common/mlx5: introduce common library
[dpdk.git] / drivers / net / mlx5 / mlx5_rxtx_vec_altivec.h
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2017 6WIND S.A.
3  * Copyright 2017 Mellanox Technologies, Ltd
4  */
5
6 #ifndef RTE_PMD_MLX5_RXTX_VEC_ALTIVEC_H_
7 #define RTE_PMD_MLX5_RXTX_VEC_ALTIVEC_H_
8
9 #include <assert.h>
10 #include <stdint.h>
11 #include <string.h>
12 #include <stdlib.h>
13
14 #include <altivec.h>
15
16 #include <rte_mbuf.h>
17 #include <rte_mempool.h>
18 #include <rte_prefetch.h>
19
20 #include <mlx5_prm.h>
21
22 #include "mlx5_defs.h"
23 #include "mlx5.h"
24 #include "mlx5_utils.h"
25 #include "mlx5_rxtx.h"
26 #include "mlx5_rxtx_vec.h"
27 #include "mlx5_autoconf.h"
28
29 #ifndef __INTEL_COMPILER
30 #pragma GCC diagnostic ignored "-Wcast-qual"
31 #pragma GCC diagnostic ignored "-Wstrict-aliasing"
32 #endif
33
34 /**
35  * Store free buffers to RX SW ring.
36  *
37  * @param rxq
38  *   Pointer to RX queue structure.
39  * @param pkts
40  *   Pointer to array of packets to be stored.
41  * @param pkts_n
42  *   Number of packets to be stored.
43  */
44 static inline void
45 rxq_copy_mbuf_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t n)
46 {
47         const uint16_t q_mask = (1 << rxq->elts_n) - 1;
48         struct rte_mbuf **elts = &(*rxq->elts)[rxq->rq_pi & q_mask];
49         unsigned int pos;
50         uint16_t p = n & -2;
51
52         for (pos = 0; pos < p; pos += 2) {
53                 vector unsigned char mbp;
54
55                 mbp = (vector unsigned char)vec_vsx_ld(0,
56                                 (signed int const *)&elts[pos]);
57                 *(vector unsigned char *)&pkts[pos] = mbp;
58         }
59         if (n & 1)
60                 pkts[pos] = elts[pos];
61 }
62
63 /**
64  * Decompress a compressed completion and fill in mbufs in RX SW ring with data
65  * extracted from the title completion descriptor.
66  *
67  * @param rxq
68  *   Pointer to RX queue structure.
69  * @param cq
70  *   Pointer to completion array having a compressed completion at first.
71  * @param elts
72  *   Pointer to SW ring to be filled. The first mbuf has to be pre-built from
73  *   the title completion descriptor to be copied to the rest of mbufs.
74  *
75  * @return
76  *   Number of mini-CQEs successfully decompressed.
77  */
78 static inline uint16_t
79 rxq_cq_decompress_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
80                     struct rte_mbuf **elts)
81 {
82         volatile struct mlx5_mini_cqe8 *mcq = (void *)&(cq + 1)->pkt_info;
83         struct rte_mbuf *t_pkt = elts[0]; /* Title packet is pre-built. */
84         const vector unsigned char zero = (vector unsigned char){0};
85         /* Mask to shuffle from extracted mini CQE to mbuf. */
86         const vector unsigned char shuf_mask1 = (vector unsigned char){
87                         -1, -1, -1, -1,   /* skip packet_type */
88                          7,  6, -1, -1,   /* bswap16, pkt_len */
89                          7,  6,           /* bswap16, data_len */
90                         -1, -1,           /* skip vlan_tci */
91                          3,  2,  1,  0};  /* bswap32, rss */
92         const vector unsigned char shuf_mask2 = (vector unsigned char){
93                         -1, -1, -1, -1,   /* skip packet_type */
94                         15, 14, -1, -1,   /* bswap16, pkt_len */
95                         15, 14,           /* data_len, bswap16 */
96                         -1, -1,           /* skip vlan_tci */
97                         11, 10,  9,  8};  /* bswap32, rss */
98         /* Restore the compressed count. Must be 16 bits. */
99         const uint16_t mcqe_n = t_pkt->data_len +
100                 (rxq->crc_present * RTE_ETHER_CRC_LEN);
101         const vector unsigned char rearm =
102                 (vector unsigned char)vec_vsx_ld(0,
103                 (signed int const *)&t_pkt->rearm_data);
104         const vector unsigned char rxdf =
105                 (vector unsigned char)vec_vsx_ld(0,
106                 (signed int const *)&t_pkt->rx_descriptor_fields1);
107         const vector unsigned char crc_adj =
108                 (vector unsigned char)(vector unsigned short){
109                         0, 0, rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
110                         rxq->crc_present * RTE_ETHER_CRC_LEN, 0, 0, 0};
111         const vector unsigned short rxdf_sel_mask =
112                 (vector unsigned short){
113                         0xffff, 0xffff, 0, 0, 0, 0xffff, 0, 0};
114         const uint32_t flow_tag = t_pkt->hash.fdir.hi;
115         unsigned int pos;
116         unsigned int i;
117         unsigned int inv = 0;
118
119 #ifdef MLX5_PMD_SOFT_COUNTERS
120         const vector unsigned char ones = vec_splat_u8(-1);
121         uint32_t rcvd_byte = 0;
122         /* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
123         const vector unsigned char len_shuf_mask = (vector unsigned char){
124                  3,  2, 11, 10,
125                  7,  6, 15, 14,
126                 -1, -1, -1, -1,
127                 -1, -1, -1, -1};
128 #endif
129
130         /*
131          * A. load mCQEs into a 128bit register.
132          * B. store rearm data to mbuf.
133          * C. combine data from mCQEs with rx_descriptor_fields1.
134          * D. store rx_descriptor_fields1.
135          * E. store flow tag (rte_flow mark).
