mlx5: support Rx VLAN stripping
[dpdk.git] / drivers / net / mlx5 / mlx5_rxtx.c
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright 2015 6WIND S.A.
5  *   Copyright 2015 Mellanox.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
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
16  *       distribution.
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.
20  *
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.
32  */
33
34 #include <assert.h>
35 #include <stdint.h>
36 #include <string.h>
37 #include <stdlib.h>
38
39 /* Verbs header. */
40 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
41 #ifdef PEDANTIC
42 #pragma GCC diagnostic ignored "-pedantic"
43 #endif
44 #include <infiniband/verbs.h>
45 #ifdef PEDANTIC
46 #pragma GCC diagnostic error "-pedantic"
47 #endif
48
49 /* DPDK headers don't like -pedantic. */
50 #ifdef PEDANTIC
51 #pragma GCC diagnostic ignored "-pedantic"
52 #endif
53 #include <rte_mbuf.h>
54 #include <rte_mempool.h>
55 #include <rte_prefetch.h>
56 #include <rte_common.h>
57 #include <rte_branch_prediction.h>
58 #ifdef PEDANTIC
59 #pragma GCC diagnostic error "-pedantic"
60 #endif
61
62 #include "mlx5.h"
63 #include "mlx5_utils.h"
64 #include "mlx5_rxtx.h"
65 #include "mlx5_autoconf.h"
66 #include "mlx5_defs.h"
67
68 /**
69  * Manage TX completions.
70  *
71  * When sending a burst, mlx5_tx_burst() posts several WRs.
72  * To improve performance, a completion event is only required once every
73  * MLX5_PMD_TX_PER_COMP_REQ sends. Doing so discards completion information
74  * for other WRs, but this information would not be used anyway.
75  *
76  * @param txq
77  *   Pointer to TX queue structure.
78  *
79  * @return
80  *   0 on success, -1 on failure.
81  */
82 static int
83 txq_complete(struct txq *txq)
84 {
85         unsigned int elts_comp = txq->elts_comp;
86         unsigned int elts_tail = txq->elts_tail;
87         const unsigned int elts_n = txq->elts_n;
88         int wcs_n;
89
90         if (unlikely(elts_comp == 0))
91                 return 0;
92 #ifdef DEBUG_SEND
93         DEBUG("%p: processing %u work requests completions",
94               (void *)txq, elts_comp);
95 #endif
96         wcs_n = txq->if_cq->poll_cnt(txq->cq, elts_comp);
97         if (unlikely(wcs_n == 0))
98                 return 0;
99         if (unlikely(wcs_n < 0)) {
100                 DEBUG("%p: ibv_poll_cq() failed (wcs_n=%d)",
101                       (void *)txq, wcs_n);
102                 return -1;
103         }
104         elts_comp -= wcs_n;
105         assert(elts_comp <= txq->elts_comp);
106         /*
107          * Assume WC status is successful as nothing can be done about it
108          * anyway.
109          */
110         elts_tail += wcs_n * txq->elts_comp_cd_init;
111         if (elts_tail >= elts_n)
112                 elts_tail -= elts_n;
113         txq->elts_tail = elts_tail;
114         txq->elts_comp = elts_comp;
115         return 0;
116 }
117
118 /**
119  * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which
120  * the cloned mbuf is allocated is returned instead.
121  *
122  * @param buf
123  *   Pointer to mbuf.
124  *
125  * @return
126  *   Memory pool where data is located for given mbuf.
127  */
128 static struct rte_mempool *
129 txq_mb2mp(struct rte_mbuf *buf)
130 {
131         if (unlikely(RTE_MBUF_INDIRECT(buf)))
132                 return rte_mbuf_from_indirect(buf)->pool;
133         return buf->pool;
134 }
135
136 /**
137  * Get Memory Region (MR) <-> Memory Pool (MP) association from txq->mp2mr[].
138  * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
139  * remove an entry first.
140  *
141  * @param txq
142  *   Pointer to TX queue structure.
143  * @param[in] mp
144  *   Memory Pool for which a Memory Region lkey must be returned.
145  *
146  * @return
147  *   mr->lkey on success, (uint32_t)-1 on failure.
