d55642e57a31fc4b0147813b14ee454aa69018e3
[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 <stdint.h>
10 #include <string.h>
11 #include <stdlib.h>
12
13 #include <altivec.h>
14
15 #include <rte_mbuf.h>
16 #include <rte_mempool.h>
17 #include <rte_prefetch.h>
18
19 #include <mlx5_prm.h>
20
21 #include "mlx5_defs.h"
22 #include "mlx5.h"
23 #include "mlx5_utils.h"
24 #include "mlx5_rxtx.h"
25 #include "mlx5_rxtx_vec.h"
26 #include "mlx5_autoconf.h"
27
28 #ifndef __INTEL_COMPILER
29 #pragma GCC diagnostic ignored "-Wcast-qual"
30 #pragma GCC diagnostic ignored "-Wstrict-aliasing"
31 #endif
32
33 /**
34  * Store free buffers to RX SW ring.
35  *
36  * @param rxq
37  *   Pointer to RX queue structure.
38  * @param pkts
39  *   Pointer to array of packets to be stored.
40  * @param pkts_n
41  *   Number of packets to be stored.
42  */
43 static inline void
44 rxq_copy_mbuf_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t n)
45 {
46         const uint16_t q_mask = (1 << rxq->elts_n) - 1;
47         struct rte_mbuf **elts = &(*rxq->elts)[rxq->rq_pi & q_mask];
48         unsigned int pos;
49         uint16_t p = n & -2;
50
51         for (pos = 0; pos < p; pos += 2) {
52                 vector unsigned char mbp;
53
54                 mbp = (vector unsigned char)vec_vsx_ld(0,
55                                 (signed int const *)&elts[pos]);
56                 *(vector unsigned char *)&pkts[pos] = mbp;
57         }
58         if (n & 1)
59                 pkts[pos] = elts[pos];
60 }
61
62 /**
63  * Decompress a compressed completion and fill in mbufs in RX SW ring with data
64  * extracted from the title completion descriptor.
65  *
66  * @param rxq
67  *   Pointer to RX queue structure.
68  * @param cq
69  *   Pointer to completion array having a compressed completion at first.
70  * @param elts
71  *   Pointer to SW ring to be filled. The first mbuf has to be pre-built from
72  *   the title completion descriptor to be copied to the rest of mbufs.
73  *
74  * @return
75  *   Number of mini-CQEs successfully decompressed.
76  */
77 static inline uint16_t
78 rxq_cq_decompress_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
79                     struct rte_mbuf **elts)
80 {
81         volatile struct mlx5_mini_cqe8 *mcq = (void *)&(cq + 1)->pkt_info;
82         struct rte_mbuf *t_pkt = elts[0]; /* Title packet is pre-built. */
83         const vector unsigned char zero = (vector unsigned char){0};
84         /* Mask to shuffle from extracted mini CQE to mbuf. */
85         const vector unsigned char shuf_mask1 = (vector unsigned char){
86                         -1, -1, -1, -1,   /* skip packet_type */
87                          7,  6, -1, -1,   /* bswap16, pkt_len */
88                          7,  6,           /* bswap16, data_len */
89                         -1, -1,           /* skip vlan_tci */
90                          3,  2,  1,  0};  /* bswap32, rss */
91         const vector unsigned char shuf_mask2 = (vector unsigned char){
92                         -1, -1, -1, -1,   /* skip packet_type */
93                         15, 14, -1, -1,   /* bswap16, pkt_len */
94                         15, 14,           /* data_len, bswap16 */
95                         -1, -1,           /* skip vlan_tci */
96                         11, 10,  9,  8};  /* bswap32, rss */
97         /* Restore the compressed count. Must be 16 bits. */
98         const uint16_t mcqe_n = t_pkt->data_len +
99                 (rxq->crc_present * RTE_ETHER_CRC_LEN);
100         const vector unsigned char rearm =
101                 (vector unsigned char)vec_vsx_ld(0,
102                 (signed int const *)&t_pkt->rearm_data);
103         const vector unsigned char rxdf =
104                 (vector unsigned char)vec_vsx_ld(0,
105                 (signed int const *)&t_pkt->rx_descriptor_fields1);
106         const vector unsigned char crc_adj =
107                 (vector unsigned char)(vector unsigned short){
108                         0, 0, rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
109                         rxq->crc_present * RTE_ETHER_CRC_LEN, 0, 0, 0};
110         const vector unsigned short rxdf_sel_mask =
111                 (vector unsigned short){
112                         0xffff, 0xffff, 0, 0, 0, 0xffff, 0, 0};
113         const uint32_t flow_tag = t_pkt->hash.fdir.hi;
114         unsigned int pos;
115         unsigned int i;
116         unsigned int inv = 0;
117
118 #ifdef MLX5_PMD_SOFT_COUNTERS
119         const vector unsigned char ones = vec_splat_u8(-1);
120         uint32_t rcvd_byte = 0;
121         /* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
122         const vector unsigned char len_shuf_mask = (vector unsigned char){
123                  3,  2, 11, 10,
124                  7,  6, 15, 14,
125                 -1, -1, -1, -1,
126                 -1, -1, -1, -1};
127 #endif
128
129         /*
130          * A. load mCQEs into a 128bit register.
131          * B. store rearm data to mbuf.
132          * C. combine data from mCQEs with rx_descriptor_fields1.
133          * D. store rx_descriptor_fields1.
134          * E. store flow tag (rte_flow mark).
