4 * Copyright 2015 6WIND S.A.
5 * Copyright 2015 Mellanox.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * * Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * * Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * * Neither the name of 6WIND S.A. nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
34 #ifndef RTE_PMD_MLX5_RXTX_H_
35 #define RTE_PMD_MLX5_RXTX_H_
41 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
43 #pragma GCC diagnostic ignored "-Wpedantic"
45 #include <infiniband/verbs.h>
46 #include <infiniband/mlx5_hw.h>
48 #pragma GCC diagnostic error "-Wpedantic"
52 #include <rte_mempool.h>
53 #include <rte_common.h>
54 #include <rte_hexdump.h>
56 #include "mlx5_utils.h"
58 #include "mlx5_autoconf.h"
59 #include "mlx5_defs.h"
62 struct mlx5_rxq_stats {
63 unsigned int idx; /**< Mapping index. */
64 #ifdef MLX5_PMD_SOFT_COUNTERS
65 uint64_t ipackets; /**< Total of successfully received packets. */
66 uint64_t ibytes; /**< Total of successfully received bytes. */
68 uint64_t idropped; /**< Total of packets dropped when RX ring full. */
69 uint64_t rx_nombuf; /**< Total of RX mbuf allocation failures. */
72 struct mlx5_txq_stats {
73 unsigned int idx; /**< Mapping index. */
74 #ifdef MLX5_PMD_SOFT_COUNTERS
75 uint64_t opackets; /**< Total of successfully sent packets. */
76 uint64_t obytes; /**< Total of successfully sent bytes. */
78 uint64_t oerrors; /**< Total number of failed transmitted packets. */
81 /* Flow director queue structure. */
83 struct ibv_qp *qp; /* Associated RX QP. */
84 struct ibv_exp_rwq_ind_table *ind_table; /* Indirection table. */
85 struct ibv_exp_wq *wq; /* Work queue. */
86 struct ibv_cq *cq; /* Completion queue. */
91 /* Compressed CQE context. */
93 uint16_t ai; /* Array index. */
94 uint16_t ca; /* Current array index. */
95 uint16_t na; /* Next array index. */
96 uint16_t cq_ci; /* The next CQE. */
97 uint32_t cqe_cnt; /* Number of CQEs. */
100 /* RX queue descriptor. */
102 unsigned int csum:1; /* Enable checksum offloading. */
103 unsigned int csum_l2tun:1; /* Same for L2 tunnels. */
104 unsigned int vlan_strip:1; /* Enable VLAN stripping. */
105 unsigned int crc_present:1; /* CRC must be subtracted. */
106 unsigned int sges_n:2; /* Log 2 of SGEs (max buffers per packet). */
107 unsigned int cqe_n:4; /* Log 2 of CQ elements. */
108 unsigned int elts_n:4; /* Log 2 of Mbufs. */
109 unsigned int port_id:8;
110 unsigned int rss_hash:1; /* RSS hash result is enabled. */
111 unsigned int mark:1; /* Marked flow available on the queue. */
112 unsigned int pending_err:1; /* CQE error needs to be handled. */
113 unsigned int :7; /* Remaining bits. */
114 volatile uint32_t *rq_db;
115 volatile uint32_t *cq_db;
119 volatile struct mlx5_wqe_data_seg(*wqes)[];
120 volatile struct mlx5_cqe(*cqes)[];
121 struct rxq_zip zip; /* Compressed context. */
122 struct rte_mbuf *(*elts)[];
123 struct rte_mempool *mp;
124 struct mlx5_rxq_stats stats;
125 uint64_t mbuf_initializer; /* Default rearm_data for vectorized Rx. */
126 struct rte_mbuf fake_mbuf; /* elts padding for vectorized Rx. */
127 } __rte_cache_aligned;
129 /* RX queue control descriptor. */
131 struct priv *priv; /* Back pointer to private data. */
132 struct ibv_cq *cq; /* Completion Queue. */
133 struct ibv_exp_wq *wq; /* Work Queue. */
134 struct fdir_queue *fdir_queue; /* Flow director queue. */
135 struct ibv_mr *mr; /* Memory Region (for mp). */
136 struct ibv_comp_channel *channel;
137 unsigned int socket; /* CPU socket ID for allocations. */
138 struct rxq rxq; /* Data path structure. */
141 /* Hash RX queue types. */
152 /* Flow structure with Ethernet specification. It is packed to prevent padding
153 * between attr and spec as this layout is expected by libibverbs. */