136          */
137         for (pos = 0; pos < mcqe_n; ) {
138                 vector unsigned char mcqe1, mcqe2;
139                 vector unsigned char rxdf1, rxdf2;
140 #ifdef MLX5_PMD_SOFT_COUNTERS
141                 const vector unsigned short mcqe_sel_mask =
142                         (vector unsigned short){0, 0, 0xffff, 0xffff,
143                         0, 0, 0xfff, 0xffff};
144                 const vector unsigned char lower_half = {
145                         0, 1, 4, 5, 8, 9, 12, 13, 16,
146                         17, 20, 21, 24, 25, 28, 29};
147                 const vector unsigned char upper_half = {
148                         2, 3, 6, 7, 10, 11, 14, 15,
149                         18, 19, 22, 23, 26, 27, 30, 31};
150                 vector unsigned short left, right;
151                 vector unsigned char byte_cnt, invalid_mask;
152                 vector unsigned long lshift;
153                 __attribute__((altivec(vector__)))
154                         __attribute__((altivec(bool__)))
155                         unsigned long long shmask;
156                 const vector unsigned long shmax = {64, 64};
157 #endif
158
159                 if (!(pos & 0x7) && pos + 8 < mcqe_n)
160                         rte_prefetch0((void *)(cq + pos + 8));
161
162                 /* A.1 load mCQEs into a 128bit register. */
163                 mcqe1 = (vector unsigned char)vec_vsx_ld(0,
164                         (signed int const *)&mcq[pos % 8]);
165                 mcqe2 = (vector unsigned char)vec_vsx_ld(0,
166                         (signed int const *)&mcq[pos % 8 + 2]);
167
168                 /* B.1 store rearm data to mbuf. */
169                 *(vector unsigned char *)
170                         &elts[pos]->rearm_data = rearm;
171                 *(vector unsigned char *)
172                         &elts[pos + 1]->rearm_data = rearm;
173
174                 /* C.1 combine data from mCQEs with rx_descriptor_fields1. */
175                 rxdf1 = vec_perm(mcqe1, zero, shuf_mask1);
176                 rxdf2 = vec_perm(mcqe1, zero, shuf_mask2);
177                 rxdf1 = (vector unsigned char)
178                         ((vector unsigned short)rxdf1 -
179                         (vector unsigned short)crc_adj);
180                 rxdf2 = (vector unsigned char)
181                         ((vector unsigned short)rxdf2 -
182                         (vector unsigned short)crc_adj);
183                 rxdf1 = (vector unsigned char)
184                         vec_sel((vector unsigned short)rxdf1,
185                         (vector unsigned short)rxdf, rxdf_sel_mask);
186                 rxdf2 = (vector unsigned char)
187                         vec_sel((vector unsigned short)rxdf2,
188                         (vector unsigned short)rxdf, rxdf_sel_mask);
189
190                 /* D.1 store rx_descriptor_fields1. */
191                 *(vector unsigned char *)
192                         &elts[pos]->rx_descriptor_fields1 = rxdf1;
193                 *(vector unsigned char *)
194                         &elts[pos + 1]->rx_descriptor_fields1 = rxdf2;
195
196                 /* B.1 store rearm data to mbuf. */
197                 *(vector unsigned char *)
198                         &elts[pos + 2]->rearm_data = rearm;
199                 *(vector unsigned char *)
200                         &elts[pos + 3]->rearm_data = rearm;
201
202                 /* C.1 combine data from mCQEs with rx_descriptor_fields1. */
203                 rxdf1 = vec_perm(mcqe2, zero, shuf_mask1);
204                 rxdf2 = vec_perm(mcqe2, zero, shuf_mask2);
205                 rxdf1 = (vector unsigned char)
206                         ((vector unsigned short)rxdf1 -
207                         (vector unsigned short)crc_adj);
208                 rxdf2 = (vector unsigned char)
209                         ((vector unsigned short)rxdf2 -
210                         (vector unsigned short)crc_adj);
211                 rxdf1 = (vector unsigned char)
212                         vec_sel((vector unsigned short)rxdf1,
213                         (vector unsigned short)rxdf, rxdf_sel_mask);
214                 rxdf2 = (vector unsigned char)
215                         vec_sel((vector unsigned short)rxdf2,
216                         (vector unsigned short)rxdf, rxdf_sel_mask);
217
218                 /* D.1 store rx_descriptor_fields1. */
219                 *(vector unsigned char *)
220                         &elts[pos + 2]->rx_descriptor_fields1 = rxdf1;
221                 *(vector unsigned char *)
222                         &elts[pos + 3]->rx_descriptor_fields1 = rxdf2;
223
224 #ifdef MLX5_PMD_SOFT_COUNTERS
225                 invalid_mask = (vector unsigned char)(vector unsigned long){
226                         (mcqe_n - pos) * sizeof(uint16_t) * 8, 0};
227
228                 lshift =
229                         vec_splat((vector unsigned long)invalid_mask, 0);
230                 shmask = vec_cmpgt(shmax, lshift);
231                 invalid_mask = (vector unsigned char)
232                         vec_sl((vector unsigned long)ones, lshift);
233                 invalid_mask = (vector unsigned char)
234                         vec_sel((vector unsigned long)shmask,
235                         (vector unsigned long)invalid_mask, shmask);
236
237                 mcqe1 = (vector unsigned char)
238                         vec_sro((vector unsigned short)mcqe1,
239                         (vector unsigned char){32}),
240                 byte_cnt = (vector unsigned char)
241                         vec_sel((vector unsigned short)mcqe1,
242                         (vector unsigned short)mcqe2, mcqe_sel_mask);
243                 byte_cnt = vec_perm(byte_cnt, zero, len_shuf_mask);
244                 byte_cnt = (vector unsigned char)
245                         vec_andc((vector unsigned long)byte_cnt,
246                         (vector unsigned long)invalid_mask);
247                 left = vec_perm((vector unsigned short)byte_cnt,
248                         (vector unsigned short)zero, lower_half);
249                 right = vec_perm((vector unsigned short)byte_cnt,
250                         (vector unsigned short)zero, upper_half);
251                 byte_cnt = (vector unsigned char)vec_add(left, right);
252                 left = vec_perm((vector unsigned short)byte_cnt,
253                         (vector unsigned short)zero, lower_half);
254                 right = vec_perm((vector unsigned short)byte_cnt,
255                         (vector unsigned short)zero, upper_half);
256                 byte_cnt = (vector unsigned char)vec_add(left, right);
257                 rcvd_byte += ((vector unsigned long)byte_cnt)[0];
258 #endif
259
260                 if (rxq->mark) {
261                         /* E.1 store flow tag (rte_flow mark). */
262                         elts[pos]->hash.fdir.hi = flow_tag;
263                         elts[pos + 1]->hash.fdir.hi = flow_tag;
264                         elts[pos + 2]->hash.fdir.hi = flow_tag;
265                         elts[pos + 3]->hash.fdir.hi = flow_tag;
266                 }
267
268                 pos += MLX5_VPMD_DESCS_PER_LOOP;
269                 /* Move to next CQE and invalidate consumed CQEs. */
270                 if (!(pos & 0x7) && pos < mcqe_n) {
271                         mcq = (void *)&(cq + pos)->pkt_info;
272                         for (i = 0; i < 8; ++i)
273                                 cq[inv++].op_own = MLX5_CQE_INVALIDATE;
274                 }
275         }
276
277         /* Invalidate the rest of CQEs. */
278         for (; inv < mcqe_n; ++inv)
279                 cq[inv].op_own = MLX5_CQE_INVALIDATE;
280
281 #ifdef MLX5_PMD_SOFT_COUNTERS
282         rxq->stats.ipackets += mcqe_n;
283         rxq->stats.ibytes += rcvd_byte;
284 #endif
285
286         rxq->cq_ci += mcqe_n;
287         return mcqe_n;
288 }
289
290 /**
291  * Calculate packet type and offload flag for mbuf and store it.
