xref: /dpdk/drivers/net/mlx4/mlx4_flow.c (revision e0d947a1e6c2f80aa039a4f7082a8aa16797d8b9)
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 <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
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] = RTE_ETH_RSS_IPV4,
113 		[IPV4_1] = RTE_ETH_RSS_FRAG_IPV4,
114 		[IPV4_2] = RTE_ETH_RSS_NONFRAG_IPV4_OTHER,
115 		[IPV6] = RTE_ETH_RSS_IPV6,
116 		[IPV6_1] = RTE_ETH_RSS_FRAG_IPV6,
117 		[IPV6_2] = RTE_ETH_RSS_NONFRAG_IPV6_OTHER,
118 		[IPV6_3] = RTE_ETH_RSS_IPV6_EX,
119 		[TCP] = 0,
120 		[UDP] = 0,
121 		[IPV4_TCP] = RTE_ETH_RSS_NONFRAG_IPV4_TCP,
122 		[IPV4_UDP] = RTE_ETH_RSS_NONFRAG_IPV4_UDP,
123 		[IPV6_TCP] = RTE_ETH_RSS_NONFRAG_IPV6_TCP,
124 		[IPV6_TCP_1] = RTE_ETH_RSS_IPV6_TCP_EX,
125 		[IPV6_UDP] = RTE_ETH_RSS_NONFRAG_IPV6_UDP,
126 		[IPV6_UDP_1] = RTE_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
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->hdr.dst_addr.addr_bytes); ++i) {
211 			sum_dst += mask->hdr.dst_addr.addr_bytes[i];
212 			sum_src += mask->hdr.src_addr.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->hdr.dst_addr.addr_bytes[0] == 1) {
220 			if (!(spec->hdr.dst_addr.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->hdr.dst_addr.addr_bytes, RTE_ETHER_ADDR_LEN);
255 	memcpy(eth->mask.dst_mac, mask->hdr.dst_addr.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
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->hdr.vlan_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->hdr.vlan_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->hdr.vlan_tci;
314 	eth->mask.vlan_tag = mask->hdr.vlan_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
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
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
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
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 			.hdr.dst_addr.addr_bytes = { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff },
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 			.hdr.vlan_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
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 ((item->type != RTE_FLOW_ITEM_TYPE_VLAN && flow->promisc) ||
717 		    flow->allmulti) {
718 			msg = "mlx4 does not support additional matching"
719 				" criteria combined with indiscriminate"
720 				" matching on Ethernet headers";
721 			goto exit_item_not_supported;
722 		}
723 		for (i = 0; proc->next_item && proc->next_item[i]; ++i) {
724 			if (proc->next_item[i] == item->type) {
725 				next = &mlx4_flow_proc_item_list[item->type];
726 				break;
727 			}
728 		}
729 		if (!next)
730 			goto exit_item_not_supported;
731 		proc = next;
732 		/*
733 		 * Perform basic sanity checks only once, while handle is
734 		 * not allocated.
735 		 */
736 		if (flow == &temp) {
737 			err = mlx4_flow_item_check(item, proc, error);
738 			if (err)
739 				return err;
740 		}
741 		if (proc->merge) {
742 			err = proc->merge(flow, item, proc, error);
743 			if (err)
744 				return err;
745 		}
746 		flow->ibv_attr_size += proc->dst_sz;
747 	}
748 	/* Go over actions list. */
749 	for (action = actions; action->type; ++action) {
750 		/* This one may appear anywhere multiple times. */
751 		if (action->type == RTE_FLOW_ACTION_TYPE_VOID)
752 			continue;
753 		/* Fate-deciding actions may appear exactly once. */
754 		if (overlap) {
755 			msg = "cannot combine several fate-deciding actions,"
756 				" choose between DROP, QUEUE or RSS";
757 			goto exit_action_not_supported;
758 		}
759 		overlap = 1;
760 		switch (action->type) {
761 			const struct rte_flow_action_queue *queue;
762 			const struct rte_flow_action_rss *rss;
763 			const uint8_t *rss_key;
764 			uint32_t rss_key_len;
765 			uint64_t fields;
766 			unsigned int i;
767 
768 		case RTE_FLOW_ACTION_TYPE_DROP:
769 			flow->drop = 1;
770 			break;
771 		case RTE_FLOW_ACTION_TYPE_QUEUE:
772 			if (flow->rss)
773 				break;
774 			queue = action->conf;
775 			if (queue->index >= ETH_DEV(priv)->data->nb_rx_queues) {
776 				msg = "queue target index beyond number of"
777 					" configured Rx queues";
778 				goto exit_action_not_supported;
779 			}
780 			flow->rss = mlx4_rss_get
781 				(priv, 0, mlx4_rss_hash_key_default, 1,
782 				 &queue->index);
783 			if (!flow->rss) {
784 				msg = "not enough resources for additional"
785 					" single-queue RSS context";
786 				goto exit_action_not_supported;
787 			}
788 			break;
789 		case RTE_FLOW_ACTION_TYPE_RSS:
790 			if (flow->rss)
791 				break;
792 			rss = action->conf;
793 			/* Default RSS configuration if none is provided. */
794 			if (rss->key_len) {
795 				rss_key = rss->key ?