154 struct flow_attr_spec_eth {
155 struct ibv_exp_flow_attr attr;
156 struct ibv_exp_flow_spec_eth spec;
157 } __attribute__((packed));
159 /* Define a struct flow_attr_spec_eth object as an array of at least
160 * "size" bytes. Room after the first index is normally used to store
161 * extra flow specifications. */
162 #define FLOW_ATTR_SPEC_ETH(name, size) \
163 struct flow_attr_spec_eth name \
164 [((size) / sizeof(struct flow_attr_spec_eth)) + \
165 !!((size) % sizeof(struct flow_attr_spec_eth))]
167 /* Initialization data for hash RX queue. */
168 struct hash_rxq_init {
169 uint64_t hash_fields; /* Fields that participate in the hash. */
170 uint64_t dpdk_rss_hf; /* Matching DPDK RSS hash fields. */
171 unsigned int flow_priority; /* Flow priority to use. */
174 enum ibv_exp_flow_spec_type type;
177 struct ibv_exp_flow_spec_tcp_udp tcp_udp;
178 struct ibv_exp_flow_spec_ipv4 ipv4;
179 struct ibv_exp_flow_spec_ipv6 ipv6;
180 struct ibv_exp_flow_spec_eth eth;
181 } flow_spec; /* Flow specification template. */
182 const struct hash_rxq_init *underlayer; /* Pointer to underlayer. */
185 /* Initialization data for indirection table. */
186 struct ind_table_init {
187 unsigned int max_size; /* Maximum number of WQs. */
188 /* Hash RX queues using this table. */
189 unsigned int hash_types;
190 unsigned int hash_types_n;
193 /* Initialization data for special flows. */
194 struct special_flow_init {
195 uint8_t dst_mac_val[6];
196 uint8_t dst_mac_mask[6];
197 unsigned int hash_types;
198 unsigned int per_vlan:1;
201 enum hash_rxq_flow_type {
202 HASH_RXQ_FLOW_TYPE_PROMISC,
203 HASH_RXQ_FLOW_TYPE_ALLMULTI,
204 HASH_RXQ_FLOW_TYPE_BROADCAST,
205 HASH_RXQ_FLOW_TYPE_IPV6MULTI,
206 HASH_RXQ_FLOW_TYPE_MAC,
210 static inline const char *
211 hash_rxq_flow_type_str(enum hash_rxq_flow_type flow_type)
214 case HASH_RXQ_FLOW_TYPE_PROMISC:
215 return "promiscuous";
216 case HASH_RXQ_FLOW_TYPE_ALLMULTI:
217 return "allmulticast";
218 case HASH_RXQ_FLOW_TYPE_BROADCAST:
220 case HASH_RXQ_FLOW_TYPE_IPV6MULTI:
221 return "IPv6 multicast";
222 case HASH_RXQ_FLOW_TYPE_MAC:
230 struct priv *priv; /* Back pointer to private data. */
231 struct ibv_qp *qp; /* Hash RX QP. */
232 enum hash_rxq_type type; /* Hash RX queue type. */
233 /* MAC flow steering rules, one per VLAN ID. */
234 struct ibv_exp_flow *mac_flow
235 [MLX5_MAX_MAC_ADDRESSES][MLX5_MAX_VLAN_IDS];
236 struct ibv_exp_flow *special_flow
237 [MLX5_MAX_SPECIAL_FLOWS][MLX5_MAX_VLAN_IDS];
240 /* TX queue descriptor. */
243 uint16_t elts_head; /* Current counter in (*elts)[]. */
244 uint16_t elts_tail; /* Counter of first element awaiting completion. */
245 uint16_t elts_comp; /* Counter since last completion request. */
246 uint16_t mpw_comp; /* WQ index since last completion request. */
247 uint16_t cq_ci; /* Consumer index for completion queue. */
248 uint16_t cq_pi; /* Producer index for completion queue. */
249 uint16_t wqe_ci; /* Consumer index for work queue. */
250 uint16_t wqe_pi; /* Producer index for work queue. */
251 uint16_t elts_n:4; /* (*elts)[] length (in log2). */
252 uint16_t cqe_n:4; /* Number of CQ elements (in log2). */
253 uint16_t wqe_n:4; /* Number of of WQ elements (in log2). */
254 uint16_t inline_en:1; /* When set inline is enabled. */
255 uint16_t tso_en:1; /* When set hardware TSO is enabled. */
256 uint16_t tunnel_en:1;
257 /* When set TX offload for tunneled packets are supported. */
258 uint16_t mpw_hdr_dseg:1; /* Enable DSEGs in the title WQEBB. */
259 uint16_t max_inline; /* Multiple of RTE_CACHE_LINE_SIZE to inline. */
260 uint16_t inline_max_packet_sz; /* Max packet size for inlining. */
261 uint32_t qp_num_8s; /* QP number shifted by 8. */
262 uint32_t flags; /* Flags for Tx Queue. */
263 volatile struct mlx5_cqe (*cqes)[]; /* Completion queue. */
