]>
Commit | Line | Data |
---|---|---|
8a9562d2 PS |
1 | /* |
2 | * Copyright (c) 2014 Nicira, Inc. | |
3 | * | |
4 | * Licensed under the Apache License, Version 2.0 (the "License"); | |
5 | * you may not use this file except in compliance with the License. | |
6 | * You may obtain a copy of the License at: | |
7 | * | |
8 | * http://www.apache.org/licenses/LICENSE-2.0 | |
9 | * | |
10 | * Unless required by applicable law or agreed to in writing, software | |
11 | * distributed under the License is distributed on an "AS IS" BASIS, | |
12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | |
13 | * See the License for the specific language governing permissions and | |
14 | * limitations under the License. | |
15 | */ | |
16 | ||
17 | #include <config.h> | |
18 | ||
19 | #include <stdio.h> | |
20 | #include <string.h> | |
21 | #include <signal.h> | |
22 | #include <stdlib.h> | |
23 | #include <pthread.h> | |
24 | #include <config.h> | |
25 | #include <errno.h> | |
26 | #include <sched.h> | |
27 | #include <stdlib.h> | |
28 | #include <unistd.h> | |
29 | #include <stdio.h> | |
30 | ||
8a9562d2 PS |
31 | #include "dpif-netdev.h" |
32 | #include "list.h" | |
33 | #include "netdev-dpdk.h" | |
34 | #include "netdev-provider.h" | |
35 | #include "netdev-vport.h" | |
36 | #include "odp-util.h" | |
37 | #include "ofp-print.h" | |
38 | #include "ofpbuf.h" | |
39 | #include "ovs-thread.h" | |
40 | #include "ovs-rcu.h" | |
91088554 | 41 | #include "packet-dpif.h" |
8a9562d2 PS |
42 | #include "packets.h" |
43 | #include "shash.h" | |
8a9562d2 PS |
44 | #include "sset.h" |
45 | #include "unaligned.h" | |
46 | #include "timeval.h" | |
47 | #include "unixctl.h" | |
48 | #include "vlog.h" | |
49 | ||
50 | VLOG_DEFINE_THIS_MODULE(dpdk); | |
51 | static struct vlog_rate_limit rl = VLOG_RATE_LIMIT_INIT(5, 20); | |
52 | ||
53 | #define DPDK_PORT_WATCHDOG_INTERVAL 5 | |
54 | ||
55 | #define OVS_CACHE_LINE_SIZE CACHE_LINE_SIZE | |
56 | #define OVS_VPORT_DPDK "ovs_dpdk" | |
57 | ||
58 | /* | |
59 | * need to reserve tons of extra space in the mbufs so we can align the | |
60 | * DMA addresses to 4KB. | |
61 | */ | |
62 | ||
63 | #define MTU_TO_MAX_LEN(mtu) ((mtu) + ETHER_HDR_LEN + ETHER_CRC_LEN) | |
64 | #define MBUF_SIZE(mtu) (MTU_TO_MAX_LEN(mtu) + (512) + \ | |
65 | sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM) | |
66 | ||
645b8934 | 67 | /* XXX: mempool size should be based on system resources. */ |
8a9562d2 PS |
68 | #define NB_MBUF (4096 * 64) |
69 | #define MP_CACHE_SZ (256 * 2) | |
70 | #define SOCKET0 0 | |
71 | ||
72 | #define NON_PMD_THREAD_TX_QUEUE 0 | |
73 | ||
79f5354c PM |
74 | #define NIC_PORT_RX_Q_SIZE 2048 /* Size of Physical NIC RX Queue, Max (n+32<=4096)*/ |
75 | #define NIC_PORT_TX_Q_SIZE 2048 /* Size of Physical NIC TX Queue, Max (n+32<=4096)*/ | |
76 | ||
645b8934 | 77 | /* XXX: Needs per NIC value for these constants. */ |
8a9562d2 PS |
78 | #define RX_PTHRESH 32 /* Default values of RX prefetch threshold reg. */ |
79 | #define RX_HTHRESH 32 /* Default values of RX host threshold reg. */ | |
80 | #define RX_WTHRESH 16 /* Default values of RX write-back threshold reg. */ | |
81 | ||
82 | #define TX_PTHRESH 36 /* Default values of TX prefetch threshold reg. */ | |
83 | #define TX_HTHRESH 0 /* Default values of TX host threshold reg. */ | |
84 | #define TX_WTHRESH 0 /* Default values of TX write-back threshold reg. */ | |
85 | ||
86 | static const struct rte_eth_conf port_conf = { | |
a28ddd11 DDP |
87 | .rxmode = { |
88 | .mq_mode = ETH_MQ_RX_RSS, | |
89 | .split_hdr_size = 0, | |
90 | .header_split = 0, /* Header Split disabled */ | |
91 | .hw_ip_checksum = 0, /* IP checksum offload disabled */ | |
92 | .hw_vlan_filter = 0, /* VLAN filtering disabled */ | |
93 | .jumbo_frame = 0, /* Jumbo Frame Support disabled */ | |
94 | .hw_strip_crc = 0, | |
95 | }, | |
96 | .rx_adv_conf = { | |
97 | .rss_conf = { | |
98 | .rss_key = NULL, | |
61a2647e DDP |
99 | .rss_hf = ETH_RSS_IPV4_TCP | ETH_RSS_IPV4 | ETH_RSS_IPV6 |
100 | | ETH_RSS_IPV4_UDP | ETH_RSS_IPV6_TCP | ETH_RSS_IPV6_UDP, | |
8a9562d2 | 101 | }, |
a28ddd11 DDP |
102 | }, |
103 | .txmode = { | |
104 | .mq_mode = ETH_MQ_TX_NONE, | |
105 | }, | |
8a9562d2 PS |
106 | }; |
107 | ||
108 | static const struct rte_eth_rxconf rx_conf = { | |
a28ddd11 DDP |
109 | .rx_thresh = { |
110 | .pthresh = RX_PTHRESH, | |
111 | .hthresh = RX_HTHRESH, | |
112 | .wthresh = RX_WTHRESH, | |
113 | }, | |
8a9562d2 PS |
114 | }; |
115 | ||
116 | static const struct rte_eth_txconf tx_conf = { | |
a28ddd11 DDP |
117 | .tx_thresh = { |
118 | .pthresh = TX_PTHRESH, | |
119 | .hthresh = TX_HTHRESH, | |
120 | .wthresh = TX_WTHRESH, | |
121 | }, | |
122 | .tx_free_thresh = 0, | |
123 | .tx_rs_thresh = 0, | |
94777510 | 124 | .txq_flags = ETH_TXQ_FLAGS_NOMULTSEGS|ETH_TXQ_FLAGS_NOOFFLOADS, |
8a9562d2 PS |
125 | }; |
126 | ||
3a100265 DDP |
127 | enum { MAX_RX_QUEUE_LEN = 192 }; |
128 | enum { MAX_TX_QUEUE_LEN = 384 }; | |
58f7c37b DDP |
129 | enum { DPDK_RING_SIZE = 256 }; |
130 | BUILD_ASSERT_DECL(IS_POW2(DPDK_RING_SIZE)); | |
8a9562d2 PS |
131 | enum { DRAIN_TSC = 200000ULL }; |
132 | ||
133 | static int rte_eal_init_ret = ENODEV; | |
134 | ||
135 | static struct ovs_mutex dpdk_mutex = OVS_MUTEX_INITIALIZER; | |
136 | ||
137 | /* Contains all 'struct dpdk_dev's. */ | |
138 | static struct list dpdk_list OVS_GUARDED_BY(dpdk_mutex) | |
139 | = LIST_INITIALIZER(&dpdk_list); | |
140 | ||
141 | static struct list dpdk_mp_list OVS_GUARDED_BY(dpdk_mutex) | |
142 | = LIST_INITIALIZER(&dpdk_mp_list); | |
143 | ||
db73f716 DDP |
144 | /* This mutex must be used by non pmd threads when allocating or freeing |
145 | * mbufs through mempools. Since dpdk_queue_pkts() and dpdk_queue_flush() may | |
146 | * use mempools, a non pmd thread should hold this mutex while calling them */ | |
147 | struct ovs_mutex nonpmd_mempool_mutex = OVS_MUTEX_INITIALIZER; | |
148 | ||
8a9562d2 PS |
149 | struct dpdk_mp { |
150 | struct rte_mempool *mp; | |
151 | int mtu; | |
152 | int socket_id; | |
153 | int refcount; | |
154 | struct list list_node OVS_GUARDED_BY(dpdk_mutex); | |
155 | }; | |
156 | ||
157 | struct dpdk_tx_queue { | |
158 | rte_spinlock_t tx_lock; | |
159 | int count; | |
160 | uint64_t tsc; | |
161 | struct rte_mbuf *burst_pkts[MAX_TX_QUEUE_LEN]; | |
162 | }; | |
163 | ||
95fb793a | 164 | /* dpdk has no way to remove dpdk ring ethernet devices |