135          */
136         for (pos = 0; pos < mcqe_n; ) {
137                 vector unsigned char mcqe1, mcqe2;
138                 vector unsigned char rxdf1, rxdf2;
139 #ifdef MLX5_PMD_SOFT_COUNTERS
140                 const vector unsigned short mcqe_sel_mask =
141                         (vector unsigned short){0, 0, 0xffff, 0xffff,
142                         0, 0, 0xfff, 0xffff};
143                 const vector unsigned char lower_half = {
144                         0, 1, 4, 5, 8, 9, 12, 13, 16,
145                         17, 20, 21, 24, 25, 28, 29};
146                 const vector unsigned char upper_half = {
147                         2, 3, 6, 7, 10, 11, 14, 15,
148                         18, 19, 22, 23, 26, 27, 30, 31};
149                 vector unsigned short left, right;
150                 vector unsigned char byte_cnt, invalid_mask;
151                 vector unsigned long lshift;
152                 __attribute__((altivec(vector__)))
153                         __attribute__((altivec(bool__)))
154                         unsigned long long shmask;
155                 const vector unsigned long shmax = {64, 64};
156 #endif
157
158                 if (!(pos & 0x7) && pos + 8 < mcqe_n)
159                         rte_prefetch0((void *)(cq + pos + 8));
160
161                 /* A.1 load mCQEs into a 128bit register. */
162                 mcqe1 = (vector unsigned char)vec_vsx_ld(0,
163                         (signed int const *)&mcq[pos % 8]);
164                 mcqe2 = (vector unsigned char)vec_vsx_ld(0,
165                         (signed int const *)&mcq[pos % 8 + 2]);
166
167                 /* B.1 store rearm data to mbuf. */
168                 *(vector unsigned char *)
169                         &elts[pos]->rearm_data = rearm;
170                 *(vector unsigned char *)
171                         &elts[pos + 1]->rearm_data = rearm;
172
173                 /* C.1 combine data from mCQEs with rx_descriptor_fields1. */
174                 rxdf1 = vec_perm(mcqe1, zero, shuf_mask1);
175                 rxdf2 = vec_perm(mcqe1, zero, shuf_mask2);
176                 rxdf1 = (vector unsigned char)
177                         ((vector unsigned short)rxdf1 -
178                         (vector unsigned short)crc_adj);
179                 rxdf2 = (vector unsigned char)
180                         ((vector unsigned short)rxdf2 -
181                         (vector unsigned short)crc_adj);
182                 rxdf1 = (vector unsigned char)
183                         vec_sel((vector unsigned short)rxdf1,
184                         (vector unsigned short)rxdf, rxdf_sel_mask);
185                 rxdf2 = (vector unsigned char)
186                         vec_sel((vector unsigned short)rxdf2,
187                         (vector unsigned short)rxdf, rxdf_sel_mask);
188
189                 /* D.1 store rx_descriptor_fields1. */
190                 *(vector unsigned char *)
191                         &elts[pos]->rx_descriptor_fields1 = rxdf1;
192                 *(vector unsigned char *)
193                         &elts[pos + 1]->rx_descriptor_fields1 = rxdf2;
194
195                 /* B.1 store rearm data to mbuf. */
196                 *(vector unsigned char *)
197                         &elts[pos + 2]->rearm_data = rearm;
198                 *(vector unsigned char *)
199                         &elts[pos + 3]->rearm_data = rearm;
200
201                 /* C.1 combine data from mCQEs with rx_descriptor_fields1. */
202                 rxdf1 = vec_perm(mcqe2, zero, shuf_mask1);
203                 rxdf2 = vec_perm(mcqe2, zero, shuf_mask2);
204                 rxdf1 = (vector unsigned char)
205                         ((vector unsigned short)rxdf1 -
206                         (vector unsigned short)crc_adj);
207                 rxdf2 = (vector unsigned char)
208                         ((vector unsigned short)rxdf2 -
209                         (vector unsigned short)crc_adj);
210                 rxdf1 = (vector unsigned char)
211                         vec_sel((vector unsigned short)rxdf1,
212                         (vector unsigned short)rxdf, rxdf_sel_mask);
213                 rxdf2 = (vector unsigned char)
214                         vec_sel((vector unsigned short)rxdf2,
215                         (vector unsigned short)rxdf, rxdf_sel_mask);
216
217                 /* D.1 store rx_descriptor_fields1. */
218                 *(vector unsigned char *)
219                         &elts[pos + 2]->rx_descriptor_fields1 = rxdf1;
220                 *(vector unsigned char *)
221                         &elts[pos + 3]->rx_descriptor_fields1 = rxdf2;
222
223 #ifdef MLX5_PMD_SOFT_COUNTERS
224                 invalid_mask = (vector unsigned char)(vector unsigned long){
225                         (mcqe_n - pos) * sizeof(uint16_t) * 8, 0};
226
227                 lshift =
228                         vec_splat((vector unsigned long)invalid_mask, 0);
229                 shmask = vec_cmpgt(shmax, lshift);
230                 invalid_mask = (vector unsigned char)
231                         vec_sl((vector unsigned long)ones, lshift);
232                 invalid_mask = (vector unsigned char)
233                         vec_sel((vector unsigned long)shmask,
234                         (vector unsigned long)invalid_mask, shmask);
235
236                 mcqe1 = (vector unsigned char)
237                         vec_sro((vector unsigned short)mcqe1,
238                         (vector unsigned char){32}),
239                 byte_cnt = (vector unsigned char)
240                         vec_sel((vector unsigned short)mcqe1,
241                         (vector unsigned short)mcqe2, mcqe_sel_mask);
242                 byte_cnt = vec_perm(byte_cnt, zero, len_shuf_mask);
243                 byte_cnt = (vector unsigned char)
244                         vec_andc((vector unsigned long)byte_cnt,
245                         (vector unsigned long)invalid_mask);
246                 left = vec_perm((vector unsigned short)byte_cnt,
247                         (vector unsigned short)zero, lower_half);
248                 right = vec_perm((vector unsigned short)byte_cnt,
249                         (vector unsigned short)zero, upper_half);
250                 byte_cnt = (vector unsigned char)vec_add(left, right);
251                 left = vec_perm((vector unsigned short)byte_cnt,
252                         (vector unsigned short)zero, lower_half);
253                 right = vec_perm((vector unsigned short)byte_cnt,
254                         (vector unsigned short)zero, upper_half);
255                 byte_cnt = (vector unsigned char)vec_add(left, right);
256                 rcvd_byte += ((vector unsigned long)byte_cnt)[0];
257 #endif
258
259                 if (rxq->mark) {
260                         /* E.1 store flow tag (rte_flow mark). */
261                         elts[pos]->hash.fdir.hi = flow_tag;
262                         elts[pos + 1]->hash.fdir.hi = flow_tag;
263                         elts[pos + 2]->hash.fdir.hi = flow_tag;
264                         elts[pos + 3]->hash.fdir.hi = flow_tag;
265                 }
266                 if (rte_flow_dynf_metadata_avail()) {
267                         const uint32_t meta = *RTE_FLOW_DYNF_METADATA(t_pkt);
268
269                         /* Check if title packet has valid metadata. */
270                         if (meta) {
271                                 MLX5_ASSERT(t_pkt->ol_flags &
272                                             PKT_RX_DYNF_METADATA);
273                                 *RTE_FLOW_DYNF_METADATA(elts[pos]) = meta;
274                                 *RTE_FLOW_DYNF_METADATA(elts[pos + 1]) = meta;
275                                 *RTE_FLOW_DYNF_METADATA(elts[pos + 2]) = meta;
276                                 *RTE_FLOW_DYNF_METADATA(elts[pos + 3]) = meta;
277                         }
278                 }
279
280                 pos += MLX5_VPMD_DESCS_PER_LOOP;
281                 /* Move to next CQE and invalidate consumed CQEs. */
282                 if (!(pos & 0x7) && pos < mcqe_n) {
283                         mcq = (void *)&(cq + pos)->pkt_info;
284                         for (i = 0; i < 8; ++i)
285                                 cq[inv++].op_own = MLX5_CQE_INVALIDATE;
286                 }
287         }
288
289         /* Invalidate the rest of CQEs. */
290         for (; inv < mcqe_n; ++inv)
291                 cq[inv].op_own = MLX5_CQE_INVALIDATE;
292
293 #ifdef MLX5_PMD_SOFT_COUNTERS
294         rxq->stats.ipackets += mcqe_n;
295         rxq->stats.ibytes += rcvd_byte;
296 #endif
297
298         rxq->cq_ci += mcqe_n;
299         return mcqe_n;
300 }
301
302 /**
303  * Calculate packet type and offload flag for mbuf and store it.
