1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2017 6WIND S.A.
3 * Copyright 2017 Mellanox Technologies, Ltd
4 */
5
6 /**
7 * @file
8 * Flow API operations for mlx4 driver.
9 */
10
11 #include <arpa/inet.h>
12 #include <errno.h>
13 #include <stdalign.h>
14 #include <stddef.h>
15 #include <stdint.h>
16 #include <string.h>
17 #include <sys/queue.h>
18
19 /* Verbs headers do not support -pedantic. */
20 #ifdef PEDANTIC
21 #pragma GCC diagnostic ignored "-Wpedantic"
22 #endif
23 #include <infiniband/verbs.h>
24 #ifdef PEDANTIC
25 #pragma GCC diagnostic error "-Wpedantic"
26 #endif
27
28 #include <rte_byteorder.h>
29 #include <rte_errno.h>
30 #include <rte_ethdev_driver.h>
31 #include <rte_ether.h>
32 #include <rte_flow.h>
33 #include <rte_flow_driver.h>
34 #include <rte_malloc.h>
35
36 /* PMD headers. */
37 #include "mlx4.h"
38 #include "mlx4_glue.h"
39 #include "mlx4_flow.h"
40 #include "mlx4_rxtx.h"
41 #include "mlx4_utils.h"
42
43 /** Static initializer for a list of subsequent item types. */
44 #define NEXT_ITEM(...) \
45 (const enum rte_flow_item_type []){ \
46 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
47 }
48
49 /** Processor structure associated with a flow item. */
50 struct mlx4_flow_proc_item {
51 /** Bit-mask for fields supported by this PMD. */
52 const void *mask_support;
53 /** Bit-mask to use when @p item->mask is not provided. */
54 const void *mask_default;
55 /** Size in bytes for @p mask_support and @p mask_default. */
56 const unsigned int mask_sz;
57 /** Merge a pattern item into a flow rule handle. */
58 int (*merge)(struct rte_flow *flow,
59 const struct rte_flow_item *item,
60 const struct mlx4_flow_proc_item *proc,
61 struct rte_flow_error *error);
62 /** Size in bytes of the destination structure. */
63 const unsigned int dst_sz;
64 /** List of possible subsequent items. */
65 const enum rte_flow_item_type *const next_item;
66 };
67
68 /** Shared resources for drop flow rules. */
69 struct mlx4_drop {
70 struct ibv_qp *qp; /**< QP target. */
71 struct ibv_cq *cq; /**< CQ associated with above QP. */
72 struct mlx4_priv *priv; /**< Back pointer to private data. */
73 uint32_t refcnt; /**< Reference count. */
74 };
75
76 /**
77 * Convert supported RSS hash field types between DPDK and Verbs formats.
78 *
79 * This function returns the supported (default) set when @p types has
80 * special value 0.
81 *
82 * @param priv
83 * Pointer to private structure.
84 * @param types
85 * Depending on @p verbs_to_dpdk, hash types in either DPDK (see struct
86 * rte_eth_rss_conf) or Verbs format.
87 * @param verbs_to_dpdk
88 * A zero value converts @p types from DPDK to Verbs, a nonzero value
89 * performs the reverse operation.
90 *
91 * @return
92 * Converted RSS hash fields on success, (uint64_t)-1 otherwise and
93 * rte_errno is set.
94 */
95 uint64_t
mlx4_conv_rss_types(struct mlx4_priv * priv,uint64_t types,int verbs_to_dpdk)96 mlx4_conv_rss_types(struct mlx4_priv *priv, uint64_t types, int verbs_to_dpdk)
97 {
98 enum {
99 INNER,
100 IPV4, IPV4_1, IPV4_2, IPV6, IPV6_1, IPV6_2, IPV6_3,
101 TCP, UDP,
102 IPV4_TCP, IPV4_UDP, IPV6_TCP, IPV6_TCP_1, IPV6_UDP, IPV6_UDP_1,
103 };
104 enum {
105 VERBS_IPV4 = IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4,
106 VERBS_IPV6 = IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6,
107 VERBS_TCP = IBV_RX_HASH_SRC_PORT_TCP | IBV_RX_HASH_DST_PORT_TCP,
108 VERBS_UDP = IBV_RX_HASH_SRC_PORT_UDP | IBV_RX_HASH_DST_PORT_UDP,
109 };
110 static const uint64_t dpdk[] = {
111 [INNER] = 0,
112 [IPV4] = ETH_RSS_IPV4,
113 [IPV4_1] = ETH_RSS_FRAG_IPV4,
114 [IPV4_2] = ETH_RSS_NONFRAG_IPV4_OTHER,
115 [IPV6] = ETH_RSS_IPV6,
116 [IPV6_1] = ETH_RSS_FRAG_IPV6,
117 [IPV6_2] = ETH_RSS_NONFRAG_IPV6_OTHER,
118 [IPV6_3] = ETH_RSS_IPV6_EX,
119 [TCP] = 0,
120 [UDP] = 0,
121 [IPV4_TCP] = ETH_RSS_NONFRAG_IPV4_TCP,
122 [IPV4_UDP] = ETH_RSS_NONFRAG_IPV4_UDP,
123 [IPV6_TCP] = ETH_RSS_NONFRAG_IPV6_TCP,
124 [IPV6_TCP_1] = ETH_RSS_IPV6_TCP_EX,
125 [IPV6_UDP] = ETH_RSS_NONFRAG_IPV6_UDP,
126 [IPV6_UDP_1] = ETH_RSS_IPV6_UDP_EX,
127 };
128 static const uint64_t verbs[RTE_DIM(dpdk)] = {
129 [INNER] = IBV_RX_HASH_INNER,
130 [IPV4] = VERBS_IPV4,
131 [IPV4_1] = VERBS_IPV4,
132 [IPV4_2] = VERBS_IPV4,
133 [IPV6] = VERBS_IPV6,
134 [IPV6_1] = VERBS_IPV6,
135 [IPV6_2] = VERBS_IPV6,
136 [IPV6_3] = VERBS_IPV6,
137 [TCP] = VERBS_TCP,
138 [UDP] = VERBS_UDP,
139 [IPV4_TCP] = VERBS_IPV4 | VERBS_TCP,
140 [IPV4_UDP] = VERBS_IPV4 | VERBS_UDP,
141 [IPV6_TCP] = VERBS_IPV6 | VERBS_TCP,
142 [IPV6_TCP_1] = VERBS_IPV6 | VERBS_TCP,
143 [IPV6_UDP] = VERBS_IPV6 | VERBS_UDP,
144 [IPV6_UDP_1] = VERBS_IPV6 | VERBS_UDP,
145 };
146 const uint64_t *in = verbs_to_dpdk ? verbs : dpdk;
147 const uint64_t *out = verbs_to_dpdk ? dpdk : verbs;
148 uint64_t seen = 0;
149 uint64_t conv = 0;
150 unsigned int i;
151
152 if (!types) {
153 if (!verbs_to_dpdk)
154 return priv->hw_rss_sup;
155 types = priv->hw_rss_sup;
156 }
157 for (i = 0; i != RTE_DIM(dpdk); ++i)
158 if (in[i] && (types & in[i]) == in[i]) {
159 seen |= types & in[i];
160 conv |= out[i];
161 }
162 if ((verbs_to_dpdk || (conv & priv->hw_rss_sup) == conv) &&
163 !(types & ~seen))
164 return conv;
165 rte_errno = ENOTSUP;
166 return (uint64_t)-1;
167 }
168
169 /**
170 * Merge Ethernet pattern item into flow rule handle.
171 *
172 * Additional mlx4-specific constraints on supported fields:
173 *
174 * - No support for partial masks, except in the specific case of matching
175 * all multicast traffic (@p spec->dst and @p mask->dst equal to
176 * 01:00:00:00:00:00).