148  */
149 static uint32_t
150 txq_mp2mr(struct txq *txq, const struct rte_mempool *mp)
151 {
152         unsigned int i;
153         struct ibv_mr *mr;
154
155         for (i = 0; (i != RTE_DIM(txq->mp2mr)); ++i) {
156                 if (unlikely(txq->mp2mr[i].mp == NULL)) {
157                         /* Unknown MP, add a new MR for it. */
158                         break;
159                 }
160                 if (txq->mp2mr[i].mp == mp) {
161                         assert(txq->mp2mr[i].lkey != (uint32_t)-1);
162                         assert(txq->mp2mr[i].mr->lkey == txq->mp2mr[i].lkey);
163                         return txq->mp2mr[i].lkey;
164                 }
165         }
166         /* Add a new entry, register MR first. */
167         DEBUG("%p: discovered new memory pool \"%s\" (%p)",
168               (void *)txq, mp->name, (const void *)mp);
169         mr = ibv_reg_mr(txq->priv->pd,
170                         (void *)mp->elt_va_start,
171                         (mp->elt_va_end - mp->elt_va_start),
172                         (IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE));
173         if (unlikely(mr == NULL)) {
174                 DEBUG("%p: unable to configure MR, ibv_reg_mr() failed.",
175                       (void *)txq);
176                 return (uint32_t)-1;
177         }
178         if (unlikely(i == RTE_DIM(txq->mp2mr))) {
179                 /* Table is full, remove oldest entry. */
180                 DEBUG("%p: MR <-> MP table full, dropping oldest entry.",
181                       (void *)txq);
182                 --i;
183                 claim_zero(ibv_dereg_mr(txq->mp2mr[0].mr));
184                 memmove(&txq->mp2mr[0], &txq->mp2mr[1],
185                         (sizeof(txq->mp2mr) - sizeof(txq->mp2mr[0])));
186         }
187         /* Store the new entry. */
188         txq->mp2mr[i].mp = mp;
189         txq->mp2mr[i].mr = mr;
190         txq->mp2mr[i].lkey = mr->lkey;
191         DEBUG("%p: new MR lkey for MP \"%s\" (%p): 0x%08" PRIu32,
192               (void *)txq, mp->name, (const void *)mp, txq->mp2mr[i].lkey);
193         return txq->mp2mr[i].lkey;
194 }
195
196 struct txq_mp2mr_mbuf_check_data {
197         const struct rte_mempool *mp;
198         int ret;
199 };
200
201 /**
202  * Callback function for rte_mempool_obj_iter() to check whether a given
203  * mempool object looks like a mbuf.
204  *
205  * @param[in, out] arg
206  *   Context data (struct txq_mp2mr_mbuf_check_data). Contains mempool pointer
207  *   and return value.
208  * @param[in] start
209  *   Object start address.
210  * @param[in] end
211  *   Object end address.
212  * @param index
213  *   Unused.
214  *
215  * @return
216  *   Nonzero value when object is not a mbuf.
217  */
218 static void
219 txq_mp2mr_mbuf_check(void *arg, void *start, void *end,
220                      uint32_t index __rte_unused)
221 {
222         struct txq_mp2mr_mbuf_check_data *data = arg;
223         struct rte_mbuf *buf =
224                 (void *)((uintptr_t)start + data->mp->header_size);
225
226         (void)index;
227         /* Check whether mbuf structure fits element size and whether mempool
228          * pointer is valid. */
229         if (((uintptr_t)end >= (uintptr_t)(buf + 1)) &&
230             (buf->pool == data->mp))
231                 data->ret = 0;
232         else
233                 data->ret = -1;
234 }
235
236 /**
237  * Iterator function for rte_mempool_walk() to register existing mempools and
238  * fill the MP to MR cache of a TX queue.
239  *
240  * @param[in] mp
241  *   Memory Pool to register.
242  * @param *arg
243  *   Pointer to TX queue structure.
244  */
245 void
246 txq_mp2mr_iter(const struct rte_mempool *mp, void *arg)
247 {
248         struct txq *txq = arg;
249         struct txq_mp2mr_mbuf_check_data data = {
250                 .mp = mp,
251                 .ret = -1,
252         };
253
254         /* Discard empty mempools. */
255         if (mp->size == 0)
256                 return;
257         /* Register mempool only if the first element looks like a mbuf. */
258         rte_mempool_obj_iter((void *)mp->elt_va_start,
259                              1,
260                              mp->header_size + mp->elt_size + mp->trailer_size,
261                              1,
262                              mp->elt_pa,
263                              mp->pg_num,
264                              mp->pg_shift,
265                              txq_mp2mr_mbuf_check,
266                              &data);
267         if (data.ret)
268                 return;
269         txq_mp2mr(txq, mp);
270 }
271
272 #if MLX5_PMD_SGE_WR_N > 1
273
274 /**
275  * Copy scattered mbuf contents to a single linear buffer.
276  *
277  * @param[out] linear
278  *   Linear output buffer.
279  * @param[in] buf
280  *   Scattered input buffer.
281  *
282  * @return
283  *   Number of bytes copied to the output buffer or 0 if not large enough.
284  */
285 static unsigned int
286 linearize_mbuf(linear_t *linear, struct rte_mbuf *buf)
287 {
288         unsigned int size = 0;
289         unsigned int offset;
290
291         do {
292                 unsigned int len = DATA_LEN(buf);
293
294                 offset = size;
295                 size += len;
296                 if (unlikely(size > sizeof(*linear)))
297                         return 0;
298                 memcpy(&(*linear)[offset],
299                        rte_pktmbuf_mtod(buf, uint8_t *),
300                        len);
301                 buf = NEXT(buf);
302         } while (buf != NULL);
303         return size;
304 }
305
306 /**
307  * Handle scattered buffers for mlx5_tx_burst().
308  *
309  * @param txq
310  *   TX queue structure.
311  * @param segs
312  *   Number of segments in buf.
313  * @param elt
314  *   TX queue element to fill.
315  * @param[in] buf
316  *   Buffer to process.
