FD.io VPP  v17.07.01-10-g3be13f0
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 <vlib/vlib.h>
28 #include <vlib/unix/unix.h>
29 #include <vnet/ip/ip.h>
30 #include <vnet/ethernet/ethernet.h>
31 
33 
34 #define AF_PACKET_DEBUG_SOCKET 0
35 
36 #define AF_PACKET_TX_FRAMES_PER_BLOCK 1024
37 #define AF_PACKET_TX_FRAME_SIZE (2048 * 5)
38 #define AF_PACKET_TX_BLOCK_NR 1
39 #define AF_PACKET_TX_FRAME_NR (AF_PACKET_TX_BLOCK_NR * \
40  AF_PACKET_TX_FRAMES_PER_BLOCK)
41 #define AF_PACKET_TX_BLOCK_SIZE (AF_PACKET_TX_FRAME_SIZE * \
42  AF_PACKET_TX_FRAMES_PER_BLOCK)
43 
44 #define AF_PACKET_RX_FRAMES_PER_BLOCK 1024
45 #define AF_PACKET_RX_FRAME_SIZE (2048 * 5)
46 #define AF_PACKET_RX_BLOCK_NR 1
47 #define AF_PACKET_RX_FRAME_NR (AF_PACKET_RX_BLOCK_NR * \
48  AF_PACKET_RX_FRAMES_PER_BLOCK)
49 #define AF_PACKET_RX_BLOCK_SIZE (AF_PACKET_RX_FRAME_SIZE * \
50  AF_PACKET_RX_FRAMES_PER_BLOCK)
51 
52 #if AF_PACKET_DEBUG_SOCKET == 1
53 #define DBG_SOCK(args...) clib_warning(args);
54 #else
55 #define DBG_SOCK(args...)
56 #endif
57 
58 /*defined in net/if.h but clashes with dpdk headers */
59 unsigned int if_nametoindex (const char *ifname);
60 
61 typedef struct tpacket_req tpacket_req_t;
62 
63 static u32
65  u32 flags)
66 {
67  clib_error_t *error;
68  u8 *s;
70  af_packet_if_t *apif =
72 
74  {
75  s = format (0, "/sys/class/net/%s/mtu%c", apif->host_if_name, 0);
76 
77  error = vlib_sysfs_write ((char *) s, "%d", hi->max_packet_bytes);
78  vec_free (s);
79 
80  if (error)
81  {
82  clib_error_report (error);
83  return VNET_API_ERROR_SYSCALL_ERROR_1;
84  }
85  }
86 
87  return 0;
88 }
89 
90 static clib_error_t *
92 {
94  vnet_main_t *vnm = vnet_get_main ();
95  u32 idx = uf->private_data;
96  af_packet_if_t *apif = pool_elt_at_index (apm->interfaces, idx);
97 
100 
101  /* Schedule the rx node */
103 
104  return 0;
105 }
106 
107 static int
108 is_bridge (const u8 * host_if_name)
109 {
110  u8 *s;
111  DIR *dir = NULL;
112 
113  s = format (0, "/sys/class/net/%s/bridge%c", host_if_name, 0);
114  dir = opendir ((char *) s);
115  vec_free (s);
116 
117  if (dir)
118  {
119  closedir (dir);
120  return 0;
121  }
122 
123  return -1;
124 }
125 
126 static int
127 create_packet_v2_sock (int host_if_index, tpacket_req_t * rx_req,
128  tpacket_req_t * tx_req, int *fd, u8 ** ring)
129 {
130  int ret, err;
131  struct sockaddr_ll sll;
132  int ver = TPACKET_V2;
133  socklen_t req_sz = sizeof (struct tpacket_req);
134  u32 ring_sz = rx_req->tp_block_size * rx_req->tp_block_nr +
135  tx_req->tp_block_size * tx_req->tp_block_nr;
136 
137  if ((*fd = socket (AF_PACKET, SOCK_RAW, htons (ETH_P_ALL))) < 0)
138  {
139  DBG_SOCK ("Failed to create socket");
140  ret = VNET_API_ERROR_SYSCALL_ERROR_1;
141  goto error;
142  }
143 
144  if ((err =
145  setsockopt (*fd, SOL_PACKET, PACKET_VERSION, &ver, sizeof (ver))) < 0)
146  {
147  DBG_SOCK ("Failed to set rx packet interface version");
148  ret = VNET_API_ERROR_SYSCALL_ERROR_1;
149  goto error;
150  }
151 
152  int opt = 1;
153  if ((err =
154  setsockopt (*fd, SOL_PACKET, PACKET_LOSS, &opt, sizeof (opt))) < 0)
155  {
156  DBG_SOCK ("Failed to set packet tx ring error handling option");
157  ret = VNET_API_ERROR_SYSCALL_ERROR_1;
158  goto error;
159  }
160 
161  if ((err =
162  setsockopt (*fd, SOL_PACKET, PACKET_RX_RING, rx_req, req_sz)) < 0)
163  {
164  DBG_SOCK ("Failed to set packet rx ring options");
165  ret = VNET_API_ERROR_SYSCALL_ERROR_1;
166  goto error;
167  }
168 
169  if ((err =
170  setsockopt (*fd, SOL_PACKET, PACKET_TX_RING, tx_req, req_sz)) < 0)
171  {
172  DBG_SOCK ("Failed to set packet rx ring options");
