net/mlx5: prefetch CQEs for a faster decompression
[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                 for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
159                         if (likely(pos + i < mcqe_n))
160                                 rte_prefetch0((void *)(cq + pos + i));
161
162                 /* A.1 load mCQEs into a 128bit register. */
163                 mcqe1 = (vector unsigned char)vec_vsx_ld(0,
164                         (signed int const *)&mcq[pos % 8]);
165                 mcqe2 = (vector unsigned char)vec_vsx_ld(0,
166                         (signed int const *)&mcq[pos % 8 + 2]);
167
168                 /* B.1 store rearm data to mbuf. */
169                 *(vector unsigned char *)
170                         &elts[pos]->rearm_data = rearm;
171                 *(vector unsigned char *)
172                         &elts[pos + 1]->rearm_data = rearm;
173
174                 /* C.1 combine data from mCQEs with rx_descriptor_fields1. */
175                 rxdf1 = vec_perm(mcqe1, zero, shuf_mask1);
176                 rxdf2 = vec_perm(mcqe1, zero, shuf_mask2);
177                 rxdf1 = (vector unsigned char)
178                         ((vector unsigned short)rxdf1 -
179                         (vector unsigned short)crc_adj);
180                 rxdf2 = (vector unsigned char)
181                         ((vector unsigned short)rxdf2 -
182                         (vector unsigned short)crc_adj);
183                 rxdf1 = (vector unsigned char)
184                         vec_sel((vector unsigned short)rxdf1,
185                         (vector unsigned short)rxdf, rxdf_sel_mask);
186                 rxdf2 = (vector unsigned char)
187                         vec_sel((vector unsigned short)rxdf2,
188                         (vector unsigned short)rxdf, rxdf_sel_mask);
189
190                 /* D.1 store rx_descriptor_fields1. */
191                 *(vector unsigned char *)
192                         &elts[pos]->rx_descriptor_fields1 = rxdf1;
193                 *(vector unsigned char *)
194                         &elts[pos + 1]->rx_descriptor_fields1 = rxdf2;
195
196                 /* B.1 store rearm data to mbuf. */
197                 *(vector unsigned char *)
198                         &elts[pos + 2]->rearm_data = rearm;
199                 *(vector unsigned char *)
200                         &elts[pos + 3]->rearm_data = rearm;
201
202                 /* C.1 combine data from mCQEs with rx_descriptor_fields1. */
203                 rxdf1 = vec_perm(mcqe2, zero, shuf_mask1);
204                 rxdf2 = vec_perm(mcqe2, zero, shuf_mask2);
205                 rxdf1 = (vector unsigned char)
206                         ((vector unsigned short)rxdf1 -
207                         (vector unsigned short)crc_adj);
208                 rxdf2 = (vector unsigned char)
209                         ((vector unsigned short)rxdf2 -
210                         (vector unsigned short)crc_adj);
211                 rxdf1 = (vector unsigned char)
212                         vec_sel((vector unsigned short)rxdf1,
213                         (vector unsigned short)rxdf, rxdf_sel_mask);
214                 rxdf2 = (vector unsigned char)
215                         vec_sel((vector unsigned short)rxdf2,
216                         (vector unsigned short)rxdf, rxdf_sel_mask);
217
218                 /* D.1 store rx_descriptor_fields1. */
219                 *(vector unsigned char *)
220                         &elts[pos + 2]->rx_descriptor_fields1 = rxdf1;
221                 *(vector unsigned char *)
222                         &elts[pos + 3]->rx_descriptor_fields1 = rxdf2;
223
224 #ifdef MLX5_PMD_SOFT_COUNTERS
225                 invalid_mask = (vector unsigned char)(vector unsigned long){
226                         (mcqe_n - pos) * sizeof(uint16_t) * 8, 0};
227
228                 lshift =
229                         vec_splat((vector unsigned long)invalid_mask, 0);
230                 shmask = vec_cmpgt(shmax, lshift);
231                 invalid_mask = (vector unsigned char)
232                         vec_sl((vector unsigned long)ones, lshift);
233                 invalid_mask = (vector unsigned char)
234                         vec_sel((vector unsigned long)shmask,
235                         (vector unsigned long)invalid_mask, shmask);
236
237                 mcqe1 = (vector unsigned char)
238                         vec_sro((vector unsigned short)mcqe1,
239                         (vector unsigned char){32}),
240                 byte_cnt = (vector unsigned char)
241                         vec_sel((vector unsigned short)mcqe1,
242                         (vector unsigned short)mcqe2, mcqe_sel_mask);
243                 byte_cnt = vec_perm(byte_cnt, zero, len_shuf_mask);
244                 byte_cnt = (vector unsigned char)
245                         vec_andc((vector unsigned long)byte_cnt,
246                         (vector unsigned long)invalid_mask);
247                 left = vec_perm((vector unsigned short)byte_cnt,
248                         (vector unsigned short)zero, lower_half);
249                 right = vec_perm((vector unsigned short)byte_cnt,
250                         (vector unsigned short)zero, upper_half);
251                 byte_cnt = (vector unsigned char)vec_add(left, right);
252                 left = vec_perm((vector unsigned short)byte_cnt,
253                         (vector unsigned short)zero, lower_half);
254                 right = vec_perm((vector unsigned short)byte_cnt,
255                         (vector unsigned short)zero, upper_half);
256                 byte_cnt = (vector unsigned char)vec_add(left, right);
257                 rcvd_byte += ((vector unsigned long)byte_cnt)[0];
258 #endif
259
260                 if (rxq->mark) {
261                         /* E.1 store flow tag (rte_flow mark). */
262                         elts[pos]->hash.fdir.hi = flow_tag;
263                         elts[pos + 1]->hash.fdir.hi = flow_tag;
264                         elts[pos + 2]->hash.fdir.hi = flow_tag;
265                         elts[pos + 3]->hash.fdir.hi = flow_tag;
266                 }
267                 if (rte_flow_dynf_metadata_avail()) {
268                         const uint32_t meta = *RTE_FLOW_DYNF_METADATA(t_pkt);
269
270                         /* Check if title packet has valid metadata. */
271                         if (meta) {
272                                 MLX5_ASSERT(t_pkt->ol_flags &
273                                             PKT_RX_DYNF_METADATA);
274                                 *RTE_FLOW_DYNF_METADATA(elts[pos]) = meta;
275                                 *RTE_FLOW_DYNF_METADATA(elts[pos + 1]) = meta;
276                                 *RTE_FLOW_DYNF_METADATA(elts[pos + 2]) = meta;
277                                 *RTE_FLOW_DYNF_METADATA(elts[pos + 3]) = meta;
278                         }
279                 }
280
281                 pos += MLX5_VPMD_DESCS_PER_LOOP;
282                 /* Move to next CQE and invalidate consumed CQEs. */
283                 if (!(pos & 0x7) && pos < mcqe_n) {
284                         mcq = (void *)&(cq + pos)->pkt_info;
285                         for (i = 0; i < 8; ++i)
286                                 cq[inv++].op_own = MLX5_CQE_INVALIDATE;
287                 }
288         }
289
290         /* Invalidate the rest of CQEs. */
291         for (; inv < mcqe_n; ++inv)
292                 cq[inv].op_own = MLX5_CQE_INVALIDATE;
293
294 #ifdef MLX5_PMD_SOFT_COUNTERS
295         rxq->stats.ipackets += mcqe_n;
296         rxq->stats.ibytes += rcvd_byte;
297 #endif
298
299         rxq->cq_ci += mcqe_n;
300         return mcqe_n;
301 }
302
303 /**
304  * Calculate packet type and offload flag for mbuf and store it.