317  * @param elts_head
318  *   Index of the linear buffer to use if necessary (normally txq->elts_head).
319  * @param[out] sges
320  *   Array filled with SGEs on success.
321  *
322  * @return
323  *   A structure containing the processed packet size in bytes and the
324  *   number of SGEs. Both fields are set to (unsigned int)-1 in case of
325  *   failure.
326  */
327 static struct tx_burst_sg_ret {
328         unsigned int length;
329         unsigned int num;
330 }
331 tx_burst_sg(struct txq *txq, unsigned int segs, struct txq_elt *elt,
332             struct rte_mbuf *buf, unsigned int elts_head,
333             struct ibv_sge (*sges)[MLX5_PMD_SGE_WR_N])
334 {
335         unsigned int sent_size = 0;
336         unsigned int j;
337         int linearize = 0;
338
339         /* When there are too many segments, extra segments are
340          * linearized in the last SGE. */
341         if (unlikely(segs > RTE_DIM(*sges))) {
342                 segs = (RTE_DIM(*sges) - 1);
343                 linearize = 1;
344         }
345         /* Update element. */
346         elt->buf = buf;
347         /* Register segments as SGEs. */
348         for (j = 0; (j != segs); ++j) {
349                 struct ibv_sge *sge = &(*sges)[j];
350                 uint32_t lkey;
351
352                 /* Retrieve Memory Region key for this memory pool. */
353                 lkey = txq_mp2mr(txq, txq_mb2mp(buf));
354                 if (unlikely(lkey == (uint32_t)-1)) {
355                         /* MR does not exist. */
356                         DEBUG("%p: unable to get MP <-> MR association",
357                               (void *)txq);
358                         /* Clean up TX element. */
359                         elt->buf = NULL;
360                         goto stop;
361                 }
362                 /* Update SGE. */
363                 sge->addr = rte_pktmbuf_mtod(buf, uintptr_t);
364                 if (txq->priv->vf)
365                         rte_prefetch0((volatile void *)
366                                       (uintptr_t)sge->addr);
367                 sge->length = DATA_LEN(buf);
368                 sge->lkey = lkey;
369                 sent_size += sge->length;
370                 buf = NEXT(buf);
371         }
372         /* If buf is not NULL here and is not going to be linearized,
373          * nb_segs is not valid. */
374         assert(j == segs);
375         assert((buf == NULL) || (linearize));
376         /* Linearize extra segments. */
377         if (linearize) {
378                 struct ibv_sge *sge = &(*sges)[segs];
379                 linear_t *linear = &(*txq->elts_linear)[elts_head];
380                 unsigned int size = linearize_mbuf(linear, buf);
381
382                 assert(segs == (RTE_DIM(*sges) - 1));
383                 if (size == 0) {
384                         /* Invalid packet. */
385                         DEBUG("%p: packet too large to be linearized.",
386                               (void *)txq);
387                         /* Clean up TX element. */
388                         elt->buf = NULL;
389                         goto stop;
390                 }
391                 /* If MLX5_PMD_SGE_WR_N is 1, free mbuf immediately. */
392                 if (RTE_DIM(*sges) == 1) {
393                         do {
394                                 struct rte_mbuf *next = NEXT(buf);
395
396                                 rte_pktmbuf_free_seg(buf);
397                                 buf = next;
398                         } while (buf != NULL);
399                         elt->buf = NULL;
400                 }
401                 /* Update SGE. */
402                 sge->addr = (uintptr_t)&(*linear)[0];
403                 sge->length = size;
404                 sge->lkey = txq->mr_linear->lkey;
405                 sent_size += size;
406                 /* Include last segment. */
407                 segs++;
408         }
409         return (struct tx_burst_sg_ret){
410                 .length = sent_size,
411                 .num = segs,
412         };
413 stop:
414         return (struct tx_burst_sg_ret){
415                 .length = -1,
416                 .num = -1,
417         };
418 }
419
420 #endif /* MLX5_PMD_SGE_WR_N > 1 */
421
422 /**
423  * DPDK callback for TX.
424  *
425  * @param dpdk_txq
426  *   Generic pointer to TX queue structure.
427  * @param[in] pkts
428  *   Packets to transmit.
429  * @param pkts_n
430  *   Number of packets in array.
431  *
432  * @return
433  *   Number of packets successfully transmitted (<= pkts_n).