165 | so we have to keep them around once they've been created | |
166 | */ | |
167 | ||
168 | static struct list dpdk_ring_list OVS_GUARDED_BY(dpdk_mutex) | |
169 | = LIST_INITIALIZER(&dpdk_ring_list); | |
170 | ||
171 | struct dpdk_ring { | |
172 | /* For the client rings */ | |
173 | struct rte_ring *cring_tx; | |
174 | struct rte_ring *cring_rx; | |
175 | int user_port_id; /* User given port no, parsed from port name */ | |
176 | int eth_port_id; /* ethernet device port id */ | |
177 | struct list list_node OVS_GUARDED_BY(dpdk_mutex); | |
178 | }; | |
179 | ||
8a9562d2 PS |
180 | struct netdev_dpdk { |
181 | struct netdev up; | |
182 | int port_id; | |
183 | int max_packet_len; | |
184 | ||
185 | struct dpdk_tx_queue tx_q[NR_QUEUE]; | |
186 | ||
187 | struct ovs_mutex mutex OVS_ACQ_AFTER(dpdk_mutex); | |
188 | ||
189 | struct dpdk_mp *dpdk_mp; | |
190 | int mtu; | |
191 | int socket_id; | |
192 | int buf_size; | |
193 | struct netdev_stats stats_offset; | |
194 | struct netdev_stats stats; | |
195 | ||
196 | uint8_t hwaddr[ETH_ADDR_LEN]; | |
197 | enum netdev_flags flags; | |
198 | ||
199 | struct rte_eth_link link; | |
200 | int link_reset_cnt; | |
201 | ||
202 | /* In dpdk_list. */ | |
203 | struct list list_node OVS_GUARDED_BY(dpdk_mutex); | |
204 | }; | |
205 | ||
206 | struct netdev_rxq_dpdk { | |
207 | struct netdev_rxq up; | |
208 | int port_id; | |
209 | }; | |
210 | ||
db73f716 DDP |
211 | static bool thread_is_pmd(void); |
212 | ||
8a9562d2 PS |
213 | static int netdev_dpdk_construct(struct netdev *); |
214 | ||
215 | static bool | |
216 | is_dpdk_class(const struct netdev_class *class) | |
217 | { | |
218 | return class->construct == netdev_dpdk_construct; | |
219 | } | |
220 | ||
645b8934 | 221 | /* XXX: use dpdk malloc for entire OVS. infact huge page shld be used |
8a9562d2 PS |
222 | * for all other sengments data, bss and text. */ |
223 | ||
224 | static void * | |
225 | dpdk_rte_mzalloc(size_t sz) | |
226 | { | |
227 | void *ptr; | |
228 | ||
229 | ptr = rte_zmalloc(OVS_VPORT_DPDK, sz, OVS_CACHE_LINE_SIZE); | |
230 | if (ptr == NULL) { | |
231 | out_of_memory(); | |
232 | } | |
233 | return ptr; | |
234 | } | |
235 | ||
db73f716 DDP |
236 | /* XXX this function should be called only by pmd threads (or by non pmd |
237 | * threads holding the nonpmd_mempool_mutex) */ | |
8a9562d2 | 238 | void |
91088554 | 239 | free_dpdk_buf(struct dpif_packet *p) |
8a9562d2 | 240 | { |
db73f716 | 241 | struct rte_mbuf *pkt = (struct rte_mbuf *) p; |
8a9562d2 | 242 | |
db73f716 | 243 | rte_pktmbuf_free_seg(pkt); |
8a9562d2 PS |
244 | } |
245 | ||
b3cd9f9d PS |
246 | static void |
247 | __rte_pktmbuf_init(struct rte_mempool *mp, | |
248 | void *opaque_arg OVS_UNUSED, | |
249 | void *_m, | |
250 | unsigned i OVS_UNUSED) | |
251 | { | |
252 | struct rte_mbuf *m = _m; | |
91088554 | 253 | uint32_t buf_len = mp->elt_size - sizeof(struct dpif_packet); |
b3cd9f9d | 254 | |
91088554 | 255 | RTE_MBUF_ASSERT(mp->elt_size >= sizeof(struct dpif_packet)); |
b3cd9f9d PS |
256 | |
257 | memset(m, 0, mp->elt_size); | |
258 | ||
259 | /* start of buffer is just after mbuf structure */ | |
91088554 | 260 | m->buf_addr = (char *)m + sizeof(struct dpif_packet); |
b3cd9f9d | 261 | m->buf_physaddr = rte_mempool_virt2phy(mp, m) + |
91088554 | 262 | sizeof(struct dpif_packet); |
b3cd9f9d PS |
263 | m->buf_len = (uint16_t)buf_len; |
264 | ||
265 | /* keep some headroom between start of buffer and data */ | |
266 | m->pkt.data = (char*) m->buf_addr + RTE_MIN(RTE_PKTMBUF_HEADROOM, m->buf_len); | |
267 | ||
268 | /* init some constant fields */ | |
269 | m->type = RTE_MBUF_PKT; | |
270 | m->pool = mp; | |
271 | m->pkt.nb_segs = 1; | |
272 | m->pkt.in_port = 0xff; | |
273 | } | |
274 | ||
275 | static void | |
276 | ovs_rte_pktmbuf_init(struct rte_mempool *mp, | |
277 | void *opaque_arg OVS_UNUSED, | |
278 | void *_m, | |
279 | unsigned i OVS_UNUSED) | |
280 | { | |
281 | struct rte_mbuf *m = _m; | |
282 | ||
283 | __rte_pktmbuf_init(mp, opaque_arg, _m, i); | |
284 | ||
285 | ofpbuf_init_dpdk((struct ofpbuf *) m, m->buf_len); | |
286 | } | |
287 | ||
8a9562d2 PS |
288 | static struct dpdk_mp * |
289 | dpdk_mp_get(int socket_id, int mtu) OVS_REQUIRES(dpdk_mutex) | |
290 | { | |
291 | struct dpdk_mp *dmp = NULL; | |
292 | char mp_name[RTE_MEMPOOL_NAMESIZE]; | |
293 | ||
294 | LIST_FOR_EACH (dmp, list_node, &dpdk_mp_list) { | |
295 | if (dmp->socket_id == socket_id && dmp->mtu == mtu) { | |
296 | dmp->refcount++; | |
297 | return dmp; | |
298 | } | |
299 | } | |
300 | ||
301 | dmp = dpdk_rte_mzalloc(sizeof *dmp); | |
302 | dmp->socket_id = socket_id; | |
303 | dmp->mtu = mtu; | |
304 | dmp->refcount = 1; | |
305 | ||
34631d72 AW |
306 | if (snprintf(mp_name, RTE_MEMPOOL_NAMESIZE, "ovs_mp_%d_%d", dmp->mtu, |
307 | dmp->socket_id) < 0) { | |
95fb793a | 308 | return NULL; |
309 | } | |
310 | ||
8a9562d2 PS |
311 | dmp->mp = rte_mempool_create(mp_name, NB_MBUF, MBUF_SIZE(mtu), |
312 | MP_CACHE_SZ, | |
313 | sizeof(struct rte_pktmbuf_pool_private), | |
314 | rte_pktmbuf_pool_init, NULL, | |
b3cd9f9d | 315 | ovs_rte_pktmbuf_init, NULL, |
8a9562d2 PS |
316 | socket_id, 0); |
317 | ||
318 | if (dmp->mp == NULL) { | |
319 | return NULL; | |
320 | } | |
321 | ||
322 | list_push_back(&dpdk_mp_list, &dmp->list_node); | |
323 | return dmp; | |
324 | } | |
325 | ||
326 | static void | |
327 | dpdk_mp_put(struct dpdk_mp *dmp) | |
328 | { | |
329 | ||
330 | if (!dmp) { | |
331 | return; | |
332 | } | |
333 | ||
334 | dmp->refcount--; | |
335 | ovs_assert(dmp->refcount >= 0); | |
336 | ||
337 | #if 0 | |
338 | /* I could not find any API to destroy mp. */ | |
339 | if (dmp->refcount == 0) { | |
340 | list_delete(dmp->list_node); | |
341 | /* destroy mp-pool. */ | |
342 | } | |
343 | #endif | |
344 | } | |
345 | ||
346 | static void | |
347 | check_link_status(struct netdev_dpdk *dev) | |
348 | { | |
349 | struct rte_eth_link link; | |
350 | ||
351 | rte_eth_link_get_nowait(dev->port_id, &link); | |
352 | ||
353 | if (dev->link.link_status != link.link_status) { | |
3e912ffc | 354 | netdev_change_seq_changed(&dev->up); |
8a9562d2 PS |
355 | |
356 | dev->link_reset_cnt++; | |
357 | dev->link = link; | |
358 | if (dev->link.link_status) { | |
359 | VLOG_DBG_RL(&rl, "Port %d Link Up - speed %u Mbps - %s", | |
360 | dev->port_id, (unsigned)dev->link.link_speed, | |
