eal/ppc: fix bool type after altivec include
[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 <rte_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 (!!rxq->flow_meta_mask) {
268                         int32_t offs = rxq->flow_meta_offset;
269                         const uint32_t meta =
270                                 *RTE_MBUF_DYNFIELD(t_pkt, offs, uint32_t *);
271
272                         /* Check if title packet has valid metadata. */
273                         if (meta) {
274                                 MLX5_ASSERT(t_pkt->ol_flags & offs);
275                                 *RTE_MBUF_DYNFIELD(elts[pos], offs,
276                                                         uint32_t *) = meta;
277                                 *RTE_MBUF_DYNFIELD(elts[pos + 1], offs,
278                                                         uint32_t *) = meta;
279                                 *RTE_MBUF_DYNFIELD(elts[pos + 2], offs,
280                                                         uint32_t *) = meta;
281                                 *RTE_MBUF_DYNFIELD(elts[pos + 3], offs,
282                                                         uint32_t *) = meta;
283                         }
284                 }
285
286                 pos += MLX5_VPMD_DESCS_PER_LOOP;
287                 /* Move to next CQE and invalidate consumed CQEs. */
288                 if (!(pos & 0x7) && pos < mcqe_n) {
289                         mcq = (void *)&(cq + pos)->pkt_info;
290                         for (i = 0; i < 8; ++i)
291                                 cq[inv++].op_own = MLX5_CQE_INVALIDATE;
292                 }
293         }
294
295         /* Invalidate the rest of CQEs. */
296         for (; inv < mcqe_n; ++inv)
297                 cq[inv].op_own = MLX5_CQE_INVALIDATE;
298
299 #ifdef MLX5_PMD_SOFT_COUNTERS
300         rxq->stats.ipackets += mcqe_n;
301         rxq->stats.ibytes += rcvd_byte;
302 #endif
303
304         rxq->cq_ci += mcqe_n;
305         return mcqe_n;
306 }
307
308 /**
309  * Calculate packet type and offload flag for mbuf and store it.
310  *
311  * @param rxq
312  *   Pointer to RX queue structure.
313  * @param cqes[4]
314  *   Array of four 16bytes completions extracted from the original completion
315  *   descriptor.
316  * @param op_err
317  *   Opcode vector having responder error status. Each field is 4B.
318  * @param pkts
319  *   Pointer to array of packets to be filled.
320  */
321 static inline void
322 rxq_cq_to_ptype_oflags_v(struct mlx5_rxq_data *rxq,
323                 vector unsigned char cqes[4], vector unsigned char op_err,
324                 struct rte_mbuf **pkts)
325 {
326         vector unsigned char pinfo0, pinfo1;
327         vector unsigned char pinfo, ptype;
328         vector unsigned char ol_flags = (vector unsigned char)
329                 (vector unsigned int){
330                         rxq->rss_hash * PKT_RX_RSS_HASH |
331                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP,
332                         rxq->rss_hash * PKT_RX_RSS_HASH |
333                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP,
334                         rxq->rss_hash * PKT_RX_RSS_HASH |
335                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP,
336                         rxq->rss_hash * PKT_RX_RSS_HASH |
337                                 rxq->hw_timestamp * PKT_RX_TIMESTAMP};
338         vector unsigned char cv_flags;
339         const vector unsigned char zero = (vector unsigned char){0};
340         const vector unsigned char ptype_mask =
341                 (vector unsigned char)(vector unsigned int){
342                 0x0000fd06, 0x0000fd06, 0x0000fd06, 0x0000fd06};
343         const vector unsigned char ptype_ol_mask =
344                 (vector unsigned char)(vector unsigned int){
345                 0x00000106, 0x00000106, 0x00000106, 0x00000106};
346         const vector unsigned char pinfo_mask =
347                 (vector unsigned char)(vector unsigned int){
348                 0x00000003, 0x00000003, 0x00000003, 0x00000003};
349         const vector unsigned char cv_flag_sel = (vector unsigned char){