434  */
435 uint16_t
436 mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
437 {
438         struct txq *txq = (struct txq *)dpdk_txq;
439         unsigned int elts_head = txq->elts_head;
440         const unsigned int elts_n = txq->elts_n;
441         unsigned int elts_comp_cd = txq->elts_comp_cd;
442         unsigned int elts_comp = 0;
443         unsigned int i;
444         unsigned int max;
445         int err;
446
447         assert(elts_comp_cd != 0);
448         txq_complete(txq);
449         max = (elts_n - (elts_head - txq->elts_tail));
450         if (max > elts_n)
451                 max -= elts_n;
452         assert(max >= 1);
453         assert(max <= elts_n);
454         /* Always leave one free entry in the ring. */
455         --max;
456         if (max == 0)
457                 return 0;
458         if (max > pkts_n)
459                 max = pkts_n;
460         for (i = 0; (i != max); ++i) {
461                 struct rte_mbuf *buf = pkts[i];
462                 unsigned int elts_head_next =
463                         (((elts_head + 1) == elts_n) ? 0 : elts_head + 1);
464                 struct txq_elt *elt_next = &(*txq->elts)[elts_head_next];
465                 struct txq_elt *elt = &(*txq->elts)[elts_head];
466                 unsigned int segs = NB_SEGS(buf);
467 #ifdef MLX5_PMD_SOFT_COUNTERS
468                 unsigned int sent_size = 0;
469 #endif
470                 uint32_t send_flags = 0;
471
472                 /* Clean up old buffer. */
473                 if (likely(elt->buf != NULL)) {
474                         struct rte_mbuf *tmp = elt->buf;
475
476                         /* Faster than rte_pktmbuf_free(). */
477                         do {
478                                 struct rte_mbuf *next = NEXT(tmp);
479
480                                 rte_pktmbuf_free_seg(tmp);
481                                 tmp = next;
482                         } while (tmp != NULL);
483                 }
484                 /* Request TX completion. */
485                 if (unlikely(--elts_comp_cd == 0)) {
486                         elts_comp_cd = txq->elts_comp_cd_init;
487                         ++elts_comp;
488                         send_flags |= IBV_EXP_QP_BURST_SIGNALED;
489                 }
490                 /* Should we enable HW CKSUM offload */
491                 if (buf->ol_flags &
492                     (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
493                         send_flags |= IBV_EXP_QP_BURST_IP_CSUM;
494                         /* HW does not support checksum offloads at arbitrary
495                          * offsets but automatically recognizes the packet
496                          * type. For inner L3/L4 checksums, only VXLAN (UDP)
497                          * tunnels are currently supported. */
498                         if (RTE_ETH_IS_TUNNEL_PKT(buf->packet_type))
499                                 send_flags |= IBV_EXP_QP_BURST_TUNNEL;
500                 }
501                 if (likely(segs == 1)) {
502                         uintptr_t addr;
503                         uint32_t length;
504                         uint32_t lkey;
505
506                         /* Retrieve buffer information. */
507                         addr = rte_pktmbuf_mtod(buf, uintptr_t);
508                         length = DATA_LEN(buf);
509                         /* Retrieve Memory Region key for this memory pool. */
510                         lkey = txq_mp2mr(txq, txq_mb2mp(buf));
511                         if (unlikely(lkey == (uint32_t)-1)) {
512                                 /* MR does not exist. */
513                                 DEBUG("%p: unable to get MP <-> MR"
514                                       " association", (void *)txq);
515                                 /* Clean up TX element. */
516                                 elt->buf = NULL;
517                                 goto stop;
518                         }
519                         /* Update element. */
520                         elt->buf = buf;
521                         if (txq->priv->vf)
522                                 rte_prefetch0((volatile void *)
523                                               (uintptr_t)addr);
524                         RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
525                         /* Put packet into send queue. */
526 #if MLX5_PMD_MAX_INLINE > 0
527                         if (length <= txq->max_inline)
528                                 err = txq->if_qp->send_pending_inline
529                                         (txq->qp,
530                                          (void *)addr,
531                                          length,
532                                          send_flags);
533                         else
534 #endif
535                                 err = txq->if_qp->send_pending
536                                         (txq->qp,
537                                          addr,
538                                          length,
539                                          lkey,
540                                          send_flags);
541                         if (unlikely(err))
542                                 goto stop;
543 #ifdef MLX5_PMD_SOFT_COUNTERS
544                         sent_size += length;
545 #endif
546                 } else {
547 #if MLX5_PMD_SGE_WR_N > 1
548                         struct ibv_sge sges[MLX5_PMD_SGE_WR_N];
549                         struct tx_burst_sg_ret ret;
550
551                         ret = tx_burst_sg(txq, segs, elt, buf, elts_head,
552                                           &sges);
553                         if (ret.length == (unsigned int)-1)
554                                 goto stop;
555                         RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
556                         /* Put SG list into send queue. */
557                         err = txq->if_qp->send_pending_sg_list
558                                 (txq->qp,
559                                  sges,
560                                  ret.num,
561                                  send_flags);
562                         if (unlikely(err))
563                                 goto stop;
564 #ifdef MLX5_PMD_SOFT_COUNTERS
565                         sent_size += ret.length;
566 #endif
567 #else /* MLX5_PMD_SGE_WR_N > 1 */
568                         DEBUG("%p: TX scattered buffers support not"
569                               " compiled in", (void *)txq);
570                         goto stop;
571 #endif /* MLX5_PMD_SGE_WR_N > 1 */
572                 }
573                 elts_head = elts_head_next;
574 #ifdef MLX5_PMD_SOFT_COUNTERS
575                 /* Increment sent bytes counter. */
576                 txq->stats.obytes += sent_size;
577 #endif
578         }
579 stop:
580         /* Take a shortcut if nothing must be sent. */
581         if (unlikely(i == 0))
582                 return 0;
583 #ifdef MLX5_PMD_SOFT_COUNTERS
584         /* Increment sent packets counter. */
585         txq->stats.opackets += i;
586 #endif
587         /* Ring QP doorbell. */
588         err = txq->if_qp->send_flush(txq->qp);
589         if (unlikely(err)) {
590                 /* A nonzero value is not supposed to be returned.