305  *
306  * @param rxq
307  *   Pointer to RX queue structure.
308  * @param cqes[4]
309  *   Array of four 16bytes completions extracted from the original completion
310  *   descriptor.
311  * @param op_err
312  *   Opcode vector having responder error status. Each field is 4B.
313  * @param pkts
314  *   Pointer to array of packets to be filled.
315  */
316 static inline void
317 rxq_cq_to_ptype_oflags_v(struct mlx5_rxq_data *rxq,
318                 vector unsigned char cqes[4], vector unsigned char op_err,
319                 struct rte_mbuf **pkts)
320 {
321         vector unsigned char pinfo0, pinfo1;
322         vector unsigned char pinfo, ptype;
323         vector unsigned char ol_flags = (vector unsigned char)
324                 (vector unsigned int){
325                         rxq->rss_hash * PKT_RX_RSS_HASH |
326                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP,
327                         rxq->rss_hash * PKT_RX_RSS_HASH |
328                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP,
329                         rxq->rss_hash * PKT_RX_RSS_HASH |
330                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP,
331                         rxq->rss_hash * PKT_RX_RSS_HASH |
332                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP};
333         vector unsigned char cv_flags;
334         const vector unsigned char zero = (vector unsigned char){0};
335         const vector unsigned char ptype_mask =
336                 (vector unsigned char)(vector unsigned int){
337                 0x0000fd06, 0x0000fd06, 0x0000fd06, 0x0000fd06};
338         const vector unsigned char ptype_ol_mask =
339                 (vector unsigned char)(vector unsigned int){
340                 0x00000106, 0x00000106, 0x00000106, 0x00000106};
341         const vector unsigned char pinfo_mask =
342                 (vector unsigned char)(vector unsigned int){
343                 0x00000003, 0x00000003, 0x00000003, 0x00000003};
344         const vector unsigned char cv_flag_sel = (vector unsigned char){
345                 0, (uint8_t)(PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED),
346                 (uint8_t)(PKT_RX_IP_CKSUM_GOOD >> 1), 0,
347                 (uint8_t)(PKT_RX_L4_CKSUM_GOOD >> 1), 0,
348                 (uint8_t)((PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1),
349                 0, 0, 0, 0, 0, 0, 0, 0, 0};
350         const vector unsigned char cv_mask =
351                 (vector unsigned char)(vector unsigned int){
352                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
353                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
354                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
355                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
356                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
357                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
358                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
359                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED};
360         const vector unsigned char mbuf_init =
361                 (vector unsigned char)vec_vsx_ld
362                         (0, (vector unsigned char *)&rxq->mbuf_initializer);
363         const vector unsigned short rearm_sel_mask =
364                 (vector unsigned short){0, 0, 0, 0, 0xffff, 0xffff, 0, 0};
365         vector unsigned char rearm0, rearm1, rearm2, rearm3;
366         uint8_t pt_idx0, pt_idx1, pt_idx2, pt_idx3;
367
368         /* Extract pkt_info field. */
369         pinfo0 = (vector unsigned char)
370                 vec_mergeh((vector unsigned int)cqes[0],
371                 (vector unsigned int)cqes[1]);
372         pinfo1 = (vector unsigned char)
373                 vec_mergeh((vector unsigned int)cqes[2],
374                 (vector unsigned int)cqes[3]);
375         pinfo = (vector unsigned char)
376                 vec_mergeh((vector unsigned long)pinfo0,
377                 (vector unsigned long)pinfo1);
378
379         /* Extract hdr_type_etc field. */
380         pinfo0 = (vector unsigned char)
381                 vec_mergel((vector unsigned int)cqes[0],
382                 (vector unsigned int)cqes[1]);
383         pinfo1 = (vector unsigned char)
384                 vec_mergel((vector unsigned int)cqes[2],
385                 (vector unsigned int)cqes[3]);
386         ptype = (vector unsigned char)
387                 vec_mergeh((vector unsigned long)pinfo0,
388                 (vector unsigned long)pinfo1);
389
390         if (rxq->mark) {
391                 const vector unsigned char pinfo_ft_mask =
392                         (vector unsigned char)(vector unsigned int){
393                         0xffffff00, 0xffffff00, 0xffffff00, 0xffffff00};
394                 const vector unsigned char fdir_flags =
395                         (vector unsigned char)(vector unsigned int){
396                         PKT_RX_FDIR, PKT_RX_FDIR,
397                         PKT_RX_FDIR, PKT_RX_FDIR};
398                 vector unsigned char fdir_id_flags =
399                         (vector unsigned char)(vector unsigned int){
400                         PKT_RX_FDIR_ID, PKT_RX_FDIR_ID,
401                         PKT_RX_FDIR_ID, PKT_RX_FDIR_ID};
402                 vector unsigned char flow_tag, invalid_mask;
403
404                 flow_tag = (vector unsigned char)
405                         vec_and((vector unsigned long)pinfo,
406                         (vector unsigned long)pinfo_ft_mask);
407
408                 /* Check if flow tag is non-zero then set PKT_RX_FDIR. */
409                 invalid_mask = (vector unsigned char)
410                         vec_cmpeq((vector unsigned int)flow_tag,
411                         (vector unsigned int)zero);
412                 ol_flags = (vector unsigned char)
413                         vec_or((vector unsigned long)ol_flags,
414                         (vector unsigned long)
415                         vec_andc((vector unsigned long)fdir_flags,
416                         (vector unsigned long)invalid_mask));
417
418                 /* Mask out invalid entries. */