304  *
305  * @param rxq
306  *   Pointer to RX queue structure.
307  * @param cqes[4]
308  *   Array of four 16bytes completions extracted from the original completion
309  *   descriptor.
310  * @param op_err
311  *   Opcode vector having responder error status. Each field is 4B.
312  * @param pkts
313  *   Pointer to array of packets to be filled.
314  */
315 static inline void
316 rxq_cq_to_ptype_oflags_v(struct mlx5_rxq_data *rxq,
317                 vector unsigned char cqes[4], vector unsigned char op_err,
318                 struct rte_mbuf **pkts)
319 {
320         vector unsigned char pinfo0, pinfo1;
321         vector unsigned char pinfo, ptype;
322         vector unsigned char ol_flags = (vector unsigned char)
323                 (vector unsigned int){
324                         rxq->rss_hash * PKT_RX_RSS_HASH |
325                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP,
326                         rxq->rss_hash * PKT_RX_RSS_HASH |
327                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP,
328                         rxq->rss_hash * PKT_RX_RSS_HASH |
329                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP,
330                         rxq->rss_hash * PKT_RX_RSS_HASH |
331                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP};
332         vector unsigned char cv_flags;
333         const vector unsigned char zero = (vector unsigned char){0};
334         const vector unsigned char ptype_mask =
335                 (vector unsigned char)(vector unsigned int){
336                 0x0000fd06, 0x0000fd06, 0x0000fd06, 0x0000fd06};
337         const vector unsigned char ptype_ol_mask =
338                 (vector unsigned char)(vector unsigned int){
339                 0x00000106, 0x00000106, 0x00000106, 0x00000106};
340         const vector unsigned char pinfo_mask =
341                 (vector unsigned char)(vector unsigned int){
342                 0x00000003, 0x00000003, 0x00000003, 0x00000003};
343         const vector unsigned char cv_flag_sel = (vector unsigned char){
344                 0, (uint8_t)(PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED),
345                 (uint8_t)(PKT_RX_IP_CKSUM_GOOD >> 1), 0,
346                 (uint8_t)(PKT_RX_L4_CKSUM_GOOD >> 1), 0,
347                 (uint8_t)((PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1),
348                 0, 0, 0, 0, 0, 0, 0, 0, 0};
349         const vector unsigned char cv_mask =
350                 (vector unsigned char)(vector unsigned int){
351                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
352                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
353                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
354                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
355                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
356                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
357                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
358                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED};
359         const vector unsigned char mbuf_init =
360                 (vector unsigned char)vec_vsx_ld
361                         (0, (vector unsigned char *)&rxq->mbuf_initializer);
362         const vector unsigned short rearm_sel_mask =
363                 (vector unsigned short){0, 0, 0, 0, 0xffff, 0xffff, 0, 0};
364         vector unsigned char rearm0, rearm1, rearm2, rearm3;
365         uint8_t pt_idx0, pt_idx1, pt_idx2, pt_idx3;
366
367         /* Extract pkt_info field. */
368         pinfo0 = (vector unsigned char)
369                 vec_mergeh((vector unsigned int)cqes[0],
370                 (vector unsigned int)cqes[1]);
371         pinfo1 = (vector unsigned char)
372                 vec_mergeh((vector unsigned int)cqes[2],
373                 (vector unsigned int)cqes[3]);
374         pinfo = (vector unsigned char)
375                 vec_mergeh((vector unsigned long)pinfo0,
376                 (vector unsigned long)pinfo1);
377
378         /* Extract hdr_type_etc field. */
379         pinfo0 = (vector unsigned char)
380                 vec_mergel((vector unsigned int)cqes[0],
381                 (vector unsigned int)cqes[1]);
382         pinfo1 = (vector unsigned char)
383                 vec_mergel((vector unsigned int)cqes[2],
384                 (vector unsigned int)cqes[3]);
385         ptype = (vector unsigned char)
386                 vec_mergeh((vector unsigned long)pinfo0,
387                 (vector unsigned long)pinfo1);
388
389         if (rxq->mark) {
390                 const vector unsigned char pinfo_ft_mask =
391                         (vector unsigned char)(vector unsigned int){
392                         0xffffff00, 0xffffff00, 0xffffff00, 0xffffff00};
393                 const vector unsigned char fdir_flags =
394                         (vector unsigned char)(vector unsigned int){
395                         PKT_RX_FDIR, PKT_RX_FDIR,
396                         PKT_RX_FDIR, PKT_RX_FDIR};
397                 vector unsigned char fdir_id_flags =
398                         (vector unsigned char)(vector unsigned int){
399                         PKT_RX_FDIR_ID, PKT_RX_FDIR_ID,
400                         PKT_RX_FDIR_ID, PKT_RX_FDIR_ID};
401                 vector unsigned char flow_tag, invalid_mask;
402
403                 flow_tag = (vector unsigned char)
404                         vec_and((vector unsigned long)pinfo,
405                         (vector unsigned long)pinfo_ft_mask);
406
407                 /* Check if flow tag is non-zero then set PKT_RX_FDIR. */
408                 invalid_mask = (vector unsigned char)
409                         vec_cmpeq((vector unsigned int)flow_tag,
410                         (vector unsigned int)zero);
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_flags,
415                         (vector unsigned long)invalid_mask));
416
417                 /* Mask out invalid entries. */
418                 fdir_id_flags = (vector unsigned char)
419                         vec_andc((vector unsigned long)fdir_id_flags,
420                         (vector unsigned long)invalid_mask);
421
422                 /* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */
423                 ol_flags = (vector unsigned char)
