PER_BLOCK)
46 
47 #define AF_PACKET_RX_FRAMES_PER_BLOCK 1024
48 #define AF_PACKET_RX_FRAME_SIZE (2048 * 5)
49 #define AF_PACKET_RX_BLOCK_NR 1
50 #define AF_PACKET_RX_FRAME_NR (AF_PACKET_RX_BLOCK_NR * \
51  AF_PACKET_RX_FRAMES_PER_BLOCK)
52 #define AF_PACKET_RX_BLOCK_SIZE (AF_PACKET_RX_FRAME_SIZE * \
53  AF_PACKET_RX_FRAMES_PER_BLOCK)
54 
55 #if AF_PACKET_DEBUG_SOCKET == 1
56 #define DBG_SOCK(args...) clib_warning(args);
57 #else
58 #define DBG_SOCK(args...)
59 #endif
60 
61 /*defined in net/if.h but clashes with dpdk headers */
62 unsigned int if_nametoindex (const char *ifname);
63 
64 typedef struct tpacket_req tpacket_req_t;
65 
66 static u32
68  u32 flags)
69 {
70  clib_error_t *error;
71  u8 *s;
73  af_packet_if_t *apif =
75 
77  {
78  s = format (0, "/sys/class/net/%s/mtu%c", apif->host_if_name, 0);
79 
80  error = clib_sysfs_write ((char *) s, "%d", hi->max_packet_bytes);
81  vec_free (s);
82 
83  if (error)
84  {
85  clib_error_report (error);
86  return VNET_API_ERROR_SYSCALL_ERROR_1;
87  }
88  }
89 
90  return 0;
91 }
92 
93 static clib_error_t *
95 {
97  vnet_main_t *vnm = vnet_get_main ();
98  u32 idx = uf->private_data;
99  af_packet_if_t *apif = pool_elt_at_index (apm->interfaces, idx);
100 
101  apm->pending_input_bitmap =
102  clib_bitmap_set (apm->pending_input_bitmap, idx, 1);
103 
104  /* Schedule the rx node */
106 
107  return 0;
108 }
109 
110 static int
111 is_bridge (const u8 * host_if_name)
112 {
113  u8 *s;
114  DIR *dir = NULL;
115 
116  s = format (0, "/sys/class/net/%s/bridge%c", host_if_name, 0);
117  dir = opendir ((char *) s);
118  vec_free (s);
119 
120  if (dir)
121  {
122  closedir (dir);
123  return 0;
124  }
125 
126  return -1;
127 }
128 
129 static int
130 create_packet_v2_sock (int host_if_index, tpacket_req_t * rx_req,
131  tpacket_req_t * tx_req, int *fd, u8 ** ring)
132 {
133  int ret, err;
134  struct sockaddr_ll sll;
135  int ver = TPACKET_V2;
136  socklen_t req_sz = sizeof (struct tpacket_req);
137  u32 ring_sz = rx_req->tp_block_size * rx_req->tp_block_nr +
138  tx_req->tp_block_size * tx_req->tp_block_nr;
139 
140  if ((*fd = socket (AF_PACKET, SOCK_RAW, htons (ETH_P_ALL))) < 0)
141  {
142  DBG_SOCK ("Failed to create socket");
143  ret = VNET_API_ERROR_SYSCALL_ERROR_1;
144  goto error;
145  }
146 
147  /* bind before rx ring is cfged so we don't receive packets from other interfaces */
148  memset (&sll, 0, sizeof (sll));
149  sll.sll_family = PF_PACKET;
150  sll.sll_protocol = htons (ETH_P_ALL);
151  sll.sll_ifindex = host_if_index;
152  if ((err = bind (*fd, (struct sockaddr *) &sll, sizeof (sll))) < 0)
153  {
154  DBG_SOCK ("Failed to bind rx packet socket (error %d)", err);
155  ret = VNET_API_ERROR_SYSCALL_ERROR_1;
156  goto error;
157  }
158 
159  if ((err =
160  setsockopt (*fd, SOL_PACKET, PACKET_VERSION, &ver, sizeof (ver))) < 0)
161  {
162  DBG_SOCK ("Failed to set rx packet interface version");
163  ret = VNET_API_ERROR_SYSCALL_ERROR_1;
164  goto error;
165  }
166 
167  int opt = 1;
168  if ((err =
169  setsockopt (*fd, SOL_PACKET, PACKET_LOSS, &opt, sizeof (opt))) < 0)
170  {
171  DBG_SOCK ("Failed to set packet tx ring error handling option");
172  ret = VNET_API_ERROR_SYSCALL_ERROR_1;
173  goto error;
174  }
175 
176  if ((err =
177  setsockopt (*fd, SOL_PACKET, PACKET_RX_RING, rx_req, req_sz)) < 0)