173  ret = VNET_API_ERROR_SYSCALL_ERROR_1;
174  goto error;
175  }
176 
177  *ring =
178  mmap (NULL, ring_sz, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_LOCKED, *fd,
179  0);
180  if (*ring == MAP_FAILED)
181  {
182  DBG_SOCK ("mmap failure");
183  ret = VNET_API_ERROR_SYSCALL_ERROR_1;
184  goto error;
185  }
186 
187  memset (&sll, 0, sizeof (sll));
188  sll.sll_family = PF_PACKET;
189  sll.sll_protocol = htons (ETH_P_ALL);
190  sll.sll_ifindex = host_if_index;
191 
192  if ((err = bind (*fd, (struct sockaddr *) &sll, sizeof (sll))) < 0)
193  {
194  DBG_SOCK ("Failed to bind rx packet socket (error %d)", err);
195  ret = VNET_API_ERROR_SYSCALL_ERROR_1;
196  goto error;
197  }
198 
199  return 0;
200 error:
201  if (*fd >= 0)
202  close (*fd);
203  *fd = -1;
204  return ret;
205 }
206 
207 int
208 af_packet_create_if (vlib_main_t * vm, u8 * host_if_name, u8 * hw_addr_set,
209  u32 * sw_if_index)
210 {
212  int ret, fd = -1;
213  struct tpacket_req *rx_req = 0;
214  struct tpacket_req *tx_req = 0;
215  u8 *ring = 0;
216  af_packet_if_t *apif = 0;
217  u8 hw_addr[6];
218  clib_error_t *error;
222  vnet_main_t *vnm = vnet_get_main ();
223  uword *p;
224  uword if_index;
225  u8 *host_if_name_dup = vec_dup (host_if_name);
226  int host_if_index = -1;
227 
228  p = mhash_get (&apm->if_index_by_host_if_name, host_if_name);
229  if (p)
230  {
231  return VNET_API_ERROR_SUBIF_ALREADY_EXISTS;
232  }
233 
234  vec_validate (rx_req, 0);
235  rx_req->tp_block_size = AF_PACKET_RX_BLOCK_SIZE;
236  rx_req->tp_frame_size = AF_PACKET_RX_FRAME_SIZE;
237  rx_req->tp_block_nr = AF_PACKET_RX_BLOCK_NR;
238  rx_req->tp_frame_nr = AF_PACKET_RX_FRAME_NR;
239 
240  vec_validate (tx_req, 0);
241  tx_req->tp_block_size = AF_PACKET_TX_BLOCK_SIZE;
242  tx_req->tp_frame_size = AF_PACKET_TX_FRAME_SIZE;
243  tx_req->tp_block_nr = AF_PACKET_TX_BLOCK_NR;
244  tx_req->tp_frame_nr = AF_PACKET_TX_FRAME_NR;
245 
246  host_if_index = if_nametoindex ((const char *) host_if_name);
247 
248  if (!host_if_index)
249  {
250  DBG_SOCK ("Wrong host interface name");
251  return VNET_API_ERROR_INVALID_INTERFACE;
252  }
253 
254  ret = create_packet_v2_sock (host_if_index, rx_req, tx_req, &fd, &ring);
255 
256  if (ret != 0)
257  goto error;
258 
259  ret = is_bridge (host_if_name);
260 
261  if (ret == 0) /* is a bridge, ignore state */
262  host_if_index = -1;
263 
264  /* So far everything looks good, let's create interface */
265  pool_get (apm->interfaces, apif);
266  if_index = apif - apm->interfaces;
267 
268  apif->host_if_index = host_if_index;
269  apif->fd = fd;
270  apif->rx_ring = ring;
271  apif->tx_ring = ring + rx_req->tp_block_size * rx_req->tp_block_nr;
272  apif->rx_req = rx_req;
273  apif->tx_req = tx_req;
274  apif->host_if_name = host_if_name_dup;
275  apif->per_interface_next_index = ~0;
276  apif->next_tx_frame = 0;
277  apif->next_rx_frame = 0;
278 
279  if (tm->n_vlib_mains > 1)
280  clib_spinlock_init (&apif->lockp);
281 
282  {
283  unix_file_t template = { 0 };
285  template.file_descriptor = fd;
286  template.private_data = if_index;
288  apif->unix_file_index = unix_file_add (&unix_main, &template);
289  }
290 
291  /*use configured or generate random MAC address */
292  if (hw_addr_set)
293  clib_memcpy (hw_addr, hw_addr_set, 6);
294  else
295  {
296  f64 now = vlib_time_now (vm);
297  u32 rnd;
298  rnd = (u32) (now * 1e6);
299  rnd = random_u32 (&rnd);
300 
301  clib_memcpy (hw_addr + 2, &rnd, sizeof (rnd));
302  hw_addr[0] = 2;
303  hw_addr[1] = 0xfe;
304  }
305 
307  if_index, hw_addr, &apif->hw_if_index,
309 
310  if (error)
311  {