177 * - Not providing @p item->spec or providing an empty @p mask->dst is
178 * *only* supported if the rule doesn't specify additional matching
179 * criteria (i.e. rule is promiscuous-like).
180 *
181 * @param[in, out] flow
182 * Flow rule handle to update.
183 * @param[in] item
184 * Pattern item to merge.
185 * @param[in] proc
186 * Associated item-processing object.
187 * @param[out] error
188 * Perform verbose error reporting if not NULL.
189 *
190 * @return
191 * 0 on success, a negative errno value otherwise and rte_errno is set.
192 */
193 static int
mlx4_flow_merge_eth(struct rte_flow * flow,const struct rte_flow_item * item,const struct mlx4_flow_proc_item * proc,struct rte_flow_error * error)194 mlx4_flow_merge_eth(struct rte_flow *flow,
195 const struct rte_flow_item *item,
196 const struct mlx4_flow_proc_item *proc,
197 struct rte_flow_error *error)
198 {
199 const struct rte_flow_item_eth *spec = item->spec;
200 const struct rte_flow_item_eth *mask =
201 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
202 struct ibv_flow_spec_eth *eth;
203 const char *msg;
204 unsigned int i;
205
206 if (mask) {
207 uint32_t sum_dst = 0;
208 uint32_t sum_src = 0;
209
210 for (i = 0; i != sizeof(mask->dst.addr_bytes); ++i) {
211 sum_dst += mask->dst.addr_bytes[i];
212 sum_src += mask->src.addr_bytes[i];
213 }
214 if (sum_src) {
215 msg = "mlx4 does not support source MAC matching";
216 goto error;
217 } else if (!sum_dst) {
218 flow->promisc = 1;
219 } else if (sum_dst == 1 && mask->dst.addr_bytes[0] == 1) {
220 if (!(spec->dst.addr_bytes[0] & 1)) {
221 msg = "mlx4 does not support the explicit"
222 " exclusion of all multicast traffic";
223 goto error;
224 }
225 flow->allmulti = 1;
226 } else if (sum_dst != (UINT8_C(0xff) * RTE_ETHER_ADDR_LEN)) {
227 msg = "mlx4 does not support matching partial"
228 " Ethernet fields";
229 goto error;
230 }
231 }
232 if (!flow->ibv_attr)
233 return 0;
234 if (flow->promisc) {
235 flow->ibv_attr->type = IBV_FLOW_ATTR_ALL_DEFAULT;
236 return 0;
237 }
238 if (flow->allmulti) {
239 flow->ibv_attr->type = IBV_FLOW_ATTR_MC_DEFAULT;
240 return 0;
241 }
242 ++flow->ibv_attr->num_of_specs;
243 eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
244 *eth = (struct ibv_flow_spec_eth) {
245 .type = IBV_FLOW_SPEC_ETH,
246 .size = sizeof(*eth),
247 };
248 if (!mask) {
249 eth->val.dst_mac[0] = 0xff;
250 flow->ibv_attr->type = IBV_FLOW_ATTR_ALL_DEFAULT;
251 flow->promisc = 1;
252 return 0;
253 }
254 memcpy(eth->val.dst_mac, spec->dst.addr_bytes, RTE_ETHER_ADDR_LEN);
255 memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, RTE_ETHER_ADDR_LEN);
256 /* Remove unwanted bits from values. */
257 for (i = 0; i < RTE_ETHER_ADDR_LEN; ++i)
258 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
259
260 return 0;
261 error:
262 return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
263 item, msg);
264 }
265
266 /**
267 * Merge VLAN pattern item into flow rule handle.
268 *
269 * Additional mlx4-specific constraints on supported fields:
270 *
271 * - Matching *all* VLAN traffic by omitting @p item->spec or providing an
272 * empty @p item->mask would also include non-VLAN traffic. Doing so is
273 * therefore unsupported.
274 * - No support for partial masks.
275 *
276 * @param[in, out] flow
277 * Flow rule handle to update.
278 * @param[in] item
279 * Pattern item to merge.
280 * @param[in] proc
281 * Associated item-processing object.
282 * @param[out] error
283 * Perform verbose error reporting if not NULL.
284 *
285 * @return
286 * 0 on success, a negative errno value otherwise and rte_errno is set.
287 */
288 static int
mlx4_flow_merge_vlan(struct rte_flow * flow,const struct rte_flow_item * item,const struct mlx4_flow_proc_item * proc,struct rte_flow_error * error)289 mlx4_flow_merge_vlan(struct rte_flow *flow,
290 const struct rte_flow_item *item,
291 const struct mlx4_flow_proc_item *proc,
292 struct rte_flow_error *error)
293 {
294 const struct rte_flow_item_vlan *spec = item->spec;
295 const struct rte_flow_item_vlan *mask =
296 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
297 struct ibv_flow_spec_eth *eth;
298 const char *msg;
299
300 if (!mask || !mask->tci) {
301 msg = "mlx4 cannot match all VLAN traffic while excluding"
302 " non-VLAN traffic, TCI VID must be specified";
303 goto error;
304 }
305 if (mask->tci != RTE_BE16(0x0fff)) {
306 msg = "mlx4 does not support partial TCI VID matching";
307 goto error;
308 }
309 if (!flow->ibv_attr)
310 return 0;
311 eth = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size -
312 sizeof(*eth));
313 eth->val.vlan_tag = spec->tci;
314 eth->mask.vlan_tag = mask->tci;
315 eth->val.vlan_tag &= eth->mask.vlan_tag;
316 if (flow->ibv_attr->type == IBV_FLOW_ATTR_ALL_DEFAULT)
317 flow->ibv_attr->type = IBV_FLOW_ATTR_NORMAL;
318 return 0;
319 error:
320 return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
321 item, msg);
322 }
323
324 /**
325 * Merge IPv4 pattern item into flow rule handle.
326 *
327 * Additional mlx4-specific constraints on supported fields:
328 *
329 * - No support for partial masks.
330 *
331 * @param[in, out] flow
332 * Flow rule handle to update.
333 * @param[in] item
334 * Pattern item to merge.
335 * @param[in] proc
336 * Associated item-processing object.
337 * @param[out] error
338 * Perform verbose error reporting if not NULL.
339 *
340 * @return
341 * 0 on success, a negative errno value otherwise and rte_errno is set.
342 */
343 static int
mlx4_flow_merge_ipv4(struct rte_flow * flow,const struct rte_flow_item * item,const struct mlx4_flow_proc_item * proc,struct rte_flow_error * error)344 mlx4_flow_merge_ipv4(struct rte_flow *flow,
345 const struct rte_flow_item *item,
346 const struct mlx4_flow_proc_item *proc,
347 struct rte_flow_error *error)
348 {
349 const struct rte_flow_item_ipv4 *spec = item->spec;
350 const struct rte_flow_item_ipv4 *mask =
351 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
352 struct ibv_flow_spec_ipv4 *ipv4;
353 const char *msg;
354
355 if (mask &&
356 ((uint32_t)(mask->hdr.src_addr + 1) > UINT32_C(1) ||
357 (uint32_t)(mask->hdr.dst_addr + 1) > UINT32_C(1))) {
358 msg = "mlx4 does not support matching partial IPv4 fields";
359 goto error;
360 }
361 if (!flow->ibv_attr)
362 return 0;
363 ++flow->ibv_attr->num_of_specs;
364 ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
365 *ipv4 = (struct ibv_flow_spec_ipv4) {
366 .type = IBV_FLOW_SPEC_IPV4,
367 .size = sizeof(*ipv4),
368 };
369 if (!spec)
370 return 0;
371 ipv4->val = (struct ibv_flow_ipv4_filter) {
372 .src_ip = spec->hdr.src_addr,
373 .dst_ip = spec->hdr.dst_addr,
374 };
375 ipv4->mask = (struct ibv_flow_ipv4_filter) {
376 .src_ip = mask->hdr.src_addr,
377 .dst_ip = mask->hdr.dst_addr,
378 };
379 /* Remove unwanted bits from values. */
380 ipv4->val.src_ip &= ipv4->mask.src_ip;
381 ipv4->val.dst_ip &= ipv4->mask.dst_ip;
382 return 0;
383 error:
384 return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
385 item, msg);
386 }
387
388 /**
389 * Merge UDP pattern item into flow rule handle.