591                  * Nothing can be done about it. */
592                 DEBUG("%p: send_flush() failed with error %d",
593                       (void *)txq, err);
594         }
595         txq->elts_head = elts_head;
596         txq->elts_comp += elts_comp;
597         txq->elts_comp_cd = elts_comp_cd;
598         return i;
599 }
600
601 /**
602  * Translate RX completion flags to packet type.
603  *
604  * @param flags
605  *   RX completion flags returned by poll_length_flags().
606  *
607  * @return
608  *   Packet type for struct rte_mbuf.
609  */
610 static inline uint32_t
611 rxq_cq_to_pkt_type(uint32_t flags)
612 {
613         uint32_t pkt_type;
614
615         if (flags & IBV_EXP_CQ_RX_TUNNEL_PACKET)
616                 pkt_type =
617                         TRANSPOSE(flags,
618                                   IBV_EXP_CQ_RX_OUTER_IPV4_PACKET,
619                                   RTE_PTYPE_L3_IPV4) |
620                         TRANSPOSE(flags,
621                                   IBV_EXP_CQ_RX_OUTER_IPV6_PACKET,
622                                   RTE_PTYPE_L3_IPV6) |
623                         TRANSPOSE(flags,
624                                   IBV_EXP_CQ_RX_IPV4_PACKET,
625                                   RTE_PTYPE_INNER_L3_IPV4) |
626                         TRANSPOSE(flags,
627                                   IBV_EXP_CQ_RX_IPV6_PACKET,
628                                   RTE_PTYPE_INNER_L3_IPV6);
629         else
630                 pkt_type =
631                         TRANSPOSE(flags,
632                                   IBV_EXP_CQ_RX_IPV4_PACKET,
633                                   RTE_PTYPE_L3_IPV4) |
634                         TRANSPOSE(flags,
635                                   IBV_EXP_CQ_RX_IPV6_PACKET,
636                                   RTE_PTYPE_L3_IPV6);
637         return pkt_type;
638 }
639
640 /**
641  * Translate RX completion flags to offload flags.
642  *
643  * @param[in] rxq
644  *   Pointer to RX queue structure.
645  * @param flags
646  *   RX completion flags returned by poll_length_flags().
647  *
648  * @return
649  *   Offload flags (ol_flags) for struct rte_mbuf.
650  */
651 static inline uint32_t
652 rxq_cq_to_ol_flags(const struct rxq *rxq, uint32_t flags)
653 {
654         uint32_t ol_flags = 0;
655
656         if (rxq->csum)
657                 ol_flags |=
658                         TRANSPOSE(~flags,
659                                   IBV_EXP_CQ_RX_IP_CSUM_OK,
660                                   PKT_RX_IP_CKSUM_BAD) |
661                         TRANSPOSE(~flags,
662                                   IBV_EXP_CQ_RX_TCP_UDP_CSUM_OK,
663                                   PKT_RX_L4_CKSUM_BAD);
664         /*
665          * PKT_RX_IP_CKSUM_BAD and PKT_RX_L4_CKSUM_BAD are used in place
666          * of PKT_RX_EIP_CKSUM_BAD because the latter is not functional
667          * (its value is 0).
668          */
669         if ((flags & IBV_EXP_CQ_RX_TUNNEL_PACKET) && (rxq->csum_l2tun))
670                 ol_flags |=
671                         TRANSPOSE(~flags,
672                                   IBV_EXP_CQ_RX_OUTER_IP_CSUM_OK,
673                                   PKT_RX_IP_CKSUM_BAD) |
674                         TRANSPOSE(~flags,
675                                   IBV_EXP_CQ_RX_OUTER_TCP_UDP_CSUM_OK,
676                                   PKT_RX_L4_CKSUM_BAD);
677         return ol_flags;
678 }
679
680 /**
681  * DPDK callback for RX with scattered packets support.
682  *
683  * @param dpdk_rxq
684  *   Generic pointer to RX queue structure.
685  * @param[out] pkts
686  *   Array to store received packets.
687  * @param pkts_n
688  *   Maximum number of packets in array.
689  *
690  * @return
691  *   Number of packets successfully received (<= pkts_n).