419                 fdir_id_flags = (vector unsigned char)
420                         vec_andc((vector unsigned long)fdir_id_flags,
421                         (vector unsigned long)invalid_mask);
422
423                 /* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */
424                 ol_flags = (vector unsigned char)
425                         vec_or((vector unsigned long)ol_flags,
426                         (vector unsigned long)
427                         vec_andc((vector unsigned long)fdir_id_flags,
428                         (vector unsigned long)
429                         vec_cmpeq((vector unsigned int)flow_tag,
430                         (vector unsigned int)pinfo_ft_mask)));
431         }
432         /*
433          * Merge the two fields to generate the following:
434          * bit[1]     = l3_ok
435          * bit[2]     = l4_ok
436          * bit[8]     = cv
437          * bit[11:10] = l3_hdr_type
438          * bit[14:12] = l4_hdr_type
439          * bit[15]    = ip_frag
440          * bit[16]    = tunneled
441          * bit[17]    = outer_l3_type
442          */
443         ptype = (vector unsigned char)
444                 vec_and((vector unsigned long)ptype,
445                 (vector unsigned long)ptype_mask);
446         pinfo = (vector unsigned char)
447                 vec_and((vector unsigned long)pinfo,
448                 (vector unsigned long)pinfo_mask);
449         pinfo = (vector unsigned char)
450                 vec_sl((vector unsigned int)pinfo,
451                 (vector unsigned int){16, 16, 16, 16});
452
453         /* Make pinfo has merged fields for ol_flags calculation. */
454         pinfo = (vector unsigned char)
455                 vec_or((vector unsigned long)ptype,
456                 (vector unsigned long)pinfo);
457         ptype = (vector unsigned char)
458                 vec_sr((vector unsigned int)pinfo,
459                 (vector unsigned int){10, 10, 10, 10});
460         ptype = (vector unsigned char)
461                 vec_packs((vector unsigned int)ptype,
462                 (vector unsigned int)zero);
463
464         /* Errored packets will have RTE_PTYPE_ALL_MASK. */
465         op_err = (vector unsigned char)
466                 vec_sr((vector unsigned short)op_err,
467                 (vector unsigned short){8, 8, 8, 8, 8, 8, 8, 8});
468         ptype = (vector unsigned char)
469                 vec_or((vector unsigned long)ptype,
470                 (vector unsigned long)op_err);
471
472         pt_idx0 = (uint8_t)((vector unsigned char)ptype)[0];
473         pt_idx1 = (uint8_t)((vector unsigned char)ptype)[2];
474         pt_idx2 = (uint8_t)((vector unsigned char)ptype)[4];
475         pt_idx3 = (uint8_t)((vector unsigned char)ptype)[6];
476
477         pkts[0]->packet_type = mlx5_ptype_table[pt_idx0] |
478                 !!(pt_idx0 & (1 << 6)) * rxq->tunnel;
479         pkts[1]->packet_type = mlx5_ptype_table[pt_idx1] |
480                 !!(pt_idx1 & (1 << 6)) * rxq->tunnel;
481         pkts[2]->packet_type = mlx5_ptype_table[pt_idx2] |
482                 !!(pt_idx2 & (1 << 6)) * rxq->tunnel;
483         pkts[3]->packet_type = mlx5_ptype_table[pt_idx3] |
484                 !!(pt_idx3 & (1 << 6)) * rxq->tunnel;
485
486         /* Fill flags for checksum and VLAN. */
487         pinfo = (vector unsigned char)
488                 vec_and((vector unsigned long)pinfo,
489                 (vector unsigned long)ptype_ol_mask);
490         pinfo = vec_perm(cv_flag_sel, zero, pinfo);
491
492         /* Locate checksum flags at byte[2:1] and merge with VLAN flags. */
493         cv_flags = (vector unsigned char)
494                 vec_sl((vector unsigned int)pinfo,
495                 (vector unsigned int){9, 9, 9, 9});
496         cv_flags = (vector unsigned char)
497                 vec_or((vector unsigned long)pinfo,
498                 (vector unsigned long)cv_flags);
499
500         /* Move back flags to start from byte[0]. */
501         cv_flags = (vector unsigned char)
502                 vec_sr((vector unsigned int)cv_flags,
503                 (vector unsigned int){8, 8, 8, 8});
504
505         /* Mask out garbage bits. */
506         cv_flags = (vector unsigned char)
507                 vec_and((vector unsigned long)cv_flags,
508                 (vector unsigned long)cv_mask);
509
510         /* Merge to ol_flags. */
511         ol_flags = (vector unsigned char)
512                 vec_or((vector unsigned long)ol_flags,
513                 (vector unsigned long)cv_flags);
514
515         /* Merge mbuf_init and ol_flags. */
516         rearm0 = (vector unsigned char)
517                 vec_sel((vector unsigned short)mbuf_init,
518                 (vector unsigned short)
519                 vec_slo((vector unsigned short)ol_flags,
520                 (vector unsigned char){64}), rearm_sel_mask);
521         rearm1 = (vector unsigned char)
522                 vec_sel((vector unsigned short)mbuf_init,
523                 (vector unsigned short)
524                 vec_slo((vector unsigned short)ol_flags,
525                 (vector unsigned char){32}), rearm_sel_mask);
526         rearm2 = (vector unsigned char)
527                 vec_sel((vector unsigned short)mbuf_init,
528                 (vector unsigned short)ol_flags, rearm_sel_mask);
529         rearm3 = (vector unsigned char)
530                 vec_sel((vector unsigned short)mbuf_init,
531                 (vector unsigned short)
532                 vec_sro((vector unsigned short)ol_flags,
533                 (vector unsigned char){32}), rearm_sel_mask);
534
535         /* Write 8B rearm_data and 8B ol_flags. */
536         vec_vsx_st(rearm0, 0,
537                 (vector unsigned char *)&pkts[0]->rearm_data);
538         vec_vsx_st(rearm1, 0,
539                 (vector unsigned char *)&pkts[1]->rearm_data);
540         vec_vsx_st(rearm2, 0,
541                 (vector unsigned char *)&pkts[2]->rearm_data);
542         vec_vsx_st(rearm3, 0,
543                 (vector unsigned char *)&pkts[3]->rearm_data);
544 }
545
546
547 /**
548  * Receive burst of packets. An errored completion also consumes a mbuf, but the
549  * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
550  * before returning to application.