361 | (dev->link.link_duplex == ETH_LINK_FULL_DUPLEX) ? | |
362 | ("full-duplex") : ("half-duplex")); | |
363 | } else { | |
364 | VLOG_DBG_RL(&rl, "Port %d Link Down", dev->port_id); | |
365 | } | |
366 | } | |
367 | } | |
368 | ||
369 | static void * | |
370 | dpdk_watchdog(void *dummy OVS_UNUSED) | |
371 | { | |
372 | struct netdev_dpdk *dev; | |
373 | ||
374 | pthread_detach(pthread_self()); | |
375 | ||
376 | for (;;) { | |
377 | ovs_mutex_lock(&dpdk_mutex); | |
378 | LIST_FOR_EACH (dev, list_node, &dpdk_list) { | |
379 | ovs_mutex_lock(&dev->mutex); | |
380 | check_link_status(dev); | |
381 | ovs_mutex_unlock(&dev->mutex); | |
382 | } | |
383 | ovs_mutex_unlock(&dpdk_mutex); | |
384 | xsleep(DPDK_PORT_WATCHDOG_INTERVAL); | |
385 | } | |
386 | ||
387 | return NULL; | |
388 | } | |
389 | ||
390 | static int | |
391 | dpdk_eth_dev_init(struct netdev_dpdk *dev) OVS_REQUIRES(dpdk_mutex) | |
392 | { | |
393 | struct rte_pktmbuf_pool_private *mbp_priv; | |
394 | struct ether_addr eth_addr; | |
395 | int diag; | |
396 | int i; | |
397 | ||
398 | if (dev->port_id < 0 || dev->port_id >= rte_eth_dev_count()) { | |
95fb793a | 399 | return ENODEV; |
8a9562d2 PS |
400 | } |
401 | ||
402 | diag = rte_eth_dev_configure(dev->port_id, NR_QUEUE, NR_QUEUE, &port_conf); | |
403 | if (diag) { | |
404 | VLOG_ERR("eth dev config error %d",diag); | |
95fb793a | 405 | return -diag; |
8a9562d2 PS |
406 | } |
407 | ||
408 | for (i = 0; i < NR_QUEUE; i++) { | |
79f5354c | 409 | diag = rte_eth_tx_queue_setup(dev->port_id, i, NIC_PORT_TX_Q_SIZE, |
d221ffa1 | 410 | dev->socket_id, &tx_conf); |
8a9562d2 PS |
411 | if (diag) { |
412 | VLOG_ERR("eth dev tx queue setup error %d",diag); | |
95fb793a | 413 | return -diag; |
8a9562d2 PS |
414 | } |
415 | } | |
416 | ||
417 | for (i = 0; i < NR_QUEUE; i++) { | |
79f5354c | 418 | diag = rte_eth_rx_queue_setup(dev->port_id, i, NIC_PORT_RX_Q_SIZE, |
d221ffa1 | 419 | dev->socket_id, |
a715f600 | 420 | &rx_conf, dev->dpdk_mp->mp); |
8a9562d2 PS |
421 | if (diag) { |
422 | VLOG_ERR("eth dev rx queue setup error %d",diag); | |
95fb793a | 423 | return -diag; |
8a9562d2 PS |
424 | } |
425 | } | |
426 | ||
427 | diag = rte_eth_dev_start(dev->port_id); | |
428 | if (diag) { | |
429 | VLOG_ERR("eth dev start error %d",diag); | |
95fb793a | 430 | return -diag; |
8a9562d2 PS |
431 | } |
432 | ||
433 | rte_eth_promiscuous_enable(dev->port_id); | |
434 | rte_eth_allmulticast_enable(dev->port_id); | |
435 | ||
436 | memset(ð_addr, 0x0, sizeof(eth_addr)); | |
437 | rte_eth_macaddr_get(dev->port_id, ð_addr); | |
438 | VLOG_INFO_RL(&rl, "Port %d: "ETH_ADDR_FMT"", | |
439 | dev->port_id, ETH_ADDR_ARGS(eth_addr.addr_bytes)); | |
440 | ||
441 | memcpy(dev->hwaddr, eth_addr.addr_bytes, ETH_ADDR_LEN); | |
442 | rte_eth_link_get_nowait(dev->port_id, &dev->link); | |
443 | ||
444 | mbp_priv = rte_mempool_get_priv(dev->dpdk_mp->mp); | |
445 | dev->buf_size = mbp_priv->mbuf_data_room_size - RTE_PKTMBUF_HEADROOM; | |
446 | ||
447 | dev->flags = NETDEV_UP | NETDEV_PROMISC; | |
448 | return 0; | |
449 | } | |
450 | ||
451 | static struct netdev_dpdk * | |
452 | netdev_dpdk_cast(const struct netdev *netdev) | |
453 | { | |
454 | return CONTAINER_OF(netdev, struct netdev_dpdk, up); | |
455 | } | |
456 | ||
457 | static struct netdev * | |
458 | netdev_dpdk_alloc(void) | |
459 | { | |
460 | struct netdev_dpdk *netdev = dpdk_rte_mzalloc(sizeof *netdev); | |
461 | return &netdev->up; | |
462 | } | |
463 | ||
464 | static int | |
95fb793a | 465 | netdev_dpdk_init(struct netdev *netdev_, unsigned int port_no) OVS_REQUIRES(dpdk_mutex) |
8a9562d2 PS |
466 | { |
467 | struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_); | |
95fb793a | 468 | int err = 0; |
8a9562d2 PS |
469 | int i; |
470 | ||
95fb793a | 471 | ovs_mutex_init(&netdev->mutex); |
8a9562d2 | 472 | |
95fb793a | 473 | ovs_mutex_lock(&netdev->mutex); |
8a9562d2 PS |
474 | |
475 | for (i = 0; i < NR_QUEUE; i++) { | |
476 | rte_spinlock_init(&netdev->tx_q[i].tx_lock); | |
477 | } | |
478 | ||
95fb793a | 479 | netdev->port_id = port_no; |
8a9562d2 | 480 | |
8a9562d2 | 481 | netdev->flags = 0; |
8a9562d2 PS |
482 | netdev->mtu = ETHER_MTU; |
483 | netdev->max_packet_len = MTU_TO_MAX_LEN(netdev->mtu); | |
484 | ||
645b8934 | 485 | /* XXX: need to discover device node at run time. */ |
8a9562d2 | 486 | netdev->socket_id = SOCKET0; |
8a9562d2 PS |
487 | |
488 | netdev->dpdk_mp = dpdk_mp_get(netdev->socket_id, netdev->mtu); | |
489 | if (!netdev->dpdk_mp) { | |
490 | err = ENOMEM; | |
95fb793a | 491 | goto unlock; |
8a9562d2 PS |
492 | } |
493 | ||
494 | err = dpdk_eth_dev_init(netdev); | |
495 | if (err) { | |
95fb793a | 496 | goto unlock; |
8a9562d2 | 497 | } |
f00fa8cb | 498 | netdev_->n_txq = NR_QUEUE; |
8a9562d2 PS |
499 | netdev_->n_rxq = NR_QUEUE; |
500 | ||
501 | list_push_back(&dpdk_list, &netdev->list_node); | |
502 | ||
95fb793a | 503 | unlock: |
8a9562d2 | 504 | ovs_mutex_unlock(&netdev->mutex); |
95fb793a | 505 | return err; |
506 | } | |
507 | ||
508 | static int | |
509 | dpdk_dev_parse_name(const char dev_name[], const char prefix[], | |
510 | unsigned int *port_no) | |
511 | { | |
512 | const char *cport; | |
513 | ||
514 | if (strncmp(dev_name, prefix, strlen(prefix))) { | |
515 | return ENODEV; | |
516 | } | |
517 | ||
518 | cport = dev_name + strlen(prefix); | |
519 | *port_no = strtol(cport, 0, 0); /* string must be null terminated */ | |
520 | return 0; | |
521 | } | |
522 | ||
523 | static int | |
524 | netdev_dpdk_construct(struct netdev *netdev) | |
525 | { | |
526 | unsigned int port_no; | |
527 | int err; | |
528 | ||
529 | if (rte_eal_init_ret) { | |
530 | return rte_eal_init_ret; | |
531 | } | |
532 | ||
533 | /* Names always start with "dpdk" */ | |
534 | err = dpdk_dev_parse_name(netdev->name, "dpdk", &port_no); | |
535 | if (err) { | |
536 | return err; | |
537 | } | |
538 | ||
539 | ovs_mutex_lock(&dpdk_mutex); | |
540 | err = netdev_dpdk_init(netdev, port_no); | |
8a9562d2 PS |
541 | ovs_mutex_unlock(&dpdk_mutex); |
542 | return err; | |
543 | } | |
544 | ||
545 | static void | |
546 | netdev_dpdk_destruct(struct netdev *netdev_) | |
547 | { | |
548 | struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_); | |
549 | ||
550 | ovs_mutex_lock(&dev->mutex); | |
551 | rte_eth_dev_stop(dev->port_id); | |
552 | ovs_mutex_unlock(&dev->mutex); | |
553 | ||