390 *
391 * Additional mlx4-specific constraints on supported fields:
392 *
393 * - No support for partial masks.
394 * - Due to HW/FW limitation, flow rule priority is not taken into account
395 * when matching UDP destination ports, doing is therefore only supported
396 * at the highest priority level (0).
397 *
398 * @param[in, out] flow
399 * Flow rule handle to update.
400 * @param[in] item
401 * Pattern item to merge.
402 * @param[in] proc
403 * Associated item-processing object.
404 * @param[out] error
405 * Perform verbose error reporting if not NULL.
406 *
407 * @return
408 * 0 on success, a negative errno value otherwise and rte_errno is set.
409 */
410 static int
mlx4_flow_merge_udp(struct rte_flow * flow,const struct rte_flow_item * item,const struct mlx4_flow_proc_item * proc,struct rte_flow_error * error)411 mlx4_flow_merge_udp(struct rte_flow *flow,
412 const struct rte_flow_item *item,
413 const struct mlx4_flow_proc_item *proc,
414 struct rte_flow_error *error)
415 {
416 const struct rte_flow_item_udp *spec = item->spec;
417 const struct rte_flow_item_udp *mask =
418 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
419 struct ibv_flow_spec_tcp_udp *udp;
420 const char *msg;
421
422 if (mask &&
423 ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
424 (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
425 msg = "mlx4 does not support matching partial UDP fields";
426 goto error;
427 }
428 if (mask && mask->hdr.dst_port && flow->priority) {
429 msg = "combining UDP destination port matching with a nonzero"
430 " priority level is not supported";
431 goto error;
432 }
433 if (!flow->ibv_attr)
434 return 0;
435 ++flow->ibv_attr->num_of_specs;
436 udp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
437 *udp = (struct ibv_flow_spec_tcp_udp) {
438 .type = IBV_FLOW_SPEC_UDP,
439 .size = sizeof(*udp),
440 };
441 if (!spec)
442 return 0;
443 udp->val.dst_port = spec->hdr.dst_port;
444 udp->val.src_port = spec->hdr.src_port;
445 udp->mask.dst_port = mask->hdr.dst_port;
446 udp->mask.src_port = mask->hdr.src_port;
447 /* Remove unwanted bits from values. */
448 udp->val.src_port &= udp->mask.src_port;
449 udp->val.dst_port &= udp->mask.dst_port;
450 return 0;
451 error:
452 return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
453 item, msg);
454 }
455
456 /**
457 * Merge TCP pattern item into flow rule handle.
458 *
459 * Additional mlx4-specific constraints on supported fields:
460 *
461 * - No support for partial masks.
462 *
463 * @param[in, out] flow
464 * Flow rule handle to update.
465 * @param[in] item
466 * Pattern item to merge.
467 * @param[in] proc
468 * Associated item-processing object.
469 * @param[out] error
470 * Perform verbose error reporting if not NULL.
471 *
472 * @return
473 * 0 on success, a negative errno value otherwise and rte_errno is set.
474 */
475 static int
mlx4_flow_merge_tcp(struct rte_flow * flow,const struct rte_flow_item * item,const struct mlx4_flow_proc_item * proc,struct rte_flow_error * error)476 mlx4_flow_merge_tcp(struct rte_flow *flow,
477 const struct rte_flow_item *item,
478 const struct mlx4_flow_proc_item *proc,
479 struct rte_flow_error *error)
480 {
481 const struct rte_flow_item_tcp *spec = item->spec;
482 const struct rte_flow_item_tcp *mask =
483 spec ? (item->mask ? item->mask : proc->mask_default) : NULL;
484 struct ibv_flow_spec_tcp_udp *tcp;
485 const char *msg;
486
487 if (mask &&
488 ((uint16_t)(mask->hdr.src_port + 1) > UINT16_C(1) ||
489 (uint16_t)(mask->hdr.dst_port + 1) > UINT16_C(1))) {
490 msg = "mlx4 does not support matching partial TCP fields";
491 goto error;
492 }
493 if (!flow->ibv_attr)
494 return 0;
495 ++flow->ibv_attr->num_of_specs;
496 tcp = (void *)((uintptr_t)flow->ibv_attr + flow->ibv_attr_size);
497 *tcp = (struct ibv_flow_spec_tcp_udp) {
498 .type = IBV_FLOW_SPEC_TCP,
499 .size = sizeof(*tcp),
500 };
501 if (!spec)
502 return 0;
503 tcp->val.dst_port = spec->hdr.dst_port;
504 tcp->val.src_port = spec->hdr.src_port;
505 tcp->mask.dst_port = mask->hdr.dst_port;
506 tcp->mask.src_port = mask->hdr.src_port;
507 /* Remove unwanted bits from values. */
508 tcp->val.src_port &= tcp->mask.src_port;
509 tcp->val.dst_port &= tcp->mask.dst_port;
510 return 0;
511 error:
512 return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
513 item, msg);
514 }
515
516 /**
517 * Perform basic sanity checks on a pattern item.
518 *
519 * @param[in] item
520 * Item specification.
521 * @param[in] proc
522 * Associated item-processing object.
523 * @param[out] error
524 * Perform verbose error reporting if not NULL.
525 *
526 * @return
527 * 0 on success, a negative errno value otherwise and rte_errno is set.
528 */
529 static int
mlx4_flow_item_check(const struct rte_flow_item * item,const struct mlx4_flow_proc_item * proc,struct rte_flow_error * error)530 mlx4_flow_item_check(const struct rte_flow_item *item,
531 const struct mlx4_flow_proc_item *proc,
532 struct rte_flow_error *error)
533 {
534 const uint8_t *mask;
535 unsigned int i;
536
537 /* item->last and item->mask cannot exist without item->spec. */
538 if (!item->spec && (item->mask || item->last))
539 return rte_flow_error_set
540 (error, EINVAL, RTE_FLOW_ERROR_TYPE_ITEM, item,
541 "\"mask\" or \"last\" field provided without a"
542 " corresponding \"spec\"");
543 /* No spec, no mask, no problem. */
544 if (!item->spec)
545 return 0;
546 mask = item->mask ?
547 (const uint8_t *)item->mask :
548 (const uint8_t *)proc->mask_default;
549 MLX4_ASSERT(mask);
550 /*
551 * Single-pass check to make sure that:
552 * - Mask is supported, no bits are set outside proc->mask_support.
553 * - Both item->spec and item->last are included in mask.