424                         vec_or((vector unsigned long)ol_flags,
425                         (vector unsigned long)
426                         vec_andc((vector unsigned long)fdir_id_flags,
427                         (vector unsigned long)
428                         vec_cmpeq((vector unsigned int)flow_tag,
429                         (vector unsigned int)pinfo_ft_mask)));
430         }
431         /*
432          * Merge the two fields to generate the following:
433          * bit[1]     = l3_ok
434          * bit[2]     = l4_ok
435          * bit[8]     = cv
436          * bit[11:10] = l3_hdr_type
437          * bit[14:12] = l4_hdr_type
438          * bit[15]    = ip_frag
439          * bit[16]    = tunneled
440          * bit[17]    = outer_l3_type
441          */
442         ptype = (vector unsigned char)
443                 vec_and((vector unsigned long)ptype,
444                 (vector unsigned long)ptype_mask);
445         pinfo = (vector unsigned char)
446                 vec_and((vector unsigned long)pinfo,
447                 (vector unsigned long)pinfo_mask);
448         pinfo = (vector unsigned char)
449                 vec_sl((vector unsigned int)pinfo,
450                 (vector unsigned int){16, 16, 16, 16});
451
452         /* Make pinfo has merged fields for ol_flags calculation. */
453         pinfo = (vector unsigned char)
454                 vec_or((vector unsigned long)ptype,
455                 (vector unsigned long)pinfo);
456         ptype = (vector unsigned char)
457                 vec_sr((vector unsigned int)pinfo,
458                 (vector unsigned int){10, 10, 10, 10});
459         ptype = (vector unsigned char)
460                 vec_packs((vector unsigned int)ptype,
461                 (vector unsigned int)zero);
462
463         /* Errored packets will have RTE_PTYPE_ALL_MASK. */
464         op_err = (vector unsigned char)
465                 vec_sr((vector unsigned short)op_err,
466                 (vector unsigned short){8, 8, 8, 8, 8, 8, 8, 8});
467         ptype = (vector unsigned char)
468                 vec_or((vector unsigned long)ptype,
469                 (vector unsigned long)op_err);
470
471         pt_idx0 = (uint8_t)((vector unsigned char)ptype)[0];
472         pt_idx1 = (uint8_t)((vector unsigned char)ptype)[2];
473         pt_idx2 = (uint8_t)((vector unsigned char)ptype)[4];
474         pt_idx3 = (uint8_t)((vector unsigned char)ptype)[6];
475
476         pkts[0]->packet_type = mlx5_ptype_table[pt_idx0] |
477                 !!(pt_idx0 & (1 << 6)) * rxq->tunnel;
478         pkts[1]->packet_type = mlx5_ptype_table[pt_idx1] |
479                 !!(pt_idx1 & (1 << 6)) * rxq->tunnel;
480         pkts[2]->packet_type = mlx5_ptype_table[pt_idx2] |
481                 !!(pt_idx2 & (1 << 6)) * rxq->tunnel;
482         pkts[3]->packet_type = mlx5_ptype_table[pt_idx3] |
483                 !!(pt_idx3 & (1 << 6)) * rxq->tunnel;
484
485         /* Fill flags for checksum and VLAN. */
486         pinfo = (vector unsigned char)
487                 vec_and((vector unsigned long)pinfo,
488                 (vector unsigned long)ptype_ol_mask);
489         pinfo = vec_perm(cv_flag_sel, zero, pinfo);
490
491         /* Locate checksum flags at byte[2:1] and merge with VLAN flags. */
492         cv_flags = (vector unsigned char)
493                 vec_sl((vector unsigned int)pinfo,
494                 (vector unsigned int){9, 9, 9, 9});
495         cv_flags = (vector unsigned char)
496                 vec_or((vector unsigned long)pinfo,
497                 (vector unsigned long)cv_flags);
498
499         /* Move back flags to start from byte[0]. */
500         cv_flags = (vector unsigned char)
501                 vec_sr((vector unsigned int)cv_flags,
502                 (vector unsigned int){8, 8, 8, 8});
503
504         /* Mask out garbage bits. */
505         cv_flags = (vector unsigned char)
506                 vec_and((vector unsigned long)cv_flags,
507                 (vector unsigned long)cv_mask);
508
509         /* Merge to ol_flags. */
510         ol_flags = (vector unsigned char)
511                 vec_or((vector unsigned long)ol_flags,
512                 (vector unsigned long)cv_flags);
513
514         /* Merge mbuf_init and ol_flags. */
515         rearm0 = (vector unsigned char)
516                 vec_sel((vector unsigned short)mbuf_init,
517                 (vector unsigned short)
518                 vec_slo((vector unsigned short)ol_flags,
519                 (vector unsigned char){64}), rearm_sel_mask);
520         rearm1 = (vector unsigned char)
521                 vec_sel((vector unsigned short)mbuf_init,
522                 (vector unsigned short)
523                 vec_slo((vector unsigned short)ol_flags,
524                 (vector unsigned char){32}), rearm_sel_mask);
525         rearm2 = (vector unsigned char)
526                 vec_sel((vector unsigned short)mbuf_init,
527                 (vector unsigned short)ol_flags, rearm_sel_mask);
528         rearm3 = (vector unsigned char)
529                 vec_sel((vector unsigned short)mbuf_init,
530                 (vector unsigned short)
531                 vec_sro((vector unsigned short)ol_flags,
532                 (vector unsigned char){32}), rearm_sel_mask);
533
534         /* Write 8B rearm_data and 8B ol_flags. */
535         vec_vsx_st(rearm0, 0,
536                 (vector unsigned char *)&pkts[0]->rearm_data);
537         vec_vsx_st(rearm1, 0,
538                 (vector unsigned char *)&pkts[1]->rearm_data);
539         vec_vsx_st(rearm2, 0,
540                 (vector unsigned char *)&pkts[2]->rearm_data);
541         vec_vsx_st(rearm3, 0,
542                 (vector unsigned char *)&pkts[3]->rearm_data);
543 }
544
545
546 /**
547  * Receive burst of packets. An errored completion also consumes a mbuf, but the
548  * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
549  * before returning to application.
550  *
551  * @param rxq
552  *   Pointer to RX queue structure.
553  * @param[out] pkts
554  *   Array to store received packets.
555  * @param pkts_n
556  *   Maximum number of packets in array.
557  * @param[out] err
558  *   Pointer to a flag. Set non-zero value if pkts array has at least one error
559  *   packet to handle.
560  *
561  * @return
562  *   Number of packets received including errors (<= pkts_n).
