net/mlx5: add Altivec Rx
[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.h"
21 #include "mlx5_utils.h"
22 #include "mlx5_rxtx.h"
23 #include "mlx5_rxtx_vec.h"
24 #include "mlx5_autoconf.h"
25 #include "mlx5_defs.h"
26 #include "mlx5_prm.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
267                 pos += MLX5_VPMD_DESCS_PER_LOOP;
268                 /* Move to next CQE and invalidate consumed CQEs. */
269                 if (!(pos & 0x7) && pos < mcqe_n) {
270                         mcq = (void *)&(cq + pos)->pkt_info;
271                         for (i = 0; i < 8; ++i)
272                                 cq[inv++].op_own = MLX5_CQE_INVALIDATE;
273                 }
274         }
275
276         /* Invalidate the rest of CQEs. */
277         for (; inv < mcqe_n; ++inv)
278                 cq[inv].op_own = MLX5_CQE_INVALIDATE;
279
280 #ifdef MLX5_PMD_SOFT_COUNTERS
281         rxq->stats.ipackets += mcqe_n;
282         rxq->stats.ibytes += rcvd_byte;
283 #endif
284
285         rxq->cq_ci += mcqe_n;
286         return mcqe_n;
287 }
288
289 /**
290  * Calculate packet type and offload flag for mbuf and store it.
291  *
292  * @param rxq
293  *   Pointer to RX queue structure.
294  * @param cqes[4]
295  *   Array of four 16bytes completions extracted from the original completion
296  *   descriptor.
297  * @param op_err
298  *   Opcode vector having responder error status. Each field is 4B.
299  * @param pkts
300  *   Pointer to array of packets to be filled.
301  */
302 static inline void
303 rxq_cq_to_ptype_oflags_v(struct mlx5_rxq_data *rxq,
304                 vector unsigned char cqes[4], vector unsigned char op_err,
305                 struct rte_mbuf **pkts)
306 {
307         vector unsigned char pinfo0, pinfo1;
308         vector unsigned char pinfo, ptype;
309         vector unsigned char ol_flags = (vector unsigned char)
310                 (vector unsigned int){
311                         rxq->rss_hash * PKT_RX_RSS_HASH |
312                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP,
313                         rxq->rss_hash * PKT_RX_RSS_HASH |
314                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP,
315                         rxq->rss_hash * PKT_RX_RSS_HASH |
316                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP,
317                         rxq->rss_hash * PKT_RX_RSS_HASH |
318                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP};
319         vector unsigned char cv_flags;
320         const vector unsigned char zero = (vector unsigned char){0};
321         const vector unsigned char ptype_mask =
322                 (vector unsigned char)(vector unsigned int){
323                 0x0000fd06, 0x0000fd06, 0x0000fd06, 0x0000fd06};
324         const vector unsigned char ptype_ol_mask =
325                 (vector unsigned char)(vector unsigned int){
326                 0x00000106, 0x00000106, 0x00000106, 0x00000106};
327         const vector unsigned char pinfo_mask =
328                 (vector unsigned char)(vector unsigned int){
329                 0x00000003, 0x00000003, 0x00000003, 0x00000003};
330         const vector unsigned char cv_flag_sel = (vector unsigned char){
331                 0, (uint8_t)(PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED),
332                 (uint8_t)(PKT_RX_IP_CKSUM_GOOD >> 1), 0,
333                 (uint8_t)(PKT_RX_L4_CKSUM_GOOD >> 1), 0,
334                 (uint8_t)((PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1),
335                 0, 0, 0, 0, 0, 0, 0, 0, 0};
336         const vector unsigned char cv_mask =
337                 (vector unsigned char)(vector unsigned int){
338                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
339                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
340                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
341                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
342                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
343                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
344                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
345                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED};
346         const vector unsigned char mbuf_init =
347                 (vector unsigned char)(vector unsigned long){
348                 *(__attribute__((__aligned__(8))) unsigned long *)
349                 &rxq->mbuf_initializer, 0LL};
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         /*
421          * Merge the two fields to generate the following:
422          * bit[1]     = l3_ok
423          * bit[2]     = l4_ok
424          * bit[8]     = cv
425          * bit[11:10] = l3_hdr_type
426          * bit[14:12] = l4_hdr_type
427          * bit[15]    = ip_frag
428          * bit[16]    = tunneled
429          * bit[17]    = outer_l3_type
430          */
431         ptype = (vector unsigned char)
432                 vec_and((vector unsigned long)ptype,