554 */
555 for (i = 0; i != proc->mask_sz; ++i) {
556 if (!mask[i])
557 continue;
558 if ((mask[i] | ((const uint8_t *)proc->mask_support)[i]) !=
559 ((const uint8_t *)proc->mask_support)[i])
560 return rte_flow_error_set
561 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
562 item, "unsupported field found in \"mask\"");
563 if (item->last &&
564 (((const uint8_t *)item->spec)[i] & mask[i]) !=
565 (((const uint8_t *)item->last)[i] & mask[i]))
566 return rte_flow_error_set
567 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
568 item,
569 "range between \"spec\" and \"last\""
570 " is larger than \"mask\"");
571 }
572 return 0;
573 }
574
575 /** Graph of supported items and associated actions. */
576 static const struct mlx4_flow_proc_item mlx4_flow_proc_item_list[] = {
577 [RTE_FLOW_ITEM_TYPE_END] = {
578 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_ETH),
579 },
580 [RTE_FLOW_ITEM_TYPE_ETH] = {
581 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_VLAN,
582 RTE_FLOW_ITEM_TYPE_IPV4),
583 .mask_support = &(const struct rte_flow_item_eth){
584 /* Only destination MAC can be matched. */
585 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
586 },
587 .mask_default = &rte_flow_item_eth_mask,
588 .mask_sz = sizeof(struct rte_flow_item_eth),
589 .merge = mlx4_flow_merge_eth,
590 .dst_sz = sizeof(struct ibv_flow_spec_eth),
591 },
592 [RTE_FLOW_ITEM_TYPE_VLAN] = {
593 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_IPV4),
594 .mask_support = &(const struct rte_flow_item_vlan){
595 /* Only TCI VID matching is supported. */
596 .tci = RTE_BE16(0x0fff),
597 },
598 .mask_default = &rte_flow_item_vlan_mask,
599 .mask_sz = sizeof(struct rte_flow_item_vlan),
600 .merge = mlx4_flow_merge_vlan,
601 .dst_sz = 0,
602 },
603 [RTE_FLOW_ITEM_TYPE_IPV4] = {
604 .next_item = NEXT_ITEM(RTE_FLOW_ITEM_TYPE_UDP,
605 RTE_FLOW_ITEM_TYPE_TCP),
606 .mask_support = &(const struct rte_flow_item_ipv4){
607 .hdr = {
608 .src_addr = RTE_BE32(0xffffffff),
609 .dst_addr = RTE_BE32(0xffffffff),
610 },
611 },
612 .mask_default = &rte_flow_item_ipv4_mask,
613 .mask_sz = sizeof(struct rte_flow_item_ipv4),
614 .merge = mlx4_flow_merge_ipv4,
615 .dst_sz = sizeof(struct ibv_flow_spec_ipv4),
616 },
617 [RTE_FLOW_ITEM_TYPE_UDP] = {
618 .mask_support = &(const struct rte_flow_item_udp){
619 .hdr = {
620 .src_port = RTE_BE16(0xffff),
621 .dst_port = RTE_BE16(0xffff),
622 },
623 },
624 .mask_default = &rte_flow_item_udp_mask,
625 .mask_sz = sizeof(struct rte_flow_item_udp),
626 .merge = mlx4_flow_merge_udp,
627 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
628 },
629 [RTE_FLOW_ITEM_TYPE_TCP] = {
630 .mask_support = &(const struct rte_flow_item_tcp){
631 .hdr = {
632 .src_port = RTE_BE16(0xffff),
633 .dst_port = RTE_BE16(0xffff),
634 },
635 },
636 .mask_default = &rte_flow_item_tcp_mask,
637 .mask_sz = sizeof(struct rte_flow_item_tcp),
638 .merge = mlx4_flow_merge_tcp,
639 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
640 },
641 };
642
643 /**
644 * Make sure a flow rule is supported and initialize associated structure.
645 *
646 * @param priv
647 * Pointer to private structure.
648 * @param[in] attr
649 * Flow rule attributes.
650 * @param[in] pattern
651 * Pattern specification (list terminated by the END pattern item).
652 * @param[in] actions
653 * Associated actions (list terminated by the END action).
654 * @param[out] error
655 * Perform verbose error reporting if not NULL.
656 * @param[in, out] addr
657 * Buffer where the resulting flow rule handle pointer must be stored.
658 * If NULL, stop processing after validation stage.
659 *
660 * @return
661 * 0 on success, a negative errno value otherwise and rte_errno is set.
662 */
663 static int
mlx4_flow_prepare(struct mlx4_priv * priv,const struct rte_flow_attr * attr,const struct rte_flow_item pattern[],const struct rte_flow_action actions[],struct rte_flow_error * error,struct rte_flow ** addr)664 mlx4_flow_prepare(struct mlx4_priv *priv,
665 const struct rte_flow_attr *attr,
666 const struct rte_flow_item pattern[],
667 const struct rte_flow_action actions[],
668 struct rte_flow_error *error,
669 struct rte_flow **addr)
670 {
671 const struct rte_flow_item *item;
672 const struct rte_flow_action *action;
673 const struct mlx4_flow_proc_item *proc;
674 struct rte_flow temp = { .ibv_attr_size = sizeof(*temp.ibv_attr) };
675 struct rte_flow *flow = &temp;
676 const char *msg = NULL;
677 int overlap;
678
679 if (attr->group)
680 return rte_flow_error_set
681 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
682 NULL, "groups are not supported");
683 if (attr->priority > MLX4_FLOW_PRIORITY_LAST)
684 return rte_flow_error_set
685 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
686 NULL, "maximum priority level is "
687 MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST));
688 if (attr->egress)
689 return rte_flow_error_set
690 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
691 NULL, "egress is not supported");
692 if (attr->transfer)
693 return rte_flow_error_set
694 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_TRANSFER,
695 NULL, "transfer is not supported");
696 if (!attr->ingress)
697 return rte_flow_error_set
698 (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
699 NULL, "only ingress is supported");
700 fill:
701 overlap = 0;
702 proc = mlx4_flow_proc_item_list;
703 flow->priority = attr->priority;
704 /* Go over pattern. */
705 for (item = pattern; item->type; ++item) {
706 const struct mlx4_flow_proc_item *next = NULL;
707 unsigned int i;
708 int err;
709
710 if (item->type == RTE_FLOW_ITEM_TYPE_VOID)
711 continue;
712 if (item->type == MLX4_FLOW_ITEM_TYPE_INTERNAL) {
713 flow->internal = 1;
714 continue;
715 }
716 if (flow->promisc || flow->allmulti) {
717 msg = "mlx4 does not support additional matching"
718 " criteria combined with indiscriminate"
719 " matching on Ethernet headers";
720 goto exit_item_not_supported;
721 }
722 for (i = 0; proc->next_item && proc->next_item[i]; ++i) {
723 if (proc->next_item[i] == item->type) {
724 next = &mlx4_flow_proc_item_list[item->type];
725 break;
726 }
727 }
728 if (!next)
729 goto exit_item_not_supported;
730 proc = next;
731 /*
732 * Perform basic sanity checks only once, while handle is
733 * not allocated.