796 					  rss->key : mlx4_rss_hash_key_default;
797 				rss_key_len = rss->key_len;
798 			} else {
799 				rss_key = mlx4_rss_hash_key_default;
800 				rss_key_len = MLX4_RSS_HASH_KEY_SIZE;
801 			}
802 			/* Sanity checks. */
803 			for (i = 0; i < rss->queue_num; ++i)
804 				if (rss->queue[i] >=
805 				    ETH_DEV(priv)->data->nb_rx_queues)
806 					break;
807 			if (i != rss->queue_num) {
808 				msg = "queue index target beyond number of"
809 					" configured Rx queues";
810 				goto exit_action_not_supported;
811 			}
812 			if (!rte_is_power_of_2(rss->queue_num)) {
813 				msg = "for RSS, mlx4 requires the number of"
814 					" queues to be a power of two";
815 				goto exit_action_not_supported;
816 			}
817 			if (rss_key_len != sizeof(flow->rss->key)) {
818 				msg = "mlx4 supports exactly one RSS hash key"
819 					" length: "
820 					MLX4_STR_EXPAND(MLX4_RSS_HASH_KEY_SIZE);
821 				goto exit_action_not_supported;
822 			}
823 			for (i = 1; i < rss->queue_num; ++i)
824 				if (rss->queue[i] - rss->queue[i - 1] != 1)
825 					break;
826 			if (i != rss->queue_num) {
827 				msg = "mlx4 requires RSS contexts to use"
828 					" consecutive queue indices only";
829 				goto exit_action_not_supported;
830 			}
831 			if (rss->queue[0] % rss->queue_num) {
832 				msg = "mlx4 requires the first queue of a RSS"
833 					" context to be aligned on a multiple"
834 					" of the context size";
835 				goto exit_action_not_supported;
836 			}
837 			if (rss->func &&
838 			    rss->func != RTE_ETH_HASH_FUNCTION_TOEPLITZ) {
839 				msg = "the only supported RSS hash function"
840 					" is Toeplitz";
841 				goto exit_action_not_supported;
842 			}
843 			if (rss->level) {
844 				msg = "a nonzero RSS encapsulation level is"
845 					" not supported";
846 				goto exit_action_not_supported;
847 			}
848 			rte_errno = 0;
849 			fields = mlx4_conv_rss_types(priv, rss->types, 0);
850 			if (fields == (uint64_t)-1 && rte_errno) {
851 				msg = "unsupported RSS hash type requested";
852 				goto exit_action_not_supported;
853 			}
854 			flow->rss = mlx4_rss_get
855 				(priv, fields, rss_key, rss->queue_num,
856 				 rss->queue);
857 			if (!flow->rss) {
858 				msg = "either invalid parameters or not enough"
859 					" resources for additional multi-queue"
860 					" RSS context";
861 				goto exit_action_not_supported;
862 			}
863 			break;
864 		default:
865 			goto exit_action_not_supported;
866 		}
867 	}
868 	/* When fate is unknown, drop traffic. */
869 	if (!overlap)
870 		flow->drop = 1;
871 	/* Validation ends here. */
872 	if (!addr) {
873 		if (flow->rss)
874 			mlx4_rss_put(flow->rss);
875 		return 0;
876 	}
877 	if (flow == &temp) {
878 		/* Allocate proper handle based on collected data. */
879 		const struct mlx4_malloc_vec vec[] = {
880 			{
881 				.align = alignof(struct rte_flow),
882 				.size = sizeof(*flow),
883 				.addr = (void **)&flow,
884 			},
885 			{
886 				.align = alignof(struct ibv_flow_attr),
887 				.size = temp.ibv_attr_size,
888 				.addr = (void **)&temp.ibv_attr,
889 			},
890 		};
891 
892 		if (!mlx4_zmallocv(__func__, vec, RTE_DIM(vec))) {
893 			if (temp.rss)
894 				mlx4_rss_put(temp.rss);
895 			return rte_flow_error_set
896 				(error, -rte_errno,
897 				 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
898 				 "flow rule handle allocation failure");
899 		}
900 		/* Most fields will be updated by second pass. */
901 		*flow = (struct rte_flow){
902 			.ibv_attr = temp.ibv_attr,
903 			.ibv_attr_size = sizeof(*flow->ibv_attr),
904 			.rss = temp.rss,
905 		};
906 		*flow->ibv_attr = (struct ibv_flow_attr){
907 			.type = IBV_FLOW_ATTR_NORMAL,
908 			.size = sizeof(*flow->ibv_attr),
909 			.priority = attr->priority,
910 			.port = priv->port,
911 		};
912 		goto fill;
913 	}
914 	*addr = flow;
915 	return 0;
916 exit_item_not_supported:
917 	return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
918 				  item, msg ? msg : "item not supported");
919 exit_action_not_supported:
920 	return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
921 				  action, msg ? msg : "action not supported");
922 }
923 
924 /**
925  * Validate a flow supported by the NIC.