433                 (vector unsigned long)ptype_mask);
434         pinfo = (vector unsigned char)
435                 vec_and((vector unsigned long)pinfo,
436                 (vector unsigned long)pinfo_mask);
437         pinfo = (vector unsigned char)
438                 vec_sl((vector unsigned int)pinfo,
439                 (vector unsigned int){16, 16, 16, 16});
440
441         /* Make pinfo has merged fields for ol_flags calculation. */
442         pinfo = (vector unsigned char)
443                 vec_or((vector unsigned long)ptype,
444                 (vector unsigned long)pinfo);
445         ptype = (vector unsigned char)
446                 vec_sr((vector unsigned int)pinfo,
447                 (vector unsigned int){10, 10, 10, 10});
448         ptype = (vector unsigned char)
449                 vec_packs((vector unsigned int)ptype,
450                 (vector unsigned int)zero);
451
452         /* Errored packets will have RTE_PTYPE_ALL_MASK. */
453         op_err = (vector unsigned char)
454                 vec_sr((vector unsigned short)op_err,
455                 (vector unsigned short){8, 8, 8, 8, 8, 8, 8, 8});
456         ptype = (vector unsigned char)
457                 vec_or((vector unsigned long)ptype,
458                 (vector unsigned long)op_err);
459
460         pt_idx0 = (uint8_t)((vector unsigned char)ptype)[0];
461         pt_idx1 = (uint8_t)((vector unsigned char)ptype)[2];
462         pt_idx2 = (uint8_t)((vector unsigned char)ptype)[4];
463         pt_idx3 = (uint8_t)((vector unsigned char)ptype)[6];
464
465         pkts[0]->packet_type = mlx5_ptype_table[pt_idx0] |
466                 !!(pt_idx0 & (1 << 6)) * rxq->tunnel;
467         pkts[1]->packet_type = mlx5_ptype_table[pt_idx1] |
468                 !!(pt_idx1 & (1 << 6)) * rxq->tunnel;
469         pkts[2]->packet_type = mlx5_ptype_table[pt_idx2] |
470                 !!(pt_idx2 & (1 << 6)) * rxq->tunnel;
471         pkts[3]->packet_type = mlx5_ptype_table[pt_idx3] |
472                 !!(pt_idx3 & (1 << 6)) * rxq->tunnel;
473
474         /* Fill flags for checksum and VLAN. */
475         pinfo = (vector unsigned char)
476                 vec_and((vector unsigned long)pinfo,
477                 (vector unsigned long)ptype_ol_mask);
478         pinfo = vec_perm(cv_flag_sel, zero, pinfo);
479
480         /* Locate checksum flags at byte[2:1] and merge with VLAN flags. */
481         cv_flags = (vector unsigned char)
482                 vec_sl((vector unsigned int)pinfo,
483                 (vector unsigned int){9, 9, 9, 9});
484         cv_flags = (vector unsigned char)
485                 vec_or((vector unsigned long)pinfo,
486                 (vector unsigned long)cv_flags);
487
488         /* Move back flags to start from byte[0]. */
489         cv_flags = (vector unsigned char)
490                 vec_sr((vector unsigned int)cv_flags,
491                 (vector unsigned int){8, 8, 8, 8});
492
493         /* Mask out garbage bits. */
494         cv_flags = (vector unsigned char)
495                 vec_and((vector unsigned long)cv_flags,
496                 (vector unsigned long)cv_mask);
497
498         /* Merge to ol_flags. */
499         ol_flags = (vector unsigned char)
500                 vec_or((vector unsigned long)ol_flags,
501                 (vector unsigned long)cv_flags);
502
503         /* Merge mbuf_init and ol_flags. */
504         rearm0 = (vector unsigned char)
505                 vec_sel((vector unsigned short)mbuf_init,
506                 (vector unsigned short)
507                 vec_slo((vector unsigned short)ol_flags,
508                 (vector unsigned char){64}), rearm_sel_mask);
509         rearm1 = (vector unsigned char)
510                 vec_sel((vector unsigned short)mbuf_init,
511                 (vector unsigned short)
512                 vec_slo((vector unsigned short)ol_flags,
513                 (vector unsigned char){32}), rearm_sel_mask);
514         rearm2 = (vector unsigned char)
515                 vec_sel((vector unsigned short)mbuf_init,
516                 (vector unsigned short)ol_flags, rearm_sel_mask);
517         rearm3 = (vector unsigned char)
518                 vec_sel((vector unsigned short)mbuf_init,
519                 (vector unsigned short)
520                 vec_sro((vector unsigned short)ol_flags,
521                 (vector unsigned char){32}), rearm_sel_mask);
522
523         /* Write 8B rearm_data and 8B ol_flags. */
524         vec_vsx_st(rearm0, 0,
525                 (vector unsigned char *)&pkts[0]->rearm_data);
526         vec_vsx_st(rearm1, 0,
527                 (vector unsigned char *)&pkts[1]->rearm_data);
528         vec_vsx_st(rearm2, 0,
529                 (vector unsigned char *)&pkts[2]->rearm_data);
530         vec_vsx_st(rearm3, 0,
531                 (vector unsigned char *)&pkts[3]->rearm_data);
532 }
533
534
535 /**
536  * Receive burst of packets. An errored completion also consumes a mbuf, but the
537  * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
538  * before returning to application.