292  *
293  * @param rxq
294  *   Pointer to RX queue structure.
295  * @param cqes[4]
296  *   Array of four 16bytes completions extracted from the original completion
297  *   descriptor.
298  * @param op_err
299  *   Opcode vector having responder error status. Each field is 4B.
300  * @param pkts
301  *   Pointer to array of packets to be filled.
302  */
303 static inline void
304 rxq_cq_to_ptype_oflags_v(struct mlx5_rxq_data *rxq,
305                 vector unsigned char cqes[4], vector unsigned char op_err,
306                 struct rte_mbuf **pkts)
307 {
308         vector unsigned char pinfo0, pinfo1;
309         vector unsigned char pinfo, ptype;
310         vector unsigned char ol_flags = (vector unsigned char)
311                 (vector unsigned int){
312                         rxq->rss_hash * PKT_RX_RSS_HASH |
313                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP,
314                         rxq->rss_hash * PKT_RX_RSS_HASH |
315                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP,
316                         rxq->rss_hash * PKT_RX_RSS_HASH |
317                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP,
318                         rxq->rss_hash * PKT_RX_RSS_HASH |
319                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP};
320         vector unsigned char cv_flags;
321         const vector unsigned char zero = (vector unsigned char){0};
322         const vector unsigned char ptype_mask =
323                 (vector unsigned char)(vector unsigned int){
324                 0x0000fd06, 0x0000fd06, 0x0000fd06, 0x0000fd06};
325         const vector unsigned char ptype_ol_mask =
326                 (vector unsigned char)(vector unsigned int){
327                 0x00000106, 0x00000106, 0x00000106, 0x00000106};
328         const vector unsigned char pinfo_mask =
329                 (vector unsigned char)(vector unsigned int){
330                 0x00000003, 0x00000003, 0x00000003, 0x00000003};
331         const vector unsigned char cv_flag_sel = (vector unsigned char){
332                 0, (uint8_t)(PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED),
333                 (uint8_t)(PKT_RX_IP_CKSUM_GOOD >> 1), 0,
334                 (uint8_t)(PKT_RX_L4_CKSUM_GOOD >> 1), 0,
335                 (uint8_t)((PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1),
336                 0, 0, 0, 0, 0, 0, 0, 0, 0};
337         const vector unsigned char cv_mask =
338                 (vector unsigned char)(vector unsigned int){
339                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
340                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
341                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
342                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
343                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
344                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
345                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
346                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED};
347         const vector unsigned char mbuf_init =
348                 (vector unsigned char)vec_vsx_ld
349                         (0, (vector unsigned char *)&rxq->mbuf_initializer);
350         const vector unsigned short rearm_sel_mask =
351                 (vector unsigned short){0, 0, 0, 0, 0xffff, 0xffff, 0, 0};
352         vector unsigned char rearm0, rearm1, rearm2, rearm3;
353         uint8_t pt_idx0, pt_idx1, pt_idx2, pt_idx3;
354
355         /* Extract pkt_info field. */
356         pinfo0 = (vector unsigned char)
357                 vec_mergeh((vector unsigned int)cqes[0],
358                 (vector unsigned int)cqes[1]);
359         pinfo1 = (vector unsigned char)
360                 vec_mergeh((vector unsigned int)cqes[2],
361                 (vector unsigned int)cqes[3]);
362         pinfo = (vector unsigned char)
363                 vec_mergeh((vector unsigned long)pinfo0,
364                 (vector unsigned long)pinfo1);
365
366         /* Extract hdr_type_etc field. */
367         pinfo0 = (vector unsigned char)
368                 vec_mergel((vector unsigned int)cqes[0],
369                 (vector unsigned int)cqes[1]);
370         pinfo1 = (vector unsigned char)
371                 vec_mergel((vector unsigned int)cqes[2],
372                 (vector unsigned int)cqes[3]);
373         ptype = (vector unsigned char)
374                 vec_mergeh((vector unsigned long)pinfo0,
375                 (vector unsigned long)pinfo1);
376
377         if (rxq->mark) {
378                 const vector unsigned char pinfo_ft_mask =
379                         (vector unsigned char)(vector unsigned int){
380                         0xffffff00, 0xffffff00, 0xffffff00, 0xffffff00};
381                 const vector unsigned char fdir_flags =
382                         (vector unsigned char)(vector unsigned int){
383                         PKT_RX_FDIR, PKT_RX_FDIR,
384                         PKT_RX_FDIR, PKT_RX_FDIR};
385                 vector unsigned char fdir_id_flags =
386                         (vector unsigned char)(vector unsigned int){
387                         PKT_RX_FDIR_ID, PKT_RX_FDIR_ID,
388                         PKT_RX_FDIR_ID, PKT_RX_FDIR_ID};
389                 vector unsigned char flow_tag, invalid_mask;
390
391                 flow_tag = (vector unsigned char)
392                         vec_and((vector unsigned long)pinfo,
393                         (vector unsigned long)pinfo_ft_mask);
394
395                 /* Check if flow tag is non-zero then set PKT_RX_FDIR. */
396                 invalid_mask = (vector unsigned char)
397                         vec_cmpeq((vector unsigned int)flow_tag,
398                         (vector unsigned int)zero);
399                 ol_flags = (vector unsigned char)
400                         vec_or((vector unsigned long)ol_flags,
401                         (vector unsigned long)
402                         vec_andc((vector unsigned long)fdir_flags,
403                         (vector unsigned long)invalid_mask));
404
405                 /* Mask out invalid entries. */
406                 fdir_id_flags = (vector unsigned char)
407                         vec_andc((vector unsigned long)fdir_id_flags,
408                         (vector unsigned long)invalid_mask);
409