554 | ovs_mutex_lock(&dpdk_mutex); | |
555 | list_remove(&dev->list_node); | |
556 | dpdk_mp_put(dev->dpdk_mp); | |
557 | ovs_mutex_unlock(&dpdk_mutex); | |
558 | ||
559 | ovs_mutex_destroy(&dev->mutex); | |
560 | } | |
561 | ||
562 | static void | |
563 | netdev_dpdk_dealloc(struct netdev *netdev_) | |
564 | { | |
565 | struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_); | |
566 | ||
567 | rte_free(netdev); | |
568 | } | |
569 | ||
570 | static int | |
571 | netdev_dpdk_get_config(const struct netdev *netdev_, struct smap *args) | |
572 | { | |
573 | struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_); | |
574 | ||
575 | ovs_mutex_lock(&dev->mutex); | |
576 | ||
645b8934 | 577 | /* XXX: Allow to configure number of queues. */ |
8a9562d2 PS |
578 | smap_add_format(args, "configured_rx_queues", "%u", netdev_->n_rxq); |
579 | smap_add_format(args, "configured_tx_queues", "%u", netdev_->n_rxq); | |
580 | ovs_mutex_unlock(&dev->mutex); | |
581 | ||
582 | return 0; | |
583 | } | |
584 | ||
7dec44fe AW |
585 | static int |
586 | netdev_dpdk_get_numa_id(const struct netdev *netdev_) | |
587 | { | |
588 | struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_); | |
589 | ||
590 | return netdev->socket_id; | |
591 | } | |
592 | ||
8a9562d2 PS |
593 | static struct netdev_rxq * |
594 | netdev_dpdk_rxq_alloc(void) | |
595 | { | |
596 | struct netdev_rxq_dpdk *rx = dpdk_rte_mzalloc(sizeof *rx); | |
597 | ||
598 | return &rx->up; | |
599 | } | |
600 | ||
601 | static struct netdev_rxq_dpdk * | |
602 | netdev_rxq_dpdk_cast(const struct netdev_rxq *rx) | |
603 | { | |
604 | return CONTAINER_OF(rx, struct netdev_rxq_dpdk, up); | |
605 | } | |
606 | ||
607 | static int | |
608 | netdev_dpdk_rxq_construct(struct netdev_rxq *rxq_) | |
609 | { | |
610 | struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_); | |
611 | struct netdev_dpdk *netdev = netdev_dpdk_cast(rx->up.netdev); | |
612 | ||
613 | ovs_mutex_lock(&netdev->mutex); | |
614 | rx->port_id = netdev->port_id; | |
615 | ovs_mutex_unlock(&netdev->mutex); | |
616 | ||
617 | return 0; | |
618 | } | |
619 | ||
620 | static void | |
621 | netdev_dpdk_rxq_destruct(struct netdev_rxq *rxq_ OVS_UNUSED) | |
622 | { | |
623 | } | |
624 | ||
625 | static void | |
626 | netdev_dpdk_rxq_dealloc(struct netdev_rxq *rxq_) | |
627 | { | |
628 | struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_); | |
629 | ||
630 | rte_free(rx); | |
631 | } | |
632 | ||
b170db2a RW |
633 | static inline void |
634 | dpdk_queue_flush__(struct netdev_dpdk *dev, int qid) | |
8a9562d2 PS |
635 | { |
636 | struct dpdk_tx_queue *txq = &dev->tx_q[qid]; | |
1304f1f8 DDP |
637 | uint32_t nb_tx = 0; |
638 | ||
639 | while (nb_tx != txq->count) { | |
640 | uint32_t ret; | |
641 | ||
642 | ret = rte_eth_tx_burst(dev->port_id, qid, txq->burst_pkts + nb_tx, | |
643 | txq->count - nb_tx); | |
644 | if (!ret) { | |
645 | break; | |
646 | } | |
647 | ||
648 | nb_tx += ret; | |
649 | } | |
8a9562d2 | 650 | |
b170db2a | 651 | if (OVS_UNLIKELY(nb_tx != txq->count)) { |
db73f716 DDP |
652 | /* free buffers, which we couldn't transmit, one at a time (each |
653 | * packet could come from a different mempool) */ | |
654 | int i; | |
655 | ||
656 | for (i = nb_tx; i < txq->count; i++) { | |
657 | rte_pktmbuf_free_seg(txq->burst_pkts[i]); | |
658 | } | |
1304f1f8 DDP |
659 | ovs_mutex_lock(&dev->mutex); |
660 | dev->stats.tx_dropped += txq->count-nb_tx; | |
661 | ovs_mutex_unlock(&dev->mutex); | |
8a9562d2 | 662 | } |
1304f1f8 | 663 | |
8a9562d2 | 664 | txq->count = 0; |
844f2d74 | 665 | txq->tsc = rte_get_timer_cycles(); |
b170db2a RW |
666 | } |
667 | ||
668 | static inline void | |
669 | dpdk_queue_flush(struct netdev_dpdk *dev, int qid) | |
670 | { | |
671 | struct dpdk_tx_queue *txq = &dev->tx_q[qid]; | |
672 | ||
673 | if (txq->count == 0) { | |
674 | return; | |
675 | } | |
676 | rte_spinlock_lock(&txq->tx_lock); | |
677 | dpdk_queue_flush__(dev, qid); | |
8a9562d2 PS |
678 | rte_spinlock_unlock(&txq->tx_lock); |
679 | } | |
680 | ||
8a9562d2 | 681 | static int |
91088554 DDP |
682 | netdev_dpdk_rxq_recv(struct netdev_rxq *rxq_, struct dpif_packet **packets, |
683 | int *c) | |
8a9562d2 PS |
684 | { |
685 | struct netdev_rxq_dpdk *rx = netdev_rxq_dpdk_cast(rxq_); | |
686 | struct netdev *netdev = rx->up.netdev; | |
687 | struct netdev_dpdk *dev = netdev_dpdk_cast(netdev); | |
8a9562d2 | 688 | int nb_rx; |
8a9562d2 PS |
689 | |
690 | dpdk_queue_flush(dev, rxq_->queue_id); | |
691 | ||
692 | nb_rx = rte_eth_rx_burst(rx->port_id, rxq_->queue_id, | |
7d08d53e DDP |
693 | (struct rte_mbuf **) packets, |
694 | MIN((int)NETDEV_MAX_RX_BATCH, | |
695 | (int)MAX_RX_QUEUE_LEN)); | |
8a9562d2 PS |
696 | if (!nb_rx) { |
697 | return EAGAIN; | |
698 | } | |
699 | ||
8a9562d2 PS |
700 | *c = nb_rx; |
701 | ||
702 | return 0; | |
703 | } | |
704 | ||
705 | inline static void | |
f4fd623c DDP |
706 | dpdk_queue_pkts(struct netdev_dpdk *dev, int qid, |
707 | struct rte_mbuf **pkts, int cnt) | |
8a9562d2 PS |
708 | { |
709 | struct dpdk_tx_queue *txq = &dev->tx_q[qid]; | |
710 | uint64_t diff_tsc; | |
8a9562d2 | 711 | |
f4fd623c DDP |
712 | int i = 0; |
713 | ||
8a9562d2 | 714 | rte_spinlock_lock(&txq->tx_lock); |
f4fd623c DDP |
715 | while (i < cnt) { |
716 | int freeslots = MAX_TX_QUEUE_LEN - txq->count; | |
717 | int tocopy = MIN(freeslots, cnt-i); | |
8a9562d2 | 718 | |
f4fd623c DDP |
719 | memcpy(&txq->burst_pkts[txq->count], &pkts[i], |
720 | tocopy * sizeof (struct rte_mbuf *)); | |
721 | ||
722 | txq->count += tocopy; | |
723 | i += tocopy; | |
724 | ||
725 | if (txq->count == MAX_TX_QUEUE_LEN) { | |
b170db2a | 726 | dpdk_queue_flush__(dev, qid); |
f4fd623c | 727 | } |
844f2d74 | 728 | diff_tsc = rte_get_timer_cycles() - txq->tsc; |
f4fd623c | 729 | if (diff_tsc >= DRAIN_TSC) { |
b170db2a | 730 | dpdk_queue_flush__(dev, qid); |
f4fd623c | 731 | } |
8a9562d2 | 732 | } |
8a9562d2 PS |
733 | rte_spinlock_unlock(&txq->tx_lock); |
734 | } | |
735 | ||
736 | /* Tx function. Transmit packets indefinitely */ | |
737 | static void | |
f4fd623c | 738 | dpdk_do_tx_copy(struct netdev *netdev, struct dpif_packet ** pkts, int cnt) |
db73f716 | 739 | OVS_NO_THREAD_SAFETY_ANALYSIS |
8a9562d2 PS |
740 | { |
741 | struct netdev_dpdk *dev = netdev_dpdk_cast(netdev); | |
f4fd623c | 742 | struct rte_mbuf *mbufs[cnt]; |
175cf4de RW |