539  *
540  * @param rxq
541  *   Pointer to RX queue structure.
542  * @param[out] pkts
543  *   Array to store received packets.
544  * @param pkts_n
545  *   Maximum number of packets in array.
546  * @param[out] err
547  *   Pointer to a flag. Set non-zero value if pkts array has at least one error
548  *   packet to handle.
549  *
550  * @return
551  *   Number of packets received including errors (<= pkts_n).
552  */
553 static inline uint16_t
554 rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
555             uint64_t *err)
556 {
557         const uint16_t q_n = 1 << rxq->cqe_n;
558         const uint16_t q_mask = q_n - 1;
559         volatile struct mlx5_cqe *cq;
560         struct rte_mbuf **elts;
561         unsigned int pos;
562         uint64_t n;
563         uint16_t repl_n;
564         uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
565         uint16_t nocmp_n = 0;
566         uint16_t rcvd_pkt = 0;
567         unsigned int cq_idx = rxq->cq_ci & q_mask;
568         unsigned int elts_idx;
569         unsigned int ownership = !!(rxq->cq_ci & (q_mask + 1));
570         const vector unsigned char zero = (vector unsigned char){0};
571         const vector unsigned char ones = vec_splat_u8(-1);
572         const vector unsigned char owner_check =
573                 (vector unsigned char)(vector unsigned long){
574                 0x0100000001000000LL, 0x0100000001000000LL};
575         const vector unsigned char opcode_check =
576                 (vector unsigned char)(vector unsigned long){
577                 0xf0000000f0000000LL, 0xf0000000f0000000LL};
578         const vector unsigned char format_check =
579                 (vector unsigned char)(vector unsigned long){
580                 0x0c0000000c000000LL, 0x0c0000000c000000LL};
581         const vector unsigned char resp_err_check =
582                 (vector unsigned char)(vector unsigned long){
583                 0xe0000000e0000000LL, 0xe0000000e0000000LL};
584 #ifdef MLX5_PMD_SOFT_COUNTERS
585         uint32_t rcvd_byte = 0;
586         /* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
587         const vector unsigned char len_shuf_mask = (vector unsigned char){
588                  1,  0,  5,  4,
589                  9,  8, 13, 12,
590                 -1, -1, -1, -1,
591                 -1, -1, -1, -1};
592 #endif
593         /* Mask to shuffle from extracted CQE to mbuf. */
594         const vector unsigned char shuf_mask = (vector unsigned char){
595                  5,  4,           /* bswap16, pkt_len */
596                 -1, -1,           /* zero out 2nd half of pkt_len */
597                  5,  4,           /* bswap16, data_len */
598                 11, 10,           /* bswap16, vlan+tci */
599                 15, 14, 13, 12,   /* bswap32, rss */
600                  1,  2,  3, -1};  /* fdir.hi */
601         /* Mask to blend from the last Qword to the first DQword. */
602         /* Mask to blend from the last Qword to the first DQword. */
603         const vector unsigned char blend_mask = (vector unsigned char){
604                 -1,  0,  0,  0,
605                  0,  0,  0,  0,
606                 -1, -1, -1, -1,
607                 -1, -1, -1, -1};
608         const vector unsigned char crc_adj =
609                 (vector unsigned char)(vector unsigned short){
610                 rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
611                 rxq->crc_present * RTE_ETHER_CRC_LEN, 0, 0, 0, 0, 0};
612         const vector unsigned char flow_mark_adj =
613                 (vector unsigned char)(vector unsigned int){
614                 0, 0, 0, rxq->mark * (-1)};
615         const vector unsigned short cqe_sel_mask1 =
616                 (vector unsigned short){0, 0, 0, 0, 0xffff, 0xffff, 0, 0};
617         const vector unsigned short cqe_sel_mask2 =
618                 (vector unsigned short){0, 0, 0xffff, 0, 0, 0, 0, 0};
619
620         assert(rxq->sges_n == 0);
621         assert(rxq->cqe_n == rxq->elts_n);
622         cq = &(*rxq->cqes)[cq_idx];
623         rte_prefetch0(cq);
624         rte_prefetch0(cq + 1);
625         rte_prefetch0(cq + 2);
626         rte_prefetch0(cq + 3);
627         pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
628
629         repl_n = q_n - (rxq->rq_ci - rxq->rq_pi);
630         if (repl_n >= rxq->rq_repl_thresh)
631                 mlx5_rx_replenish_bulk_mbuf(rxq, repl_n);
632         /* See if there're unreturned mbufs from compressed CQE. */
633         rcvd_pkt = rxq->decompressed;
634         if (rcvd_pkt > 0) {
635                 rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
636                 rxq_copy_mbuf_v(rxq, pkts, rcvd_pkt);