264 volatile void *wqes; /* Work queue (use volatile to write into). */
265 volatile uint32_t *qp_db; /* Work queue doorbell. */
266 volatile uint32_t *cq_db; /* Completion queue doorbell. */
267 volatile void *bf_reg; /* Blueflame register. */
269 uintptr_t start; /* Start address of MR */
270 uintptr_t end; /* End address of MR */
271 struct ibv_mr *mr; /* Memory Region (for mp). */
272 uint32_t lkey; /* rte_cpu_to_be_32(mr->lkey) */
273 } mp2mr[MLX5_PMD_TX_MP_CACHE]; /* MP to MR translation table. */
274 uint16_t mr_cache_idx; /* Index of last hit entry. */
275 struct rte_mbuf *(*elts)[]; /* TX elements. */
276 struct mlx5_txq_stats stats; /* TX queue counters. */
277 } __rte_cache_aligned;
279 /* TX queue control descriptor. */
281 struct priv *priv; /* Back pointer to private data. */
282 struct ibv_cq *cq; /* Completion Queue. */
283 struct ibv_qp *qp; /* Queue Pair. */
284 unsigned int socket; /* CPU socket ID for allocations. */
285 struct txq txq; /* Data path structure. */
290 extern const struct hash_rxq_init hash_rxq_init[];
291 extern const unsigned int hash_rxq_init_n;
293 extern uint8_t rss_hash_default_key[];
294 extern const size_t rss_hash_default_key_len;
296 size_t priv_flow_attr(struct priv *, struct ibv_exp_flow_attr *,
297 size_t, enum hash_rxq_type);
298 int priv_create_hash_rxqs(struct priv *);
299 void priv_destroy_hash_rxqs(struct priv *);
300 int priv_allow_flow_type(struct priv *, enum hash_rxq_flow_type);
301 int priv_rehash_flows(struct priv *);
302 void rxq_cleanup(struct rxq_ctrl *);
303 int mlx5_rx_queue_setup(struct rte_eth_dev *, uint16_t, uint16_t, unsigned int,
304 const struct rte_eth_rxconf *, struct rte_mempool *);
305 void mlx5_rx_queue_release(void *);
306 int priv_rx_intr_vec_enable(struct priv *priv);
307 void priv_rx_intr_vec_disable(struct priv *priv);
308 #ifdef HAVE_UPDATE_CQ_CI
309 int mlx5_rx_intr_enable(struct rte_eth_dev *dev, uint16_t rx_queue_id);
310 int mlx5_rx_intr_disable(struct rte_eth_dev *dev, uint16_t rx_queue_id);
311 #endif /* HAVE_UPDATE_CQ_CI */
315 void txq_cleanup(struct txq_ctrl *);
316 int txq_ctrl_setup(struct rte_eth_dev *, struct txq_ctrl *, uint16_t,
317 unsigned int, const struct rte_eth_txconf *);
318 int mlx5_tx_queue_setup(struct rte_eth_dev *, uint16_t, uint16_t, unsigned int,
319 const struct rte_eth_txconf *);
320 void mlx5_tx_queue_release(void *);
324 extern uint32_t mlx5_ptype_table[];
326 void mlx5_set_ptype_table(void);
327 uint16_t mlx5_tx_burst(void *, struct rte_mbuf **, uint16_t);
328 uint16_t mlx5_tx_burst_mpw(void *, struct rte_mbuf **, uint16_t);
329 uint16_t mlx5_tx_burst_mpw_inline(void *, struct rte_mbuf **, uint16_t);
330 uint16_t mlx5_tx_burst_empw(void *, struct rte_mbuf **, uint16_t);
331 uint16_t mlx5_rx_burst(void *, struct rte_mbuf **, uint16_t);
332 uint16_t removed_tx_burst(void *, struct rte_mbuf **, uint16_t);
333 uint16_t removed_rx_burst(void *, struct rte_mbuf **, uint16_t);
334 int mlx5_rx_descriptor_status(void *, uint16_t);
335 int mlx5_tx_descriptor_status(void *, uint16_t);
337 /* Vectorized version of mlx5_rxtx.c */
338 int priv_check_raw_vec_tx_support(struct priv *);
339 int priv_check_vec_tx_support(struct priv *);
340 int rxq_check_vec_support(struct rxq *);
341 int priv_check_vec_rx_support(struct priv *);
342 uint16_t mlx5_tx_burst_raw_vec(void *, struct rte_mbuf **, uint16_t);
343 uint16_t mlx5_tx_burst_vec(void *, struct rte_mbuf **, uint16_t);
344 uint16_t mlx5_rx_burst_vec(void *, struct rte_mbuf **, uint16_t);
348 struct ibv_mr *mlx5_mp2mr(struct ibv_pd *, struct rte_mempool *);
349 void txq_mp2mr_iter(struct rte_mempool *, void *);
350 uint32_t txq_mp2mr_reg(struct txq *, struct rte_mempool *, unsigned int);
354 * Verify or set magic value in CQE.