926  *
927  * @see rte_flow_validate()
928  * @see rte_flow_ops
929  */
930 static int
931 mlx4_flow_validate(struct rte_eth_dev *dev,
932 		   const struct rte_flow_attr *attr,
933 		   const struct rte_flow_item pattern[],
934 		   const struct rte_flow_action actions[],
935 		   struct rte_flow_error *error)
936 {
937 	struct mlx4_priv *priv = dev->data->dev_private;
938 
939 	return mlx4_flow_prepare(priv, attr, pattern, actions, error, NULL);
940 }
941 
942 /**
943  * Get a drop flow rule resources instance.
944  *
945  * @param priv
946  *   Pointer to private structure.
947  *
948  * @return
949  *   Pointer to drop flow resources on success, NULL otherwise and rte_errno
950  *   is set.
951  */
952 static struct mlx4_drop *
953 mlx4_drop_get(struct mlx4_priv *priv)
954 {
955 	struct mlx4_drop *drop = priv->drop;
956 
957 	if (drop) {
958 		MLX4_ASSERT(drop->refcnt);
959 		MLX4_ASSERT(drop->priv == priv);
960 		++drop->refcnt;
961 		return drop;
962 	}
963 	drop = rte_malloc(__func__, sizeof(*drop), 0);
964 	if (!drop)
965 		goto error;
966 	*drop = (struct mlx4_drop){
967 		.priv = priv,
968 		.refcnt = 1,
969 	};
970 	drop->cq = mlx4_glue->create_cq(priv->ctx, 1, NULL, NULL, 0);
971 	if (!drop->cq)
972 		goto error;
973 	drop->qp = mlx4_glue->create_qp
974 		(priv->pd,
975 		 &(struct ibv_qp_init_attr){
976 			.send_cq = drop->cq,
977 			.recv_cq = drop->cq,
978 			.qp_type = IBV_QPT_RAW_PACKET,
979 		 });
980 	if (!drop->qp)
981 		goto error;
982 	priv->drop = drop;
983 	return drop;
984 error:
985 	if (drop) {
986 		if (drop->qp)
987 			claim_zero(mlx4_glue->destroy_qp(drop->qp));
988 		if (drop->cq)
989 			claim_zero(mlx4_glue->destroy_cq(drop->cq));
990 		rte_free(drop);
991 	}
992 	rte_errno = ENOMEM;
993 	return NULL;
994 }
995 
996 /**
997  * Give back a drop flow rule resources instance.
998  *
999  * @param drop
1000  *   Pointer to drop flow rule resources.
1001  */
1002 static void
1003 mlx4_drop_put(struct mlx4_drop *drop)
1004 {
1005 	MLX4_ASSERT(drop->refcnt);
1006 	if (--drop->refcnt)
1007 		return;
1008 	drop->priv->drop = NULL;
1009 	claim_zero(mlx4_glue->destroy_qp(drop->qp));
1010 	claim_zero(mlx4_glue->destroy_cq(drop->cq));
1011 	rte_free(drop);
1012 }
1013 
1014 /**
1015  * Toggle a configured flow rule.
1016  *
1017  * @param priv
1018  *   Pointer to private structure.
1019  * @param flow
1020  *   Flow rule handle to toggle.
1021  * @param enable
1022  *   Whether associated Verbs flow must be created or removed.
1023  * @param[out] error
1024  *   Perform verbose error reporting if not NULL.
1025  *
1026  * @return
1027  *   0 on success, a negative errno value otherwise and rte_errno is set.