551  *
552  * @param rxq
553  *   Pointer to RX queue structure.
554  * @param[out] pkts
555  *   Array to store received packets.
556  * @param pkts_n
557  *   Maximum number of packets in array.
558  * @param[out] err
559  *   Pointer to a flag. Set non-zero value if pkts array has at least one error
560  *   packet to handle.
561  *
562  * @return
563  *   Number of packets received including errors (<= pkts_n).
564  */
565 static inline uint16_t
566 rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
567             uint64_t *err)
568 {
569         const uint16_t q_n = 1 << rxq->cqe_n;
570         const uint16_t q_mask = q_n - 1;
571         volatile struct mlx5_cqe *cq;
572         struct rte_mbuf **elts;
573         unsigned int pos;
574         uint64_t n;
575         uint16_t repl_n;
576         uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
577         uint16_t nocmp_n = 0;
578         uint16_t rcvd_pkt = 0;
579         unsigned int cq_idx = rxq->cq_ci & q_mask;
580         unsigned int elts_idx;
581         unsigned int ownership = !!(rxq->cq_ci & (q_mask + 1));
582         const vector unsigned char zero = (vector unsigned char){0};
583         const vector unsigned char ones = vec_splat_u8(-1);
584         const vector unsigned char owner_check =
585                 (vector unsigned char)(vector unsigned long){
586                 0x0100000001000000LL, 0x0100000001000000LL};
587         const vector unsigned char opcode_check =
588                 (vector unsigned char)(vector unsigned long){
589                 0xf0000000f0000000LL, 0xf0000000f0000000LL};
590         const vector unsigned char format_check =
591                 (vector unsigned char)(vector unsigned long){
592                 0x0c0000000c000000LL, 0x0c0000000c000000LL};
593         const vector unsigned char resp_err_check =
594                 (vector unsigned char)(vector unsigned long){
595                 0xe0000000e0000000LL, 0xe0000000e0000000LL};
596 #ifdef MLX5_PMD_SOFT_COUNTERS
597         uint32_t rcvd_byte = 0;
598         /* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
599         const vector unsigned char len_shuf_mask = (vector unsigned char){
600                  1,  0,  5,  4,
601                  9,  8, 13, 12,
602                 -1, -1, -1, -1,
603                 -1, -1, -1, -1};
604 #endif
605         /* Mask to shuffle from extracted CQE to mbuf. */
606         const vector unsigned char shuf_mask = (vector unsigned char){
607                  5,  4,           /* bswap16, pkt_len */
608                 -1, -1,           /* zero out 2nd half of pkt_len */
609                  5,  4,           /* bswap16, data_len */
610                 11, 10,           /* bswap16, vlan+tci */
611                 15, 14, 13, 12,   /* bswap32, rss */
612                  1,  2,  3, -1};  /* fdir.hi */
613         /* Mask to blend from the last Qword to the first DQword. */
614         /* Mask to blend from the last Qword to the first DQword. */
615         const vector unsigned char blend_mask = (vector unsigned char){
616                 -1,  0,  0,  0,
617                  0,  0,  0,  0,
618                 -1, -1, -1, -1,
619                 -1, -1, -1, -1};
620         const vector unsigned char crc_adj =
621                 (vector unsigned char)(vector unsigned short){
622                 rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
623                 rxq->crc_present * RTE_ETHER_CRC_LEN, 0, 0, 0, 0, 0};
624         const vector unsigned char flow_mark_adj =
625                 (vector unsigned char)(vector unsigned int){
626                 0, 0, 0, rxq->mark * (-1)};
627         const vector unsigned short cqe_sel_mask1 =
628                 (vector unsigned short){0, 0, 0, 0, 0xffff, 0xffff, 0, 0};
629         const vector unsigned short cqe_sel_mask2 =
630                 (vector unsigned short){0, 0, 0xffff, 0, 0, 0, 0, 0};
631
632         MLX5_ASSERT(rxq->sges_n == 0);
633         MLX5_ASSERT(rxq->cqe_n == rxq->elts_n);
634         cq = &(*rxq->cqes)[cq_idx];
635         rte_prefetch0(cq);
636         rte_prefetch0(cq + 1);
637         rte_prefetch0(cq + 2);
638         rte_prefetch0(cq + 3);
639         pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
640
641         repl_n = q_n - (rxq->rq_ci - rxq->rq_pi);
642         if (repl_n >= rxq->rq_repl_thresh)
643                 mlx5_rx_replenish_bulk_mbuf(rxq, repl_n);
644         /* See if there're unreturned mbufs from compressed CQE. */
645         rcvd_pkt = rxq->decompressed;
646         if (rcvd_pkt > 0) {
647                 rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
648                 rxq_copy_mbuf_v(rxq, pkts, rcvd_pkt);
649                 rxq->rq_pi += rcvd_pkt;
650                 rxq->decompressed -= rcvd_pkt;
651                 pkts += rcvd_pkt;
652         }
653         elts_idx = rxq->rq_pi & q_mask;
654         elts = &(*rxq->elts)[elts_idx];
655         /* Not to overflow pkts array. */
656         pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
657         /* Not to cross queue end. */
658         pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
659         pkts_n = RTE_MIN(pkts_n, q_n - cq_idx);
660         if (!pkts_n)
661                 return rcvd_pkt;
662         /* At this point, there shouldn't be any remaining packets. */
663         MLX5_ASSERT(rxq->decompressed == 0);
664
665         /*
666          * A. load first Qword (8bytes) in one loop.