410                 /* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */
411                 ol_flags = (vector unsigned char)
412                         vec_or((vector unsigned long)ol_flags,
413                         (vector unsigned long)
414                         vec_andc((vector unsigned long)fdir_id_flags,
415                         (vector unsigned long)
416                         vec_cmpeq((vector unsigned int)flow_tag,
417                         (vector unsigned int)pinfo_ft_mask)));
418         }
419         /*
420          * Merge the two fields to generate the following:
421          * bit[1]     = l3_ok
422          * bit[2]     = l4_ok
423          * bit[8]     = cv
424          * bit[11:10] = l3_hdr_type
425          * bit[14:12] = l4_hdr_type
426          * bit[15]    = ip_frag
427          * bit[16]    = tunneled
428          * bit[17]    = outer_l3_type
429          */
430         ptype = (vector unsigned char)
431                 vec_and((vector unsigned long)ptype,
432                 (vector unsigned long)ptype_mask);
433         pinfo = (vector unsigned char)
434                 vec_and((vector unsigned long)pinfo,
435                 (vector unsigned long)pinfo_mask);
436         pinfo = (vector unsigned char)
437                 vec_sl((vector unsigned int)pinfo,
438                 (vector unsigned int){16, 16, 16, 16});
439
440         /* Make pinfo has merged fields for ol_flags calculation. */
441         pinfo = (vector unsigned char)
442                 vec_or((vector unsigned long)ptype,
443                 (vector unsigned long)pinfo);
444         ptype = (vector unsigned char)
445                 vec_sr((vector unsigned int)pinfo,
446                 (vector unsigned int){10, 10, 10, 10});
447         ptype = (vector unsigned char)
448                 vec_packs((vector unsigned int)ptype,
449                 (vector unsigned int)zero);
450
451         /* Errored packets will have RTE_PTYPE_ALL_MASK. */
452         op_err = (vector unsigned char)
453                 vec_sr((vector unsigned short)op_err,
454                 (vector unsigned short){8, 8, 8, 8, 8, 8, 8, 8});
455         ptype = (vector unsigned char)
456                 vec_or((vector unsigned long)ptype,
457                 (vector unsigned long)op_err);
458
459         pt_idx0 = (uint8_t)((vector unsigned char)ptype)[0];
460         pt_idx1 = (uint8_t)((vector unsigned char)ptype)[2];
461         pt_idx2 = (uint8_t)((vector unsigned char)ptype)[4];
462         pt_idx3 = (uint8_t)((vector unsigned char)ptype)[6];
463
464         pkts[0]->packet_type = mlx5_ptype_table[pt_idx0] |
465                 !!(pt_idx0 & (1 << 6)) * rxq->tunnel;
466         pkts[1]->packet_type = mlx5_ptype_table[pt_idx1] |
467                 !!(pt_idx1 & (1 << 6)) * rxq->tunnel;
468         pkts[2]->packet_type = mlx5_ptype_table[pt_idx2] |
469                 !!(pt_idx2 & (1 << 6)) * rxq->tunnel;
470         pkts[3]->packet_type = mlx5_ptype_table[pt_idx3] |
471                 !!(pt_idx3 & (1 << 6)) * rxq->tunnel;
472
473         /* Fill flags for checksum and VLAN. */
474         pinfo = (vector unsigned char)
475                 vec_and((vector unsigned long)pinfo,
476                 (vector unsigned long)ptype_ol_mask);
477         pinfo = vec_perm(cv_flag_sel, zero, pinfo);
478
479         /* Locate checksum flags at byte[2:1] and merge with VLAN flags. */
480         cv_flags = (vector unsigned char)
481                 vec_sl((vector unsigned int)pinfo,
482                 (vector unsigned int){9, 9, 9, 9});
483         cv_flags = (vector unsigned char)
484                 vec_or((vector unsigned long)pinfo,
485                 (vector unsigned long)cv_flags);
486
487         /* Move back flags to start from byte[0]. */
488         cv_flags = (vector unsigned char)
489                 vec_sr((vector unsigned int)cv_flags,
490                 (vector unsigned int){8, 8, 8, 8});
491
492         /* Mask out garbage bits. */
493         cv_flags = (vector unsigned char)
494                 vec_and((vector unsigned long)cv_flags,
495                 (vector unsigned long)cv_mask);
496
497         /* Merge to ol_flags. */
498         ol_flags = (vector unsigned char)
499                 vec_or((vector unsigned long)ol_flags,
500                 (vector unsigned long)cv_flags);
501
502         /* Merge mbuf_init and ol_flags. */
503         rearm0 = (vector unsigned char)
504                 vec_sel((vector unsigned short)mbuf_init,
505                 (vector unsigned short)
506                 vec_slo((vector unsigned short)ol_flags,
507                 (vector unsigned char){64}), rearm_sel_mask);
508         rearm1 = (vector unsigned char)
509                 vec_sel((vector unsigned short)mbuf_init,
510                 (vector unsigned short)
511                 vec_slo((vector unsigned short)ol_flags,
512                 (vector unsigned char){32}), rearm_sel_mask);
513         rearm2 = (vector unsigned char)
514                 vec_sel((vector unsigned short)mbuf_init,
515                 (vector unsigned short)ol_flags, rearm_sel_mask);
516         rearm3 = (vector unsigned char)
517                 vec_sel((vector unsigned short)mbuf_init,
518                 (vector unsigned short)
519                 vec_sro((vector unsigned short)ol_flags,
520                 (vector unsigned char){32}), rearm_sel_mask);
521
522         /* Write 8B rearm_data and 8B ol_flags. */
523         vec_vsx_st(rearm0, 0,
524                 (vector unsigned char *)&pkts[0]->rearm_data);
525         vec_vsx_st(rearm1, 0,
526                 (vector unsigned char *)&pkts[1]->rearm_data);
527         vec_vsx_st(rearm2, 0,
528                 (vector unsigned char *)&pkts[2]->rearm_data);
529         vec_vsx_st(rearm3, 0,
530                 (vector unsigned char *)&pkts[3]->rearm_data);
531 }
532
533
534 /**
535  * Receive burst of packets. An errored completion also consumes a mbuf, but the
536  * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
537  * before returning to application.
538  *
539  * @param rxq
540  *   Pointer to RX queue structure.
541  * @param[out] pkts
542  *   Array to store received packets.
543  * @param pkts_n
544  *   Maximum number of packets in array.
545  * @param[out] err
546  *   Pointer to a flag. Set non-zero value if pkts array has at least one error
547  *   packet to handle.
548  *
549  * @return
550  *   Number of packets received including errors (<= pkts_n).