637                 rxq->rq_pi += rcvd_pkt;
638                 rxq->decompressed -= rcvd_pkt;
639                 pkts += rcvd_pkt;
640         }
641         elts_idx = rxq->rq_pi & q_mask;
642         elts = &(*rxq->elts)[elts_idx];
643         /* Not to overflow pkts array. */
644         pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
645         /* Not to cross queue end. */
646         pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
647         pkts_n = RTE_MIN(pkts_n, q_n - cq_idx);
648         if (!pkts_n)
649                 return rcvd_pkt;
650         /* At this point, there shouldn't be any remaining packets. */
651         assert(rxq->decompressed == 0);
652
653         /*
654          * A. load first Qword (8bytes) in one loop.
655          * B. copy 4 mbuf pointers from elts ring to returing pkts.
656          * C. load remaining CQE data and extract necessary fields.
657          *    Final 16bytes cqes[] extracted from original 64bytes CQE has the
658          *    following structure:
659          *        struct {
660          *          uint8_t  pkt_info;
661          *          uint8_t  flow_tag[3];
662          *          uint16_t byte_cnt;
663          *          uint8_t  rsvd4;
664          *          uint8_t  op_own;
665          *          uint16_t hdr_type_etc;
666          *          uint16_t vlan_info;
667          *          uint32_t rx_has_res;
668          *        } c;
669          * D. fill in mbuf.
670          * E. get valid CQEs.
671          * F. find compressed CQE.
672          */
673         for (pos = 0;
674              pos < pkts_n;
675              pos += MLX5_VPMD_DESCS_PER_LOOP) {
676                 vector unsigned char cqes[MLX5_VPMD_DESCS_PER_LOOP];
677                 vector unsigned char cqe_tmp1, cqe_tmp2;
678                 vector unsigned char pkt_mb0, pkt_mb1, pkt_mb2, pkt_mb3;
679                 vector unsigned char op_own, op_own_tmp1, op_own_tmp2;
680                 vector unsigned char opcode, owner_mask, invalid_mask;
681                 vector unsigned char comp_mask;
682                 vector unsigned char mask;
683 #ifdef MLX5_PMD_SOFT_COUNTERS
684                 const vector unsigned char lower_half = {
685                         0, 1, 4, 5, 8, 9, 12, 13,
686                         16, 17, 20, 21, 24, 25, 28, 29};
687                 const vector unsigned char upper_half = {
688                         2, 3, 6, 7, 10, 11, 14, 15,
689                         18, 19, 22, 23, 26, 27, 30, 31};
690                 const vector unsigned long shmax = {64, 64};
691                 vector unsigned char byte_cnt;
692                 vector unsigned short left, right;
693                 vector unsigned long lshift;
694                 vector __attribute__((altivec(bool__)))
695                         unsigned long shmask;
696 #endif
697                 vector unsigned char mbp1, mbp2;
698                 vector unsigned char p =
699                         (vector unsigned char)(vector unsigned short){
700                                 0, 1, 2, 3, 0, 0, 0, 0};
701                 unsigned int p1, p2, p3;
702
703                 /* Prefetch next 4 CQEs. */
704                 if (pkts_n - pos >= 2 * MLX5_VPMD_DESCS_PER_LOOP) {
705                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP]);
706                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 1]);
707                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 2]);
708                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 3]);
709                 }
710
711                 /* A.0 do not cross the end of CQ. */
712                 mask = (vector unsigned char)(vector unsigned long){
713                         (pkts_n - pos) * sizeof(uint16_t) * 8, 0};
714
715                 {
716                         vector unsigned long lshift;
717                         vector __attribute__((altivec(bool__)))
718                                 unsigned long shmask;
719                         const vector unsigned long shmax = {64, 64};
720
721                         lshift = vec_splat((vector unsigned long)mask, 0);
722                         shmask = vec_cmpgt(shmax, lshift);
723                         mask = (vector unsigned char)
724                                 vec_sl((vector unsigned long)ones, lshift);
725                         mask = (vector unsigned char)
726                                 vec_sel((vector unsigned long)shmask,
727                                 (vector unsigned long)mask, shmask);
728                 }
729
730                 p = (vector unsigned char)
731                         vec_andc((vector unsigned long)p,