178  {
179  DBG_SOCK ("Failed to set packet rx ring options");
180  ret = VNET_API_ERROR_SYSCALL_ERROR_1;
181  goto error;
182  }
183 
184  if ((err =
185  setsockopt (*fd, SOL_PACKET, PACKET_TX_RING, tx_req, req_sz)) < 0)
186  {
187  DBG_SOCK ("Failed to set packet tx ring options");
188  ret = VNET_API_ERROR_SYSCALL_ERROR_1;
189  goto error;
190  }
191 
192  *ring =
193  mmap (NULL, ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_LOCKED, *fd,
194  0);
195  if (*ring == MAP_FAILED)
196  {
197  DBG_SOCK ("mmap failure");
198  ret = VNET_API_ERROR_SYSCALL_ERROR_1;
199  goto error;
200  }
201 
202  return 0;
203 error:
204  if (*fd >= 0)
205  close (*fd);
206  *fd = -1;
207  return ret;
208 }
209 
210 int
211 af_packet_create_if (vlib_main_t * vm, u8 * host_if_name, u8 * hw_addr_set,
212  u32 * sw_if_index)
213 {
215  int ret, fd = -1;
216  struct tpacket_req *rx_req = 0;
217  struct tpacket_req *tx_req = 0;
218  u8 *ring = 0;
219  af_packet_if_t *apif = 0;
220  u8 hw_addr[6];
221  clib_error_t *error;
225  vnet_main_t *vnm = vnet_get_main ();
226  uword *p;
227  uword if_index;
228  u8 *host_if_name_dup = vec_dup (host_if_name);
229  int host_if_index = -1;
230 
231  p = mhash_get (&apm->if_index_by_host_if_name, host_if_name);
232  if (p)
233  {
234  apif = vec_elt_at_index (apm->interfaces, p[0]);
235  *sw_if_index = apif->sw_if_index;
236  return VNET_API_ERROR_IF_ALREADY_EXISTS;
237  }
238 
239  vec_validate (rx_req, 0);
240  rx_req->tp_block_size = AF_PACKET_RX_BLOCK_SIZE;
241  rx_req->tp_frame_size = AF_PACKET_RX_FRAME_SIZE;
242  rx_req->tp_block_nr = AF_PACKET_RX_BLOCK_NR;
243  rx_req->tp_frame_nr = AF_PACKET_RX_FRAME_NR;
244 
245  vec_validate (tx_req, 0);
246  tx_req->tp_block_size = AF_PACKET_TX_BLOCK_SIZE;
247  tx_req->tp_frame_size = AF_PACKET_TX_FRAME_SIZE;
248  tx_req->tp_block_nr = AF_PACKET_TX_BLOCK_NR;
249  tx_req->tp_frame_nr = AF_PACKET_TX_FRAME_NR;
250 
251  host_if_index = if_nametoindex ((const char *) host_if_name);
252 
253  if (!host_if_index)
254  {
255  DBG_SOCK ("Wrong host interface name");
256  return VNET_API_ERROR_INVALID_INTERFACE;
257  }
258 
259  ret = create_packet_v2_sock (host_if_index, rx_req, tx_req, &fd, &ring);
260 
261  if (ret != 0)
262  goto error;
263 
264  ret = is_bridge (host_if_name);
265 
266  if (ret == 0) /* is a bridge, ignore state */
267  host_if_index = -1;
268 
269  /* So far everything looks good, let's create interface */
270  pool_get (apm->interfaces, apif);
271  if_index = apif - apm->interfaces;
272 
273  apif->host_if_index = host_if_index;
274  apif->fd = fd;
275  apif->rx_ring = ring;
276  apif->tx_ring = ring + rx_req->tp_block_size * rx_req->tp_block_nr;
277  apif->rx_req = rx_req;
278  apif->tx_req = tx_req;
279  apif->host_if_name = host_if_name_dup;
280  apif->per_interface_next_index = ~0;
281  apif->next_tx_frame = 0;
282  apif->next_rx_frame = 0;
283 
284  if (tm->n_vlib_mains > 1)
285  clib_spinlock_init (&apif->lockp);
286 
287  {
288  clib_file_t template = { 0 };
290  template.file_descriptor = fd;
291  template.private_data = if_index;
293  template.description = format (0, "%U", format_af_packet_device_name,
294  if_index);
295  apif->clib_file_index = clib_file_add (&file_main, &template);
296  }
297 