563  */
564 static inline uint16_t
565 rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
566             uint64_t *err)
567 {
568         const uint16_t q_n = 1 << rxq->cqe_n;
569         const uint16_t q_mask = q_n - 1;
570         volatile struct mlx5_cqe *cq;
571         struct rte_mbuf **elts;
572         unsigned int pos;
573         uint64_t n;
574         uint16_t repl_n;
575         uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
576         uint16_t nocmp_n = 0;
577         uint16_t rcvd_pkt = 0;
578         unsigned int cq_idx = rxq->cq_ci & q_mask;
579         unsigned int elts_idx;
580         unsigned int ownership = !!(rxq->cq_ci & (q_mask + 1));
581         const vector unsigned char zero = (vector unsigned char){0};
582         const vector unsigned char ones = vec_splat_u8(-1);
583         const vector unsigned char owner_check =
584                 (vector unsigned char)(vector unsigned long){
585                 0x0100000001000000LL, 0x0100000001000000LL};
586         const vector unsigned char opcode_check =
587                 (vector unsigned char)(vector unsigned long){
588                 0xf0000000f0000000LL, 0xf0000000f0000000LL};
589         const vector unsigned char format_check =
590                 (vector unsigned char)(vector unsigned long){
591                 0x0c0000000c000000LL, 0x0c0000000c000000LL};
592         const vector unsigned char resp_err_check =
593                 (vector unsigned char)(vector unsigned long){
594                 0xe0000000e0000000LL, 0xe0000000e0000000LL};
595 #ifdef MLX5_PMD_SOFT_COUNTERS
596         uint32_t rcvd_byte = 0;
597         /* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
598         const vector unsigned char len_shuf_mask = (vector unsigned char){
599                  1,  0,  5,  4,
600                  9,  8, 13, 12,
601                 -1, -1, -1, -1,
602                 -1, -1, -1, -1};
603 #endif
604         /* Mask to shuffle from extracted CQE to mbuf. */
605         const vector unsigned char shuf_mask = (vector unsigned char){
606                  5,  4,           /* bswap16, pkt_len */
607                 -1, -1,           /* zero out 2nd half of pkt_len */
608                  5,  4,           /* bswap16, data_len */
609                 11, 10,           /* bswap16, vlan+tci */
610                 15, 14, 13, 12,   /* bswap32, rss */
611                  1,  2,  3, -1};  /* fdir.hi */
612         /* Mask to blend from the last Qword to the first DQword. */
613         /* Mask to blend from the last Qword to the first DQword. */
614         const vector unsigned char blend_mask = (vector unsigned char){
615                 -1,  0,  0,  0,
616                  0,  0,  0,  0,
617                 -1, -1, -1, -1,
618                 -1, -1, -1, -1};
619         const vector unsigned char crc_adj =
620                 (vector unsigned char)(vector unsigned short){
621                 rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
622                 rxq->crc_present * RTE_ETHER_CRC_LEN, 0, 0, 0, 0, 0};
623         const vector unsigned char flow_mark_adj =
624                 (vector unsigned char)(vector unsigned int){
625                 0, 0, 0, rxq->mark * (-1)};
626         const vector unsigned short cqe_sel_mask1 =
627                 (vector unsigned short){0, 0, 0, 0, 0xffff, 0xffff, 0, 0};
628         const vector unsigned short cqe_sel_mask2 =
629                 (vector unsigned short){0, 0, 0xffff, 0, 0, 0, 0, 0};
630
631         MLX5_ASSERT(rxq->sges_n == 0);
632         MLX5_ASSERT(rxq->cqe_n == rxq->elts_n);
633         cq = &(*rxq->cqes)[cq_idx];
634         rte_prefetch0(cq);
635         rte_prefetch0(cq + 1);
636         rte_prefetch0(cq + 2);
637         rte_prefetch0(cq + 3);
638         pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
639
640         repl_n = q_n - (rxq->rq_ci - rxq->rq_pi);
641         if (repl_n >= rxq->rq_repl_thresh)
642                 mlx5_rx_replenish_bulk_mbuf(rxq, repl_n);
643         /* See if there're unreturned mbufs from compressed CQE. */
644         rcvd_pkt = rxq->decompressed;
645         if (rcvd_pkt > 0) {
646                 rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
647                 rxq_copy_mbuf_v(rxq, pkts, rcvd_pkt);
648                 rxq->rq_pi += rcvd_pkt;
649                 rxq->decompressed -= rcvd_pkt;
650                 pkts += rcvd_pkt;
651         }
652         elts_idx = rxq->rq_pi & q_mask;
653         elts = &(*rxq->elts)[elts_idx];
654         /* Not to overflow pkts array. */
655         pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
656         /* Not to cross queue end. */
657         pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
658         pkts_n = RTE_MIN(pkts_n, q_n - cq_idx);
659         if (!pkts_n)
660                 return rcvd_pkt;
661         /* At this point, there shouldn't be any remaining packets. */
662         MLX5_ASSERT(rxq->decompressed == 0);
663
664         /*
665          * A. load first Qword (8bytes) in one loop.
666          * B. copy 4 mbuf pointers from elts ring to returing pkts.
667          * C. load remaining CQE data and extract necessary fields.
668          *    Final 16bytes cqes[] extracted from original 64bytes CQE has the
669          *    following structure:
670          *        struct {
671          *          uint8_t  pkt_info;
672          *          uint8_t  flow_tag[3];
673          *          uint16_t byte_cnt;
674          *          uint8_t  rsvd4;
675          *          uint8_t  op_own;
676          *          uint16_t hdr_type_etc;
677          *          uint16_t vlan_info;
678          *          uint32_t rx_has_res;
679          *        } c;
680          * D. fill in mbuf.
681          * E. get valid CQEs.
682          * F. find compressed CQE.