551  */
552 static inline uint16_t
553 rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
554             uint64_t *err)
555 {
556         const uint16_t q_n = 1 << rxq->cqe_n;
557         const uint16_t q_mask = q_n - 1;
558         volatile struct mlx5_cqe *cq;
559         struct rte_mbuf **elts;
560         unsigned int pos;
561         uint64_t n;
562         uint16_t repl_n;
563         uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
564         uint16_t nocmp_n = 0;
565         uint16_t rcvd_pkt = 0;
566         unsigned int cq_idx = rxq->cq_ci & q_mask;
567         unsigned int elts_idx;
568         unsigned int ownership = !!(rxq->cq_ci & (q_mask + 1));
569         const vector unsigned char zero = (vector unsigned char){0};
570         const vector unsigned char ones = vec_splat_u8(-1);
571         const vector unsigned char owner_check =
572                 (vector unsigned char)(vector unsigned long){
573                 0x0100000001000000LL, 0x0100000001000000LL};
574         const vector unsigned char opcode_check =
575                 (vector unsigned char)(vector unsigned long){
576                 0xf0000000f0000000LL, 0xf0000000f0000000LL};
577         const vector unsigned char format_check =
578                 (vector unsigned char)(vector unsigned long){
579                 0x0c0000000c000000LL, 0x0c0000000c000000LL};
580         const vector unsigned char resp_err_check =
581                 (vector unsigned char)(vector unsigned long){
582                 0xe0000000e0000000LL, 0xe0000000e0000000LL};
583 #ifdef MLX5_PMD_SOFT_COUNTERS
584         uint32_t rcvd_byte = 0;
585         /* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
586         const vector unsigned char len_shuf_mask = (vector unsigned char){
587                  1,  0,  5,  4,
588                  9,  8, 13, 12,
589                 -1, -1, -1, -1,
590                 -1, -1, -1, -1};
591 #endif
592         /* Mask to shuffle from extracted CQE to mbuf. */
593         const vector unsigned char shuf_mask = (vector unsigned char){
594                  5,  4,           /* bswap16, pkt_len */
595                 -1, -1,           /* zero out 2nd half of pkt_len */
596                  5,  4,           /* bswap16, data_len */
597                 11, 10,           /* bswap16, vlan+tci */
598                 15, 14, 13, 12,   /* bswap32, rss */
599                  1,  2,  3, -1};  /* fdir.hi */
600         /* Mask to blend from the last Qword to the first DQword. */
601         /* Mask to blend from the last Qword to the first DQword. */
602         const vector unsigned char blend_mask = (vector unsigned char){
603                 -1,  0,  0,  0,
604                  0,  0,  0,  0,
605                 -1, -1, -1, -1,
606                 -1, -1, -1, -1};
607         const vector unsigned char crc_adj =
608                 (vector unsigned char)(vector unsigned short){
609                 rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
610                 rxq->crc_present * RTE_ETHER_CRC_LEN, 0, 0, 0, 0, 0};
611         const vector unsigned char flow_mark_adj =
612                 (vector unsigned char)(vector unsigned int){
613                 0, 0, 0, rxq->mark * (-1)};
614         const vector unsigned short cqe_sel_mask1 =
615                 (vector unsigned short){0, 0, 0, 0, 0xffff, 0xffff, 0, 0};
616         const vector unsigned short cqe_sel_mask2 =
617                 (vector unsigned short){0, 0, 0xffff, 0, 0, 0, 0, 0};
618
619         assert(rxq->sges_n == 0);
620         assert(rxq->cqe_n == rxq->elts_n);
621         cq = &(*rxq->cqes)[cq_idx];
622         rte_prefetch0(cq);
623         rte_prefetch0(cq + 1);
624         rte_prefetch0(cq + 2);
625         rte_prefetch0(cq + 3);
626         pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
627
628         repl_n = q_n - (rxq->rq_ci - rxq->rq_pi);
629         if (repl_n >= rxq->rq_repl_thresh)
630                 mlx5_rx_replenish_bulk_mbuf(rxq, repl_n);
631         /* See if there're unreturned mbufs from compressed CQE. */
632         rcvd_pkt = rxq->decompressed;
633         if (rcvd_pkt > 0) {
634                 rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
635                 rxq_copy_mbuf_v(rxq, pkts, rcvd_pkt);
636                 rxq->rq_pi += rcvd_pkt;
637                 rxq->decompressed -= rcvd_pkt;
638                 pkts += rcvd_pkt;
639         }
640         elts_idx = rxq->rq_pi & q_mask;
641         elts = &(*rxq->elts)[elts_idx];
642         /* Not to overflow pkts array. */
643         pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
644         /* Not to cross queue end. */
645         pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
646         pkts_n = RTE_MIN(pkts_n, q_n - cq_idx);
647         if (!pkts_n)
648                 return rcvd_pkt;
649         /* At this point, there shouldn't be any remaining packets. */
650         assert(rxq->decompressed == 0);
651
652         /*
653          * A. load first Qword (8bytes) in one loop.
654          * B. copy 4 mbuf pointers from elts ring to returing pkts.
655          * C. load remaining CQE data and extract necessary fields.
656          *    Final 16bytes cqes[] extracted from original 64bytes CQE has the
657          *    following structure:
658          *        struct {
659          *          uint8_t  pkt_info;
660          *          uint8_t  flow_tag[3];
661          *          uint16_t byte_cnt;
662          *          uint8_t  rsvd4;
663          *          uint8_t  op_own;
664          *          uint16_t hdr_type_etc;
665          *          uint16_t vlan_info;
666          *          uint32_t rx_has_res;
667          *        } c;
668          * D. fill in mbuf.
669          * E. get valid CQEs.
670          * F. find compressed CQE.