298  /*use configured or generate random MAC address */
299  if (hw_addr_set)
300  clib_memcpy (hw_addr, hw_addr_set, 6);
301  else
302  {
303  f64 now = vlib_time_now (vm);
304  u32 rnd;
305  rnd = (u32) (now * 1e6);
306  rnd = random_u32 (&rnd);
307 
308  clib_memcpy (hw_addr + 2, &rnd, sizeof (rnd));
309  hw_addr[0] = 2;
310  hw_addr[1] = 0xfe;
311  }
312 
314  if_index, hw_addr, &apif->hw_if_index,
316 
317  if (error)
318  {
319  memset (apif, 0, sizeof (*apif));
320  pool_put (apm->interfaces, apif);
321  clib_error_report (error);
322  ret = VNET_API_ERROR_SYSCALL_ERROR_1;
323  goto error;
324  }
325 
326  sw = vnet_get_hw_sw_interface (vnm, apif->hw_if_index);
327  hw = vnet_get_hw_interface (vnm, apif->hw_if_index);
328  apif->sw_if_index = sw->sw_if_index;
330  af_packet_input_node.index);
331 
332  vnet_hw_interface_assign_rx_thread (vnm, apif->hw_if_index, 0, /* queue */
333  ~0 /* any cpu */ );
334 
338 
341 
342  mhash_set_mem (&apm->if_index_by_host_if_name, host_if_name_dup, &if_index,
343  0);
344  if (sw_if_index)
345  *sw_if_index = apif->sw_if_index;
346 
347  return 0;
348 
349 error:
350  vec_free (host_if_name_dup);
351  vec_free (rx_req);
352  vec_free (tx_req);
353  return ret;
354 }
355 
356 int
358 {
359  vnet_main_t *vnm = vnet_get_main ();
361  af_packet_if_t *apif;
362  uword *p;
363  uword if_index;
364  u32 ring_sz;
365 
366  p = mhash_get (&apm->if_index_by_host_if_name, host_if_name);
367  if (p == NULL)
368  {
369  clib_warning ("Host interface %s does not exist", host_if_name);
370  return VNET_API_ERROR_SYSCALL_ERROR_1;
371  }
372  apif = pool_elt_at_index (apm->interfaces, p[0]);
373  if_index = apif - apm->interfaces;
374 
375  /* bring down the interface */
378 
379  /* clean up */
380  if (apif->clib_file_index != ~0)
381  {
383  apif->clib_file_index = ~0;
384  }
385  else
386  close (apif->fd);
387 
388  ring_sz = apif->rx_req->tp_block_size * apif->rx_req->tp_block_nr +
389  apif->tx_req->tp_block_size * apif->tx_req->tp_block_nr;
390  if (munmap (apif->rx_ring, ring_sz))
391  clib_warning ("Host interface %s could not free rx/tx ring",
392  host_if_name);
393  apif->rx_ring = NULL;
394  apif->tx_ring = NULL;
395  apif->fd = -1;
396 
397  vec_free (apif->rx_req);
398  apif->rx_req = NULL;
399  vec_free (apif->tx_req);
400  apif->tx_req = NULL;
401 
402  vec_free (apif->host_if_name);
403  apif->host_if_name = NULL;
404  apif->host_if_index = -1;
405 
406  mhash_unset (&apm->if_index_by_host_if_name, host_if_name, &if_index);
407 
409 
410  pool_put (apm->interfaces, apif);
411 
412  return 0;
413 }
414 
415 int
417 {
418  vnet_main_t *vnm = vnet_get_main ();
420 
421  hw = vnet_get_sup_hw_interface (vnm, sw_if_index);
422 
423  if (hw->dev_class_index != af_packet_device_class.index)
424  return VNET_API_ERROR_INVALID_INTERFACE;
425 
426  if (set)
428  else
430 
431  return 0;
432 }
433 
434 static clib_error_t *
436 {
439 
440  memset (apm, 0, sizeof (af_packet_main_t));
441 
443 
446 
447  return 0;
448 }
449 
451 
452 /*
453  * fd.io coding-style-patch-verification: ON
454  *
455  * Local Variables:
456  * eval: (c-set-style "gnu")
457  * End:
458  */
u32 ** rx_buffers
Definition: af_packet.h:53
#define vec_validate(V, I)
Make sure vector is long enough for given index (no header, unspecified alignment) ...