350                 0, (uint8_t)(PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED),
351                 (uint8_t)(PKT_RX_IP_CKSUM_GOOD >> 1), 0,
352                 (uint8_t)(PKT_RX_L4_CKSUM_GOOD >> 1), 0,
353                 (uint8_t)((PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1),
354                 0, 0, 0, 0, 0, 0, 0, 0, 0};
355         const vector unsigned char cv_mask =
356                 (vector unsigned char)(vector unsigned int){
357                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
358                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
359                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
360                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
361                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
362                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
363                 PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
364                 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED};
365         const vector unsigned char mbuf_init =
366                 (vector unsigned char)vec_vsx_ld
367                         (0, (vector unsigned char *)&rxq->mbuf_initializer);
368         const vector unsigned short rearm_sel_mask =
369                 (vector unsigned short){0, 0, 0, 0, 0xffff, 0xffff, 0, 0};
370         vector unsigned char rearm0, rearm1, rearm2, rearm3;
371         uint8_t pt_idx0, pt_idx1, pt_idx2, pt_idx3;
372
373         /* Extract pkt_info field. */
374         pinfo0 = (vector unsigned char)
375                 vec_mergeh((vector unsigned int)cqes[0],
376                 (vector unsigned int)cqes[1]);
377         pinfo1 = (vector unsigned char)
378                 vec_mergeh((vector unsigned int)cqes[2],
379                 (vector unsigned int)cqes[3]);
380         pinfo = (vector unsigned char)
381                 vec_mergeh((vector unsigned long)pinfo0,
382                 (vector unsigned long)pinfo1);
383
384         /* Extract hdr_type_etc field. */
385         pinfo0 = (vector unsigned char)
386                 vec_mergel((vector unsigned int)cqes[0],
387                 (vector unsigned int)cqes[1]);
388         pinfo1 = (vector unsigned char)
389                 vec_mergel((vector unsigned int)cqes[2],
390                 (vector unsigned int)cqes[3]);
391         ptype = (vector unsigned char)
392                 vec_mergeh((vector unsigned long)pinfo0,
393                 (vector unsigned long)pinfo1);
394
395         if (rxq->mark) {
396                 const vector unsigned char pinfo_ft_mask =
397                         (vector unsigned char)(vector unsigned int){
398                         0xffffff00, 0xffffff00, 0xffffff00, 0xffffff00};
399                 const vector unsigned char fdir_flags =
400                         (vector unsigned char)(vector unsigned int){
401                         PKT_RX_FDIR, PKT_RX_FDIR,
402                         PKT_RX_FDIR, PKT_RX_FDIR};
403                 vector unsigned char fdir_id_flags =
404                         (vector unsigned char)(vector unsigned int){
405                         PKT_RX_FDIR_ID, PKT_RX_FDIR_ID,
406                         PKT_RX_FDIR_ID, PKT_RX_FDIR_ID};
407                 vector unsigned char flow_tag, invalid_mask;
408
409                 flow_tag = (vector unsigned char)
410                         vec_and((vector unsigned long)pinfo,
411                         (vector unsigned long)pinfo_ft_mask);
412
413                 /* Check if flow tag is non-zero then set PKT_RX_FDIR. */
414                 invalid_mask = (vector unsigned char)
415                         vec_cmpeq((vector unsigned int)flow_tag,
416                         (vector unsigned int)zero);
417                 ol_flags = (vector unsigned char)
418                         vec_or((vector unsigned long)ol_flags,
419                         (vector unsigned long)
420                         vec_andc((vector unsigned long)fdir_flags,
421                         (vector unsigned long)invalid_mask));
422
423                 /* Mask out invalid entries. */
424                 fdir_id_flags = (vector unsigned char)