732                         (vector unsigned long)mask);
733
734                 /* A.1 load cqes. */
735                 p3 = (unsigned int)((vector unsigned short)p)[3];
736                 cqes[3] = (vector unsigned char)(vector unsigned long){
737                         *(__attribute__((__aligned__(8))) unsigned long *)
738                         &cq[pos + p3].sop_drop_qpn, 0LL};
739                 rte_compiler_barrier();
740
741                 p2 = (unsigned int)((vector unsigned short)p)[2];
742                 cqes[2] = (vector unsigned char)(vector unsigned long){
743                         *(__attribute__((__aligned__(8))) unsigned long *)
744                         &cq[pos + p2].sop_drop_qpn, 0LL};
745                 rte_compiler_barrier();
746
747                 /* B.1 load mbuf pointers. */
748                 mbp1 = (vector unsigned char)vec_vsx_ld(0,
749                         (signed int const *)&elts[pos]);
750                 mbp2 = (vector unsigned char)vec_vsx_ld(0,
751                         (signed int const *)&elts[pos + 2]);
752
753                 /* A.1 load a block having op_own. */
754                 p1 = (unsigned int)((vector unsigned short)p)[1];
755                 cqes[1] = (vector unsigned char)(vector unsigned long){
756                         *(__attribute__((__aligned__(8))) unsigned long *)
757                         &cq[pos + p1].sop_drop_qpn, 0LL};
758                 rte_compiler_barrier();
759
760                 cqes[0] = (vector unsigned char)(vector unsigned long){
761                         *(__attribute__((__aligned__(8))) unsigned long *)
762                         &cq[pos].sop_drop_qpn, 0LL};
763                 rte_compiler_barrier();
764
765                 /* B.2 copy mbuf pointers. */
766                 *(vector unsigned char *)&pkts[pos] = mbp1;
767                 *(vector unsigned char *)&pkts[pos + 2] = mbp2;
768                 rte_cio_rmb();
769
770                 /* C.1 load remaining CQE data and extract necessary fields. */
771                 cqe_tmp2 = *(vector unsigned char *)
772                         &cq[pos + p3].pkt_info;
773                 cqe_tmp1 = *(vector unsigned char *)
774                         &cq[pos + p2].pkt_info;
775                 cqes[3] = vec_sel(cqes[3], cqe_tmp2, blend_mask);
776                 cqes[2] = vec_sel(cqes[2], cqe_tmp1, blend_mask);
777                 cqe_tmp2 = (vector unsigned char)vec_vsx_ld(0,
778                         (signed int const *)&cq[pos + p3].csum);
779                 cqe_tmp1 = (vector unsigned char)vec_vsx_ld(0,
780                         (signed int const *)&cq[pos + p2].csum);
781                 cqes[3] = (vector unsigned char)
782                         vec_sel((vector unsigned short)cqes[3],
783                         (vector unsigned short)cqe_tmp2, cqe_sel_mask1);
784                 cqes[2] = (vector unsigned char)
785                         vec_sel((vector unsigned short)cqes[2],
786                         (vector unsigned short)cqe_tmp1, cqe_sel_mask1);
787                 cqe_tmp2 = (vector unsigned char)(vector unsigned long){
788                         *(__attribute__((__aligned__(8))) unsigned long *)
789                         &cq[pos + p3].rsvd3[9], 0LL};
790                 cqe_tmp1 = (vector unsigned char)(vector unsigned long){
791                         *(__attribute__((__aligned__(8))) unsigned long *)
792                         &cq[pos + p2].rsvd3[9], 0LL};
793                 cqes[3] = (vector unsigned char)
794                         vec_sel((vector unsigned short)cqes[3],
795                         (vector unsigned short)cqe_tmp2,
796                         (vector unsigned short)cqe_sel_mask2);
797                 cqes[2] = (vector unsigned char)
798                         vec_sel((vector unsigned short)cqes[2],
799                         (vector unsigned short)cqe_tmp1,
800                         (vector unsigned short)cqe_sel_mask2);
801
802                 /* C.2 generate final structure for mbuf with swapping bytes. */
803                 pkt_mb3 = vec_perm(cqes[3], zero, shuf_mask);
804                 pkt_mb2 = vec_perm(cqes[2], zero, shuf_mask);
805
806                 /* C.3 adjust CRC length. */
807                 pkt_mb3 = (vector unsigned char)
808                         ((vector unsigned short)pkt_mb3 -
809                         (vector unsigned short)crc_adj);
810                 pkt_mb2 = (vector unsigned char)
811                         ((vector unsigned short)pkt_mb2 -
812                         (vector unsigned short)crc_adj);
813
814                 /* C.4 adjust flow mark. */
815                 pkt_mb3 = (vector unsigned char)