Definition: vec.h:434
#define UNIX_FILE_EVENT_EDGE_TRIGGERED
Definition: file.h:58
vmrglw vmrglh hi
static void clib_file_del(clib_file_main_t *um, clib_file_t *f)
Definition: file.h:109
clib_error_t * vnet_hw_interface_set_flags(vnet_main_t *vnm, u32 hw_if_index, u32 flags)
Definition: interface.c:554
uword * pending_input_bitmap
Definition: af_packet.h:50
void ethernet_delete_interface(vnet_main_t *vnm, u32 hw_if_index)
Definition: interface.c:321
int af_packet_set_l4_cksum_offload(vlib_main_t *vm, u32 sw_if_index, u8 set)
Definition: af_packet.c:416
vnet_main_t * vnet_get_main(void)
Definition: misc.c:47
static vnet_hw_interface_t * vnet_get_sup_hw_interface(vnet_main_t *vnm, u32 sw_if_index)
#define AF_PACKET_TX_BLOCK_NR
Definition: af_packet.c:41
#define NULL
Definition: clib.h:55
uword mhash_unset(mhash_t *h, void *key, uword *old_value)
Definition: mhash.c:353
static f64 vlib_time_now(vlib_main_t *vm)
Definition: main.h:227
struct tpacket_req * tx_req
Definition: af_packet.h:30
static vnet_hw_interface_t * vnet_get_hw_interface(vnet_main_t *vnm, u32 hw_if_index)
#define AF_PACKET_RX_FRAME_NR
Definition: af_packet.c:50
static void mhash_init_vec_string(mhash_t *h, uword n_value_bytes)
Definition: mhash.h:84
static uword * clib_bitmap_set(uword *ai, uword i, uword value)
Sets the ith bit of a bitmap to new_value Removes trailing zeros from the bitmap. ...
Definition: bitmap.h:167
u8 * format(u8 *s, const char *fmt,...)
Definition: format.c:419
struct tpacket_req * rx_req
Definition: af_packet.h:29
#define VNET_HW_INTERFACE_FLAG_LINK_UP
Definition: interface.h:390
#define vec_validate_aligned(V, I, A)
Make sure vector is long enough for given index (no header, specified alignment)
Definition: vec.h:445
#define pool_get(P, E)
Allocate an object E from a pool P (unspecified alignment).
Definition: pool.h:227
clib_error_t * clib_sysfs_write(char *file_name, char *fmt,...)
Definition: sysfs.c:26
clib_file_function_t * read_function
Definition: file.h:67
static vnet_sw_interface_t * vnet_get_hw_sw_interface(vnet_main_t *vnm, u32 hw_if_index)
u8 * host_if_name
Definition: af_packet.h:26
clib_file_t * file_pool
Definition: file.h:88
af_packet_if_t * interfaces
Definition: af_packet.h:47
uword flags
Definition: clib_error.h:29
#define AF_PACKET_TX_BLOCK_SIZE
Definition: af_packet.c:44
clib_spinlock_t lockp
Definition: af_packet.h:25
#define VLIB_INIT_FUNCTION(x)
Definition: init.h:111
u32 next_rx_frame
Definition: af_packet.h:37
static_always_inline void vnet_device_input_set_interrupt_pending(vnet_main_t *vnm, u32 hw_if_index, u16 queue_id)
Definition: devices.h:136
#define vec_elt_at_index(v, i)
Get vector value at index i checking that i is in bounds.
clib_file_main_t file_main
Definition: main.c:63
static void clib_spinlock_init(clib_spinlock_t *p)
Definition: lock.h:57
#define pool_elt_at_index(p, i)
Returns pointer to element at given index.
Definition: pool.h:461
format_function_t format_af_packet_device_name
Definition: af_packet.h:69
uword mhash_set_mem(mhash_t *h, void *key, uword *new_value, uword *old_value)
Definition: mhash.c:271
#define pool_put(P, E)
Free an object E in pool P.