1028  */
1029 static int
1030 mlx4_flow_toggle(struct mlx4_priv *priv,
1031 		 struct rte_flow *flow,
1032 		 int enable,
1033 		 struct rte_flow_error *error)
1034 {
1035 	struct ibv_qp *qp = NULL;
1036 	const char *msg;
1037 	int err;
1038 
1039 	if (!enable) {
1040 		if (!flow->ibv_flow)
1041 			return 0;
1042 		claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1043 		flow->ibv_flow = NULL;
1044 		if (flow->drop)
1045 			mlx4_drop_put(priv->drop);
1046 		else if (flow->rss)
1047 			mlx4_rss_detach(flow->rss);
1048 		return 0;
1049 	}
1050 	MLX4_ASSERT(flow->ibv_attr);
1051 	if (!flow->internal &&
1052 	    !priv->isolated &&
1053 	    flow->ibv_attr->priority == MLX4_FLOW_PRIORITY_LAST) {
1054 		if (flow->ibv_flow) {
1055 			claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1056 			flow->ibv_flow = NULL;
1057 			if (flow->drop)
1058 				mlx4_drop_put(priv->drop);
1059 			else if (flow->rss)
1060 				mlx4_rss_detach(flow->rss);
1061 		}
1062 		err = EACCES;
1063 		msg = ("priority level "
1064 		       MLX4_STR_EXPAND(MLX4_FLOW_PRIORITY_LAST)
1065 		       " is reserved when not in isolated mode");
1066 		goto error;
1067 	}
1068 	if (flow->rss) {
1069 		struct mlx4_rss *rss = flow->rss;
1070 		int missing = 0;
1071 		unsigned int i;
1072 
1073 		/* Stop at the first nonexistent target queue. */
1074 		for (i = 0; i != rss->queues; ++i)
1075 			if (rss->queue_id[i] >=
1076 			    ETH_DEV(priv)->data->nb_rx_queues ||
1077 			    !ETH_DEV(priv)->data->rx_queues[rss->queue_id[i]]) {
1078 				missing = 1;
1079 				break;
1080 			}
1081 		if (flow->ibv_flow) {
1082 			if (missing ^ !flow->drop)
1083 				return 0;
1084 			/* Verbs flow needs updating. */
1085 			claim_zero(mlx4_glue->destroy_flow(flow->ibv_flow));
1086 			flow->ibv_flow = NULL;
1087 			if (flow->drop)
1088 				mlx4_drop_put(priv->drop);
1089 			else
1090 				mlx4_rss_detach(rss);
1091 		}
1092 		if (!missing) {
1093 			err = mlx4_rss_attach(rss);
1094 			if (err) {
1095 				err = -err;
1096 				msg = "cannot create indirection table or hash"
1097 					" QP to associate flow rule with";
1098 				goto error;
1099 			}
1100 			qp = rss->qp;
1101 		}
1102 		/* A missing target queue drops traffic implicitly. */
1103 		flow->drop = missing;
1104 	}
1105 	if (flow->drop) {
1106 		if (flow->ibv_flow)
1107 			return 0;
1108 		mlx4_drop_get(priv);
1109 		if (!priv->drop) {
1110 			err = rte_errno;
1111 			msg = "resources for drop flow rule cannot be created";
1112 			goto error;
1113 		}
1114 		qp = priv->drop->qp;
1115 	}
1116 	MLX4_ASSERT(qp);
1117 	if (flow->ibv_flow)
1118 		return 0;
1119 	flow->ibv_flow = mlx4_glue->create_flow(qp, flow->ibv_attr);
1120 	if (flow->ibv_flow)
1121 		return 0;
1122 	if (flow->drop)
1123 		mlx4_drop_put(priv->drop);
1124 	else if (flow->rss)
1125 		mlx4_rss_detach(flow->rss);
1126 	err = errno;
1127 	msg = "flow rule rejected by device";
1128 error:
1129 	return rte_flow_error_set
1130 		(error, err, RTE_FLOW_ERROR_TYPE_HANDLE, flow, msg);
1131 }
1132 
1133 /**
1134  * Create a flow.