312  memset (apif, 0, sizeof (*apif));
313  pool_put (apm->interfaces, apif);
314  clib_error_report (error);
315  ret = VNET_API_ERROR_SYSCALL_ERROR_1;
316  goto error;
317  }
318 
319  sw = vnet_get_hw_sw_interface (vnm, apif->hw_if_index);
320  hw = vnet_get_hw_interface (vnm, apif->hw_if_index);
321  apif->sw_if_index = sw->sw_if_index;
323  af_packet_input_node.index);
324 
325  vnet_hw_interface_assign_rx_thread (vnm, apif->hw_if_index, 0, /* queue */
326  ~0 /* any cpu */ );
327 
331 
334 
335  mhash_set_mem (&apm->if_index_by_host_if_name, host_if_name_dup, &if_index,
336  0);
337  if (sw_if_index)
338  *sw_if_index = apif->sw_if_index;
339 
340  return 0;
341 
342 error:
343  vec_free (host_if_name_dup);
344  vec_free (rx_req);
345  vec_free (tx_req);
346  return ret;
347 }
348 
349 int
350 af_packet_delete_if (vlib_main_t * vm, u8 * host_if_name)
351 {
352  vnet_main_t *vnm = vnet_get_main ();
354  af_packet_if_t *apif;
355  uword *p;
356  uword if_index;
357  u32 ring_sz;
358 
359  p = mhash_get (&apm->if_index_by_host_if_name, host_if_name);
360  if (p == NULL)
361  {
362  clib_warning ("Host interface %s does not exist", host_if_name);
363  return VNET_API_ERROR_SYSCALL_ERROR_1;
364  }
365  apif = pool_elt_at_index (apm->interfaces, p[0]);
366  if_index = apif - apm->interfaces;
367 
368  /* bring down the interface */
371 
372  /* clean up */
373  if (apif->unix_file_index != ~0)
374  {
376  apif->unix_file_index = ~0;
377  }
378  else
379  close (apif->fd);
380 
381  ring_sz = apif->rx_req->tp_block_size * apif->rx_req->tp_block_nr +
382  apif->tx_req->tp_block_size * apif->tx_req->tp_block_nr;
383  if (munmap (apif->rx_ring, ring_sz))
384  clib_warning ("Host interface %s could not free rx/tx ring",
385  host_if_name);
386  apif->rx_ring = NULL;
387  apif->tx_ring = NULL;
388  apif->fd = -1;
389 
390  vec_free (apif->rx_req);
391  apif->rx_req = NULL;
392  vec_free (apif->tx_req);
393  apif->tx_req = NULL;
394 
395  vec_free (apif->host_if_name);
396  apif->host_if_name = NULL;
397  apif->host_if_index = -1;
398 
399  mhash_unset (&apm->if_index_by_host_if_name, host_if_name, &if_index);
400 
402 
403  pool_put (apm->interfaces, apif);
404 
405  return 0;
406 }
407 
408 static clib_error_t *
410 {
413 
414  memset (apm, 0, sizeof (af_packet_main_t));
415 
417 
420 
421  return 0;
422 }
423 
425 
426 /*
427  * fd.io coding-style-patch-verification: ON
428  *
429  * Local Variables:
430  * eval: (c-set-style "gnu")
431  * End:
432  */
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:436
unix_file_t * file_pool
Definition: unix.h:89
vmrglw vmrglh hi
clib_error_t * vnet_hw_interface_set_flags(vnet_main_t *vnm, u32 hw_if_index, u32 flags)
Definition: interface.c:537
unix_file_function_t * read_function
Definition: unix.h:62
uword * pending_input_bitmap
Definition: af_packet.h:50
void ethernet_delete_interface(vnet_main_t *vnm, u32 hw_if_index)
Definition: interface.c:295
vnet_main_t * vnet_get_main(void)
Definition: misc.c:46
#define AF_PACKET_TX_BLOCK_NR
Definition: af_packet.c:38
unix_main_t unix_main
Definition: main.c:60
#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:192
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:47
u32 unix_file_index
Definition: af_packet.h:35
static clib_error_t * af_packet_fd_read_ready(unix_file_t *uf)
Definition: af_packet.c:91
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:397
#define vec_validate_aligned(V, I, A)
Make sure vector is long enough for given index (no header, specified alignment)
Definition: vec.h:447
#define pool_get(P, E)
Allocate an object E from a pool P (unspecified alignment).