Definition: pool.h:273
#define vec_dup(V)
Return copy of vector (no header, no alignment)
Definition: vec.h:370
#define AF_PACKET_RX_BLOCK_NR
Definition: af_packet.c:49
#define VNET_HW_INTERFACE_FLAG_SUPPORTS_INT_MODE
Definition: interface.h:428
struct tpacket_req tpacket_req_t
Definition: af_packet.c:64
u32 next_tx_frame
Definition: af_packet.h:38
vlib_main_t * vm
Definition: buffer.c:294
#define vec_free(V)
Free vector&#39;s memory (no header).
Definition: vec.h:336
#define clib_warning(format, args...)
Definition: error.h:59
#define clib_memcpy(a, b, c)
Definition: string.h:75
#define AF_PACKET_RX_BLOCK_SIZE
Definition: af_packet.c:52
#define ETHERNET_INTERFACE_FLAG_MTU
Definition: ethernet.h:152
#define VNET_HW_INTERFACE_FLAG_SUPPORTS_TX_L4_CKSUM_OFFLOAD
Definition: interface.h:431
u32 per_interface_next_index
Definition: af_packet.h:40
vlib_node_registration_t af_packet_input_node
(constructor) VLIB_REGISTER_NODE (af_packet_input_node)
Definition: node.c:369
int host_if_index
Definition: af_packet.h:27
u32 clib_file_index
Definition: af_packet.h:35
af_packet_main_t af_packet_main
Definition: af_packet.c:35
unsigned int u32
Definition: types.h:88
void vnet_hw_interface_assign_rx_thread(vnet_main_t *vnm, u32 hw_if_index, u16 queue_id, uword thread_index)
Definition: devices.c:138
#define DBG_SOCK(args...)
Definition: af_packet.c:58
static uword * mhash_get(mhash_t *h, const void *key)
Definition: mhash.h:110
static uword clib_file_add(clib_file_main_t *um, clib_file_t *template)
Definition: file.h:96
clib_error_t * ethernet_register_interface(vnet_main_t *vnm, u32 dev_class_index, u32 dev_instance, u8 *address, u32 *hw_if_index_return, ethernet_flag_change_function_t flag_change)
Definition: interface.c:273
#define clib_error_report(e)
Definition: error.h:113
mhash_t if_index_by_host_if_name
Definition: af_packet.h:56
unsigned int if_nametoindex(const char *ifname)
vnet_device_class_t af_packet_device_class
u64 uword
Definition: types.h:112
static clib_error_t * af_packet_fd_read_ready(clib_file_t *uf)
Definition: af_packet.c:94
#define AF_PACKET_TX_FRAME_SIZE
Definition: af_packet.c:40
static int create_packet_v2_sock(int host_if_index, tpacket_req_t *rx_req, tpacket_req_t *tx_req, int *fd, u8 **ring)
Definition: af_packet.c:130
double f64
Definition: types.h:142
unsigned char u8
Definition: types.h:56
#define AF_PACKET_RX_FRAME_SIZE
Definition: af_packet.c:48
int vnet_hw_interface_unassign_rx_thread(vnet_main_t *vnm, u32 hw_if_index, u16 queue_id)
Definition: devices.c:187
static u32 random_u32(u32 *seed)
32-bit random number generator
Definition: random.h:69
static vlib_thread_main_t * vlib_get_thread_main()
Definition: global_funcs.h:32
uword private_data
Definition: file.h:64
static clib_error_t * af_packet_init(vlib_main_t *vm)
Definition: af_packet.c:435
static int is_bridge(const u8 *host_if_name)
Definition: af_packet.c:111
Definition: file.h:51
int vnet_hw_interface_set_rx_mode(vnet_main_t *vnm, u32 hw_if_index, u16 queue_id, vnet_hw_interface_rx_mode mode)
Definition: devices.c:252
static u32 af_packet_eth_flag_change(vnet_main_t *vnm, vnet_hw_interface_t *hi, u32 flags)
Definition: af_packet.c:67
int af_packet_create_if(vlib_main_t *vm, u8 *host_if_name, u8 *hw_addr_set, u32 *sw_if_index)
Definition: af_packet.c:211
u32 flags
Definition: vhost-user.h:77
#define CLIB_CACHE_LINE_BYTES
Definition: cache.h:59
static void vnet_hw_interface_set_input_node(vnet_main_t *vnm, u32 hw_if_index, u32 node_index)
Definition: devices.h:79
#define AF_PACKET_TX_FRAME_NR
Definition: af_packet.c:42
int af_packet_delete_if(vlib_main_t *vm, u8 *host_if_name)
Definition: af_packet.c:357