671          */
672         for (pos = 0;
673              pos < pkts_n;
674              pos += MLX5_VPMD_DESCS_PER_LOOP) {
675                 vector unsigned char cqes[MLX5_VPMD_DESCS_PER_LOOP];
676                 vector unsigned char cqe_tmp1, cqe_tmp2;
677                 vector unsigned char pkt_mb0, pkt_mb1, pkt_mb2, pkt_mb3;
678                 vector unsigned char op_own, op_own_tmp1, op_own_tmp2;
679                 vector unsigned char opcode, owner_mask, invalid_mask;
680                 vector unsigned char comp_mask;
681                 vector unsigned char mask;
682 #ifdef MLX5_PMD_SOFT_COUNTERS
683                 const vector unsigned char lower_half = {
684                         0, 1, 4, 5, 8, 9, 12, 13,
685                         16, 17, 20, 21, 24, 25, 28, 29};
686                 const vector unsigned char upper_half = {
687                         2, 3, 6, 7, 10, 11, 14, 15,
688                         18, 19, 22, 23, 26, 27, 30, 31};
689                 const vector unsigned long shmax = {64, 64};
690                 vector unsigned char byte_cnt;
691                 vector unsigned short left, right;
692                 vector unsigned long lshift;
693                 vector __attribute__((altivec(bool__)))
694                         unsigned long shmask;
695 #endif
696                 vector unsigned char mbp1, mbp2;
697                 vector unsigned char p =
698                         (vector unsigned char)(vector unsigned short){
699                                 0, 1, 2, 3, 0, 0, 0, 0};
700                 unsigned int p1, p2, p3;
701
702                 /* Prefetch next 4 CQEs. */
703                 if (pkts_n - pos >= 2 * MLX5_VPMD_DESCS_PER_LOOP) {
704                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP]);
705                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 1]);
706                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 2]);
707                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 3]);
708                 }
709
710                 /* A.0 do not cross the end of CQ. */
711                 mask = (vector unsigned char)(vector unsigned long){
712                         (pkts_n - pos) * sizeof(uint16_t) * 8, 0};
713
714                 {
715                         vector unsigned long lshift;
716                         vector __attribute__((altivec(bool__)))
717                                 unsigned long shmask;
718                         const vector unsigned long shmax = {64, 64};
719
720                         lshift = vec_splat((vector unsigned long)mask, 0);
721                         shmask = vec_cmpgt(shmax, lshift);
722                         mask = (vector unsigned char)
723                                 vec_sl((vector unsigned long)ones, lshift);
724                         mask = (vector unsigned char)
725                                 vec_sel((vector unsigned long)shmask,
726                                 (vector unsigned long)mask, shmask);
727                 }
728
729                 p = (vector unsigned char)
730                         vec_andc((vector unsigned long)p,
731                         (vector unsigned long)mask);
732
733                 /* A.1 load cqes. */
734                 p3 = (unsigned int)((vector unsigned short)p)[3];
735                 cqes[3] = (vector unsigned char)(vector unsigned long){
736                         *(__attribute__((__aligned__(8))) unsigned long *)
737                         &cq[pos + p3].sop_drop_qpn, 0LL};
738                 rte_compiler_barrier();
739
740                 p2 = (unsigned int)((vector unsigned short)p)[2];
741                 cqes[2] = (vector unsigned char)(vector unsigned long){
742                         *(__attribute__((__aligned__(8))) unsigned long *)
743                         &cq[pos + p2].sop_drop_qpn, 0LL};
744                 rte_compiler_barrier();
745
746                 /* B.1 load mbuf pointers. */
747                 mbp1 = (vector unsigned char)vec_vsx_ld(0,
748                         (signed int const *)&elts[pos]);
749                 mbp2 = (vector unsigned char)vec_vsx_ld(0,
750                         (signed int const *)&elts[pos + 2]);
751
752                 /* A.1 load a block having op_own. */
753                 p1 = (unsigned int)((vector unsigned short)p)[1];
754                 cqes[1] = (vector unsigned char)(vector unsigned long){
755                         *(__attribute__((__aligned__(8))) unsigned long *)
756                         &cq[pos + p1].sop_drop_qpn, 0LL};
757                 rte_compiler_barrier();
758
759                 cqes[0] = (vector unsigned char)(vector unsigned long){
760                         *(__attribute__((__aligned__(8))) unsigned long *)
761                         &cq[pos].sop_drop_qpn, 0LL};
762                 rte_compiler_barrier();
763
764                 /* B.2 copy mbuf pointers. */
765                 *(vector unsigned char *)&pkts[pos] = mbp1;
766                 *(vector unsigned char *)&pkts[pos + 2] = mbp2;
767                 rte_cio_rmb();
768
769                 /* C.1 load remaining CQE data and extract necessary fields. */
770                 cqe_tmp2 = *(vector unsigned char *)
771                         &cq[pos + p3].pkt_info;
772                 cqe_tmp1 = *(vector unsigned char *)
773                         &cq[pos + p2].pkt_info;
774                 cqes[3] = vec_sel(cqes[3], cqe_tmp2, blend_mask);
775                 cqes[2] = vec_sel(cqes[2], cqe_tmp1, blend_mask);
776                 cqe_tmp2 = (vector unsigned char)vec_vsx_ld(0,
777                         (signed int const *)&cq[pos + p3].csum);
778                 cqe_tmp1 = (vector unsigned char)vec_vsx_ld(0,
779                         (signed int const *)&cq[pos + p2].csum);
780                 cqes[3] = (vector unsigned char)
781                         vec_sel((vector unsigned short)cqes[3],
782                         (vector unsigned short)cqe_tmp2, cqe_sel_mask1);
783                 cqes[2] = (vector unsigned char)
784                         vec_sel((vector unsigned short)cqes[2],
785                         (vector unsigned short)cqe_tmp1, cqe_sel_mask1);
786                 cqe_tmp2 = (vector unsigned char)(vector unsigned long){
787                         *(__attribute__((__aligned__(8))) unsigned long *)
788                         &cq[pos + p3].rsvd3[9], 0LL};
789                 cqe_tmp1 = (vector unsigned char)(vector unsigned long){
790                         *(__attribute__((__aligned__(8))) unsigned long *)
791                         &cq[pos + p2].rsvd3[9], 0LL};
792                 cqes[3] = (vector unsigned char)
793                         vec_sel((vector unsigned short)cqes[3],
794                         (vector unsigned short)cqe_tmp2,
795                         (vector unsigned short)cqe_sel_mask2);
796                 cqes[2] = (vector unsigned char)
797                         vec_sel((vector unsigned short)cqes[2],
798                         (vector unsigned short)cqe_tmp1,
799                         (vector unsigned short)cqe_sel_mask2);
800
801                 /* C.2 generate final structure for mbuf with swapping bytes. */
802                 pkt_mb3 = vec_perm(cqes[3], zero, shuf_mask);
803                 pkt_mb2 = vec_perm(cqes[2], zero, shuf_mask);
804
805                 /* C.3 adjust CRC length. */
806                 pkt_mb3 = (vector unsigned char)
807                         ((vector unsigned short)pkt_mb3 -
808                         (vector unsigned short)crc_adj);
809                 pkt_mb2 = (vector unsigned char)
810                         ((vector unsigned short)pkt_mb2 -
811                         (vector unsigned short)crc_adj);
812
813                 /* C.4 adjust flow mark. */
814                 pkt_mb3 = (vector unsigned char)