734 */
735 if (flow == &temp) {
736 err = mlx4_flow_item_check(item, proc, error);
737 if (err)
738 return err;
739 }
740 if (proc->merge) {
741 err = proc->merge(flow, item, proc, error);
742 if (err)
743 return err;
744 }
745 flow->ibv_attr_size += proc->dst_sz;
746 }
747 /* Go over actions list. */
748 for (action = actions; action->type; ++action) {
749 /* This one may appear anywhere multiple times. */
750 if (action->type == RTE_FLOW_ACTION_TYPE_VOID)
751 continue;
752 /* Fate-deciding actions may appear exactly once. */
753 if (overlap) {
754 msg = "cannot combine several fate-deciding actions,"
755 " choose between DROP, QUEUE or RSS";
756 goto exit_action_not_supported;
757 }
758 overlap = 1;
759 switch (action->type) {
760 const struct rte_flow_action_queue *queue;
761 const struct rte_flow_action_rss *rss;
762 const uint8_t *rss_key;
763 uint32_t rss_key_len;
764 uint64_t fields;
765 unsigned int i;
766
767 case RTE_FLOW_ACTION_TYPE_DROP:
768 flow->drop = 1;
769 break;
770 case RTE_FLOW_ACTION_TYPE_QUEUE:
771 if (flow->rss)
772 break;
773 queue = action->conf;
774 if (queue->index >= ETH_DEV(priv)->data->nb_rx_queues) {
775 msg = "queue target index beyond number of"
776 " configured Rx queues";
777 goto exit_action_not_supported;
778 }
779 flow->rss = mlx4_rss_get
780 (priv, 0, mlx4_rss_hash_key_default, 1,
781 &queue->index);
782 if (!flow->rss) {
783 msg = "not enough resources for additional"
784 " single-queue RSS context";
785 goto exit_action_not_supported;
786 }
787 break;
788 case RTE_FLOW_ACTION_TYPE_RSS:
789 if (flow->rss)
790 break;
791 rss = action->conf;
792 /* Default RSS configuration if none is provided. */
793 if (rss->key_len) {
794 rss_key = rss->key;
795 rss_key_len = rss->key_len;
796 } else {
797 rss_key = mlx4_rss_hash_key_default;
798 rss_key_len = MLX4_RSS_HASH_KEY_SIZE;
799 }
800 /* Sanity checks. */
801 for (i = 0; i < rss->queue_num; ++i)
802 if (rss->queue[i] >=
803 ETH_DEV(priv)->data->nb_rx_queues)
804 break;
805 if (i != rss->queue_num) {
806 msg = "queue index target beyond number of"
807 " configured Rx queues";
808 goto exit_action_not_supported;
809 }
810 if (!rte_is_power_of_2(rss->queue_num)) {
811 msg = "for RSS, mlx4 requires the number of"
812 " queues to be a power of two";
813 goto exit_action_not_supported;
814 }
815 if (rss_key_len != sizeof(flow->rss->key)) {
816 msg = "mlx4 supports exactly one RSS hash key"
817 " length: "
818 MLX4_STR_EXPAND(MLX4_RSS_HASH_KEY_SIZE);
819 goto exit_action_not_supported;
820 }
821 for (i = 1; i < rss->queue_num; ++i)
822 if (rss->queue[i] - rss->queue[i - 1] != 1)
823 break;
824 if (i != rss->queue_num) {
825 msg = "mlx4 requires RSS contexts to use"
826 " consecutive queue indices only";
827 goto exit_action_not_supported;
828 }
829 if (rss->queue[0] % rss->queue_num) {
830 msg = "mlx4 requires the first queue of a RSS"
831 " context to be aligned on a multiple"
832 " of the context size";
833 goto exit_action_not_supported;
834 }
835 if (rss->func &&
836 rss->func != RTE_ETH_HASH_FUNCTION_TOEPLITZ) {
837 msg = "the only supported RSS hash function"
838 " is Toeplitz";
839 goto exit_action_not_supported;
840 }
841 if (rss->level) {
842 msg = "a nonzero RSS encapsulation level is"
843 " not supported";
844 goto exit_action_not_supported;
845 }
846 rte_errno = 0;
847 fields = mlx4_conv_rss_types(priv, rss->types, 0);
848 if (fields == (uint64_t)-1 && rte_errno) {
849 msg = "unsupported RSS hash type requested";
850 goto exit_action_not_supported;
851 }
852 flow->rss = mlx4_rss_get
853 (priv, fields, rss_key, rss->queue_num,
854 rss->queue);
855 if (!flow->rss) {
856 msg = "either invalid parameters or not enough"
857 " resources for additional multi-queue"
858 " RSS context";
859 goto exit_action_not_supported;
860 }
861 break;
862 default:
863 goto exit_action_not_supported;
864 }
865 }
866 /* When fate is unknown, drop traffic. */
867 if (!overlap)
868 flow->drop = 1;
869 /* Validation ends here. */
870 if (!addr) {
871 if (flow->rss)
872 mlx4_rss_put(flow->rss);
873 return 0;
874 }
875 if (flow == &temp) {
876 /* Allocate proper handle based on collected data. */
877 const struct mlx4_malloc_vec vec[] = {
878 {
879 .align = alignof(struct rte_flow),
880 .size = sizeof(*flow),
881 .addr = (void **)&flow,
882 },
883 {
884 .align = alignof(struct ibv_flow_attr),
885 .size = temp.ibv_attr_size,
886 .addr = (void **)&temp.ibv_attr,
887 },
888 };
889
890 if (!mlx4_zmallocv(__func__, vec, RTE_DIM(vec))) {
891 if (temp.rss)
892 mlx4_rss_put(temp.rss);
893 return rte_flow_error_set
894 (error, -rte_errno,
895 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
896 "flow rule handle allocation failure");
897 }
898 /* Most fields will be updated by second pass. */
899 *flow = (struct rte_flow){
900 .ibv_attr = temp.ibv_attr,
901 .ibv_attr_size = sizeof(*flow->ibv_attr),
902 .rss = temp.rss,
903 };
904 *flow->ibv_attr = (struct ibv_flow_attr){
905 .type = IBV_FLOW_ATTR_NORMAL,
906 .size = sizeof(*flow->ibv_attr),
907 .priority = attr->priority,
908 .port = priv->port,
909 };
910 goto fill;
911 }
912 *addr = flow;
913 return 0;
914 exit_item_not_supported:
915 return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
916 item, msg ? msg : "item not supported");
917 exit_action_not_supported:
918 return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
919 action, msg ? msg : "action not supported");
920 }
921
922 /**
923 * Validate a flow supported by the NIC.
924 *
925 * @see rte_flow_validate()
926 * @see rte_flow_ops
927 */
928 static int
mlx4_flow_validate(struct rte_eth_dev * dev,const struct rte_flow_attr * attr,const struct rte_flow_item pattern[],const struct rte_flow_action actions[],struct rte_flow_error * error)929 mlx4_flow_validate(struct rte_eth_dev *dev,
930 const struct rte_flow_attr *attr,
931 const struct rte_flow_item pattern[],
932 const struct rte_flow_action actions[],
933 struct rte_flow_error *error)
934 {
935 struct mlx4_priv *priv = dev->data->dev_private;
936
937 return mlx4_flow_prepare(priv, attr, pattern, actions, error, NULL);
938 }
939
940 /**
941 * Get a drop flow rule resources instance.
942 *
943 * @param priv
944 * Pointer to private structure.
945 *
946 * @return
947 * Pointer to drop flow resources on success, NULL otherwise and rte_errno
948 * is set.
949 */
950 static struct mlx4_drop *
mlx4_drop_get(struct mlx4_priv * priv)951 mlx4_drop_get(struct mlx4_priv *priv)
952 {
953 struct mlx4_drop *drop = priv->drop;
954
955 if (drop) {
956 MLX4_ASSERT(drop->refcnt);
957 MLX4_ASSERT(drop->priv == priv);
958 ++drop->refcnt;
959 return drop;
960 }
961 drop = rte_malloc(__func__, sizeof(*drop), 0);
962 if (!drop)
963 goto error;
964 *drop = (struct mlx4_drop){
965 .priv = priv,
966 .refcnt = 1,
967 };
968 drop->cq = mlx4_glue->create_cq(priv->ctx, 1, NULL, NULL, 0);
969 if (!drop->cq)
970 goto error;
971 drop->qp = mlx4_glue->create_qp
972 (priv->pd,
973 &(struct ibv_qp_init_attr){
974 .send_cq = drop->cq,
975 .recv_cq = drop->cq,
976 .qp_type = IBV_QPT_RAW_PACKET,
977 });
978 if (!drop->qp)
979 goto error;
980 priv->drop = drop;
981 return drop;
982 error:
983 if (drop) {
984 if (drop->qp)
985 claim_zero(mlx4_glue->destroy_qp(drop->qp));
986 if (drop->cq)
987 claim_zero(mlx4_glue->destroy_cq(drop->cq));
988 rte_free(drop);
989 }
990 rte_errno = ENOMEM;
991 return NULL;
992 }
993
994 /**
995 * Give back a drop flow rule resources instance.
996 *
997 * @param drop
998 * Pointer to drop flow rule resources.
999 */
1000 static void
mlx4_drop_put(struct mlx4_drop * drop)1001 mlx4_drop_put(struct mlx4_drop *drop)
1002 {
1003 MLX4_ASSERT(drop->refcnt);
1004 if (--drop->refcnt)
1005 return;
1006 drop->priv->drop = NULL;
1007 claim_zero(mlx4_glue->destroy_qp(drop->qp));
1008 claim_zero(mlx4_glue->destroy_cq(drop->cq));
1009 rte_free(drop);
1010 }
1011
1012 /**
1013 * Toggle a configured flow rule.