1135  *
1136  * @see rte_flow_create()
1137  * @see rte_flow_ops
1138  */
1139 static struct rte_flow *
1140 mlx4_flow_create(struct rte_eth_dev *dev,
1141 		 const struct rte_flow_attr *attr,
1142 		 const struct rte_flow_item pattern[],
1143 		 const struct rte_flow_action actions[],
1144 		 struct rte_flow_error *error)
1145 {
1146 	struct mlx4_priv *priv = dev->data->dev_private;
1147 	struct rte_flow *flow;
1148 	int err;
1149 
1150 	err = mlx4_flow_prepare(priv, attr, pattern, actions, error, &flow);
1151 	if (err)
1152 		return NULL;
1153 	err = mlx4_flow_toggle(priv, flow, priv->started, error);
1154 	if (!err) {
1155 		struct rte_flow *curr = LIST_FIRST(&priv->flows);
1156 
1157 		/* New rules are inserted after internal ones. */
1158 		if (!curr || !curr->internal) {
1159 			LIST_INSERT_HEAD(&priv->flows, flow, next);
1160 		} else {
1161 			while (LIST_NEXT(curr, next) &&
1162 			       LIST_NEXT(curr, next)->internal)
1163 				curr = LIST_NEXT(curr, next);
1164 			LIST_INSERT_AFTER(curr, flow, next);
1165 		}
1166 		return flow;
1167 	}
1168 	if (flow->rss)
1169 		mlx4_rss_put(flow->rss);
1170 	rte_flow_error_set(error, -err, RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1171 			   error->message);
1172 	rte_free(flow);
1173 	return NULL;
1174 }
1175 
1176 /**
1177  * Configure isolated mode.
1178  *
1179  * @see rte_flow_isolate()
1180  * @see rte_flow_ops
1181  */
1182 static int
1183 mlx4_flow_isolate(struct rte_eth_dev *dev,
1184 		  int enable,
1185 		  struct rte_flow_error *error)
1186 {
1187 	struct mlx4_priv *priv = dev->data->dev_private;
1188 
1189 	if (!!enable == !!priv->isolated)
1190 		return 0;
1191 	priv->isolated = !!enable;
1192 	if (mlx4_flow_sync(priv, error)) {
1193 		priv->isolated = !enable;
1194 		return -rte_errno;
1195 	}
1196 	return 0;
1197 }
1198 
1199 /**
1200  * Destroy a flow rule.
1201  *
1202  * @see rte_flow_destroy()
1203  * @see rte_flow_ops
1204  */
1205 static int
1206 mlx4_flow_destroy(struct rte_eth_dev *dev,
1207 		  struct rte_flow *flow,
1208 		  struct rte_flow_error *error)
1209 {
1210 	struct mlx4_priv *priv = dev->data->dev_private;
1211 	int err = mlx4_flow_toggle(priv, flow, 0, error);
1212 
1213 	if (err)
1214 		return err;
1215 	LIST_REMOVE(flow, next);
1216 	if (flow->rss)
1217 		mlx4_rss_put(flow->rss);
1218 	rte_free(flow);
1219 	return 0;
1220 }
1221 
1222 /**
1223  * Destroy user-configured flow rules.
1224  *
1225  * This function skips internal flows rules.
1226  *
1227  * @see rte_flow_flush()
1228  * @see rte_flow_ops
1229  */
1230 static int
1231 mlx4_flow_flush(struct rte_eth_dev *dev,
1232 		struct rte_flow_error *error)
1233 {
1234 	struct mlx4_priv *priv = dev->data->dev_private;
1235 	struct rte_flow *flow = LIST_FIRST(&priv->flows);
1236 
1237 	while (flow) {
1238 		struct rte_flow *next = LIST_NEXT(flow, next);
1239 
1240 		if (!flow->internal)
1241 			mlx4_flow_destroy(dev, flow, error);
1242 		flow = next;
1243 	}
1244 	return 0;
1245 }
1246 
1247 /**
1248  * Helper function to determine the next configured VLAN filter.
1249  *
1250  * @param priv
1251  *   Pointer to private structure.
1252  * @param vlan
1253  *   VLAN ID to use as a starting point.
1254  *
1255  * @return
1256  *   Next configured VLAN ID or a high value (>= 4096) if there is none.
1257  */
1258 static uint16_t
1259 mlx4_flow_internal_next_vlan(struct mlx4_priv *priv, uint16_t vlan)
1260 {
1261 	while (vlan < 4096) {
1262 		if (ETH_DEV(priv)->data->vlan_filter_conf.ids[vlan / 64] &
1263 		    (UINT64_C(1) << (vlan % 64)))
1264 			return vlan;
1265 		++vlan;
1266 	}
1267 	return vlan;
1268 }
1269 
1270 /**
1271  * Generate internal flow rules.
1272  *
1273  * Various flow rules are created depending on the mode the device is in:
1274  *
1275  * 1. Promiscuous:
1276  *       port MAC + broadcast + catch-all (VLAN filtering is ignored).
1277  * 2. All multicast:
1278  *       port MAC/VLAN + broadcast + catch-all multicast.
1279  * 3. Otherwise:
1280  *       port MAC/VLAN + broadcast MAC/VLAN.
1281  *
1282  * About MAC flow rules:
1283  *
1284  * - MAC flow rules are generated from @p dev->data->mac_addrs
1285  *   (@p priv->mac array).