743 | int dropped = 0; |
744 | int newcnt = 0; | |
745 | int i; | |
8a9562d2 | 746 | |
db73f716 DDP |
747 | /* If we are on a non pmd thread we have to use the mempool mutex, because |
748 | * every non pmd thread shares the same mempool cache */ | |
749 | ||
750 | if (!thread_is_pmd()) { | |
751 | ovs_mutex_lock(&nonpmd_mempool_mutex); | |
752 | } | |
753 | ||
f4fd623c DDP |
754 | for (i = 0; i < cnt; i++) { |
755 | int size = ofpbuf_size(&pkts[i]->ofpbuf); | |
95fb793a | 756 | |
f98d7864 | 757 | if (OVS_UNLIKELY(size > dev->max_packet_len)) { |
f4fd623c DDP |
758 | VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d", |
759 | (int)size , dev->max_packet_len); | |
760 | ||
175cf4de | 761 | dropped++; |
f4fd623c DDP |
762 | continue; |
763 | } | |
8a9562d2 | 764 | |
f4fd623c | 765 | mbufs[newcnt] = rte_pktmbuf_alloc(dev->dpdk_mp->mp); |
8a9562d2 | 766 | |
f4fd623c | 767 | if (!mbufs[newcnt]) { |
175cf4de RW |
768 | dropped += cnt - i; |
769 | break; | |
f4fd623c DDP |
770 | } |
771 | ||
772 | /* We have to do a copy for now */ | |
773 | memcpy(mbufs[newcnt]->pkt.data, ofpbuf_data(&pkts[i]->ofpbuf), size); | |
774 | ||
775 | rte_pktmbuf_data_len(mbufs[newcnt]) = size; | |
776 | rte_pktmbuf_pkt_len(mbufs[newcnt]) = size; | |
777 | ||
778 | newcnt++; | |
779 | } | |
8a9562d2 | 780 | |
f98d7864 | 781 | if (OVS_UNLIKELY(dropped)) { |
175cf4de RW |
782 | ovs_mutex_lock(&dev->mutex); |
783 | dev->stats.tx_dropped += dropped; | |
784 | ovs_mutex_unlock(&dev->mutex); | |
785 | } | |
786 | ||
f4fd623c | 787 | dpdk_queue_pkts(dev, NON_PMD_THREAD_TX_QUEUE, mbufs, newcnt); |
8a9562d2 | 788 | dpdk_queue_flush(dev, NON_PMD_THREAD_TX_QUEUE); |
db73f716 DDP |
789 | |
790 | if (!thread_is_pmd()) { | |
791 | ovs_mutex_unlock(&nonpmd_mempool_mutex); | |
792 | } | |
8a9562d2 PS |
793 | } |
794 | ||
795 | static int | |
f00fa8cb AW |
796 | netdev_dpdk_send(struct netdev *netdev, int qid, struct dpif_packet **pkts, |
797 | int cnt, bool may_steal) | |
8a9562d2 PS |
798 | { |
799 | struct netdev_dpdk *dev = netdev_dpdk_cast(netdev); | |
800 | int ret; | |
f4fd623c | 801 | int i; |
8a9562d2 | 802 | |
f4fd623c DDP |
803 | if (!may_steal || pkts[0]->ofpbuf.source != OFPBUF_DPDK) { |
804 | dpdk_do_tx_copy(netdev, pkts, cnt); | |
b3cd9f9d PS |
805 | |
806 | if (may_steal) { | |
f4fd623c DDP |
807 | for (i = 0; i < cnt; i++) { |
808 | dpif_packet_delete(pkts[i]); | |
809 | } | |
b3cd9f9d | 810 | } |
8a9562d2 | 811 | } else { |
f4fd623c DDP |
812 | int next_tx_idx = 0; |
813 | int dropped = 0; | |
8a9562d2 | 814 | |
8a9562d2 PS |
815 | qid = rte_lcore_id() % NR_QUEUE; |
816 | ||
f4fd623c DDP |
817 | for (i = 0; i < cnt; i++) { |
818 | int size = ofpbuf_size(&pkts[i]->ofpbuf); | |
819 | if (OVS_UNLIKELY(size > dev->max_packet_len)) { | |
820 | if (next_tx_idx != i) { | |
821 | dpdk_queue_pkts(dev, qid, | |
822 | (struct rte_mbuf **)&pkts[next_tx_idx], | |
823 | i-next_tx_idx); | |
1ebfe1ac | 824 | } |
f4fd623c | 825 | |
1ebfe1ac DDP |
826 | VLOG_WARN_RL(&rl, "Too big size %d max_packet_len %d", |
827 | (int)size , dev->max_packet_len); | |
f4fd623c | 828 | |
1ebfe1ac DDP |
829 | dpif_packet_delete(pkts[i]); |
830 | dropped++; | |
f4fd623c DDP |
831 | next_tx_idx = i + 1; |
832 | } | |
833 | } | |
834 | if (next_tx_idx != cnt) { | |
835 | dpdk_queue_pkts(dev, qid, | |
836 | (struct rte_mbuf **)&pkts[next_tx_idx], | |
837 | cnt-next_tx_idx); | |
838 | } | |
8a9562d2 | 839 | |
f4fd623c DDP |
840 | if (OVS_UNLIKELY(dropped)) { |
841 | ovs_mutex_lock(&dev->mutex); | |
842 | dev->stats.tx_dropped += dropped; | |
843 | ovs_mutex_unlock(&dev->mutex); | |
844 | } | |
8a9562d2 PS |
845 | } |
846 | ret = 0; | |
847 | ||
8a9562d2 PS |
848 | return ret; |
849 | } | |
850 | ||
851 | static int | |
852 | netdev_dpdk_set_etheraddr(struct netdev *netdev, | |
853 | const uint8_t mac[ETH_ADDR_LEN]) | |
854 | { | |
855 | struct netdev_dpdk *dev = netdev_dpdk_cast(netdev); | |
856 | ||
857 | ovs_mutex_lock(&dev->mutex); | |
858 | if (!eth_addr_equals(dev->hwaddr, mac)) { | |
859 | memcpy(dev->hwaddr, mac, ETH_ADDR_LEN); | |
045c0d1a | 860 | netdev_change_seq_changed(netdev); |
8a9562d2 PS |
861 | } |
862 | ovs_mutex_unlock(&dev->mutex); | |
863 | ||
864 | return 0; | |
865 | } | |
866 | ||
867 | static int | |
868 | netdev_dpdk_get_etheraddr(const struct netdev *netdev, | |
869 | uint8_t mac[ETH_ADDR_LEN]) | |
870 | { | |
871 | struct netdev_dpdk *dev = netdev_dpdk_cast(netdev); | |
872 | ||
873 | ovs_mutex_lock(&dev->mutex); | |
874 | memcpy(mac, dev->hwaddr, ETH_ADDR_LEN); | |
875 | ovs_mutex_unlock(&dev->mutex); | |
876 | ||
877 | return 0; | |
878 | } | |
879 | ||
880 | static int | |
881 | netdev_dpdk_get_mtu(const struct netdev *netdev, int *mtup) | |
882 | { | |
883 | struct netdev_dpdk *dev = netdev_dpdk_cast(netdev); | |
884 | ||
885 | ovs_mutex_lock(&dev->mutex); | |
886 | *mtup = dev->mtu; | |
887 | ovs_mutex_unlock(&dev->mutex); | |
888 | ||
889 | return 0; | |
890 | } | |
891 | ||
892 | static int | |
893 | netdev_dpdk_set_mtu(const struct netdev *netdev, int mtu) | |
894 | { | |
895 | struct netdev_dpdk *dev = netdev_dpdk_cast(netdev); | |
896 | int old_mtu, err; | |
897 | struct dpdk_mp *old_mp; | |
898 | struct dpdk_mp *mp; | |
899 | ||
900 | ovs_mutex_lock(&dpdk_mutex); | |
901 | ovs_mutex_lock(&dev->mutex); | |
902 | if (dev->mtu == mtu) { | |
903 | err = 0; | |
904 | goto out; | |
905 | } | |
906 | ||
907 | mp = dpdk_mp_get(dev->socket_id, dev->mtu); | |
908 | if (!mp) { | |
909 | err = ENOMEM; | |
910 | goto out; | |
911 | } | |
912 | ||
913 | rte_eth_dev_stop(dev->port_id); | |
914 | ||
915 | old_mtu = dev->mtu; | |
916 | old_mp = dev->dpdk_mp; | |
917 | dev->dpdk_mp = mp; | |
918 | dev->mtu = mtu; | |
919 | dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu); | |
920 | ||
921 | err = dpdk_eth_dev_init(dev); | |
922 | if (err) { | |
8a9562d2 PS |
923 | dpdk_mp_put(mp); |
924 | dev->mtu = old_mtu; | |
925 | dev->dpdk_mp = old_mp; | |
926 | dev->max_packet_len = MTU_TO_MAX_LEN(dev->mtu); | |
927 | dpdk_eth_dev_init(dev); | |
928 | goto out; | |
929 | } | |
930 | ||
931 | dpdk_mp_put(old_mp); | |
045c0d1a | 932 | netdev_change_seq_changed(netdev); |
8a9562d2 PS |
933 | out: |
934 | ovs_mutex_unlock(&dev->mutex); | |
935 | ovs_mutex_unlock(&dpdk_mutex); | |
936 | return err; | |
937 | } | |
938 | ||
939 | static int | |
940 | netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier); | |