667          * B. copy 4 mbuf pointers from elts ring to returing pkts.
668          * C. load remaining CQE data and extract necessary fields.
669          *    Final 16bytes cqes[] extracted from original 64bytes CQE has the
670          *    following structure:
671          *        struct {
672          *          uint8_t  pkt_info;
673          *          uint8_t  flow_tag[3];
674          *          uint16_t byte_cnt;
675          *          uint8_t  rsvd4;
676          *          uint8_t  op_own;
677          *          uint16_t hdr_type_etc;
678          *          uint16_t vlan_info;
679          *          uint32_t rx_has_res;
680          *        } c;
681          * D. fill in mbuf.
682          * E. get valid CQEs.
683          * F. find compressed CQE.
684          */
685         for (pos = 0;
686              pos < pkts_n;
687              pos += MLX5_VPMD_DESCS_PER_LOOP) {
688                 vector unsigned char cqes[MLX5_VPMD_DESCS_PER_LOOP];
689                 vector unsigned char cqe_tmp1, cqe_tmp2;
690                 vector unsigned char pkt_mb0, pkt_mb1, pkt_mb2, pkt_mb3;
691                 vector unsigned char op_own, op_own_tmp1, op_own_tmp2;
692                 vector unsigned char opcode, owner_mask, invalid_mask;
693                 vector unsigned char comp_mask;
694                 vector unsigned char mask;
695 #ifdef MLX5_PMD_SOFT_COUNTERS
696                 const vector unsigned char lower_half = {
697                         0, 1, 4, 5, 8, 9, 12, 13,
698                         16, 17, 20, 21, 24, 25, 28, 29};
699                 const vector unsigned char upper_half = {
700                         2, 3, 6, 7, 10, 11, 14, 15,
701                         18, 19, 22, 23, 26, 27, 30, 31};
702                 const vector unsigned long shmax = {64, 64};
703                 vector unsigned char byte_cnt;
704                 vector unsigned short left, right;
705                 vector unsigned long lshift;
706                 vector __attribute__((altivec(bool__)))
707                         unsigned long shmask;
708 #endif
709                 vector unsigned char mbp1, mbp2;
710                 vector unsigned char p =
711                         (vector unsigned char)(vector unsigned short){
712                                 0, 1, 2, 3, 0, 0, 0, 0};
713                 unsigned int p1, p2, p3;
714
715                 /* Prefetch next 4 CQEs. */
716                 if (pkts_n - pos >= 2 * MLX5_VPMD_DESCS_PER_LOOP) {
717                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP]);
718                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 1]);
719                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 2]);
720                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 3]);
721                 }
722
723                 /* A.0 do not cross the end of CQ. */
724                 mask = (vector unsigned char)(vector unsigned long){
725                         (pkts_n - pos) * sizeof(uint16_t) * 8, 0};
726
727                 {
728                         vector unsigned long lshift;
729                         vector __attribute__((altivec(bool__)))
730                                 unsigned long shmask;
731                         const vector unsigned long shmax = {64, 64};
732
733                         lshift = vec_splat((vector unsigned long)mask, 0);
734                         shmask = vec_cmpgt(shmax, lshift);
735                         mask = (vector unsigned char)
736                                 vec_sl((vector unsigned long)ones, lshift);
737                         mask = (vector unsigned char)
738                                 vec_sel((vector unsigned long)shmask,
739                                 (vector unsigned long)mask, shmask);
740                 }
741
742                 p = (vector unsigned char)
743                         vec_andc((vector unsigned long)p,
744                         (vector unsigned long)mask);
745
746                 /* A.1 load cqes. */
747                 p3 = (unsigned int)((vector unsigned short)p)[3];
748                 cqes[3] = (vector unsigned char)(vector unsigned long){
749                         *(__rte_aligned(8) unsigned long *)
750                         &cq[pos + p3].sop_drop_qpn, 0LL};
751                 rte_compiler_barrier();
752
753                 p2 = (unsigned int)((vector unsigned short)p)[2];
754                 cqes[2] = (vector unsigned char)(vector unsigned long){
755                         *(__rte_aligned(8) unsigned long *)
756                         &cq[pos + p2].sop_drop_qpn, 0LL};
757                 rte_compiler_barrier();
758
759                 /* B.1 load mbuf pointers. */
760                 mbp1 = (vector unsigned char)vec_vsx_ld(0,
761                         (signed int const *)&elts[pos]);
762                 mbp2 = (vector unsigned char)vec_vsx_ld(0,
763                         (signed int const *)&elts[pos + 2]);
764
765                 /* A.1 load a block having op_own. */
766                 p1 = (unsigned int)((vector unsigned short)p)[1];
767                 cqes[1] = (vector unsigned char)(vector unsigned long){
768                         *(__rte_aligned(8) unsigned long *)
769                         &cq[pos + p1].sop_drop_qpn, 0LL};
770                 rte_compiler_barrier();
771
772                 cqes[0] = (vector unsigned char)(vector unsigned long){
773                         *(__rte_aligned(8) unsigned long *)
774                         &cq[pos].sop_drop_qpn, 0LL};
775                 rte_compiler_barrier();
776
777                 /* B.2 copy mbuf pointers. */
778                 *(vector unsigned char *)&pkts[pos] = mbp1;
779                 *(vector unsigned char *)&pkts[pos + 2] = mbp2;
780                 rte_cio_rmb();
781
782                 /* C.1 load remaining CQE data and extract necessary fields. */
783                 cqe_tmp2 = *(vector unsigned char *)
784                         &cq[pos + p3].pkt_info;
785                 cqe_tmp1 = *(vector unsigned char *)