Definition: pool.h:200
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
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:41
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:135
static void clib_spinlock_init(clib_spinlock_t *p)
Definition: lock.h:33
static uword unix_file_add(unix_main_t *um, unix_file_t *template)
Definition: unix.h:136
#define pool_elt_at_index(p, i)
Returns pointer to element at given index.
Definition: pool.h:397
uword private_data
Definition: unix.h:59
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:241
#define vec_dup(V)
Return copy of vector (no header, no alignment)
Definition: vec.h:374
#define AF_PACKET_RX_BLOCK_NR
Definition: af_packet.c:46
#define VNET_HW_INTERFACE_FLAG_SUPPORTS_INT_MODE
Definition: interface.h:423
#define UNIX_FILE_EVENT_EDGE_TRIGGERED
Definition: unix.h:56
struct tpacket_req tpacket_req_t
Definition: af_packet.c:61
u32 next_tx_frame
Definition: af_packet.h:38
#define vec_free(V)
Free vector&#39;s memory (no header).
Definition: vec.h:340
#define clib_warning(format, args...)
Definition: error.h:59
#define clib_memcpy(a, b, c)
Definition: string.h:69
#define AF_PACKET_RX_BLOCK_SIZE
Definition: af_packet.c:49
#define ETHERNET_INTERFACE_FLAG_MTU
Definition: ethernet.h:118
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:289
int host_if_index
Definition: af_packet.h:27
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:122
#define DBG_SOCK(args...)
Definition: af_packet.c:55
static uword * mhash_get(mhash_t *h, const void *key)
Definition: mhash.h:110
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:246
af_packet_main_t af_packet_main
Definition: af_packet.h:59
#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
#define AF_PACKET_TX_FRAME_SIZE
Definition: af_packet.c:37
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:127
double f64
Definition: types.h:142
unsigned char u8
Definition: types.h:56
Definition: unix.h:49
#define AF_PACKET_RX_FRAME_SIZE
Definition: af_packet.c:45
clib_error_t * vlib_sysfs_write(char *file_name, char *fmt,...)
Definition: util.c:102
int vnet_hw_interface_unassign_rx_thread(vnet_main_t *vnm, u32 hw_if_index, u16 queue_id)
Definition: devices.c:171
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
static clib_error_t * af_packet_init(vlib_main_t *vm)
Definition: af_packet.c:409
static int is_bridge(const u8 *host_if_name)
Definition: af_packet.c:108
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:236
static u32 af_packet_eth_flag_change(vnet_main_t *vnm, vnet_hw_interface_t *hi, u32 flags)
Definition: af_packet.c:64
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:208
u32 flags
Definition: vhost-user.h:76
#define CLIB_CACHE_LINE_BYTES
Definition: cache.h:67
static void vnet_hw_interface_set_input_node(vnet_main_t *vnm, u32 hw_if_index, u32 node_index)
Definition: devices.h:78
#define AF_PACKET_TX_FRAME_NR
Definition: af_packet.c:39
static void unix_file_del(unix_main_t *um, unix_file_t *f)
Definition: unix.h:146
int af_packet_delete_if(vlib_main_t *vm, u8 *host_if_name)
Definition: af_packet.c:350