941 | ||
942 | static int | |
943 | netdev_dpdk_get_stats(const struct netdev *netdev, struct netdev_stats *stats) | |
944 | { | |
945 | struct netdev_dpdk *dev = netdev_dpdk_cast(netdev); | |
946 | struct rte_eth_stats rte_stats; | |
947 | bool gg; | |
948 | ||
949 | netdev_dpdk_get_carrier(netdev, &gg); | |
950 | ovs_mutex_lock(&dev->mutex); | |
951 | rte_eth_stats_get(dev->port_id, &rte_stats); | |
952 | ||
953 | *stats = dev->stats_offset; | |
954 | ||
955 | stats->rx_packets += rte_stats.ipackets; | |
956 | stats->tx_packets += rte_stats.opackets; | |
957 | stats->rx_bytes += rte_stats.ibytes; | |
958 | stats->tx_bytes += rte_stats.obytes; | |
959 | stats->rx_errors += rte_stats.ierrors; | |
960 | stats->tx_errors += rte_stats.oerrors; | |
961 | stats->multicast += rte_stats.imcasts; | |
962 | ||
963 | stats->tx_dropped += dev->stats.tx_dropped; | |
964 | ovs_mutex_unlock(&dev->mutex); | |
965 | ||
966 | return 0; | |
967 | } | |
968 | ||
969 | static int | |
970 | netdev_dpdk_set_stats(struct netdev *netdev, const struct netdev_stats *stats) | |
971 | { | |
972 | struct netdev_dpdk *dev = netdev_dpdk_cast(netdev); | |
973 | ||
974 | ovs_mutex_lock(&dev->mutex); | |
975 | dev->stats_offset = *stats; | |
976 | ovs_mutex_unlock(&dev->mutex); | |
977 | ||
978 | return 0; | |
979 | } | |
980 | ||
981 | static int | |
982 | netdev_dpdk_get_features(const struct netdev *netdev_, | |
983 | enum netdev_features *current, | |
984 | enum netdev_features *advertised OVS_UNUSED, | |
985 | enum netdev_features *supported OVS_UNUSED, | |
986 | enum netdev_features *peer OVS_UNUSED) | |
987 | { | |
988 | struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_); | |
989 | struct rte_eth_link link; | |
990 | ||
991 | ovs_mutex_lock(&dev->mutex); | |
992 | link = dev->link; | |
993 | ovs_mutex_unlock(&dev->mutex); | |
994 | ||
995 | if (link.link_duplex == ETH_LINK_AUTONEG_DUPLEX) { | |
996 | if (link.link_speed == ETH_LINK_SPEED_AUTONEG) { | |
997 | *current = NETDEV_F_AUTONEG; | |
998 | } | |
999 | } else if (link.link_duplex == ETH_LINK_HALF_DUPLEX) { | |
1000 | if (link.link_speed == ETH_LINK_SPEED_10) { | |
1001 | *current = NETDEV_F_10MB_HD; | |
1002 | } | |
1003 | if (link.link_speed == ETH_LINK_SPEED_100) { | |
1004 | *current = NETDEV_F_100MB_HD; | |
1005 | } | |
1006 | if (link.link_speed == ETH_LINK_SPEED_1000) { | |
1007 | *current = NETDEV_F_1GB_HD; | |
1008 | } | |
1009 | } else if (link.link_duplex == ETH_LINK_FULL_DUPLEX) { | |
1010 | if (link.link_speed == ETH_LINK_SPEED_10) { | |
1011 | *current = NETDEV_F_10MB_FD; | |
1012 | } | |
1013 | if (link.link_speed == ETH_LINK_SPEED_100) { | |
1014 | *current = NETDEV_F_100MB_FD; | |
1015 | } | |
1016 | if (link.link_speed == ETH_LINK_SPEED_1000) { | |
1017 | *current = NETDEV_F_1GB_FD; | |
1018 | } | |
1019 | if (link.link_speed == ETH_LINK_SPEED_10000) { | |
1020 | *current = NETDEV_F_10GB_FD; | |
1021 | } | |
1022 | } | |
1023 | ||
1024 | return 0; | |
1025 | } | |
1026 | ||
1027 | static int | |
1028 | netdev_dpdk_get_ifindex(const struct netdev *netdev) | |
1029 | { | |
1030 | struct netdev_dpdk *dev = netdev_dpdk_cast(netdev); | |
1031 | int ifindex; | |
1032 | ||
1033 | ovs_mutex_lock(&dev->mutex); | |
1034 | ifindex = dev->port_id; | |
1035 | ovs_mutex_unlock(&dev->mutex); | |
1036 | ||
1037 | return ifindex; | |
1038 | } | |
1039 | ||
1040 | static int | |
1041 | netdev_dpdk_get_carrier(const struct netdev *netdev_, bool *carrier) | |
1042 | { | |
1043 | struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_); | |
1044 | ||
1045 | ovs_mutex_lock(&dev->mutex); | |
1046 | check_link_status(dev); | |
1047 | *carrier = dev->link.link_status; | |
1048 | ovs_mutex_unlock(&dev->mutex); | |
1049 | ||
1050 | return 0; | |
1051 | } | |
1052 | ||
1053 | static long long int | |
1054 | netdev_dpdk_get_carrier_resets(const struct netdev *netdev_) | |
1055 | { | |
1056 | struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_); | |
1057 | long long int carrier_resets; | |
1058 | ||
1059 | ovs_mutex_lock(&dev->mutex); | |
1060 | carrier_resets = dev->link_reset_cnt; | |
1061 | ovs_mutex_unlock(&dev->mutex); | |
1062 | ||
1063 | return carrier_resets; | |
1064 | } | |
1065 | ||
1066 | static int | |
1067 | netdev_dpdk_set_miimon(struct netdev *netdev_ OVS_UNUSED, | |
1068 | long long int interval OVS_UNUSED) | |
1069 | { | |
1070 | return 0; | |
1071 | } | |
1072 | ||
1073 | static int | |
1074 | netdev_dpdk_update_flags__(struct netdev_dpdk *dev, | |
1075 | enum netdev_flags off, enum netdev_flags on, | |
95fb793a | 1076 | enum netdev_flags *old_flagsp) OVS_REQUIRES(dev->mutex) |
8a9562d2 PS |
1077 | { |
1078 | int err; | |
1079 | ||
1080 | if ((off | on) & ~(NETDEV_UP | NETDEV_PROMISC)) { | |
1081 | return EINVAL; | |
1082 | } | |
1083 | ||
1084 | *old_flagsp = dev->flags; | |
1085 | dev->flags |= on; | |
1086 | dev->flags &= ~off; | |
1087 | ||
1088 | if (dev->flags == *old_flagsp) { | |
1089 | return 0; | |
1090 | } | |
1091 | ||
1092 | if (dev->flags & NETDEV_UP) { | |
1093 | err = rte_eth_dev_start(dev->port_id); | |
1094 | if (err) | |
95fb793a | 1095 | return -err; |
8a9562d2 PS |
1096 | } |
1097 | ||
1098 | if (dev->flags & NETDEV_PROMISC) { | |
1099 | rte_eth_promiscuous_enable(dev->port_id); | |
1100 | } | |
1101 | ||
1102 | if (!(dev->flags & NETDEV_UP)) { | |
1103 | rte_eth_dev_stop(dev->port_id); | |
1104 | } | |
1105 | ||
1106 | return 0; | |
1107 | } | |
1108 | ||
1109 | static int | |
1110 | netdev_dpdk_update_flags(struct netdev *netdev_, | |
1111 | enum netdev_flags off, enum netdev_flags on, | |
1112 | enum netdev_flags *old_flagsp) | |
1113 | { | |
1114 | struct netdev_dpdk *netdev = netdev_dpdk_cast(netdev_); | |
1115 | int error; | |
1116 | ||
1117 | ovs_mutex_lock(&netdev->mutex); | |
1118 | error = netdev_dpdk_update_flags__(netdev, off, on, old_flagsp); | |
1119 | ovs_mutex_unlock(&netdev->mutex); | |
1120 | ||
1121 | return error; | |
1122 | } | |
1123 | ||
1124 | static int | |
1125 | netdev_dpdk_get_status(const struct netdev *netdev_, struct smap *args) | |
1126 | { | |
1127 | struct netdev_dpdk *dev = netdev_dpdk_cast(netdev_); | |
1128 | struct rte_eth_dev_info dev_info; | |
1129 | ||
e0a801c7 | 1130 | if (dev->port_id < 0) |
8a9562d2 PS |
1131 | return ENODEV; |
1132 | ||
1133 | ovs_mutex_lock(&dev->mutex); | |