786                         &cq[pos + p2].pkt_info;
787                 cqes[3] = vec_sel(cqes[3], cqe_tmp2, blend_mask);
788                 cqes[2] = vec_sel(cqes[2], cqe_tmp1, blend_mask);
789                 cqe_tmp2 = (vector unsigned char)vec_vsx_ld(0,
790                         (signed int const *)&cq[pos + p3].csum);
791                 cqe_tmp1 = (vector unsigned char)vec_vsx_ld(0,
792                         (signed int const *)&cq[pos + p2].csum);
793                 cqes[3] = (vector unsigned char)
794                         vec_sel((vector unsigned short)cqes[3],
795                         (vector unsigned short)cqe_tmp2, cqe_sel_mask1);
796                 cqes[2] = (vector unsigned char)
797                         vec_sel((vector unsigned short)cqes[2],
798                         (vector unsigned short)cqe_tmp1, cqe_sel_mask1);
799                 cqe_tmp2 = (vector unsigned char)(vector unsigned long){
800                         *(__rte_aligned(8) unsigned long *)
801                         &cq[pos + p3].rsvd3[9], 0LL};
802                 cqe_tmp1 = (vector unsigned char)(vector unsigned long){
803                         *(__rte_aligned(8) unsigned long *)
804                         &cq[pos + p2].rsvd3[9], 0LL};
805                 cqes[3] = (vector unsigned char)
806                         vec_sel((vector unsigned short)cqes[3],
807                         (vector unsigned short)cqe_tmp2,
808                         (vector unsigned short)cqe_sel_mask2);
809                 cqes[2] = (vector unsigned char)
810                         vec_sel((vector unsigned short)cqes[2],
811                         (vector unsigned short)cqe_tmp1,
812                         (vector unsigned short)cqe_sel_mask2);
813
814                 /* C.2 generate final structure for mbuf with swapping bytes. */
815                 pkt_mb3 = vec_perm(cqes[3], zero, shuf_mask);
816                 pkt_mb2 = vec_perm(cqes[2], zero, shuf_mask);
817
818                 /* C.3 adjust CRC length. */
819                 pkt_mb3 = (vector unsigned char)
820                         ((vector unsigned short)pkt_mb3 -
821                         (vector unsigned short)crc_adj);
822                 pkt_mb2 = (vector unsigned char)
823                         ((vector unsigned short)pkt_mb2 -
824                         (vector unsigned short)crc_adj);
825
826                 /* C.4 adjust flow mark. */
827                 pkt_mb3 = (vector unsigned char)
828                         ((vector unsigned int)pkt_mb3 +
829                         (vector unsigned int)flow_mark_adj);
830                 pkt_mb2 = (vector unsigned char)
831                         ((vector unsigned int)pkt_mb2 +
832                         (vector unsigned int)flow_mark_adj);
833
834                 /* D.1 fill in mbuf - rx_descriptor_fields1. */
835                 *(vector unsigned char *)
836                         &pkts[pos + 3]->pkt_len = pkt_mb3;
837                 *(vector unsigned char *)
838                         &pkts[pos + 2]->pkt_len = pkt_mb2;
839
840                 /* E.1 extract op_own field. */
841                 op_own_tmp2 = (vector unsigned char)
842                         vec_mergeh((vector unsigned int)cqes[2],
843                         (vector unsigned int)cqes[3]);
844
845                 /* C.1 load remaining CQE data and extract necessary fields. */
846                 cqe_tmp2 = *(vector unsigned char *)
847                         &cq[pos + p1].pkt_info;
848                 cqe_tmp1 = *(vector unsigned char *)
849                         &cq[pos].pkt_info;
850                 cqes[1] = vec_sel(cqes[1], cqe_tmp2, blend_mask);
851                 cqes[0] = vec_sel(cqes[0], cqe_tmp2, blend_mask);
852                 cqe_tmp2 = (vector unsigned char)vec_vsx_ld(0,
853                         (signed int const *)&cq[pos + p1].csum);
854                 cqe_tmp1 = (vector unsigned char)vec_vsx_ld(0,
855                         (signed int const *)&cq[pos].csum);
856                 cqes[1] = (vector unsigned char)
857                         vec_sel((vector unsigned short)cqes[1],
858                         (vector unsigned short)cqe_tmp2, cqe_sel_mask1);
859                 cqes[0] = (vector unsigned char)
860                         vec_sel((vector unsigned short)cqes[0],
861                         (vector unsigned short)cqe_tmp1, cqe_sel_mask1);
862                 cqe_tmp2 = (vector unsigned char)(vector unsigned long){
863                         *(__rte_aligned(8) unsigned long *)
864                         &cq[pos + p1].rsvd3[9], 0LL};
865                 cqe_tmp1 = (vector unsigned char)(vector unsigned long){
866                         *(__rte_aligned(8) unsigned long *)
867                         &cq[pos].rsvd3[9], 0LL};
868                 cqes[1] = (vector unsigned char)
869                         vec_sel((vector unsigned short)cqes[1],
870                         (vector unsigned short)cqe_tmp2, cqe_sel_mask2);
871                 cqes[0] = (vector unsigned char)
872                         vec_sel((vector unsigned short)cqes[0],
873                         (vector unsigned short)cqe_tmp1, cqe_sel_mask2);
874
875                 /* C.2 generate final structure for mbuf with swapping bytes. */
876                 pkt_mb1 = vec_perm(cqes[1], zero, shuf_mask);
877                 pkt_mb0 = vec_perm(cqes[0], zero, shuf_mask);
878
879                 /* C.3 adjust CRC length. */
880                 pkt_mb1 = (vector unsigned char)
881                         ((vector unsigned short)pkt_mb1 -
882                         (vector unsigned short)crc_adj);
883                 pkt_mb0 = (vector unsigned char)
884                         ((vector unsigned short)pkt_mb0 -
885                         (vector unsigned short)crc_adj);
886
887                 /* C.4 adjust flow mark. */
888                 pkt_mb1 = (vector unsigned char)
889                         ((vector unsigned int)pkt_mb1 +
890                         (vector unsigned int)flow_mark_adj);