683          */
684         for (pos = 0;
685              pos < pkts_n;
686              pos += MLX5_VPMD_DESCS_PER_LOOP) {
687                 vector unsigned char cqes[MLX5_VPMD_DESCS_PER_LOOP];
688                 vector unsigned char cqe_tmp1, cqe_tmp2;
689                 vector unsigned char pkt_mb0, pkt_mb1, pkt_mb2, pkt_mb3;
690                 vector unsigned char op_own, op_own_tmp1, op_own_tmp2;
691                 vector unsigned char opcode, owner_mask, invalid_mask;
692                 vector unsigned char comp_mask;
693                 vector unsigned char mask;
694 #ifdef MLX5_PMD_SOFT_COUNTERS
695                 const vector unsigned char lower_half = {
696                         0, 1, 4, 5, 8, 9, 12, 13,
697                         16, 17, 20, 21, 24, 25, 28, 29};
698                 const vector unsigned char upper_half = {
699                         2, 3, 6, 7, 10, 11, 14, 15,
700                         18, 19, 22, 23, 26, 27, 30, 31};
701                 const vector unsigned long shmax = {64, 64};
702                 vector unsigned char byte_cnt;
703                 vector unsigned short left, right;
704                 vector unsigned long lshift;
705                 vector __attribute__((altivec(bool__)))
706                         unsigned long shmask;
707 #endif
708                 vector unsigned char mbp1, mbp2;
709                 vector unsigned char p =
710                         (vector unsigned char)(vector unsigned short){
711                                 0, 1, 2, 3, 0, 0, 0, 0};
712                 unsigned int p1, p2, p3;
713
714                 /* Prefetch next 4 CQEs. */
715                 if (pkts_n - pos >= 2 * MLX5_VPMD_DESCS_PER_LOOP) {
716                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP]);
717                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 1]);
718                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 2]);
719                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 3]);
720                 }
721
722                 /* A.0 do not cross the end of CQ. */
723                 mask = (vector unsigned char)(vector unsigned long){
724                         (pkts_n - pos) * sizeof(uint16_t) * 8, 0};
725
726                 {
727                         vector unsigned long lshift;
728                         vector __attribute__((altivec(bool__)))
729                                 unsigned long shmask;
730                         const vector unsigned long shmax = {64, 64};
731
732                         lshift = vec_splat((vector unsigned long)mask, 0);
733                         shmask = vec_cmpgt(shmax, lshift);
734                         mask = (vector unsigned char)
735                                 vec_sl((vector unsigned long)ones, lshift);
736                         mask = (vector unsigned char)
737                                 vec_sel((vector unsigned long)shmask,
738                                 (vector unsigned long)mask, shmask);
739                 }
740
741                 p = (vector unsigned char)
742                         vec_andc((vector unsigned long)p,
743                         (vector unsigned long)mask);
744
745                 /* A.1 load cqes. */
746                 p3 = (unsigned int)((vector unsigned short)p)[3];
747                 cqes[3] = (vector unsigned char)(vector unsigned long){
748                         *(__rte_aligned(8) unsigned long *)
749                         &cq[pos + p3].sop_drop_qpn, 0LL};
750                 rte_compiler_barrier();
751
752                 p2 = (unsigned int)((vector unsigned short)p)[2];
753                 cqes[2] = (vector unsigned char)(vector unsigned long){
754                         *(__rte_aligned(8) unsigned long *)
755                         &cq[pos + p2].sop_drop_qpn, 0LL};
756                 rte_compiler_barrier();
757
758                 /* B.1 load mbuf pointers. */
759                 mbp1 = (vector unsigned char)vec_vsx_ld(0,
760                         (signed int const *)&elts[pos]);
761                 mbp2 = (vector unsigned char)vec_vsx_ld(0,
762                         (signed int const *)&elts[pos + 2]);
763
764                 /* A.1 load a block having op_own. */
765                 p1 = (unsigned int)((vector unsigned short)p)[1];
766                 cqes[1] = (vector unsigned char)(vector unsigned long){
767                         *(__rte_aligned(8) unsigned long *)
768                         &cq[pos + p1].sop_drop_qpn, 0LL};
769                 rte_compiler_barrier();
770
771                 cqes[0] = (vector unsigned char)(vector unsigned long){
772                         *(__rte_aligned(8) unsigned long *)
773                         &cq[pos].sop_drop_qpn, 0LL};
774                 rte_compiler_barrier();
775
776                 /* B.2 copy mbuf pointers. */
777                 *(vector unsigned char *)&pkts[pos] = mbp1;
778                 *(vector unsigned char *)&pkts[pos + 2] = mbp2;
779                 rte_cio_rmb();
780
781                 /* C.1 load remaining CQE data and extract necessary fields. */
782                 cqe_tmp2 = *(vector unsigned char *)
783                         &cq[pos + p3].pkt_info;
784                 cqe_tmp1 = *(vector unsigned char *)
785                         &cq[pos + p2].pkt_info;
786                 cqes[3] = vec_sel(cqes[3], cqe_tmp2, blend_mask);
787                 cqes[2] = vec_sel(cqes[2], cqe_tmp1, blend_mask);
788                 cqe_tmp2 = (vector unsigned char)vec_vsx_ld(0,
789                         (signed int const *)&cq[pos + p3].csum);
790                 cqe_tmp1 = (vector unsigned char)vec_vsx_ld(0,
791                         (signed int const *)&cq[pos + p2].csum);
792                 cqes[3] = (vector unsigned char)
793                         vec_sel((vector unsigned short)cqes[3],
794                         (vector unsigned short)cqe_tmp2, cqe_sel_mask1);
795                 cqes[2] = (vector unsigned char)
796                         vec_sel((vector unsigned short)cqes[2],
797                         (vector unsigned short)cqe_tmp1, cqe_sel_mask1);
798                 cqe_tmp2 = (vector unsigned char)(vector unsigned long){
799                         *(__rte_aligned(8) unsigned long *)
800                         &cq[pos + p3].rsvd3[9], 0LL};
801                 cqe_tmp1 = (vector unsigned char)(vector unsigned long){
802                         *(__rte_aligned(8) unsigned long *)
803                         &cq[pos + p2].rsvd3[9], 0LL};
804                 cqes[3] = (vector unsigned char)
805                         vec_sel((vector unsigned short)cqes[3],
806                         (vector unsigned short)cqe_tmp2,
807                         (vector unsigned short)cqe_sel_mask2);
808                 cqes[2] = (vector unsigned char)
809                         vec_sel((vector unsigned short)cqes[2],
810                         (vector unsigned short)cqe_tmp1,
811                         (vector unsigned short)cqe_sel_mask2);
812
813                 /* C.2 generate final structure for mbuf with swapping bytes. */
814                 pkt_mb3 = vec_perm(cqes[3], zero, shuf_mask);
815                 pkt_mb2 = vec_perm(cqes[2], zero, shuf_mask);
816
817                 /* C.3 adjust CRC length. */
818                 pkt_mb3 = (vector unsigned char)
819                         ((vector unsigned short)pkt_mb3 -
820                         (vector unsigned short)crc_adj);
821                 pkt_mb2 = (vector unsigned char)
822                         ((vector unsigned short)pkt_mb2 -
823                         (vector unsigned short)crc_adj);
824
825                 /* C.4 adjust flow mark. */
826                 pkt_mb3 = (vector unsigned char)
827                         ((vector unsigned int)pkt_mb3 +