1014 *
1015 * @param priv
1016 * Pointer to private structure.
1017 * @param flow
1018 * Flow rule handle to toggle.
1019 * @param enable
1020 * Whether associated Verbs flow must be created or removed.
1021 * @param[out] error
1022 * Perform verbose error reporting if not NULL.
1023 *
1024 * @return
1025 * 0 on success, a negative errno value otherwise and rte_errno is set.
1026 */
1027 static int
mlx4_flow_toggle(struct mlx4_priv * priv,struct rte_flow * flow,int enable,struct rte_flow_error * error)1028 mlx4_flow_toggle(struct mlx4_priv *priv,
1029 struct rte_flow *flow,
1030 int enable,
1031 struct rte_flow_error *error)
1032 {
1033 struct ibv_qp *qp = NULL;
1034 const char *msg;
1035 int err;
1036
1037 if (!enable) {
1038 if (!flow->ibv_flow)
1039 return 0;
1040 claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1041 flow->ibv_flow = NULL;
1042 if (flow->drop)
1043 mlx4_drop_put(priv->drop);
1044 else if (flow->rss)
1045 mlx4_rss_detach(flow->rss);
1046 return 0;
1047 }
1048 MLX4_ASSERT(flow->ibv_attr);
1049 if (!flow->internal &&
1050 !priv->isolated &&
1051 flow->ibv_attr->priority == MLX4_FLOW_PRIORITY_LAST) {
1052 if (flow->ibv_flow) {
1053 claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1054 flow->ibv_flow = NULL;
1055 if (flow->drop)
1056 mlx4_drop_put(priv->drop);
1057 else if (flow->rss)
1058 mlx4_rss_detach(flow->rss);
1059 }
1060 err = EACCES;
1061 msg = ("priority level "
1062 MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST)
1063 " is reserved when not in isolated mode");
1064 goto error;
1065 }
1066 if (flow->rss) {
1067 struct mlx4_rss *rss = flow->rss;
1068 int missing = 0;
1069 unsigned int i;
1070
1071 /* Stop at the first nonexistent target queue. */
1072 for (i = 0; i != rss->queues; ++i)
1073 if (rss->queue_id[i] >=
1074 ETH_DEV(priv)->data->nb_rx_queues ||
1075 !ETH_DEV(priv)->data->rx_queues[rss->queue_id[i]]) {
1076 missing = 1;
1077 break;
1078 }
1079 if (flow->ibv_flow) {
1080 if (missing ^ !flow->drop)
1081 return 0;
1082 /* Verbs flow needs updating. */
1083 claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1084 flow->ibv_flow = NULL;
1085 if (flow->drop)
1086 mlx4_drop_put(priv->drop);
1087 else
1088 mlx4_rss_detach(rss);
1089 }
1090 if (!missing) {
1091 err = mlx4_rss_attach(rss);
1092 if (err) {
1093 err = -err;
1094 msg = "cannot create indirection table or hash"
1095 " QP to associate flow rule with";
1096 goto error;
1097 }
1098 qp = rss->qp;
1099 }
1100 /* A missing target queue drops traffic implicitly. */
1101 flow->drop = missing;
1102 }
1103 if (flow->drop) {
1104 if (flow->ibv_flow)
1105 return 0;
1106 mlx4_drop_get(priv);
1107 if (!priv->drop) {
1108 err = rte_errno;
1109 msg = "resources for drop flow rule cannot be created";
1110 goto error;
1111 }
1112 qp = priv->drop->qp;
1113 }
1114 MLX4_ASSERT(qp);
1115 if (flow->ibv_flow)
1116 return 0;
1117 flow->ibv_flow = mlx4_glue->create_flow(qp, flow->ibv_attr);
1118 if (flow->ibv_flow)
1119 return 0;
1120 if (flow->drop)
1121 mlx4_drop_put(priv->drop);
1122 else if (flow->rss)
1123 mlx4_rss_detach(flow->rss);
1124 err = errno;
1125 msg = "flow rule rejected by device";
1126 error:
1127 return rte_flow_error_set
1128 (error, err, RTE_FLOW_ERROR_TYPE_HANDLE, flow, msg);
1129 }
1130
1131 /**
1132 * Create a flow.
1133 *
1134 * @see rte_flow_create()
1135 * @see rte_flow_ops
1136 */
1137 static struct rte_flow *
mlx4_flow_create(struct rte_eth_dev * dev,const struct rte_flow_attr * attr,const struct rte_flow_item pattern[],const struct rte_flow_action actions[],struct rte_flow_error * error)1138 mlx4_flow_create(struct rte_eth_dev *dev,
1139 const struct rte_flow_attr *attr,
1140 const struct rte_flow_item pattern[],
1141 const struct rte_flow_action actions[],
1142 struct rte_flow_error *error)
1143 {
1144 struct mlx4_priv *priv = dev->data->dev_private;
1145 struct rte_flow *flow;
1146 int err;
1147
1148 err = mlx4_flow_prepare(priv, attr, pattern, actions, error, &flow);
1149 if (err)
1150 return NULL;
1151 err = mlx4_flow_toggle(priv, flow, priv->started, error);
1152 if (!err) {
1153 struct rte_flow *curr = LIST_FIRST(&priv->flows);
1154
1155 /* New rules are inserted after internal ones. */
1156 if (!curr || !curr->internal) {
1157 LIST_INSERT_HEAD(&priv->flows, flow, next);
1158 } else {
1159 while (LIST_NEXT(curr, next) &&
1160 LIST_NEXT(curr, next)->internal)
1161 curr = LIST_NEXT(curr, next);
1162 LIST_INSERT_AFTER(curr, flow, next);
1163 }
1164 return flow;
1165 }
1166 if (flow->rss)
1167 mlx4_rss_put(flow->rss);
1168 rte_flow_error_set(error, -err, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1169 error->message);
1170 rte_free(flow);
1171 return NULL;
1172 }
1173
1174 /**
1175 * Configure isolated mode.
1176 *
1177 * @see rte_flow_isolate()
1178 * @see rte_flow_ops
1179 */
1180 static int
mlx4_flow_isolate(struct rte_eth_dev * dev,int enable,struct rte_flow_error * error)1181 mlx4_flow_isolate(struct rte_eth_dev *dev,
1182 int enable,
1183 struct rte_flow_error *error)
1184 {
1185 struct mlx4_priv *priv = dev->data->dev_private;
1186
1187 if (!!enable == !!priv->isolated)
1188 return 0;
1189 priv->isolated = !!enable;
1190 if (mlx4_flow_sync(priv, error)) {
1191 priv->isolated = !enable;
1192 return -rte_errno;
1193 }
1194 return 0;
1195 }
1196
1197 /**
1198 * Destroy a flow rule.
1199 *
1200 * @see rte_flow_destroy()
1201 * @see rte_flow_ops
1202 */
1203 static int
mlx4_flow_destroy(struct rte_eth_dev * dev,struct rte_flow * flow,struct rte_flow_error * error)1204 mlx4_flow_destroy(struct rte_eth_dev *dev,
1205 struct rte_flow *flow,
1206 struct rte_flow_error *error)
1207 {
1208 struct mlx4_priv *priv = dev->data->dev_private;
1209 int err = mlx4_flow_toggle(priv, flow, 0, error);
1210
1211 if (err)
1212 return err;
1213 LIST_REMOVE(flow, next);
1214 if (flow->rss)
1215 mlx4_rss_put(flow->rss);
1216 rte_free(flow);
1217 return 0;
1218 }
1219
1220 /**
1221 * Destroy user-configured flow rules.
1222 *
1223 * This function skips internal flows rules.