816                         ((vector unsigned int)pkt_mb3 +
817                         (vector unsigned int)flow_mark_adj);
818                 pkt_mb2 = (vector unsigned char)
819                         ((vector unsigned int)pkt_mb2 +
820                         (vector unsigned int)flow_mark_adj);
821
822                 /* D.1 fill in mbuf - rx_descriptor_fields1. */
823                 *(vector unsigned char *)
824                         &pkts[pos + 3]->pkt_len = pkt_mb3;
825                 *(vector unsigned char *)
826                         &pkts[pos + 2]->pkt_len = pkt_mb2;
827
828                 /* E.1 extract op_own field. */
829                 op_own_tmp2 = (vector unsigned char)
830                         vec_mergeh((vector unsigned int)cqes[2],
831                         (vector unsigned int)cqes[3]);
832
833                 /* C.1 load remaining CQE data and extract necessary fields. */
834                 cqe_tmp2 = *(vector unsigned char *)
835                         &cq[pos + p1].pkt_info;
836                 cqe_tmp1 = *(vector unsigned char *)
837                         &cq[pos].pkt_info;
838                 cqes[1] = vec_sel(cqes[1], cqe_tmp2, blend_mask);
839                 cqes[0] = vec_sel(cqes[0], cqe_tmp2, blend_mask);
840                 cqe_tmp2 = (vector unsigned char)vec_vsx_ld(0,
841                         (signed int const *)&cq[pos + p1].csum);
842                 cqe_tmp1 = (vector unsigned char)vec_vsx_ld(0,
843                         (signed int const *)&cq[pos].csum);
844                 cqes[1] = (vector unsigned char)
845                         vec_sel((vector unsigned short)cqes[1],
846                         (vector unsigned short)cqe_tmp2, cqe_sel_mask1);
847                 cqes[0] = (vector unsigned char)
848                         vec_sel((vector unsigned short)cqes[0],
849                         (vector unsigned short)cqe_tmp1, cqe_sel_mask1);
850                 cqe_tmp2 = (vector unsigned char)(vector unsigned long){
851                         *(__attribute__((__aligned__(8))) unsigned long *)
852                         &cq[pos + p1].rsvd3[9], 0LL};
853                 cqe_tmp1 = (vector unsigned char)(vector unsigned long){
854                         *(__attribute__((__aligned__(8))) unsigned long *)
855                         &cq[pos].rsvd3[9], 0LL};
856                 cqes[1] = (vector unsigned char)
857                         vec_sel((vector unsigned short)cqes[1],
858                         (vector unsigned short)cqe_tmp2, cqe_sel_mask2);
859                 cqes[0] = (vector unsigned char)
860                         vec_sel((vector unsigned short)cqes[0],
861                         (vector unsigned short)cqe_tmp1, cqe_sel_mask2);
862
863                 /* C.2 generate final structure for mbuf with swapping bytes. */
864                 pkt_mb1 = vec_perm(cqes[1], zero, shuf_mask);
865                 pkt_mb0 = vec_perm(cqes[0], zero, shuf_mask);
866
867                 /* C.3 adjust CRC length. */
868                 pkt_mb1 = (vector unsigned char)
869                         ((vector unsigned short)pkt_mb1 -
870                         (vector unsigned short)crc_adj);
871                 pkt_mb0 = (vector unsigned char)
872                         ((vector unsigned short)pkt_mb0 -
873                         (vector unsigned short)crc_adj);
874
875                 /* C.4 adjust flow mark. */
876                 pkt_mb1 = (vector unsigned char)
877                         ((vector unsigned int)pkt_mb1 +
878                         (vector unsigned int)flow_mark_adj);
879                 pkt_mb0 = (vector unsigned char)
880                         ((vector unsigned int)pkt_mb0 +
881                         (vector unsigned int)flow_mark_adj);
882
883                 /* E.1 extract op_own byte. */
884                 op_own_tmp1 = (vector unsigned char)
885                         vec_mergeh((vector unsigned int)cqes[0],
886                         (vector unsigned int)cqes[1]);
887                 op_own = (vector unsigned char)
888                         vec_mergel((vector unsigned long)op_own_tmp1,
889                         (vector unsigned long)op_own_tmp2);
890
891                 /* D.1 fill in mbuf - rx_descriptor_fields1. */
892                 *(vector unsigned char *)
893                         &pkts[pos + 1]->pkt_len = pkt_mb1;
894                 *(vector unsigned char *)
895                         &pkts[pos]->pkt_len = pkt_mb0;
896
897                 /* E.2 flip owner bit to mark CQEs from last round. */
898                 owner_mask = (vector unsigned char)
899                         vec_and((vector unsigned long)op_own,