1286  * - An additional flow rule for Ethernet broadcasts is also generated.
1287  * - All these are per-VLAN if @p RTE_ETH_RX_OFFLOAD_VLAN_FILTER
1288  *   is enabled and VLAN filters are configured.
1289  *
1290  * @param priv
1291  *   Pointer to private structure.
1292  * @param[out] error
1293  *   Perform verbose error reporting if not NULL.
1294  *
1295  * @return
1296  *   0 on success, a negative errno value otherwise and rte_errno is set.
1297  */
1298 static int
1299 mlx4_flow_internal(struct mlx4_priv *priv, struct rte_flow_error *error)
1300 {
1301 	struct rte_flow_attr attr = {
1302 		.priority = MLX4_FLOW_PRIORITY_LAST,
1303 		.ingress = 1,
1304 	};
1305 	struct rte_flow_item_eth eth_spec;
1306 	const struct rte_flow_item_eth eth_mask = {
1307 		.hdr.dst_addr.addr_bytes = { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff },
1308 	};
1309 	const struct rte_flow_item_eth eth_allmulti = {
1310 		.hdr.dst_addr.addr_bytes = { 0x01, 0x00, 0x00, 0x00, 0x00, 0x00 },
1311 	};
1312 	struct rte_flow_item_vlan vlan_spec;
1313 	const struct rte_flow_item_vlan vlan_mask = {
1314 		.hdr.vlan_tci = RTE_BE16(0x0fff),
1315 	};
1316 	struct rte_flow_item pattern[] = {
1317 		{
1318 			.type = MLX4_FLOW_ITEM_TYPE_INTERNAL,
1319 		},
1320 		{
1321 			.type = RTE_FLOW_ITEM_TYPE_ETH,
1322 			.spec = &eth_spec,
1323 			.mask = &eth_mask,
1324 		},
1325 		{
1326 			/* Replaced with VLAN if filtering is enabled. */
1327 			.type = RTE_FLOW_ITEM_TYPE_END,
1328 		},
1329 		{
1330 			.type = RTE_FLOW_ITEM_TYPE_END,
1331 		},
1332 	};
1333 	/*
1334 	 * Round number of queues down to their previous power of 2 to
1335 	 * comply with RSS context limitations. Extra queues silently do not
1336 	 * get RSS by default.
1337 	 */
1338 	uint32_t queues =
1339 		rte_align32pow2(ETH_DEV(priv)->data->nb_rx_queues + 1) >> 1;
1340 	uint16_t queue[queues];
1341 	struct rte_flow_action_rss action_rss = {
1342 		.func = RTE_ETH_HASH_FUNCTION_DEFAULT,
1343 		.level = 0,
1344 		.types = 0,
1345 		.key_len = MLX4_RSS_HASH_KEY_SIZE,
1346 		.queue_num = queues,
1347 		.key = mlx4_rss_hash_key_default,
1348 		.queue = queue,
1349 	};
1350 	struct rte_flow_action actions[] = {
1351 		{
1352 			.type = RTE_FLOW_ACTION_TYPE_RSS,
1353 			.conf = &action_rss,
1354 		},
1355 		{
1356 			.type = RTE_FLOW_ACTION_TYPE_END,
1357 		},
1358 	};
1359 	struct rte_ether_addr *rule_mac = &eth_spec.hdr.dst_addr;
1360 	rte_be16_t *rule_vlan =
1361 		(ETH_DEV(priv)->data->dev_conf.rxmode.offloads &
1362 		 RTE_ETH_RX_OFFLOAD_VLAN_FILTER) &&
1363 		!ETH_DEV(priv)->data->promiscuous ?
1364 		&vlan_spec.hdr.vlan_tci :
1365 		NULL;
1366 	uint16_t vlan = 0;
1367 	struct rte_flow *flow;
1368 	unsigned int i;
1369 	int err = 0;
1370 
1371 	/* Nothing to be done if there are no Rx queues. */
1372 	if (!queues)
1373 		goto error;
1374 	/* Prepare default RSS configuration. */
1375 	for (i = 0; i != queues; ++i)
1376 		queue[i] = i;
1377 	/*
1378 	 * Set up VLAN item if filtering is enabled and at least one VLAN
1379 	 * filter is configured.