692  */
693 uint16_t
694 mlx5_rx_burst_sp(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
695 {
696         struct rxq *rxq = (struct rxq *)dpdk_rxq;
697         struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
698         const unsigned int elts_n = rxq->elts_n;
699         unsigned int elts_head = rxq->elts_head;
700         unsigned int i;
701         unsigned int pkts_ret = 0;
702         int ret;
703
704         if (unlikely(!rxq->sp))
705                 return mlx5_rx_burst(dpdk_rxq, pkts, pkts_n);
706         if (unlikely(elts == NULL)) /* See RTE_DEV_CMD_SET_MTU. */
707                 return 0;
708         for (i = 0; (i != pkts_n); ++i) {
709                 struct rxq_elt_sp *elt = &(*elts)[elts_head];
710                 unsigned int len;
711                 unsigned int pkt_buf_len;
712                 struct rte_mbuf *pkt_buf = NULL; /* Buffer returned in pkts. */
713                 struct rte_mbuf **pkt_buf_next = &pkt_buf;
714                 unsigned int seg_headroom = RTE_PKTMBUF_HEADROOM;
715                 unsigned int j = 0;
716                 uint32_t flags;
717                 uint16_t vlan_tci;
718
719                 /* Sanity checks. */
720                 assert(elts_head < rxq->elts_n);
721                 assert(rxq->elts_head < rxq->elts_n);
722 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
723                 ret = rxq->if_cq->poll_length_flags_cvlan(rxq->cq, NULL, NULL,
724                                                           &flags, &vlan_tci);
725 #else /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
726                 ret = rxq->if_cq->poll_length_flags(rxq->cq, NULL, NULL,
727                                                     &flags);
728                 (void)vlan_tci;
729 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
730                 if (unlikely(ret < 0)) {
731                         struct ibv_wc wc;
732                         int wcs_n;
733
734                         DEBUG("rxq=%p, poll_length() failed (ret=%d)",
735                               (void *)rxq, ret);
736                         /* ibv_poll_cq() must be used in case of failure. */
737                         wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
738                         if (unlikely(wcs_n == 0))
739                                 break;
740                         if (unlikely(wcs_n < 0)) {
741                                 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
742                                       (void *)rxq, wcs_n);
743                                 break;
744                         }
745                         assert(wcs_n == 1);
746                         if (unlikely(wc.status != IBV_WC_SUCCESS)) {
747                                 /* Whatever, just repost the offending WR. */
748                                 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
749                                       " completion status (%d): %s",
750                                       (void *)rxq, wc.wr_id, wc.status,
751                                       ibv_wc_status_str(wc.status));
752 #ifdef MLX5_PMD_SOFT_COUNTERS
753                                 /* Increment dropped packets counter. */
754                                 ++rxq->stats.idropped;
755 #endif
756                                 goto repost;
757                         }
758                         ret = wc.byte_len;
759                 }
760                 if (ret == 0)
761                         break;
762                 len = ret;
763                 pkt_buf_len = len;
764                 /*
765                  * Replace spent segments with new ones, concatenate and
766                  * return them as pkt_buf.
767                  */
768                 while (1) {
769                         struct ibv_sge *sge = &elt->sges[j];
770                         struct rte_mbuf *seg = elt->bufs[j];
771                         struct rte_mbuf *rep;
772                         unsigned int seg_tailroom;
773
774                         assert(seg != NULL);
775                         /*
776                          * Fetch initial bytes of packet descriptor into a
777                          * cacheline while allocating rep.
778                          */
779                         rte_prefetch0(seg);
780                         rep = __rte_mbuf_raw_alloc(rxq->mp);
781                         if (unlikely(rep == NULL)) {
782                                 /*
783                                  * Unable to allocate a replacement mbuf,
784                                  * repost WR.
785                                  */
786                                 DEBUG("rxq=%p: can't allocate a new mbuf",
787                                       (void *)rxq);
788                                 if (pkt_buf != NULL) {
789                                         *pkt_buf_next = NULL;
790                                         rte_pktmbuf_free(pkt_buf);
791                                 }
792                                 /* Increment out of memory counters. */
793                                 ++rxq->stats.rx_nombuf;
794                                 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
795                                 goto repost;
796                         }
797 #ifndef NDEBUG
798                         /* Poison user-modifiable fields in rep. */
799                         NEXT(rep) = (void *)((uintptr_t)-1);
800                         SET_DATA_OFF(rep, 0xdead);
801                         DATA_LEN(rep) = 0xd00d;
802                         PKT_LEN(rep) = 0xdeadd00d;
803                         NB_SEGS(rep) = 0x2a;
804                         PORT(rep) = 0x2a;
805                         rep->ol_flags = -1;
806 #endif