815                         ((vector unsigned int)pkt_mb3 +
816                         (vector unsigned int)flow_mark_adj);
817                 pkt_mb2 = (vector unsigned char)
818                         ((vector unsigned int)pkt_mb2 +
819                         (vector unsigned int)flow_mark_adj);
820
821                 /* D.1 fill in mbuf - rx_descriptor_fields1. */
822                 *(vector unsigned char *)
823                         &pkts[pos + 3]->pkt_len = pkt_mb3;
824                 *(vector unsigned char *)
825                         &pkts[pos + 2]->pkt_len = pkt_mb2;
826
827                 /* E.1 extract op_own field. */
828                 op_own_tmp2 = (vector unsigned char)
829                         vec_mergeh((vector unsigned int)cqes[2],
830                         (vector unsigned int)cqes[3]);
831
832                 /* C.1 load remaining CQE data and extract necessary fields. */
833                 cqe_tmp2 = *(vector unsigned char *)
834                         &cq[pos + p1].pkt_info;
835                 cqe_tmp1 = *(vector unsigned char *)
836                         &cq[pos].pkt_info;
837                 cqes[1] = vec_sel(cqes[1], cqe_tmp2, blend_mask);
838                 cqes[0] = vec_sel(cqes[0], cqe_tmp2, blend_mask);
839                 cqe_tmp2 = (vector unsigned char)vec_vsx_ld(0,
840                         (signed int const *)&cq[pos + p1].csum);
841                 cqe_tmp1 = (vector unsigned char)vec_vsx_ld(0,
842                         (signed int const *)&cq[pos].csum);
843                 cqes[1] = (vector unsigned char)
844                         vec_sel((vector unsigned short)cqes[1],
845                         (vector unsigned short)cqe_tmp2, cqe_sel_mask1);
846                 cqes[0] = (vector unsigned char)
847                         vec_sel((vector unsigned short)cqes[0],
848                         (vector unsigned short)cqe_tmp1, cqe_sel_mask1);
849                 cqe_tmp2 = (vector unsigned char)(vector unsigned long){
850                         *(__attribute__((__aligned__(8))) unsigned long *)
851                         &cq[pos + p1].rsvd3[9], 0LL};
852                 cqe_tmp1 = (vector unsigned char)(vector unsigned long){
853                         *(__attribute__((__aligned__(8))) unsigned long *)
854                         &cq[pos].rsvd3[9], 0LL};
855                 cqes[1] = (vector unsigned char)
856                         vec_sel((vector unsigned short)cqes[1],
857                         (vector unsigned short)cqe_tmp2, cqe_sel_mask2);
858                 cqes[0] = (vector unsigned char)
859                         vec_sel((vector unsigned short)cqes[0],
860                         (vector unsigned short)cqe_tmp1, cqe_sel_mask2);
861
862                 /* C.2 generate final structure for mbuf with swapping bytes. */
863                 pkt_mb1 = vec_perm(cqes[1], zero, shuf_mask);
864                 pkt_mb0 = vec_perm(cqes[0], zero, shuf_mask);
865
866                 /* C.3 adjust CRC length. */
867                 pkt_mb1 = (vector unsigned char)
868                         ((vector unsigned short)pkt_mb1 -
869                         (vector unsigned short)crc_adj);
870                 pkt_mb0 = (vector unsigned char)
871                         ((vector unsigned short)pkt_mb0 -
872                         (vector unsigned short)crc_adj);
873
874                 /* C.4 adjust flow mark. */
875                 pkt_mb1 = (vector unsigned char)
876                         ((vector unsigned int)pkt_mb1 +
877                         (vector unsigned int)flow_mark_adj);
878                 pkt_mb0 = (vector unsigned char)
879                         ((vector unsigned int)pkt_mb0 +
880                         (vector unsigned int)flow_mark_adj);
881
882                 /* E.1 extract op_own byte. */
883                 op_own_tmp1 = (vector unsigned char)
884                         vec_mergeh((vector unsigned int)cqes[0],
885                         (vector unsigned int)cqes[1]);
886                 op_own = (vector unsigned char)
887                         vec_mergel((vector unsigned long)op_own_tmp1,
888                         (vector unsigned long)op_own_tmp2);
889
890                 /* D.1 fill in mbuf - rx_descriptor_fields1. */
891                 *(vector unsigned char *)
892                         &pkts[pos + 1]->pkt_len = pkt_mb1;
893                 *(vector unsigned char *)
894                         &pkts[pos]->pkt_len = pkt_mb0;
895
896                 /* E.2 flip owner bit to mark CQEs from last round. */
897                 owner_mask = (vector unsigned char)
898                         vec_and((vector unsigned long)op_own,
899                         (vector unsigned long)owner_check);
900                 if (ownership)
901                         owner_mask = (vector unsigned char)
902                                 vec_xor((vector unsigned long)owner_mask,
903                                 (vector unsigned long)owner_check);
904                 owner_mask = (vector unsigned char)
905                         vec_cmpeq((vector unsigned int)owner_mask,
906                         (vector unsigned int)owner_check);
907                 owner_mask = (vector unsigned char)
908                         vec_packs((vector unsigned int)owner_mask,
909                         (vector unsigned int)zero);
910
911                 /* E.3 get mask for invalidated CQEs. */
912                 opcode = (vector unsigned char)
913                         vec_and((vector unsigned long)op_own,
914                         (vector unsigned long)opcode_check);
915                 invalid_mask = (vector unsigned char)
916                         vec_cmpeq((vector unsigned int)opcode_check,
917                         (vector unsigned int)opcode);
918                 invalid_mask = (vector unsigned char)
919                         vec_packs((vector unsigned int)invalid_mask,
920                         (vector unsigned int)zero);
921
922                 /* E.4 mask out beyond boundary. */
923                 invalid_mask = (vector unsigned char)
924                         vec_or((vector unsigned long)invalid_mask,
925                         (vector unsigned long)mask);
926
927                 /* E.5 merge invalid_mask with invalid owner. */
928                 invalid_mask = (vector unsigned char)
929                         vec_or((vector unsigned long)invalid_mask,
930                         (vector unsigned long)owner_mask);
931
932                 /* F.1 find compressed CQE format. */
933                 comp_mask = (vector unsigned char)
934                         vec_and((vector unsigned long)op_own,
935                         (vector unsigned long)format_check);
936                 comp_mask = (vector unsigned char)
937                         vec_cmpeq((vector unsigned int)comp_mask,
938                         (vector unsigned int)format_check);
939                 comp_mask = (vector unsigned char)
940                         vec_packs((vector unsigned int)comp_mask,
941                         (vector unsigned int)zero);
942
943                 /* F.2 mask out invalid entries. */
944                 comp_mask = (vector unsigned char)
945                         vec_andc((vector unsigned long)comp_mask,
946                         (vector unsigned long)invalid_mask);
947                 comp_idx = ((vector unsigned long)comp_mask)[0];
948
949                 /* F.3 get the first compressed CQE. */
950                 comp_idx = comp_idx ? __builtin_ctzll(comp_idx) /
951                         (sizeof(uint16_t) * 8) : MLX5_VPMD_DESCS_PER_LOOP;
952
953                 /* E.6 mask out entries after the compressed CQE. */
954                 mask = (vector unsigned char)(vector unsigned long){