1380 	 */
1381 	if (rule_vlan) {
1382 		vlan = mlx4_flow_internal_next_vlan(priv, 0);
1383 		if (vlan < 4096) {
1384 			pattern[2] = (struct rte_flow_item){
1385 				.type = RTE_FLOW_ITEM_TYPE_VLAN,
1386 				.spec = &vlan_spec,
1387 				.mask = &vlan_mask,
1388 			};
1389 next_vlan:
1390 			*rule_vlan = rte_cpu_to_be_16(vlan);
1391 		} else {
1392 			rule_vlan = NULL;
1393 		}
1394 	}
1395 	for (i = 0; i != RTE_DIM(priv->mac) + 1; ++i) {
1396 		const struct rte_ether_addr *mac;
1397 
1398 		/* Broadcasts are handled by an extra iteration. */
1399 		if (i < RTE_DIM(priv->mac))
1400 			mac = &priv->mac[i];
1401 		else
1402 			mac = &eth_mask.hdr.dst_addr;
1403 		if (rte_is_zero_ether_addr(mac))
1404 			continue;
1405 		/* Check if MAC flow rule is already present. */
1406 		for (flow = LIST_FIRST(&priv->flows);
1407 		     flow && flow->internal;
1408 		     flow = LIST_NEXT(flow, next)) {
1409 			const struct ibv_flow_spec_eth *eth =
1410 				(const void *)((uintptr_t)flow->ibv_attr +
1411 					       sizeof(*flow->ibv_attr));
1412 			unsigned int j;
1413 
1414 			if (!flow->mac)
1415 				continue;
1416 			MLX4_ASSERT(flow->ibv_attr->type ==
1417 				    IBV_FLOW_ATTR_NORMAL);
1418 			MLX4_ASSERT(flow->ibv_attr->num_of_specs == 1);
1419 			MLX4_ASSERT(eth->type == IBV_FLOW_SPEC_ETH);
1420 			MLX4_ASSERT(flow->rss);
1421 			if (rule_vlan &&
1422 			    (eth->val.vlan_tag != *rule_vlan ||
1423 			     eth->mask.vlan_tag != RTE_BE16(0x0fff)))
1424 				continue;
1425 			if (!rule_vlan && eth->mask.vlan_tag)
1426 				continue;
1427 			for (j = 0; j != sizeof(mac->addr_bytes); ++j)
1428 				if (eth->val.dst_mac[j] != mac->addr_bytes[j] ||
1429 				    eth->mask.dst_mac[j] != UINT8_C(0xff) ||
1430 				    eth->val.src_mac[j] != UINT8_C(0x00) ||
1431 				    eth->mask.src_mac[j] != UINT8_C(0x00))
1432 					break;
1433 			if (j != sizeof(mac->addr_bytes))
1434 				continue;
1435 			if (flow->rss->queues != queues ||
1436 			    memcmp(flow->rss->queue_id, action_rss.queue,
1437 				   queues * sizeof(flow->rss->queue_id[0])))
1438 				continue;
1439 			break;
1440 		}
1441 		if (!flow || !flow->internal) {
1442 			/* Not found, create a new flow rule. */
1443 			memcpy(rule_mac, mac, sizeof(*mac));
1444 			flow = mlx4_flow_create(ETH_DEV(priv), &attr, pattern,
1445 						actions, error);
1446 			if (!flow) {
1447 				err = -rte_errno;
1448 				goto error;
1449 			}
1450 		}
1451 		flow->select = 1;
1452 		flow->mac = 1;
1453 	}
1454 	if (rule_vlan) {
1455 		vlan = mlx4_flow_internal_next_vlan(priv, vlan + 1);
1456 		if (vlan < 4096)
1457 			goto next_vlan;
1458 	}
1459 	/* Take care of promiscuous and all multicast flow rules. */
1460 	if (ETH_DEV(priv)->data->promiscuous ||
1461 	    ETH_DEV(priv)->data->all_multicast) {
1462 		for (flow = LIST_FIRST(&priv->flows);
1463 		     flow && flow->internal;
1464 		     flow = LIST_NEXT(flow, next)) {
1465 			if (ETH_DEV(priv)->data->promiscuous) {
1466 				if (flow->promisc)
1467 					break;
1468 			} else {
1469 				MLX4_ASSERT(ETH_DEV(priv)->data->all_multicast);
1470 				if (flow->allmulti)
1471 					break;
1472 			}
1473 		}
1474 		if (flow && flow->internal) {
1475 			MLX4_ASSERT(flow->rss);
1476 			if (flow->rss->queues != queues ||
1477 			    memcmp(flow->rss->queue_id, action_rss.queue,
1478 				   queues * sizeof(flow->rss->queue_id[0])))
1479 				flow = NULL;
1480 		}
1481 		if (!flow || !flow->internal) {
1482 			/* Not found, create a new flow rule. */
1483 			if (ETH_DEV(priv)->data->promiscuous) {
1484 				pattern[1].spec = NULL;
1485 				pattern[1].mask = NULL;
1486 			} else {
1487 				MLX4_ASSERT(ETH_DEV(priv)->data->all_multicast);
1488 				pattern[1].spec = &eth_allmulti;
1489 				pattern[1].mask = &eth_allmulti;
1490 			}
1491 			pattern[2] = pattern[3];
1492 			flow = mlx4_flow_create(ETH_DEV(priv), &attr, pattern,
1493 						actions, error);
1494 			if (!flow) {
1495 				err = -rte_errno;
1496 				goto error;
1497 			}
1498 		}
1499 		MLX4_ASSERT(flow->promisc || flow->allmulti);
1500 		flow->select = 1;
1501 	}
1502 error:
1503 	/* Clear selection and clean up stale internal flow rules. */
1504 	flow = LIST_FIRST(&priv->flows);
1505 	while (flow && flow->internal) {
1506 		struct rte_flow *next = LIST_NEXT(flow, next);
1507 
1508 		if (!flow->select)
1509 			claim_zero(mlx4_flow_destroy(ETH_DEV(priv), flow,
1510 						     error));
1511 		else
1512 			flow->select = 0;
1513 		flow = next;
1514 	}
1515 	return err;
1516 }
1517 
1518 /**
1519  * Synchronize flow rules.