363 check_cqe_seen(volatile struct mlx5_cqe *cqe)
365 static const uint8_t magic[] = "seen";
366 volatile uint8_t (*buf)[sizeof(cqe->rsvd0)] = &cqe->rsvd0;
370 for (i = 0; i < sizeof(magic) && i < sizeof(*buf); ++i)
371 if (!ret || (*buf)[i] != magic[i]) {
373 (*buf)[i] = magic[i];
380 * Check whether CQE is valid.
385 * Size of completion queue.
390 * 0 on success, 1 on failure.
392 static __rte_always_inline int
393 check_cqe(volatile struct mlx5_cqe *cqe,
394 unsigned int cqes_n, const uint16_t ci)
396 uint16_t idx = ci & cqes_n;
397 uint8_t op_own = cqe->op_own;
398 uint8_t op_owner = MLX5_CQE_OWNER(op_own);
399 uint8_t op_code = MLX5_CQE_OPCODE(op_own);
401 if (unlikely((op_owner != (!!(idx))) || (op_code == MLX5_CQE_INVALID)))
402 return 1; /* No CQE. */
404 if ((op_code == MLX5_CQE_RESP_ERR) ||
405 (op_code == MLX5_CQE_REQ_ERR)) {
406 volatile struct mlx5_err_cqe *err_cqe = (volatile void *)cqe;
407 uint8_t syndrome = err_cqe->syndrome;
409 if ((syndrome == MLX5_CQE_SYNDROME_LOCAL_LENGTH_ERR) ||
410 (syndrome == MLX5_CQE_SYNDROME_REMOTE_ABORTED_ERR))
412 if (!check_cqe_seen(cqe)) {
413 ERROR("unexpected CQE error %u (0x%02x)"
415 op_code, op_code, syndrome);
416 rte_hexdump(stderr, "MLX5 Error CQE:",
417 (const void *)((uintptr_t)err_cqe),
421 } else if ((op_code != MLX5_CQE_RESP_SEND) &&
422 (op_code != MLX5_CQE_REQ)) {
423 if (!check_cqe_seen(cqe)) {
424 ERROR("unexpected CQE opcode %u (0x%02x)",
426 rte_hexdump(stderr, "MLX5 CQE:",
427 (const void *)((uintptr_t)cqe),
437 * Return the address of the WQE.
440 * Pointer to TX queue structure.
442 * WQE consumer index.
447 static inline uintptr_t *
448 tx_mlx5_wqe(struct txq *txq, uint16_t ci)
450 ci &= ((1 << txq->wqe_n) - 1);
451 return (uintptr_t *)((uintptr_t)txq->wqes + ci * MLX5_WQE_SIZE);
455 * Manage TX completions.
457 * When sending a burst, mlx5_tx_burst() posts several WRs.
460 * Pointer to TX queue structure.