828                         (vector unsigned int)flow_mark_adj);
829                 pkt_mb2 = (vector unsigned char)
830                         ((vector unsigned int)pkt_mb2 +
831                         (vector unsigned int)flow_mark_adj);
832
833                 /* D.1 fill in mbuf - rx_descriptor_fields1. */
834                 *(vector unsigned char *)
835                         &pkts[pos + 3]->pkt_len = pkt_mb3;
836                 *(vector unsigned char *)
837                         &pkts[pos + 2]->pkt_len = pkt_mb2;
838
839                 /* E.1 extract op_own field. */
840                 op_own_tmp2 = (vector unsigned char)
841                         vec_mergeh((vector unsigned int)cqes[2],
842                         (vector unsigned int)cqes[3]);
843
844                 /* C.1 load remaining CQE data and extract necessary fields. */
845                 cqe_tmp2 = *(vector unsigned char *)
846                         &cq[pos + p1].pkt_info;
847                 cqe_tmp1 = *(vector unsigned char *)
848                         &cq[pos].pkt_info;
849                 cqes[1] = vec_sel(cqes[1], cqe_tmp2, blend_mask);
850                 cqes[0] = vec_sel(cqes[0], cqe_tmp2, blend_mask);
851                 cqe_tmp2 = (vector unsigned char)vec_vsx_ld(0,
852                         (signed int const *)&cq[pos + p1].csum);
853                 cqe_tmp1 = (vector unsigned char)vec_vsx_ld(0,
854                         (signed int const *)&cq[pos].csum);
855                 cqes[1] = (vector unsigned char)
856                         vec_sel((vector unsigned short)cqes[1],
857                         (vector unsigned short)cqe_tmp2, cqe_sel_mask1);
858                 cqes[0] = (vector unsigned char)
859                         vec_sel((vector unsigned short)cqes[0],
860                         (vector unsigned short)cqe_tmp1, cqe_sel_mask1);
861                 cqe_tmp2 = (vector unsigned char)(vector unsigned long){
862                         *(__rte_aligned(8) unsigned long *)
863                         &cq[pos + p1].rsvd3[9], 0LL};
864                 cqe_tmp1 = (vector unsigned char)(vector unsigned long){
865                         *(__rte_aligned(8) unsigned long *)
866                         &cq[pos].rsvd3[9], 0LL};
867                 cqes[1] = (vector unsigned char)
868                         vec_sel((vector unsigned short)cqes[1],
869                         (vector unsigned short)cqe_tmp2, cqe_sel_mask2);
870                 cqes[0] = (vector unsigned char)
871                         vec_sel((vector unsigned short)cqes[0],
872                         (vector unsigned short)cqe_tmp1, cqe_sel_mask2);
873
874                 /* C.2 generate final structure for mbuf with swapping bytes. */
875                 pkt_mb1 = vec_perm(cqes[1], zero, shuf_mask);
876                 pkt_mb0 = vec_perm(cqes[0], zero, shuf_mask);
877
878                 /* C.3 adjust CRC length. */
879                 pkt_mb1 = (vector unsigned char)
880                         ((vector unsigned short)pkt_mb1 -
881                         (vector unsigned short)crc_adj);
882                 pkt_mb0 = (vector unsigned char)
883                         ((vector unsigned short)pkt_mb0 -
884                         (vector unsigned short)crc_adj);
885
886                 /* C.4 adjust flow mark. */
887                 pkt_mb1 = (vector unsigned char)
888                         ((vector unsigned int)pkt_mb1 +
889                         (vector unsigned int)flow_mark_adj);
890                 pkt_mb0 = (vector unsigned char)
891                         ((vector unsigned int)pkt_mb0 +
892                         (vector unsigned int)flow_mark_adj);
893
894                 /* E.1 extract op_own byte. */
895                 op_own_tmp1 = (vector unsigned char)
896                         vec_mergeh((vector unsigned int)cqes[0],
897                         (vector unsigned int)cqes[1]);
898                 op_own = (vector unsigned char)
899                         vec_mergel((vector unsigned long)op_own_tmp1,
900                         (vector unsigned long)op_own_tmp2);
901
902                 /* D.1 fill in mbuf - rx_descriptor_fields1. */
903                 *(vector unsigned char *)
904                         &pkts[pos + 1]->pkt_len = pkt_mb1;
905                 *(vector unsigned char *)
906                         &pkts[pos]->pkt_len = pkt_mb0;
907
908                 /* E.2 flip owner bit to mark CQEs from last round. */
909                 owner_mask = (vector unsigned char)
910                         vec_and((vector unsigned long)op_own,
911                         (vector unsigned long)owner_check);
912                 if (ownership)
913                         owner_mask = (vector unsigned char)
914                                 vec_xor((vector unsigned long)owner_mask,
915                                 (vector unsigned long)owner_check);
916                 owner_mask = (vector unsigned char)
917                         vec_cmpeq((vector unsigned int)owner_mask,
918                         (vector unsigned int)owner_check);
919                 owner_mask = (vector unsigned char)
920                         vec_packs((vector unsigned int)owner_mask,
921                         (vector unsigned int)zero);
922
923                 /* E.3 get mask for invalidated CQEs. */
924                 opcode = (vector unsigned char)
925                         vec_and((vector unsigned long)op_own,
926                         (vector unsigned long)opcode_check);
927                 invalid_mask = (vector unsigned char)
928                         vec_cmpeq((vector unsigned int)opcode_check,
929                         (vector unsigned int)opcode);
930                 invalid_mask = (vector unsigned char)
931                         vec_packs((vector unsigned int)invalid_mask,
932                         (vector unsigned int)zero);
933
934                 /* E.4 mask out beyond boundary. */
935                 invalid_mask = (vector unsigned char)
936                         vec_or((vector unsigned long)invalid_mask,
937                         (vector unsigned long)mask);
938
939                 /* E.5 merge invalid_mask with invalid owner. */
940                 invalid_mask = (vector unsigned char)
941                         vec_or((vector unsigned long)invalid_mask,
942                         (vector unsigned long)owner_mask);
943
944                 /* F.1 find compressed CQE format. */
945                 comp_mask = (vector unsigned char)
946                         vec_and((vector unsigned long)op_own,
947                         (vector unsigned long)format_check);
948                 comp_mask = (vector unsigned char)
949                         vec_cmpeq((vector unsigned int)comp_mask,
950                         (vector unsigned int)format_check);
951                 comp_mask = (vector unsigned char)
952                         vec_packs((vector unsigned int)comp_mask,
953                         (vector unsigned int)zero);
954
955                 /* F.2 mask out invalid entries. */
956                 comp_mask = (vector unsigned char)
957                         vec_andc((vector unsigned long)comp_mask,
958                         (vector unsigned long)invalid_mask);
959                 comp_idx = ((vector unsigned long)comp_mask)[0];
960
961                 /* F.3 get the first compressed CQE. */
962                 comp_idx = comp_idx ? __builtin_ctzll(comp_idx) /
963                         (sizeof(uint16_t) * 8) : MLX5_VPMD_DESCS_PER_LOOP;
964
965                 /* E.6 mask out entries after the compressed CQE. */
966                 mask = (vector unsigned char)(vector unsigned long){
967                         (comp_idx * sizeof(uint16_t) * 8), 0};