1134 | rte_eth_dev_info_get(dev->port_id, &dev_info); | |
1135 | ovs_mutex_unlock(&dev->mutex); | |
1136 | ||
1137 | smap_add_format(args, "driver_name", "%s", dev_info.driver_name); | |
1138 | ||
95fb793a | 1139 | smap_add_format(args, "port_no", "%d", dev->port_id); |
8a9562d2 PS |
1140 | smap_add_format(args, "numa_id", "%d", rte_eth_dev_socket_id(dev->port_id)); |
1141 | smap_add_format(args, "driver_name", "%s", dev_info.driver_name); | |
1142 | smap_add_format(args, "min_rx_bufsize", "%u", dev_info.min_rx_bufsize); | |
1143 | smap_add_format(args, "max_rx_pktlen", "%u", dev_info.max_rx_pktlen); | |
1144 | smap_add_format(args, "max_rx_queues", "%u", dev_info.max_rx_queues); | |
1145 | smap_add_format(args, "max_tx_queues", "%u", dev_info.max_tx_queues); | |
1146 | smap_add_format(args, "max_mac_addrs", "%u", dev_info.max_mac_addrs); | |
1147 | smap_add_format(args, "max_hash_mac_addrs", "%u", dev_info.max_hash_mac_addrs); | |
1148 | smap_add_format(args, "max_vfs", "%u", dev_info.max_vfs); | |
1149 | smap_add_format(args, "max_vmdq_pools", "%u", dev_info.max_vmdq_pools); | |
1150 | ||
1151 | smap_add_format(args, "pci-vendor_id", "0x%u", dev_info.pci_dev->id.vendor_id); | |
1152 | smap_add_format(args, "pci-device_id", "0x%x", dev_info.pci_dev->id.device_id); | |
1153 | ||
1154 | return 0; | |
1155 | } | |
1156 | ||
1157 | static void | |
1158 | netdev_dpdk_set_admin_state__(struct netdev_dpdk *dev, bool admin_state) | |
1159 | OVS_REQUIRES(dev->mutex) | |
1160 | { | |
1161 | enum netdev_flags old_flags; | |
1162 | ||
1163 | if (admin_state) { | |
1164 | netdev_dpdk_update_flags__(dev, 0, NETDEV_UP, &old_flags); | |
1165 | } else { | |
1166 | netdev_dpdk_update_flags__(dev, NETDEV_UP, 0, &old_flags); | |
1167 | } | |
1168 | } | |
1169 | ||
1170 | static void | |
1171 | netdev_dpdk_set_admin_state(struct unixctl_conn *conn, int argc, | |
1172 | const char *argv[], void *aux OVS_UNUSED) | |
1173 | { | |
1174 | bool up; | |
1175 | ||
1176 | if (!strcasecmp(argv[argc - 1], "up")) { | |
1177 | up = true; | |
1178 | } else if ( !strcasecmp(argv[argc - 1], "down")) { | |
1179 | up = false; | |
1180 | } else { | |
1181 | unixctl_command_reply_error(conn, "Invalid Admin State"); | |
1182 | return; | |
1183 | } | |
1184 | ||
1185 | if (argc > 2) { | |
1186 | struct netdev *netdev = netdev_from_name(argv[1]); | |
1187 | if (netdev && is_dpdk_class(netdev->netdev_class)) { | |
1188 | struct netdev_dpdk *dpdk_dev = netdev_dpdk_cast(netdev); | |
1189 | ||
1190 | ovs_mutex_lock(&dpdk_dev->mutex); | |
1191 | netdev_dpdk_set_admin_state__(dpdk_dev, up); | |
1192 | ovs_mutex_unlock(&dpdk_dev->mutex); | |
1193 | ||
1194 | netdev_close(netdev); | |
1195 | } else { | |
1196 | unixctl_command_reply_error(conn, "Not a DPDK Interface"); | |
1197 | netdev_close(netdev); | |
1198 | return; | |
1199 | } | |
1200 | } else { | |
1201 | struct netdev_dpdk *netdev; | |
1202 | ||
1203 | ovs_mutex_lock(&dpdk_mutex); | |
1204 | LIST_FOR_EACH (netdev, list_node, &dpdk_list) { | |
1205 | ovs_mutex_lock(&netdev->mutex); | |
1206 | netdev_dpdk_set_admin_state__(netdev, up); | |
1207 | ovs_mutex_unlock(&netdev->mutex); | |
1208 | } | |
1209 | ovs_mutex_unlock(&dpdk_mutex); | |
1210 | } | |
1211 | unixctl_command_reply(conn, "OK"); | |
1212 | } | |
1213 | ||
033e9df2 DDP |
1214 | static void |
1215 | dpdk_common_init(void) | |
1216 | { | |
1217 | unixctl_command_register("netdev-dpdk/set-admin-state", | |
1218 | "[netdev] up|down", 1, 2, | |
1219 | netdev_dpdk_set_admin_state, NULL); | |
1220 | ||
1221 | ovs_thread_create("dpdk_watchdog", dpdk_watchdog, NULL); | |
1222 | } | |
1223 | ||
8a9562d2 PS |
1224 | static int |
1225 | dpdk_class_init(void) | |
1226 | { | |
1227 | int result; | |
1228 | ||
8a9562d2 PS |
1229 | result = rte_eal_pci_probe(); |
1230 | if (result) { | |
1231 | VLOG_ERR("Cannot probe PCI"); | |
95fb793a | 1232 | return -result; |
8a9562d2 PS |
1233 | } |
1234 | ||
8a9562d2 PS |
1235 | VLOG_INFO("Ethernet Device Count: %d", (int)rte_eth_dev_count()); |
1236 | ||
8a9562d2 PS |
1237 | return 0; |
1238 | } | |
1239 | ||
95fb793a | 1240 | /* Client Rings */ |
1241 | ||
95fb793a | 1242 | static int |
1243 | dpdk_ring_create(const char dev_name[], unsigned int port_no, | |
1244 | unsigned int *eth_port_id) | |
1245 | { | |
1246 | struct dpdk_ring *ivshmem; | |
1247 | char ring_name[10]; | |
1248 | int err; | |
1249 | ||
1250 | ivshmem = dpdk_rte_mzalloc(sizeof *ivshmem); | |
1251 | if (ivshmem == NULL) { | |
1252 | return ENOMEM; | |
1253 | } | |
1254 | ||
1255 | err = snprintf(ring_name, 10, "%s_tx", dev_name); | |
1256 | if (err < 0) { | |
1257 | return -err; | |
1258 | } | |
1259 | ||
58f7c37b | 1260 | ivshmem->cring_tx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0, 0); |
95fb793a | 1261 | if (ivshmem->cring_tx == NULL) { |
1262 | rte_free(ivshmem); | |
1263 | return ENOMEM; | |
1264 | } | |
1265 | ||
1266 | err = snprintf(ring_name, 10, "%s_rx", dev_name); | |
1267 | if (err < 0) { | |
1268 | return -err; | |
1269 | } | |
1270 | ||
58f7c37b | 1271 | ivshmem->cring_rx = rte_ring_create(ring_name, DPDK_RING_SIZE, SOCKET0, 0); |
95fb793a | 1272 | if (ivshmem->cring_rx == NULL) { |
1273 | rte_free(ivshmem); | |
1274 | return ENOMEM; | |
1275 | } | |
1276 | ||
d7310583 DDP |
1277 | err = rte_eth_from_rings(dev_name, &ivshmem->cring_rx, 1, |
1278 | &ivshmem->cring_tx, 1, SOCKET0); | |
1279 | ||
95fb793a | 1280 | if (err < 0) { |
1281 | rte_free(ivshmem); | |
1282 | return ENODEV; | |
1283 | } | |
1284 | ||
1285 | ivshmem->user_port_id = port_no; | |
1286 | ivshmem->eth_port_id = rte_eth_dev_count() - 1; | |
1287 | list_push_back(&dpdk_ring_list, &ivshmem->list_node); | |
1288 | ||
1289 | *eth_port_id = ivshmem->eth_port_id; | |
1290 | return 0; | |
1291 | } | |
1292 | ||
1293 | static int | |
1294 | dpdk_ring_open(const char dev_name[], unsigned int *eth_port_id) OVS_REQUIRES(dpdk_mutex) | |
1295 | { | |
1296 | struct dpdk_ring *ivshmem; | |
1297 | unsigned int port_no; | |
1298 | int err = 0; | |
1299 | ||
1300 | /* Names always start with "dpdkr" */ | |
1301 | err = dpdk_dev_parse_name(dev_name, "dpdkr", &port_no); | |
1302 | if (err) { | |
1303 | return err; | |
1304 | } | |
1305 | ||
1306 | /* look through our list to find the device */ | |
1307 | LIST_FOR_EACH (ivshmem, list_node, &dpdk_ring_list) { | |
1308 | if (ivshmem->user_port_id == port_no) { | |