1520  *
1521  * This function synchronizes flow rules with the state of the device by
1522  * taking into account isolated mode and whether target queues are
1523  * configured.
1524  *
1525  * @param priv
1526  *   Pointer to private structure.
1527  * @param[out] error
1528  *   Perform verbose error reporting if not NULL.
1529  *
1530  * @return
1531  *   0 on success, a negative errno value otherwise and rte_errno is set.
1532  */
1533 int
1534 mlx4_flow_sync(struct mlx4_priv *priv, struct rte_flow_error *error)
1535 {
1536 	struct rte_flow *flow;
1537 	int ret;
1538 
1539 	/* Internal flow rules are guaranteed to come first in the list. */
1540 	if (priv->isolated) {
1541 		/*
1542 		 * Get rid of them in isolated mode, stop at the first
1543 		 * non-internal rule found.
1544 		 */
1545 		for (flow = LIST_FIRST(&priv->flows);
1546 		     flow && flow->internal;
1547 		     flow = LIST_FIRST(&priv->flows))
1548 			claim_zero(mlx4_flow_destroy(ETH_DEV(priv), flow,
1549 						     error));
1550 	} else {
1551 		/* Refresh internal rules. */
1552 		ret = mlx4_flow_internal(priv, error);
1553 		if (ret)
1554 			return ret;
1555 	}
1556 	/* Toggle the remaining flow rules . */
1557 	LIST_FOREACH(flow, &priv->flows, next) {
1558 		ret = mlx4_flow_toggle(priv, flow, priv->started, error);
1559 		if (ret)
1560 			return ret;
1561 	}
1562 	if (!priv->started)
1563 		MLX4_ASSERT(!priv->drop);
1564 	return 0;
1565 }
1566 
1567 /**
1568  * Clean up all flow rules.
1569  *
1570  * Unlike mlx4_flow_flush(), this function takes care of all remaining flow
1571  * rules regardless of whether they are internal or user-configured.
1572  *
1573  * @param priv
1574  *   Pointer to private structure.
1575  */
1576 void
1577 mlx4_flow_clean(struct mlx4_priv *priv)
1578 {
1579 	struct rte_flow *flow;
1580 
1581 	while ((flow = LIST_FIRST(&priv->flows)))
1582 		mlx4_flow_destroy(ETH_DEV(priv), flow, NULL);
1583 	MLX4_ASSERT(LIST_EMPTY(&priv->rss));
1584 }
1585 
1586 static const struct rte_flow_ops mlx4_flow_ops = {
1587 	.validate = mlx4_flow_validate,
1588 	.create = mlx4_flow_create,
1589 	.destroy = mlx4_flow_destroy,
1590 	.flush = mlx4_flow_flush,
1591 	.isolate = mlx4_flow_isolate,
1592 };
1593 
1594 /**
1595  * Get rte_flow callbacks.
1596  *
1597  * @param dev
1598  *   Pointer to Ethernet device structure.
1599  * @param ops
1600  *   Pointer to operation-specific structure.
1601  *
1602  * @return 0
1603  */
1604 int
1605 mlx4_flow_ops_get(struct rte_eth_dev *dev __rte_unused,
1606 		  const struct rte_flow_ops **ops)
1607 {
1608 	*ops = &mlx4_flow_ops;
1609 	return 0;
1610 }
1611