891                 pkt_mb0 = (vector unsigned char)
892                         ((vector unsigned int)pkt_mb0 +
893                         (vector unsigned int)flow_mark_adj);
894
895                 /* E.1 extract op_own byte. */
896                 op_own_tmp1 = (vector unsigned char)
897                         vec_mergeh((vector unsigned int)cqes[0],
898                         (vector unsigned int)cqes[1]);
899                 op_own = (vector unsigned char)
900                         vec_mergel((vector unsigned long)op_own_tmp1,
901                         (vector unsigned long)op_own_tmp2);
902
903                 /* D.1 fill in mbuf - rx_descriptor_fields1. */
904                 *(vector unsigned char *)
905                         &pkts[pos + 1]->pkt_len = pkt_mb1;
906                 *(vector unsigned char *)
907                         &pkts[pos]->pkt_len = pkt_mb0;
908
909                 /* E.2 flip owner bit to mark CQEs from last round. */
910                 owner_mask = (vector unsigned char)
911                         vec_and((vector unsigned long)op_own,
912                         (vector unsigned long)owner_check);
913                 if (ownership)
914                         owner_mask = (vector unsigned char)
915                                 vec_xor((vector unsigned long)owner_mask,
916                                 (vector unsigned long)owner_check);
917                 owner_mask = (vector unsigned char)
918                         vec_cmpeq((vector unsigned int)owner_mask,
919                         (vector unsigned int)owner_check);
920                 owner_mask = (vector unsigned char)
921                         vec_packs((vector unsigned int)owner_mask,
922                         (vector unsigned int)zero);
923
924                 /* E.3 get mask for invalidated CQEs. */
925                 opcode = (vector unsigned char)
926                         vec_and((vector unsigned long)op_own,
927                         (vector unsigned long)opcode_check);
928                 invalid_mask = (vector unsigned char)
929                         vec_cmpeq((vector unsigned int)opcode_check,
930                         (vector unsigned int)opcode);
931                 invalid_mask = (vector unsigned char)
932                         vec_packs((vector unsigned int)invalid_mask,
933                         (vector unsigned int)zero);
934
935                 /* E.4 mask out beyond boundary. */
936                 invalid_mask = (vector unsigned char)
937                         vec_or((vector unsigned long)invalid_mask,
938                         (vector unsigned long)mask);
939
940                 /* E.5 merge invalid_mask with invalid owner. */
941                 invalid_mask = (vector unsigned char)
942                         vec_or((vector unsigned long)invalid_mask,
943                         (vector unsigned long)owner_mask);
944
945                 /* F.1 find compressed CQE format. */
946                 comp_mask = (vector unsigned char)
947                         vec_and((vector unsigned long)op_own,
948                         (vector unsigned long)format_check);
949                 comp_mask = (vector unsigned char)
950                         vec_cmpeq((vector unsigned int)comp_mask,
951                         (vector unsigned int)format_check);
952                 comp_mask = (vector unsigned char)
953                         vec_packs((vector unsigned int)comp_mask,
954                         (vector unsigned int)zero);
955
956                 /* F.2 mask out invalid entries. */
957                 comp_mask = (vector unsigned char)
958                         vec_andc((vector unsigned long)comp_mask,
959                         (vector unsigned long)invalid_mask);
960                 comp_idx = ((vector unsigned long)comp_mask)[0];
961
962                 /* F.3 get the first compressed CQE. */
963                 comp_idx = comp_idx ? __builtin_ctzll(comp_idx) /
964                         (sizeof(uint16_t) * 8) : MLX5_VPMD_DESCS_PER_LOOP;
965
966                 /* E.6 mask out entries after the compressed CQE. */
967                 mask = (vector unsigned char)(vector unsigned long){
968                         (comp_idx * sizeof(uint16_t) * 8), 0};
969                 lshift = vec_splat((vector unsigned long)mask, 0);
970                 shmask = vec_cmpgt(shmax, lshift);
971                 mask = (vector unsigned char)
972                         vec_sl((vector unsigned long)ones, lshift);
973                 mask = (vector unsigned char)
974                         vec_sel((vector unsigned long)shmask,
975                         (vector unsigned long)mask, shmask);
976                 invalid_mask = (vector unsigned char)
977                         vec_or((vector unsigned long)invalid_mask,
978                         (vector unsigned long)mask);
979
980                 /* E.7 count non-compressed valid CQEs. */
981                 n = ((vector unsigned long)invalid_mask)[0];
982                 n = n ? __builtin_ctzll(n) / (sizeof(uint16_t) * 8) :
983                         MLX5_VPMD_DESCS_PER_LOOP;
984                 nocmp_n += n;
985
986                 /* D.2 get the final invalid mask. */
987                 mask = (vector unsigned char)(vector unsigned long){
988                         (n * sizeof(uint16_t) * 8), 0};
989                 lshift = vec_splat((vector unsigned long)mask, 0);
990                 shmask = vec_cmpgt(shmax, lshift);
991                 mask = (vector unsigned char)
992                         vec_sl((vector unsigned long)ones, lshift);
993                 mask = (vector unsigned char)
994                         vec_sel((vector unsigned long)shmask,
995                         (vector unsigned long)mask, shmask);
996                 invalid_mask = (vector unsigned char)
997                         vec_or((vector unsigned long)invalid_mask,
998                         (vector unsigned long)mask);
999
1000                 /* D.3 check error in opcode. */