1309 | VLOG_INFO("Found dpdk ring device %s:\n", dev_name); | |
1310 | *eth_port_id = ivshmem->eth_port_id; /* really all that is needed */ | |
1311 | return 0; | |
1312 | } | |
1313 | } | |
1314 | /* Need to create the device rings */ | |
1315 | return dpdk_ring_create(dev_name, port_no, eth_port_id); | |
1316 | } | |
1317 | ||
1318 | static int | |
1319 | netdev_dpdk_ring_construct(struct netdev *netdev) | |
1320 | { | |
1321 | unsigned int port_no = 0; | |
1322 | int err = 0; | |
1323 | ||
1324 | if (rte_eal_init_ret) { | |
1325 | return rte_eal_init_ret; | |
1326 | } | |
1327 | ||
1328 | ovs_mutex_lock(&dpdk_mutex); | |
1329 | ||
1330 | err = dpdk_ring_open(netdev->name, &port_no); | |
1331 | if (err) { | |
1332 | goto unlock_dpdk; | |
1333 | } | |
1334 | ||
1335 | err = netdev_dpdk_init(netdev, port_no); | |
1336 | ||
1337 | unlock_dpdk: | |
1338 | ovs_mutex_unlock(&dpdk_mutex); | |
1339 | return err; | |
1340 | } | |
1341 | ||
1342 | #define NETDEV_DPDK_CLASS(NAME, INIT, CONSTRUCT) \ | |
1343 | { \ | |
1344 | NAME, \ | |
1345 | INIT, /* init */ \ | |
1346 | NULL, /* netdev_dpdk_run */ \ | |
1347 | NULL, /* netdev_dpdk_wait */ \ | |
1348 | \ | |
1349 | netdev_dpdk_alloc, \ | |
1350 | CONSTRUCT, \ | |
1351 | netdev_dpdk_destruct, \ | |
1352 | netdev_dpdk_dealloc, \ | |
1353 | netdev_dpdk_get_config, \ | |
1354 | NULL, /* netdev_dpdk_set_config */ \ | |
1355 | NULL, /* get_tunnel_config */ \ | |
7dec44fe | 1356 | netdev_dpdk_get_numa_id, /* get_numa_id */ \ |
95fb793a | 1357 | \ |
1358 | netdev_dpdk_send, /* send */ \ | |
1359 | NULL, /* send_wait */ \ | |
1360 | \ | |
1361 | netdev_dpdk_set_etheraddr, \ | |
1362 | netdev_dpdk_get_etheraddr, \ | |
1363 | netdev_dpdk_get_mtu, \ | |
1364 | netdev_dpdk_set_mtu, \ | |
1365 | netdev_dpdk_get_ifindex, \ | |
1366 | netdev_dpdk_get_carrier, \ | |
1367 | netdev_dpdk_get_carrier_resets, \ | |
1368 | netdev_dpdk_set_miimon, \ | |
1369 | netdev_dpdk_get_stats, \ | |
1370 | netdev_dpdk_set_stats, \ | |
1371 | netdev_dpdk_get_features, \ | |
1372 | NULL, /* set_advertisements */ \ | |
1373 | \ | |
1374 | NULL, /* set_policing */ \ | |
1375 | NULL, /* get_qos_types */ \ | |
1376 | NULL, /* get_qos_capabilities */ \ | |
1377 | NULL, /* get_qos */ \ | |
1378 | NULL, /* set_qos */ \ | |
1379 | NULL, /* get_queue */ \ | |
1380 | NULL, /* set_queue */ \ | |
1381 | NULL, /* delete_queue */ \ | |
1382 | NULL, /* get_queue_stats */ \ | |
1383 | NULL, /* queue_dump_start */ \ | |
1384 | NULL, /* queue_dump_next */ \ | |
1385 | NULL, /* queue_dump_done */ \ | |
1386 | NULL, /* dump_queue_stats */ \ | |
1387 | \ | |
1388 | NULL, /* get_in4 */ \ | |
1389 | NULL, /* set_in4 */ \ | |
1390 | NULL, /* get_in6 */ \ | |
1391 | NULL, /* add_router */ \ | |
1392 | NULL, /* get_next_hop */ \ | |
1393 | netdev_dpdk_get_status, \ | |
1394 | NULL, /* arp_lookup */ \ | |
1395 | \ | |
1396 | netdev_dpdk_update_flags, \ | |
1397 | \ | |
1398 | netdev_dpdk_rxq_alloc, \ | |
1399 | netdev_dpdk_rxq_construct, \ | |
1400 | netdev_dpdk_rxq_destruct, \ | |
1401 | netdev_dpdk_rxq_dealloc, \ | |
1402 | netdev_dpdk_rxq_recv, \ | |
1403 | NULL, /* rx_wait */ \ | |
1404 | NULL, /* rxq_drain */ \ | |
1405 | } | |
8a9562d2 PS |
1406 | |
1407 | int | |
1408 | dpdk_init(int argc, char **argv) | |
1409 | { | |
1410 | int result; | |
1411 | ||
9441caf3 | 1412 | if (argc < 2 || strcmp(argv[1], "--dpdk")) |
8a9562d2 PS |
1413 | return 0; |
1414 | ||
9441caf3 DDP |
1415 | /* Make sure program name passed to rte_eal_init() is vswitchd. */ |
1416 | argv[1] = argv[0]; | |
1417 | ||
8a9562d2 PS |
1418 | argc--; |
1419 | argv++; | |
1420 | ||
1421 | /* Make sure things are initialized ... */ | |
1422 | result = rte_eal_init(argc, argv); | |
451450fa | 1423 | if (result < 0) { |
8a9562d2 | 1424 | ovs_abort(result, "Cannot init EAL\n"); |
451450fa | 1425 | } |
8a9562d2 | 1426 | |
d7310583 | 1427 | rte_memzone_dump(stdout); |
8a9562d2 PS |
1428 | rte_eal_init_ret = 0; |
1429 | ||
451450fa | 1430 | if (argc > result) { |
9441caf3 | 1431 | argv[result] = argv[0]; |
451450fa | 1432 | } |
9441caf3 | 1433 | |
db73f716 DDP |
1434 | /* We are called from the main thread here */ |
1435 | thread_set_nonpmd(); | |
1436 | ||
9441caf3 | 1437 | return result + 1; |
8a9562d2 PS |
1438 | } |
1439 | ||
95fb793a | 1440 | const struct netdev_class dpdk_class = |
1441 | NETDEV_DPDK_CLASS( | |
1442 | "dpdk", | |
1443 | dpdk_class_init, | |
1444 | netdev_dpdk_construct); | |
1445 | ||
1446 | const struct netdev_class dpdk_ring_class = | |
1447 | NETDEV_DPDK_CLASS( | |
1448 | "dpdkr", | |
033e9df2 | 1449 | NULL, |
95fb793a | 1450 | netdev_dpdk_ring_construct); |
1451 | ||
8a9562d2 PS |
1452 | void |
1453 | netdev_dpdk_register(void) | |
1454 | { | |
95fb793a | 1455 | static struct ovsthread_once once = OVSTHREAD_ONCE_INITIALIZER; |
1456 | ||
033e9df2 DDP |
1457 | if (rte_eal_init_ret) { |
1458 | return; | |
1459 | } | |
1460 | ||
95fb793a | 1461 | if (ovsthread_once_start(&once)) { |
033e9df2 | 1462 | dpdk_common_init(); |
95fb793a | 1463 | netdev_register_provider(&dpdk_class); |
1464 | netdev_register_provider(&dpdk_ring_class); | |
1465 | ovsthread_once_done(&once); | |
1466 | } | |
8a9562d2 | 1467 | } |
8617afff PS |
1468 | |
1469 | int | |
1470 | pmd_thread_setaffinity_cpu(int cpu) | |
1471 | { | |
1472 | cpu_set_t cpuset; | |
1473 | int err; | |
1474 | ||
1475 | CPU_ZERO(&cpuset); | |
1476 | CPU_SET(cpu, &cpuset); | |
1477 | err = pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset); | |
1478 | if (err) { | |
1479 | VLOG_ERR("Thread affinity error %d",err); | |
1480 | return err; | |
1481 | } | |
db73f716 DDP |
1482 | /* lcore_id 0 is reseved for use by non pmd threads. */ |
1483 | RTE_PER_LCORE(_lcore_id) = cpu + 1; | |
8617afff PS |
1484 | |
1485 | return 0; | |
1486 | } | |
db73f716 DDP |
1487 | |
1488 | void | |
1489 | thread_set_nonpmd(void) | |
1490 | { | |
1491 | /* We cannot have RTE_MAX_LCORE pmd threads, because lcore_id 0 is reserved | |
1492 | * for non pmd threads */ | |
1493 | BUILD_ASSERT(NR_PMD_THREADS < RTE_MAX_LCORE); | |
1494 | /* We have to use 0 to allow non pmd threads to perform certain DPDK | |
1495 | * operations, like rte_eth_dev_configure(). */ | |
1496 | RTE_PER_LCORE(_lcore_id) = 0; | |
1497 | } | |
1498 | ||
1499 | static bool | |
1500 | thread_is_pmd(void) | |
1501 | { | |
1502 | return rte_lcore_id() != 0; | |
1503 | } |