807                         assert(rep->buf_len == seg->buf_len);
808                         assert(rep->buf_len == rxq->mb_len);
809                         /* Reconfigure sge to use rep instead of seg. */
810                         assert(sge->lkey == rxq->mr->lkey);
811                         sge->addr = ((uintptr_t)rep->buf_addr + seg_headroom);
812                         elt->bufs[j] = rep;
813                         ++j;
814                         /* Update pkt_buf if it's the first segment, or link
815                          * seg to the previous one and update pkt_buf_next. */
816                         *pkt_buf_next = seg;
817                         pkt_buf_next = &NEXT(seg);
818                         /* Update seg information. */
819                         seg_tailroom = (seg->buf_len - seg_headroom);
820                         assert(sge->length == seg_tailroom);
821                         SET_DATA_OFF(seg, seg_headroom);
822                         if (likely(len <= seg_tailroom)) {
823                                 /* Last segment. */
824                                 DATA_LEN(seg) = len;
825                                 PKT_LEN(seg) = len;
826                                 /* Sanity check. */
827                                 assert(rte_pktmbuf_headroom(seg) ==
828                                        seg_headroom);
829                                 assert(rte_pktmbuf_tailroom(seg) ==
830                                        (seg_tailroom - len));
831                                 break;
832                         }
833                         DATA_LEN(seg) = seg_tailroom;
834                         PKT_LEN(seg) = seg_tailroom;
835                         /* Sanity check. */
836                         assert(rte_pktmbuf_headroom(seg) == seg_headroom);
837                         assert(rte_pktmbuf_tailroom(seg) == 0);
838                         /* Fix len and clear headroom for next segments. */
839                         len -= seg_tailroom;
840                         seg_headroom = 0;
841                 }
842                 /* Update head and tail segments. */
843                 *pkt_buf_next = NULL;
844                 assert(pkt_buf != NULL);
845                 assert(j != 0);
846                 NB_SEGS(pkt_buf) = j;
847                 PORT(pkt_buf) = rxq->port_id;
848                 PKT_LEN(pkt_buf) = pkt_buf_len;
849                 pkt_buf->packet_type = rxq_cq_to_pkt_type(flags);
850                 pkt_buf->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
851 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
852                 if (flags & IBV_EXP_CQ_RX_CVLAN_STRIPPED_V1) {
853                         pkt_buf->ol_flags |= PKT_RX_VLAN_PKT;
854                         pkt_buf->vlan_tci = vlan_tci;
855                 }
856 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
857
858                 /* Return packet. */
859                 *(pkts++) = pkt_buf;
860                 ++pkts_ret;
861 #ifdef MLX5_PMD_SOFT_COUNTERS
862                 /* Increment bytes counter. */
863                 rxq->stats.ibytes += pkt_buf_len;
864 #endif
865 repost:
866                 ret = rxq->if_wq->recv_sg_list(rxq->wq,
867                                                elt->sges,
868                                                RTE_DIM(elt->sges));
869                 if (unlikely(ret)) {
870                         /* Inability to repost WRs is fatal. */
871                         DEBUG("%p: recv_sg_list(): failed (ret=%d)",
872                               (void *)rxq->priv,
873                               ret);
874                         abort();
875                 }
876                 if (++elts_head >= elts_n)
877                         elts_head = 0;
878                 continue;
879         }
880         if (unlikely(i == 0))
881                 return 0;
882         rxq->elts_head = elts_head;
883 #ifdef MLX5_PMD_SOFT_COUNTERS
884         /* Increment packets counter. */
885         rxq->stats.ipackets += pkts_ret;
886 #endif
887         return pkts_ret;
888 }
889
890 /**
891  * DPDK callback for RX.
892  *
893  * The following function is the same as mlx5_rx_burst_sp(), except it doesn't
894  * manage scattered packets. Improves performance when MRU is lower than the
895  * size of the first segment.
896  *
897  * @param dpdk_rxq
898  *   Generic pointer to RX queue structure.
899  * @param[out] pkts
900  *   Array to store received packets.
901  * @param pkts_n
902  *   Maximum number of packets in array.
903  *
904  * @return
905  *   Number of packets successfully received (<= pkts_n).
906  */
907 uint16_t
908 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
909 {
910         struct rxq *rxq = (struct rxq *)dpdk_rxq;
911         struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
912         const unsigned int elts_n = rxq->elts_n;
913         unsigned int elts_head = rxq->elts_head;
914         struct ibv_sge sges[pkts_n];
915         unsigned int i;
916         unsigned int pkts_ret = 0;
917         int ret;
918
919         if (unlikely(rxq->sp))
920                 return mlx5_rx_burst_sp(dpdk_rxq, pkts, pkts_n);
921         for (i = 0; (i != pkts_n); ++i) {
922                 struct rxq_elt *elt = &(*elts)[elts_head];
923                 unsigned int len;
924                 struct rte_mbuf *seg = elt->buf;
925                 struct rte_mbuf *rep;
926                 uint32_t flags;
927                 uint16_t vlan_tci;
928
929                 /* Sanity checks. */
930                 assert(seg != NULL);
931                 assert(elts_head < rxq->elts_n);
932                 assert(rxq->elts_head < rxq->elts_n);
933                 /*
934                  * Fetch initial bytes of packet descriptor into a
935                  * cacheline while allocating rep.