425                         vec_andc((vector unsigned long)fdir_id_flags,
426                         (vector unsigned long)invalid_mask);
427
428                 /* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */
429                 ol_flags = (vector unsigned char)
430                         vec_or((vector unsigned long)ol_flags,
431                         (vector unsigned long)
432                         vec_andc((vector unsigned long)fdir_id_flags,
433                         (vector unsigned long)
434                         vec_cmpeq((vector unsigned int)flow_tag,
435                         (vector unsigned int)pinfo_ft_mask)));
436         }
437         /*
438          * Merge the two fields to generate the following:
439          * bit[1]     = l3_ok
440          * bit[2]     = l4_ok
441          * bit[8]     = cv
442          * bit[11:10] = l3_hdr_type
443          * bit[14:12] = l4_hdr_type
444          * bit[15]    = ip_frag
445          * bit[16]    = tunneled
446          * bit[17]    = outer_l3_type
447          */
448         ptype = (vector unsigned char)
449                 vec_and((vector unsigned long)ptype,
450                 (vector unsigned long)ptype_mask);
451         pinfo = (vector unsigned char)
452                 vec_and((vector unsigned long)pinfo,
453                 (vector unsigned long)pinfo_mask);
454         pinfo = (vector unsigned char)
455                 vec_sl((vector unsigned int)pinfo,
456                 (vector unsigned int){16, 16, 16, 16});
457
458         /* Make pinfo has merged fields for ol_flags calculation. */
459         pinfo = (vector unsigned char)
460                 vec_or((vector unsigned long)ptype,
461                 (vector unsigned long)pinfo);
462         ptype = (vector unsigned char)
463                 vec_sr((vector unsigned int)pinfo,
464                 (vector unsigned int){10, 10, 10, 10});
465         ptype = (vector unsigned char)
466                 vec_packs((vector unsigned int)ptype,
467                 (vector unsigned int)zero);
468
469         /* Errored packets will have RTE_PTYPE_ALL_MASK. */
470         op_err = (vector unsigned char)
471                 vec_sr((vector unsigned short)op_err,
472                 (vector unsigned short){8, 8, 8, 8, 8, 8, 8, 8});
473         ptype = (vector unsigned char)
474                 vec_or((vector unsigned long)ptype,
475                 (vector unsigned long)op_err);
476
477         pt_idx0 = (uint8_t)((vector unsigned char)ptype)[0];
478         pt_idx1 = (uint8_t)((vector unsigned char)ptype)[2];
479         pt_idx2 = (uint8_t)((vector unsigned char)ptype)[4];
480         pt_idx3 = (uint8_t)((vector unsigned char)ptype)[6];
481
482         pkts[0]->packet_type = mlx5_ptype_table[pt_idx0] |
483                 !!(pt_idx0 & (1 << 6)) * rxq->tunnel;
484         pkts[1]->packet_type = mlx5_ptype_table[pt_idx1] |
485                 !!(pt_idx1 & (1 << 6)) * rxq->tunnel;
486         pkts[2]->packet_type = mlx5_ptype_table[pt_idx2] |
487                 !!(pt_idx2 & (1 << 6)) * rxq->tunnel;
488         pkts[3]->packet_type = mlx5_ptype_table[pt_idx3] |
489                 !!(pt_idx3 & (1 << 6)) * rxq->tunnel;
490
491         /* Fill flags for checksum and VLAN. */
492         pinfo = (vector unsigned char)
493                 vec_and((vector unsigned long)pinfo,
494                 (vector unsigned long)ptype_ol_mask);
495         pinfo = vec_perm(cv_flag_sel, zero, pinfo);
496
497         /* Locate checksum flags at byte[2:1] and merge with VLAN flags. */
498         cv_flags = (vector unsigned char)
499                 vec_sl((vector unsigned int)pinfo,
500                 (vector unsigned int){9, 9, 9, 9});
501         cv_flags = (vector unsigned char)
502                 vec_or((vector unsigned long)pinfo,
503                 (vector unsigned long)cv_flags);
504
505         /* Move back flags to start from byte[0]. */
506         cv_flags = (vector unsigned char)
507                 vec_sr((vector unsigned int)cv_flags,
508                 (vector unsigned int){8, 8, 8, 8});
509
510         /* Mask out garbage bits. */
511         cv_flags = (vector unsigned char)
512                 vec_and((vector unsigned long)cv_flags,