900                         (vector unsigned long)owner_check);
901                 if (ownership)
902                         owner_mask = (vector unsigned char)
903                                 vec_xor((vector unsigned long)owner_mask,
904                                 (vector unsigned long)owner_check);
905                 owner_mask = (vector unsigned char)
906                         vec_cmpeq((vector unsigned int)owner_mask,
907                         (vector unsigned int)owner_check);
908                 owner_mask = (vector unsigned char)
909                         vec_packs((vector unsigned int)owner_mask,
910                         (vector unsigned int)zero);
911
912                 /* E.3 get mask for invalidated CQEs. */
913                 opcode = (vector unsigned char)
914                         vec_and((vector unsigned long)op_own,
915                         (vector unsigned long)opcode_check);
916                 invalid_mask = (vector unsigned char)
917                         vec_cmpeq((vector unsigned int)opcode_check,
918                         (vector unsigned int)opcode);
919                 invalid_mask = (vector unsigned char)
920                         vec_packs((vector unsigned int)invalid_mask,
921                         (vector unsigned int)zero);
922
923                 /* E.4 mask out beyond boundary. */
924                 invalid_mask = (vector unsigned char)
925                         vec_or((vector unsigned long)invalid_mask,
926                         (vector unsigned long)mask);
927
928                 /* E.5 merge invalid_mask with invalid owner. */
929                 invalid_mask = (vector unsigned char)
930                         vec_or((vector unsigned long)invalid_mask,
931                         (vector unsigned long)owner_mask);
932
933                 /* F.1 find compressed CQE format. */
934                 comp_mask = (vector unsigned char)
935                         vec_and((vector unsigned long)op_own,
936                         (vector unsigned long)format_check);
937                 comp_mask = (vector unsigned char)
938                         vec_cmpeq((vector unsigned int)comp_mask,
939                         (vector unsigned int)format_check);
940                 comp_mask = (vector unsigned char)
941                         vec_packs((vector unsigned int)comp_mask,
942                         (vector unsigned int)zero);
943
944                 /* F.2 mask out invalid entries. */
945                 comp_mask = (vector unsigned char)
946                         vec_andc((vector unsigned long)comp_mask,
947                         (vector unsigned long)invalid_mask);
948                 comp_idx = ((vector unsigned long)comp_mask)[0];
949
950                 /* F.3 get the first compressed CQE. */
951                 comp_idx = comp_idx ? __builtin_ctzll(comp_idx) /
952                         (sizeof(uint16_t) * 8) : MLX5_VPMD_DESCS_PER_LOOP;
953
954                 /* E.6 mask out entries after the compressed CQE. */
955                 mask = (vector unsigned char)(vector unsigned long){
956                         (comp_idx * sizeof(uint16_t) * 8), 0};
957                 lshift = vec_splat((vector unsigned long)mask, 0);
958                 shmask = vec_cmpgt(shmax, lshift);
959                 mask = (vector unsigned char)
960                         vec_sl((vector unsigned long)ones, lshift);
961                 mask = (vector unsigned char)
962                         vec_sel((vector unsigned long)shmask,
963                         (vector unsigned long)mask, shmask);
964                 invalid_mask = (vector unsigned char)
965                         vec_or((vector unsigned long)invalid_mask,
966                         (vector unsigned long)mask);
967
968                 /* E.7 count non-compressed valid CQEs. */
969                 n = ((vector unsigned long)invalid_mask)[0];
970                 n = n ? __builtin_ctzll(n) / (sizeof(uint16_t) * 8) :
971                         MLX5_VPMD_DESCS_PER_LOOP;
972                 nocmp_n += n;
973
974                 /* D.2 get the final invalid mask. */
975                 mask = (vector unsigned char)(vector unsigned long){
976                         (n * sizeof(uint16_t) * 8), 0};
977                 lshift = vec_splat((vector unsigned long)mask, 0);
978                 shmask = vec_cmpgt(shmax, lshift);
979                 mask = (vector unsigned char)
980                         vec_sl((vector unsigned long)ones, lshift);
981                 mask = (vector unsigned char)
982                         vec_sel((vector unsigned long)shmask,
983                         (vector unsigned long)mask, shmask);