462 static __rte_always_inline void
463 mlx5_tx_complete(struct txq *txq)
465 const uint16_t elts_n = 1 << txq->elts_n;
466 const uint16_t elts_m = elts_n - 1;
467 const unsigned int cqe_n = 1 << txq->cqe_n;
468 const unsigned int cqe_cnt = cqe_n - 1;
469 uint16_t elts_free = txq->elts_tail;
471 uint16_t cq_ci = txq->cq_ci;
472 volatile struct mlx5_cqe *cqe = NULL;
473 volatile struct mlx5_wqe_ctrl *ctrl;
474 struct rte_mbuf *m, *free[elts_n];
475 struct rte_mempool *pool = NULL;
476 unsigned int blk_n = 0;
478 cqe = &(*txq->cqes)[cq_ci & cqe_cnt];
479 if (unlikely(check_cqe(cqe, cqe_n, cq_ci)))
482 if ((MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_RESP_ERR) ||
483 (MLX5_CQE_OPCODE(cqe->op_own) == MLX5_CQE_REQ_ERR)) {
484 if (!check_cqe_seen(cqe)) {
485 ERROR("unexpected error CQE, TX stopped");
486 rte_hexdump(stderr, "MLX5 TXQ:",
487 (const void *)((uintptr_t)txq->wqes),
495 txq->wqe_pi = rte_be_to_cpu_16(cqe->wqe_counter);
496 ctrl = (volatile struct mlx5_wqe_ctrl *)
497 tx_mlx5_wqe(txq, txq->wqe_pi);
498 elts_tail = ctrl->ctrl3;
499 assert((elts_tail & elts_m) < (1 << txq->wqe_n));
501 while (elts_free != elts_tail) {
502 m = rte_pktmbuf_prefree_seg((*txq->elts)[elts_free++ & elts_m]);
503 if (likely(m != NULL)) {
504 if (likely(m->pool == pool)) {
507 if (likely(pool != NULL))
508 rte_mempool_put_bulk(pool,
518 rte_mempool_put_bulk(pool, (void *)free, blk_n);
520 elts_free = txq->elts_tail;
522 while (elts_free != elts_tail) {
523 memset(&(*txq->elts)[elts_free & elts_m],
525 sizeof((*txq->elts)[elts_free & elts_m]));
530 txq->elts_tail = elts_tail;
531 /* Update the consumer index. */
533 *txq->cq_db = rte_cpu_to_be_32(cq_ci);
537 * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which
538 * the cloned mbuf is allocated is returned instead.
544 * Memory pool where data is located for given mbuf.
546 static struct rte_mempool *
547 mlx5_tx_mb2mp(struct rte_mbuf *buf)
549 if (unlikely(RTE_MBUF_INDIRECT(buf)))
550 return rte_mbuf_from_indirect(buf)->pool;
555 * Get Memory Region (MR) <-> rte_mbuf association from txq->mp2mr[].
556 * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
557 * remove an entry first.
560 * Pointer to TX queue structure.
562 * Memory Pool for which a Memory Region lkey must be returned.
565 * mr->lkey on success, (uint32_t)-1 on failure.
567 static __rte_always_inline uint32_t
568 mlx5_tx_mb2mr(struct txq *txq, struct rte_mbuf *mb)
570 uint16_t i = txq->mr_cache_idx;
571 uintptr_t addr = rte_pktmbuf_mtod(mb, uintptr_t);
573 assert(i < RTE_DIM(txq->mp2mr));
574 if (likely(txq->mp2mr[i].start <= addr && txq->mp2mr[i].end >= addr))
575 return txq->mp2mr[i].lkey;
576 for (i = 0; (i != RTE_DIM(txq->mp2mr)); ++i) {
577 if (unlikely(txq->mp2mr[i].mr == NULL)) {
578 /* Unknown MP, add a new MR for it. */
581 if (txq->mp2mr[i].start <= addr &&
582 txq->mp2mr[i].end >= addr) {
583 assert(txq->mp2mr[i].lkey != (uint32_t)-1);
584 assert(rte_cpu_to_be_32(txq->mp2mr[i].mr->lkey) ==
586 txq->mr_cache_idx = i;
587 return txq->mp2mr[i].lkey;
590 txq->mr_cache_idx = 0;
591 return txq_mp2mr_reg(txq, mlx5_tx_mb2mp(mb), i);
595 * Ring TX queue doorbell.
598 * Pointer to TX queue structure.
600 * Pointer to the last WQE posted in the NIC.
602 static __rte_always_inline void
603 mlx5_tx_dbrec(struct txq *txq, volatile struct mlx5_wqe *wqe)
605 uint64_t *dst = (uint64_t *)((uintptr_t)txq->bf_reg);
606 volatile uint64_t *src = ((volatile uint64_t *)wqe);
609 *txq->qp_db = rte_cpu_to_be_32(txq->wqe_ci);
610 /* Ensure ordering between DB record and BF copy. */
615 #endif /* RTE_PMD_MLX5_RXTX_H_ */