1224 *
1225 * @see rte_flow_flush()
1226 * @see rte_flow_ops
1227 */
1228 static int
mlx4_flow_flush(struct rte_eth_dev * dev,struct rte_flow_error * error)1229 mlx4_flow_flush(struct rte_eth_dev *dev,
1230 struct rte_flow_error *error)
1231 {
1232 struct mlx4_priv *priv = dev->data->dev_private;
1233 struct rte_flow *flow = LIST_FIRST(&priv->flows);
1234
1235 while (flow) {
1236 struct rte_flow *next = LIST_NEXT(flow, next);
1237
1238 if (!flow->internal)
1239 mlx4_flow_destroy(dev, flow, error);
1240 flow = next;
1241 }
1242 return 0;
1243 }
1244
1245 /**
1246 * Helper function to determine the next configured VLAN filter.
1247 *
1248 * @param priv
1249 * Pointer to private structure.
1250 * @param vlan
1251 * VLAN ID to use as a starting point.
1252 *
1253 * @return
1254 * Next configured VLAN ID or a high value (>= 4096) if there is none.
1255 */
1256 static uint16_t
mlx4_flow_internal_next_vlan(struct mlx4_priv * priv,uint16_t vlan)1257 mlx4_flow_internal_next_vlan(struct mlx4_priv *priv, uint16_t vlan)
1258 {
1259 while (vlan < 4096) {
1260 if (ETH_DEV(priv)->data->vlan_filter_conf.ids[vlan / 64] &
1261 (UINT64_C(1) << (vlan % 64)))
1262 return vlan;
1263 ++vlan;
1264 }
1265 return vlan;
1266 }
1267
1268 /**
1269 * Generate internal flow rules.
1270 *
1271 * Various flow rules are created depending on the mode the device is in:
1272 *
1273 * 1. Promiscuous:
1274 * port MAC + broadcast + catch-all (VLAN filtering is ignored).
1275 * 2. All multicast:
1276 * port MAC/VLAN + broadcast + catch-all multicast.
1277 * 3. Otherwise:
1278 * port MAC/VLAN + broadcast MAC/VLAN.
1279 *
1280 * About MAC flow rules:
1281 *
1282 * - MAC flow rules are generated from @p dev->data->mac_addrs
1283 * (@p priv->mac array).
1284 * - An additional flow rule for Ethernet broadcasts is also generated.
1285 * - All these are per-VLAN if @p DEV_RX_OFFLOAD_VLAN_FILTER
1286 * is enabled and VLAN filters are configured.
1287 *
1288 * @param priv
1289 * Pointer to private structure.
1290 * @param[out] error
1291 * Perform verbose error reporting if not NULL.
1292 *
1293 * @return
1294 * 0 on success, a negative errno value otherwise and rte_errno is set.
1295 */
1296 static int
mlx4_flow_internal(struct mlx4_priv * priv,struct rte_flow_error * error)1297 mlx4_flow_internal(struct mlx4_priv *priv, struct rte_flow_error *error)
1298 {
1299 struct rte_flow_attr attr = {
1300 .priority = MLX4_FLOW_PRIORITY_LAST,
1301 .ingress = 1,
1302 };
1303 struct rte_flow_item_eth eth_spec;
1304 const struct rte_flow_item_eth eth_mask = {
1305 .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
1306 };
1307 const struct rte_flow_item_eth eth_allmulti = {
1308 .dst.addr_bytes = "\x01\x00\x00\x00\x00\x00",
1309 };
1310 struct rte_flow_item_vlan vlan_spec;
1311 const struct rte_flow_item_vlan vlan_mask = {
1312 .tci = RTE_BE16(0x0fff),
1313 };
1314 struct rte_flow_item pattern[] = {
1315 {
1316 .type = MLX4_FLOW_ITEM_TYPE_INTERNAL,
1317 },
1318 {
1319 .type = RTE_FLOW_ITEM_TYPE_ETH,
1320 .spec = ð_spec,
1321 .mask = ð_mask,
1322 },
1323 {
1324 /* Replaced with VLAN if filtering is enabled. */
1325 .type = RTE_FLOW_ITEM_TYPE_END,
1326 },
1327 {
1328 .type = RTE_FLOW_ITEM_TYPE_END,
1329 },
1330 };
1331 /*
1332 * Round number of queues down to their previous power of 2 to
1333 * comply with RSS context limitations. Extra queues silently do not
1334 * get RSS by default.
1335 */
1336 uint32_t queues =
1337 rte_align32pow2(ETH_DEV(priv)->data->nb_rx_queues + 1) >> 1;
1338 uint16_t queue[queues];
1339 struct rte_flow_action_rss action_rss = {
1340 .func = RTE_ETH_HASH_FUNCTION_DEFAULT,
1341 .level = 0,
1342 .types = 0,
1343 .key_len = MLX4_RSS_HASH_KEY_SIZE,
1344 .queue_num = queues,
1345 .key = mlx4_rss_hash_key_default,
1346 .queue = queue,
1347 };
1348 struct rte_flow_action actions[] = {
1349 {
1350 .type = RTE_FLOW_ACTION_TYPE_RSS,
1351 .conf = &action_rss,
1352 },
1353 {
1354 .type = RTE_FLOW_ACTION_TYPE_END,
1355 },
1356 };
1357 struct rte_ether_addr *rule_mac = ð_spec.dst;
1358 rte_be16_t *rule_vlan =
1359 (ETH_DEV(priv)->data->dev_conf.rxmode.offloads &
1360 DEV_RX_OFFLOAD_VLAN_FILTER) &&
1361 !ETH_DEV(priv)->data->promiscuous ?
1362 &vlan_spec.tci :
1363 NULL;
1364 uint16_t vlan = 0;
1365 struct rte_flow *flow;
1366 unsigned int i;
1367 int err = 0;
1368
1369 /* Nothing to be done if there are no Rx queues. */
1370 if (!queues)
1371 goto error;
1372 /* Prepare default RSS configuration. */
1373 for (i = 0; i != queues; ++i)
1374 queue[i] = i;
1375 /*
1376 * Set up VLAN item if filtering is enabled and at least one VLAN
1377 * filter is configured.