1001                 opcode = (vector unsigned char)
1002                         vec_cmpeq((vector unsigned int)resp_err_check,
1003                         (vector unsigned int)opcode);
1004                 opcode = (vector unsigned char)
1005                         vec_packs((vector unsigned int)opcode,
1006                         (vector unsigned int)zero);
1007                 opcode = (vector unsigned char)
1008                         vec_andc((vector unsigned long)opcode,
1009                         (vector unsigned long)invalid_mask);
1010
1011                 /* D.4 mark if any error is set */
1012                 *err |= ((vector unsigned long)opcode)[0];
1013
1014                 /* D.5 fill in mbuf - rearm_data and packet_type. */
1015                 rxq_cq_to_ptype_oflags_v(rxq, cqes, opcode, &pkts[pos]);
1016                 if (rxq->hw_timestamp) {
1017                         pkts[pos]->timestamp =
1018                                 rte_be_to_cpu_64(cq[pos].timestamp);
1019                         pkts[pos + 1]->timestamp =
1020                                 rte_be_to_cpu_64(cq[pos + p1].timestamp);
1021                         pkts[pos + 2]->timestamp =
1022                                 rte_be_to_cpu_64(cq[pos + p2].timestamp);
1023                         pkts[pos + 3]->timestamp =
1024                                 rte_be_to_cpu_64(cq[pos + p3].timestamp);
1025                 }
1026                 if (rte_flow_dynf_metadata_avail()) {
1027                         uint64_t flag = rte_flow_dynf_metadata_mask;
1028                         int offs = rte_flow_dynf_metadata_offs;
1029                         uint32_t metadata;
1030
1031                         /* This code is subject for futher optimization. */
1032                         metadata = cq[pos].flow_table_metadata;
1033                         *RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *) =
1034                                                                 metadata;
1035                         pkts[pos]->ol_flags |= metadata ? flag : 0ULL;
1036                         metadata = cq[pos + 1].flow_table_metadata;
1037                         *RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *) =
1038                                                                 metadata;
1039                         pkts[pos + 1]->ol_flags |= metadata ? flag : 0ULL;
1040                         metadata = cq[pos + 2].flow_table_metadata;
1041                         *RTE_MBUF_DYNFIELD(pkts[pos + 2], offs, uint32_t *) =
1042                                                                 metadata;
1043                         pkts[pos + 2]->ol_flags |= metadata ? flag : 0ULL;
1044                         metadata = cq[pos + 3].flow_table_metadata;
1045                         *RTE_MBUF_DYNFIELD(pkts[pos + 3], offs, uint32_t *) =
1046                                                                 metadata;
1047                         pkts[pos + 3]->ol_flags |= metadata ? flag : 0ULL;
1048                 }
1049 #ifdef MLX5_PMD_SOFT_COUNTERS
1050                 /* Add up received bytes count. */
1051                 byte_cnt = vec_perm(op_own, zero, len_shuf_mask);
1052                 byte_cnt = (vector unsigned char)
1053                         vec_andc((vector unsigned long)byte_cnt,
1054                         (vector unsigned long)invalid_mask);
1055                 left = vec_perm((vector unsigned short)byte_cnt,
1056                         (vector unsigned short)zero, lower_half);
1057                 right = vec_perm((vector unsigned short)byte_cnt,
1058                         (vector unsigned short)zero, upper_half);
1059                 byte_cnt = (vector unsigned char)vec_add(left, right);
1060                 left = vec_perm((vector unsigned short)byte_cnt,
1061                         (vector unsigned short)zero, lower_half);
1062                 right = vec_perm((vector unsigned short)byte_cnt,
1063                         (vector unsigned short)zero, upper_half);
1064                 byte_cnt = (vector unsigned char)vec_add(left, right);
1065                 rcvd_byte += ((vector unsigned long)byte_cnt)[0];
1066 #endif
1067
1068                 /*
1069                  * Break the loop unless more valid CQE is expected, or if
1070                  * there's a compressed CQE.
1071                  */
1072                 if (n != MLX5_VPMD_DESCS_PER_LOOP)
1073                         break;
1074         }
1075         /* If no new CQE seen, return without updating cq_db. */
1076         if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP))
1077                 return rcvd_pkt;
1078         /* Update the consumer indexes for non-compressed CQEs. */
1079         MLX5_ASSERT(nocmp_n <= pkts_n);
1080         rxq->cq_ci += nocmp_n;
1081         rxq->rq_pi += nocmp_n;
1082         rcvd_pkt += nocmp_n;
1083 #ifdef MLX5_PMD_SOFT_COUNTERS
1084         rxq->stats.ipackets += nocmp_n;
1085         rxq->stats.ibytes += rcvd_byte;
1086 #endif
1087         /* Decompress the last CQE if compressed. */
1088         if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP && comp_idx == n) {
1089                 MLX5_ASSERT(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
1090                 rxq->decompressed =
1091                         rxq_cq_decompress_v(rxq, &cq[nocmp_n], &elts[nocmp_n]);
1092                 /* Return more packets if needed. */
1093                 if (nocmp_n < pkts_n) {
1094                         uint16_t n = rxq->decompressed;
1095
1096                         n = RTE_MIN(n, pkts_n - nocmp_n);
1097                         rxq_copy_mbuf_v(rxq, &pkts[nocmp_n], n);
1098                         rxq->rq_pi += n;
1099                         rcvd_pkt += n;
1100                         rxq->decompressed -= n;
1101                 }
1102         }
1103         rte_compiler_barrier();
1104         *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1105         return rcvd_pkt;
1106 }
1107
1108 #endif /* RTE_PMD_MLX5_RXTX_VEC_ALTIVEC_H_ */