936                  */
937                 rte_prefetch0(seg);
938                 rte_prefetch0(&seg->cacheline1);
939 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
940                 ret = rxq->if_cq->poll_length_flags_cvlan(rxq->cq, NULL, NULL,
941                                                           &flags, &vlan_tci);
942 #else /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
943                 ret = rxq->if_cq->poll_length_flags(rxq->cq, NULL, NULL,
944                                                     &flags);
945                 (void)vlan_tci;
946 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
947                 if (unlikely(ret < 0)) {
948                         struct ibv_wc wc;
949                         int wcs_n;
950
951                         DEBUG("rxq=%p, poll_length() failed (ret=%d)",
952                               (void *)rxq, ret);
953                         /* ibv_poll_cq() must be used in case of failure. */
954                         wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
955                         if (unlikely(wcs_n == 0))
956                                 break;
957                         if (unlikely(wcs_n < 0)) {
958                                 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
959                                       (void *)rxq, wcs_n);
960                                 break;
961                         }
962                         assert(wcs_n == 1);
963                         if (unlikely(wc.status != IBV_WC_SUCCESS)) {
964                                 /* Whatever, just repost the offending WR. */
965                                 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
966                                       " completion status (%d): %s",
967                                       (void *)rxq, wc.wr_id, wc.status,
968                                       ibv_wc_status_str(wc.status));
969 #ifdef MLX5_PMD_SOFT_COUNTERS
970                                 /* Increment dropped packets counter. */
971                                 ++rxq->stats.idropped;
972 #endif
973                                 /* Add SGE to array for repost. */
974                                 sges[i] = elt->sge;
975                                 goto repost;
976                         }
977                         ret = wc.byte_len;
978                 }
979                 if (ret == 0)
980                         break;
981                 len = ret;
982                 rep = __rte_mbuf_raw_alloc(rxq->mp);
983                 if (unlikely(rep == NULL)) {
984                         /*
985                          * Unable to allocate a replacement mbuf,
986                          * repost WR.
987                          */
988                         DEBUG("rxq=%p: can't allocate a new mbuf",
989                               (void *)rxq);
990                         /* Increment out of memory counters. */
991                         ++rxq->stats.rx_nombuf;
992                         ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
993                         goto repost;
994                 }
995
996                 /* Reconfigure sge to use rep instead of seg. */
997                 elt->sge.addr = (uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM;
998                 assert(elt->sge.lkey == rxq->mr->lkey);
999                 elt->buf = rep;
1000
1001                 /* Add SGE to array for repost. */
1002                 sges[i] = elt->sge;
1003
1004                 /* Update seg information. */
1005                 SET_DATA_OFF(seg, RTE_PKTMBUF_HEADROOM);
1006                 NB_SEGS(seg) = 1;
1007                 PORT(seg) = rxq->port_id;
1008                 NEXT(seg) = NULL;
1009                 PKT_LEN(seg) = len;
1010                 DATA_LEN(seg) = len;
1011                 seg->packet_type = rxq_cq_to_pkt_type(flags);
1012                 seg->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
1013 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
1014                 if (flags & IBV_EXP_CQ_RX_CVLAN_STRIPPED_V1) {
1015                         seg->ol_flags |= PKT_RX_VLAN_PKT;
1016                         seg->vlan_tci = vlan_tci;
1017                 }
1018 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
1019
1020                 /* Return packet. */
1021                 *(pkts++) = seg;
1022                 ++pkts_ret;
1023 #ifdef MLX5_PMD_SOFT_COUNTERS
1024                 /* Increment bytes counter. */
1025                 rxq->stats.ibytes += len;
1026 #endif
1027 repost:
1028                 if (++elts_head >= elts_n)
1029                         elts_head = 0;
1030                 continue;
1031         }
1032         if (unlikely(i == 0))
1033                 return 0;
1034         /* Repost WRs. */
1035 #ifdef DEBUG_RECV
1036         DEBUG("%p: reposting %u WRs", (void *)rxq, i);
1037 #endif
1038         ret = rxq->if_wq->recv_burst(rxq->wq, sges, i);
1039         if (unlikely(ret)) {
1040                 /* Inability to repost WRs is fatal. */
1041                 DEBUG("%p: recv_burst(): failed (ret=%d)",
1042                       (void *)rxq->priv,
1043                       ret);
1044                 abort();
1045         }
1046         rxq->elts_head = elts_head;
1047 #ifdef MLX5_PMD_SOFT_COUNTERS
1048         /* Increment packets counter. */
1049         rxq->stats.ipackets += pkts_ret;
1050 #endif
1051         return pkts_ret;
1052 }
1053
1054 /**
1055  * Dummy DPDK callback for TX.
1056  *
1057  * This function is used to temporarily replace the real callback during
1058  * unsafe control operations on the queue, or in case of error.
1059  *
1060  * @param dpdk_txq
1061  *   Generic pointer to TX queue structure.
1062  * @param[in] pkts
1063  *   Packets to transmit.
1064  * @param pkts_n
1065  *   Number of packets in array.
1066  *
1067  * @return
1068  *   Number of packets successfully transmitted (<= pkts_n).
1069  */
1070 uint16_t
1071 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1072 {
1073         (void)dpdk_txq;
1074         (void)pkts;
1075         (void)pkts_n;
1076         return 0;
1077 }
1078
1079 /**
1080  * Dummy DPDK callback for RX.
1081  *
1082  * This function is used to temporarily replace the real callback during
1083  * unsafe control operations on the queue, or in case of error.
1084  *
1085  * @param dpdk_rxq
1086  *   Generic pointer to RX queue structure.
1087  * @param[out] pkts
1088  *   Array to store received packets.
1089  * @param pkts_n
1090  *   Maximum number of packets in array.
1091  *
1092  * @return
1093  *   Number of packets successfully received (<= pkts_n).
1094  */
1095 uint16_t
1096 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1097 {
1098         (void)dpdk_rxq;
1099         (void)pkts;
1100         (void)pkts_n;
1101         return 0;
1102 }