968                 lshift = vec_splat((vector unsigned long)mask, 0);
969                 shmask = vec_cmpgt(shmax, lshift);
970                 mask = (vector unsigned char)
971                         vec_sl((vector unsigned long)ones, lshift);
972                 mask = (vector unsigned char)
973                         vec_sel((vector unsigned long)shmask,
974                         (vector unsigned long)mask, shmask);
975                 invalid_mask = (vector unsigned char)
976                         vec_or((vector unsigned long)invalid_mask,
977                         (vector unsigned long)mask);
978
979                 /* E.7 count non-compressed valid CQEs. */
980                 n = ((vector unsigned long)invalid_mask)[0];
981                 n = n ? __builtin_ctzll(n) / (sizeof(uint16_t) * 8) :
982                         MLX5_VPMD_DESCS_PER_LOOP;
983                 nocmp_n += n;
984
985                 /* D.2 get the final invalid mask. */
986                 mask = (vector unsigned char)(vector unsigned long){
987                         (n * sizeof(uint16_t) * 8), 0};
988                 lshift = vec_splat((vector unsigned long)mask, 0);
989                 shmask = vec_cmpgt(shmax, lshift);
990                 mask = (vector unsigned char)
991                         vec_sl((vector unsigned long)ones, lshift);
992                 mask = (vector unsigned char)
993                         vec_sel((vector unsigned long)shmask,
994                         (vector unsigned long)mask, shmask);
995                 invalid_mask = (vector unsigned char)
996                         vec_or((vector unsigned long)invalid_mask,
997                         (vector unsigned long)mask);
998
999                 /* D.3 check error in opcode. */
1000                 opcode = (vector unsigned char)
1001                         vec_cmpeq((vector unsigned int)resp_err_check,
1002                         (vector unsigned int)opcode);
1003                 opcode = (vector unsigned char)
1004                         vec_packs((vector unsigned int)opcode,
1005                         (vector unsigned int)zero);
1006                 opcode = (vector unsigned char)
1007                         vec_andc((vector unsigned long)opcode,
1008                         (vector unsigned long)invalid_mask);
1009
1010                 /* D.4 mark if any error is set */
1011                 *err |= ((vector unsigned long)opcode)[0];
1012
1013                 /* D.5 fill in mbuf - rearm_data and packet_type. */
1014                 rxq_cq_to_ptype_oflags_v(rxq, cqes, opcode, &pkts[pos]);
1015                 if (rxq->hw_timestamp) {
1016                         pkts[pos]->timestamp =
1017                                 rte_be_to_cpu_64(cq[pos].timestamp);
1018                         pkts[pos + 1]->timestamp =
1019                                 rte_be_to_cpu_64(cq[pos + p1].timestamp);
1020                         pkts[pos + 2]->timestamp =
1021                                 rte_be_to_cpu_64(cq[pos + p2].timestamp);
1022                         pkts[pos + 3]->timestamp =
1023                                 rte_be_to_cpu_64(cq[pos + p3].timestamp);
1024                 }
1025                 if (rte_flow_dynf_metadata_avail()) {
1026                         uint64_t flag = rte_flow_dynf_metadata_mask;
1027                         int offs = rte_flow_dynf_metadata_offs;
1028                         uint32_t metadata;
1029
1030                         /* This code is subject for futher optimization. */
1031                         metadata = cq[pos].flow_table_metadata;
1032                         *RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *) =
1033                                                                 metadata;
1034                         pkts[pos]->ol_flags |= metadata ? flag : 0ULL;
1035                         metadata = cq[pos + 1].flow_table_metadata;
1036                         *RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *) =
1037                                                                 metadata;
1038                         pkts[pos + 1]->ol_flags |= metadata ? flag : 0ULL;
1039                         metadata = cq[pos + 2].flow_table_metadata;
1040                         *RTE_MBUF_DYNFIELD(pkts[pos + 2], offs, uint32_t *) =
1041                                                                 metadata;
1042                         pkts[pos + 2]->ol_flags |= metadata ? flag : 0ULL;
1043                         metadata = cq[pos + 3].flow_table_metadata;
1044                         *RTE_MBUF_DYNFIELD(pkts[pos + 3], offs, uint32_t *) =
1045                                                                 metadata;
1046                         pkts[pos + 3]->ol_flags |= metadata ? flag : 0ULL;
1047                 }
1048 #ifdef MLX5_PMD_SOFT_COUNTERS
1049                 /* Add up received bytes count. */
1050                 byte_cnt = vec_perm(op_own, zero, len_shuf_mask);
1051                 byte_cnt = (vector unsigned char)
1052                         vec_andc((vector unsigned long)byte_cnt,
1053                         (vector unsigned long)invalid_mask);
1054                 left = vec_perm((vector unsigned short)byte_cnt,
1055                         (vector unsigned short)zero, lower_half);
1056                 right = vec_perm((vector unsigned short)byte_cnt,
1057                         (vector unsigned short)zero, upper_half);
1058                 byte_cnt = (vector unsigned char)vec_add(left, right);
1059                 left = vec_perm((vector unsigned short)byte_cnt,
1060                         (vector unsigned short)zero, lower_half);
1061                 right = vec_perm((vector unsigned short)byte_cnt,
1062                         (vector unsigned short)zero, upper_half);
1063                 byte_cnt = (vector unsigned char)vec_add(left, right);
1064                 rcvd_byte += ((vector unsigned long)byte_cnt)[0];
1065 #endif
1066
1067                 /*
1068                  * Break the loop unless more valid CQE is expected, or if
1069                  * there's a compressed CQE.
1070                  */
1071                 if (n != MLX5_VPMD_DESCS_PER_LOOP)
1072                         break;
1073         }
1074         /* If no new CQE seen, return without updating cq_db. */
1075         if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP))
1076                 return rcvd_pkt;
1077         /* Update the consumer indexes for non-compressed CQEs. */
1078         MLX5_ASSERT(nocmp_n <= pkts_n);
1079         rxq->cq_ci += nocmp_n;
1080         rxq->rq_pi += nocmp_n;
1081         rcvd_pkt += nocmp_n;
1082 #ifdef MLX5_PMD_SOFT_COUNTERS
1083         rxq->stats.ipackets += nocmp_n;
1084         rxq->stats.ibytes += rcvd_byte;
1085 #endif
1086         /* Decompress the last CQE if compressed. */
1087         if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP && comp_idx == n) {
1088                 MLX5_ASSERT(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
1089                 rxq->decompressed =
1090                         rxq_cq_decompress_v(rxq, &cq[nocmp_n], &elts[nocmp_n]);
1091                 /* Return more packets if needed. */
1092                 if (nocmp_n < pkts_n) {
1093                         uint16_t n = rxq->decompressed;
1094
1095                         n = RTE_MIN(n, pkts_n - nocmp_n);
1096                         rxq_copy_mbuf_v(rxq, &pkts[nocmp_n], n);
1097                         rxq->rq_pi += n;
1098                         rcvd_pkt += n;
1099                         rxq->decompressed -= n;
1100                 }
1101         }
1102         rte_compiler_barrier();
1103         *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1104         return rcvd_pkt;
1105 }
1106
1107 #endif /* RTE_PMD_MLX5_RXTX_VEC_ALTIVEC_H_ */