1378 */
1379 if (rule_vlan) {
1380 vlan = mlx4_flow_internal_next_vlan(priv, 0);
1381 if (vlan < 4096) {
1382 pattern[2] = (struct rte_flow_item){
1383 .type = RTE_FLOW_ITEM_TYPE_VLAN,
1384 .spec = &vlan_spec,
1385 .mask = &vlan_mask,
1386 };
1387 next_vlan:
1388 *rule_vlan = rte_cpu_to_be_16(vlan);
1389 } else {
1390 rule_vlan = NULL;
1391 }
1392 }
1393 for (i = 0; i != RTE_DIM(priv->mac) + 1; ++i) {
1394 const struct rte_ether_addr *mac;
1395
1396 /* Broadcasts are handled by an extra iteration. */
1397 if (i < RTE_DIM(priv->mac))
1398 mac = &priv->mac[i];
1399 else
1400 mac = ð_mask.dst;
1401 if (rte_is_zero_ether_addr(mac))
1402 continue;
1403 /* Check if MAC flow rule is already present. */
1404 for (flow = LIST_FIRST(&priv->flows);
1405 flow && flow->internal;
1406 flow = LIST_NEXT(flow, next)) {
1407 const struct ibv_flow_spec_eth *eth =
1408 (const void *)((uintptr_t)flow->ibv_attr +
1409 sizeof(*flow->ibv_attr));
1410 unsigned int j;
1411
1412 if (!flow->mac)
1413 continue;
1414 MLX4_ASSERT(flow->ibv_attr->type ==
1415 IBV_FLOW_ATTR_NORMAL);
1416 MLX4_ASSERT(flow->ibv_attr->num_of_specs == 1);
1417 MLX4_ASSERT(eth->type == IBV_FLOW_SPEC_ETH);
1418 MLX4_ASSERT(flow->rss);
1419 if (rule_vlan &&
1420 (eth->val.vlan_tag != *rule_vlan ||
1421 eth->mask.vlan_tag != RTE_BE16(0x0fff)))
1422 continue;
1423 if (!rule_vlan && eth->mask.vlan_tag)
1424 continue;
1425 for (j = 0; j != sizeof(mac->addr_bytes); ++j)
1426 if (eth->val.dst_mac[j] != mac->addr_bytes[j] ||
1427 eth->mask.dst_mac[j] != UINT8_C(0xff) ||
1428 eth->val.src_mac[j] != UINT8_C(0x00) ||
1429 eth->mask.src_mac[j] != UINT8_C(0x00))
1430 break;
1431 if (j != sizeof(mac->addr_bytes))
1432 continue;
1433 if (flow->rss->queues != queues ||
1434 memcmp(flow->rss->queue_id, action_rss.queue,
1435 queues * sizeof(flow->rss->queue_id[0])))
1436 continue;
1437 break;
1438 }
1439 if (!flow || !flow->internal) {
1440 /* Not found, create a new flow rule. */
1441 memcpy(rule_mac, mac, sizeof(*mac));
1442 flow = mlx4_flow_create(ETH_DEV(priv), &attr, pattern,
1443 actions, error);
1444 if (!flow) {
1445 err = -rte_errno;
1446 goto error;
1447 }
1448 }
1449 flow->select = 1;
1450 flow->mac = 1;
1451 }
1452 if (rule_vlan) {
1453 vlan = mlx4_flow_internal_next_vlan(priv, vlan + 1);
1454 if (vlan < 4096)
1455 goto next_vlan;
1456 }
1457 /* Take care of promiscuous and all multicast flow rules. */
1458 if (ETH_DEV(priv)->data->promiscuous ||
1459 ETH_DEV(priv)->data->all_multicast) {
1460 for (flow = LIST_FIRST(&priv->flows);
1461 flow && flow->internal;
1462 flow = LIST_NEXT(flow, next)) {
1463 if (ETH_DEV(priv)->data->promiscuous) {
1464 if (flow->promisc)
1465 break;
1466 } else {
1467 MLX4_ASSERT(ETH_DEV(priv)->data->all_multicast);
1468 if (flow->allmulti)
1469 break;
1470 }
1471 }
1472 if (flow && flow->internal) {
1473 MLX4_ASSERT(flow->rss);
1474 if (flow->rss->queues != queues ||
1475 memcmp(flow->rss->queue_id, action_rss.queue,
1476 queues * sizeof(flow->rss->queue_id[0])))
1477 flow = NULL;
1478 }
1479 if (!flow || !flow->internal) {
1480 /* Not found, create a new flow rule. */
1481 if (ETH_DEV(priv)->data->promiscuous) {
1482 pattern[1].spec = NULL;
1483 pattern[1].mask = NULL;
1484 } else {
1485 MLX4_ASSERT(ETH_DEV(priv)->data->all_multicast);
1486 pattern[1].spec = ð_allmulti;
1487 pattern[1].mask = ð_allmulti;
1488 }
1489 pattern[2] = pattern[3];
1490 flow = mlx4_flow_create(ETH_DEV(priv), &attr, pattern,
1491 actions, error);
1492 if (!flow) {
1493 err = -rte_errno;
1494 goto error;
1495 }
1496 }
1497 MLX4_ASSERT(flow->promisc || flow->allmulti);
1498 flow->select = 1;
1499 }
1500 error:
1501 /* Clear selection and clean up stale internal flow rules. */
1502 flow = LIST_FIRST(&priv->flows);
1503 while (flow && flow->internal) {
1504 struct rte_flow *next = LIST_NEXT(flow, next);
1505
1506 if (!flow->select)
1507 claim_zero(mlx4_flow_destroy(ETH_DEV(priv), flow,
1508 error));
1509 else
1510 flow->select = 0;
1511 flow = next;
1512 }
1513 return err;
1514 }
1515
1516 /**
1517 * Synchronize flow rules.
1518 *
1519 * This function synchronizes flow rules with the state of the device by
1520 * taking into account isolated mode and whether target queues are
1521 * configured.
1522 *
1523 * @param priv
1524 * Pointer to private structure.
1525 * @param[out] error
1526 * Perform verbose error reporting if not NULL.
1527 *
1528 * @return
1529 * 0 on success, a negative errno value otherwise and rte_errno is set.
1530 */
1531 int
mlx4_flow_sync(struct mlx4_priv * priv,struct rte_flow_error * error)1532 mlx4_flow_sync(struct mlx4_priv *priv, struct rte_flow_error *error)
1533 {
1534 struct rte_flow *flow;
1535 int ret;
1536
1537 /* Internal flow rules are guaranteed to come first in the list. */
1538 if (priv->isolated) {
1539 /*
1540 * Get rid of them in isolated mode, stop at the first
1541 * non-internal rule found.
1542 */
1543 for (flow = LIST_FIRST(&priv->flows);
1544 flow && flow->internal;
1545 flow = LIST_FIRST(&priv->flows))
1546 claim_zero(mlx4_flow_destroy(ETH_DEV(priv), flow,
1547 error));
1548 } else {
1549 /* Refresh internal rules. */
1550 ret = mlx4_flow_internal(priv, error);
1551 if (ret)
1552 return ret;
1553 }
1554 /* Toggle the remaining flow rules . */
1555 LIST_FOREACH(flow, &priv->flows, next) {
1556 ret = mlx4_flow_toggle(priv, flow, priv->started, error);
1557 if (ret)
1558 return ret;
1559 }
1560 if (!priv->started)
1561 MLX4_ASSERT(!priv->drop);
1562 return 0;
1563 }
1564
1565 /**
1566 * Clean up all flow rules.
1567 *
1568 * Unlike mlx4_flow_flush(), this function takes care of all remaining flow
1569 * rules regardless of whether they are internal or user-configured.
1570 *
1571 * @param priv
1572 * Pointer to private structure.
1573 */
1574 void
mlx4_flow_clean(struct mlx4_priv * priv)1575 mlx4_flow_clean(struct mlx4_priv *priv)
1576 {
1577 struct rte_flow *flow;
1578
1579 while ((flow = LIST_FIRST(&priv->flows)))
1580 mlx4_flow_destroy(ETH_DEV(priv), flow, NULL);
1581 MLX4_ASSERT(LIST_EMPTY(&priv->rss));
1582 }
1583
1584 static const struct rte_flow_ops mlx4_flow_ops = {
1585 .validate = mlx4_flow_validate,
1586 .create = mlx4_flow_create,
1587 .destroy = mlx4_flow_destroy,
1588 .flush = mlx4_flow_flush,
1589 .isolate = mlx4_flow_isolate,
1590 };
1591
1592 /**
1593 * Manage filter operations.
1594 *
1595 * @param dev
1596 * Pointer to Ethernet device structure.
1597 * @param filter_type
1598 * Filter type.
1599 * @param filter_op
1600 * Operation to perform.
1601 * @param arg
1602 * Pointer to operation-specific structure.
1603 *
1604 * @return
1605 * 0 on success, negative errno value otherwise and rte_errno is set.
1606 */
1607 int
mlx4_filter_ctrl(struct rte_eth_dev * dev,enum rte_filter_type filter_type,enum rte_filter_op filter_op,void * arg)1608 mlx4_filter_ctrl(struct rte_eth_dev *dev,
1609 enum rte_filter_type filter_type,
1610 enum rte_filter_op filter_op,
1611 void *arg)
1612 {
1613 switch (filter_type) {
1614 case RTE_ETH_FILTER_GENERIC:
1615 if (filter_op != RTE_ETH_FILTER_GET)
1616 break;
1617 *(const void **)arg = &mlx4_flow_ops;
1618 return 0;
1619 default:
1620 ERROR("%p: filter type (%d) not supported",
1621 (void *)dev, filter_type);
1622 break;
1623 }
1624 rte_errno = ENOTSUP;
1625 return -rte_errno;
1626 }
1627