984                 invalid_mask = (vector unsigned char)
985                         vec_or((vector unsigned long)invalid_mask,
986                         (vector unsigned long)mask);
987
988                 /* D.3 check error in opcode. */
989                 opcode = (vector unsigned char)
990                         vec_cmpeq((vector unsigned int)resp_err_check,
991                         (vector unsigned int)opcode);
992                 opcode = (vector unsigned char)
993                         vec_packs((vector unsigned int)opcode,
994                         (vector unsigned int)zero);
995                 opcode = (vector unsigned char)
996                         vec_andc((vector unsigned long)opcode,
997                         (vector unsigned long)invalid_mask);
998
999                 /* D.4 mark if any error is set */
1000                 *err |= ((vector unsigned long)opcode)[0];
1001
1002                 /* D.5 fill in mbuf - rearm_data and packet_type. */
1003                 rxq_cq_to_ptype_oflags_v(rxq, cqes, opcode, &pkts[pos]);
1004                 if (rxq->hw_timestamp) {
1005                         pkts[pos]->timestamp =
1006                                 rte_be_to_cpu_64(cq[pos].timestamp);
1007                         pkts[pos + 1]->timestamp =
1008                                 rte_be_to_cpu_64(cq[pos + p1].timestamp);
1009                         pkts[pos + 2]->timestamp =
1010                                 rte_be_to_cpu_64(cq[pos + p2].timestamp);
1011                         pkts[pos + 3]->timestamp =
1012                                 rte_be_to_cpu_64(cq[pos + p3].timestamp);
1013                 }
1014
1015 #ifdef MLX5_PMD_SOFT_COUNTERS
1016                 /* Add up received bytes count. */
1017                 byte_cnt = vec_perm(op_own, zero, len_shuf_mask);
1018                 byte_cnt = (vector unsigned char)
1019                         vec_andc((vector unsigned long)byte_cnt,
1020                         (vector unsigned long)invalid_mask);
1021                 left = vec_perm((vector unsigned short)byte_cnt,
1022                         (vector unsigned short)zero, lower_half);
1023                 right = vec_perm((vector unsigned short)byte_cnt,
1024                         (vector unsigned short)zero, upper_half);
1025                 byte_cnt = (vector unsigned char)vec_add(left, right);
1026                 left = vec_perm((vector unsigned short)byte_cnt,
1027                         (vector unsigned short)zero, lower_half);
1028                 right = vec_perm((vector unsigned short)byte_cnt,
1029                         (vector unsigned short)zero, upper_half);
1030                 byte_cnt = (vector unsigned char)vec_add(left, right);
1031                 rcvd_byte += ((vector unsigned long)byte_cnt)[0];
1032 #endif
1033
1034                 /*
1035                  * Break the loop unless more valid CQE is expected, or if
1036                  * there's a compressed CQE.
1037                  */
1038                 if (n != MLX5_VPMD_DESCS_PER_LOOP)
1039                         break;
1040         }
1041         /* If no new CQE seen, return without updating cq_db. */
1042         if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP))
1043                 return rcvd_pkt;
1044         /* Update the consumer indexes for non-compressed CQEs. */
1045         assert(nocmp_n <= pkts_n);
1046         rxq->cq_ci += nocmp_n;
1047         rxq->rq_pi += nocmp_n;
1048         rcvd_pkt += nocmp_n;
1049 #ifdef MLX5_PMD_SOFT_COUNTERS
1050         rxq->stats.ipackets += nocmp_n;
1051         rxq->stats.ibytes += rcvd_byte;
1052 #endif
1053         /* Decompress the last CQE if compressed. */
1054         if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP && comp_idx == n) {
1055                 assert(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
1056                 rxq->decompressed =
1057                         rxq_cq_decompress_v(rxq, &cq[nocmp_n], &elts[nocmp_n]);
1058                 /* Return more packets if needed. */
1059                 if (nocmp_n < pkts_n) {
1060                         uint16_t n = rxq->decompressed;
1061
1062                         n = RTE_MIN(n, pkts_n - nocmp_n);
1063                         rxq_copy_mbuf_v(rxq, &pkts[nocmp_n], n);
1064                         rxq->rq_pi += n;
1065                         rcvd_pkt += n;
1066                         rxq->decompressed -= n;
1067                 }
1068         }
1069         rte_compiler_barrier();
1070         *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1071         return rcvd_pkt;
1072 }
1073
1074 #endif /* RTE_PMD_MLX5_RXTX_VEC_ALTIVEC_H_ */