955                         (comp_idx * sizeof(uint16_t) * 8), 0};
956                 lshift = vec_splat((vector unsigned long)mask, 0);
957                 shmask = vec_cmpgt(shmax, lshift);
958                 mask = (vector unsigned char)
959                         vec_sl((vector unsigned long)ones, lshift);
960                 mask = (vector unsigned char)
961                         vec_sel((vector unsigned long)shmask,
962                         (vector unsigned long)mask, shmask);
963                 invalid_mask = (vector unsigned char)
964                         vec_or((vector unsigned long)invalid_mask,
965                         (vector unsigned long)mask);
966
967                 /* E.7 count non-compressed valid CQEs. */
968                 n = ((vector unsigned long)invalid_mask)[0];
969                 n = n ? __builtin_ctzll(n) / (sizeof(uint16_t) * 8) :
970                         MLX5_VPMD_DESCS_PER_LOOP;
971                 nocmp_n += n;
972
973                 /* D.2 get the final invalid mask. */
974                 mask = (vector unsigned char)(vector unsigned long){
975                         (n * sizeof(uint16_t) * 8), 0};
976                 lshift = vec_splat((vector unsigned long)mask, 0);
977                 shmask = vec_cmpgt(shmax, lshift);
978                 mask = (vector unsigned char)
979                         vec_sl((vector unsigned long)ones, lshift);
980                 mask = (vector unsigned char)
981                         vec_sel((vector unsigned long)shmask,
982                         (vector unsigned long)mask, shmask);
983                 invalid_mask = (vector unsigned char)
984                         vec_or((vector unsigned long)invalid_mask,
985                         (vector unsigned long)mask);
986
987                 /* D.3 check error in opcode. */
988                 opcode = (vector unsigned char)
989                         vec_cmpeq((vector unsigned int)resp_err_check,
990                         (vector unsigned int)opcode);
991                 opcode = (vector unsigned char)
992                         vec_packs((vector unsigned int)opcode,
993                         (vector unsigned int)zero);
994                 opcode = (vector unsigned char)
995                         vec_andc((vector unsigned long)opcode,
996                         (vector unsigned long)invalid_mask);
997
998                 /* D.4 mark if any error is set */
999                 *err |= ((vector unsigned long)opcode)[0];
1000
1001                 /* D.5 fill in mbuf - rearm_data and packet_type. */
1002                 rxq_cq_to_ptype_oflags_v(rxq, cqes, opcode, &pkts[pos]);
1003                 if (rxq->hw_timestamp) {
1004                         pkts[pos]->timestamp =
1005                                 rte_be_to_cpu_64(cq[pos].timestamp);
1006                         pkts[pos + 1]->timestamp =
1007                                 rte_be_to_cpu_64(cq[pos + p1].timestamp);
1008                         pkts[pos + 2]->timestamp =
1009                                 rte_be_to_cpu_64(cq[pos + p2].timestamp);
1010                         pkts[pos + 3]->timestamp =
1011                                 rte_be_to_cpu_64(cq[pos + p3].timestamp);
1012                 }
1013                 if (rte_flow_dynf_metadata_avail()) {
1014                         uint64_t flag = rte_flow_dynf_metadata_mask;
1015                         int offs = rte_flow_dynf_metadata_offs;
1016                         uint32_t metadata;
1017
1018                         /* This code is subject for futher optimization. */
1019                         metadata = cq[pos].flow_table_metadata;
1020                         *RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *) =
1021                                                                 metadata;
1022                         pkts[pos]->ol_flags |= metadata ? flag : 0ULL;
1023                         metadata = cq[pos + 1].flow_table_metadata;
1024                         *RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *) =
1025                                                                 metadata;
1026                         pkts[pos + 1]->ol_flags |= metadata ? flag : 0ULL;
1027                         metadata = cq[pos + 2].flow_table_metadata;
1028                         *RTE_MBUF_DYNFIELD(pkts[pos + 2], offs, uint32_t *) =
1029                                                                 metadata;
1030                         pkts[pos + 2]->ol_flags |= metadata ? flag : 0ULL;
1031                         metadata = cq[pos + 3].flow_table_metadata;
1032                         *RTE_MBUF_DYNFIELD(pkts[pos + 3], offs, uint32_t *) =
1033                                                                 metadata;
1034                         pkts[pos + 3]->ol_flags |= metadata ? flag : 0ULL;
1035                 }
1036 #ifdef MLX5_PMD_SOFT_COUNTERS
1037                 /* Add up received bytes count. */
1038                 byte_cnt = vec_perm(op_own, zero, len_shuf_mask);
1039                 byte_cnt = (vector unsigned char)
1040                         vec_andc((vector unsigned long)byte_cnt,
1041                         (vector unsigned long)invalid_mask);
1042                 left = vec_perm((vector unsigned short)byte_cnt,
1043                         (vector unsigned short)zero, lower_half);
1044                 right = vec_perm((vector unsigned short)byte_cnt,
1045                         (vector unsigned short)zero, upper_half);
1046                 byte_cnt = (vector unsigned char)vec_add(left, right);
1047                 left = vec_perm((vector unsigned short)byte_cnt,
1048                         (vector unsigned short)zero, lower_half);
1049                 right = vec_perm((vector unsigned short)byte_cnt,
1050                         (vector unsigned short)zero, upper_half);
1051                 byte_cnt = (vector unsigned char)vec_add(left, right);
1052                 rcvd_byte += ((vector unsigned long)byte_cnt)[0];
1053 #endif
1054
1055                 /*
1056                  * Break the loop unless more valid CQE is expected, or if
1057                  * there's a compressed CQE.
1058                  */
1059                 if (n != MLX5_VPMD_DESCS_PER_LOOP)
1060                         break;
1061         }
1062         /* If no new CQE seen, return without updating cq_db. */
1063         if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP))
1064                 return rcvd_pkt;
1065         /* Update the consumer indexes for non-compressed CQEs. */
1066         assert(nocmp_n <= pkts_n);
1067         rxq->cq_ci += nocmp_n;
1068         rxq->rq_pi += nocmp_n;
1069         rcvd_pkt += nocmp_n;
1070 #ifdef MLX5_PMD_SOFT_COUNTERS
1071         rxq->stats.ipackets += nocmp_n;
1072         rxq->stats.ibytes += rcvd_byte;
1073 #endif
1074         /* Decompress the last CQE if compressed. */
1075         if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP && comp_idx == n) {
1076                 assert(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
1077                 rxq->decompressed =
1078                         rxq_cq_decompress_v(rxq, &cq[nocmp_n], &elts[nocmp_n]);
1079                 /* Return more packets if needed. */
1080                 if (nocmp_n < pkts_n) {
1081                         uint16_t n = rxq->decompressed;
1082
1083                         n = RTE_MIN(n, pkts_n - nocmp_n);
1084                         rxq_copy_mbuf_v(rxq, &pkts[nocmp_n], n);
1085                         rxq->rq_pi += n;
1086                         rcvd_pkt += n;
1087                         rxq->decompressed -= n;
1088                 }
1089         }
1090         rte_compiler_barrier();
1091         *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1092         return rcvd_pkt;
1093 }
1094
1095 #endif /* RTE_PMD_MLX5_RXTX_VEC_ALTIVEC_H_ */