513                 (vector unsigned long)cv_mask);
514
515         /* Merge to ol_flags. */
516         ol_flags = (vector unsigned char)
517                 vec_or((vector unsigned long)ol_flags,
518                 (vector unsigned long)cv_flags);
519
520         /* Merge mbuf_init and ol_flags. */
521         rearm0 = (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){64}), rearm_sel_mask);
526         rearm1 = (vector unsigned char)
527                 vec_sel((vector unsigned short)mbuf_init,
528                 (vector unsigned short)
529                 vec_slo((vector unsigned short)ol_flags,
530                 (vector unsigned char){32}), rearm_sel_mask);
531         rearm2 = (vector unsigned char)
532                 vec_sel((vector unsigned short)mbuf_init,
533                 (vector unsigned short)ol_flags, rearm_sel_mask);
534         rearm3 = (vector unsigned char)
535                 vec_sel((vector unsigned short)mbuf_init,
536                 (vector unsigned short)
537                 vec_sro((vector unsigned short)ol_flags,
538                 (vector unsigned char){32}), rearm_sel_mask);
539
540         /* Write 8B rearm_data and 8B ol_flags. */
541         vec_vsx_st(rearm0, 0,
542                 (vector unsigned char *)&pkts[0]->rearm_data);
543         vec_vsx_st(rearm1, 0,
544                 (vector unsigned char *)&pkts[1]->rearm_data);
545         vec_vsx_st(rearm2, 0,
546                 (vector unsigned char *)&pkts[2]->rearm_data);
547         vec_vsx_st(rearm3, 0,
548                 (vector unsigned char *)&pkts[3]->rearm_data);
549 }
550
551
552 /**
553  * Receive burst of packets. An errored completion also consumes a mbuf, but the
554  * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
555  * before returning to application.
556  *
557  * @param rxq
558  *   Pointer to RX queue structure.
559  * @param[out] pkts
560  *   Array to store received packets.
561  * @param pkts_n
562  *   Maximum number of packets in array.
563  * @param[out] err
564  *   Pointer to a flag. Set non-zero value if pkts array has at least one error
565  *   packet to handle.
566  *
567  * @return
568  *   Number of packets received including errors (<= pkts_n).
569  */
570 static inline uint16_t
571 rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
572             uint64_t *err)
573 {
574         const uint16_t q_n = 1 << rxq->cqe_n;
575         const uint16_t q_mask = q_n - 1;
576         volatile struct mlx5_cqe *cq;
577         struct rte_mbuf **elts;
578         unsigned int pos;
579         uint64_t n;
580         uint16_t repl_n;
581         uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
582         uint16_t nocmp_n = 0;
583         uint16_t rcvd_pkt = 0;
584         unsigned int cq_idx = rxq->cq_ci & q_mask;
585         unsigned int elts_idx;
586         unsigned int ownership = !!(rxq->cq_ci & (q_mask + 1));
587         const vector unsigned char zero = (vector unsigned char){0};
588         const vector unsigned char ones = vec_splat_u8(-1);
589         const vector unsigned char owner_check =
590                 (vector unsigned char)(vector unsigned long){
591                 0x0100000001000000LL, 0x0100000001000000LL};
592         const vector unsigned char opcode_check =
593                 (vector unsigned char)(vector unsigned long){
594                 0xf0000000f0000000LL, 0xf0000000f0000000LL};
595         const vector unsigned char format_check =
596                 (vector unsigned char)(vector unsigned long){
597                 0x0c0000000c000000LL, 0x0c0000000c000000LL};
598         const vector unsigned char resp_err_check =
599                 (vector unsigned char)(vector unsigned long){
600                 0xe0000000e0000000LL, 0xe0000000e0000000LL};
601 #ifdef MLX5_PMD_SOFT_COUNTERS
602         uint32_t rcvd_byte = 0;
603         /* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
604         const vector unsigned char len_shuf_mask = (vector unsigned char){
605                  1,  0,  5,  4,
606                  9,  8, 13, 12,
607                 -1, -1, -1, -1,
608                 -1, -1, -1, -1};
609 #endif
610         /* Mask to shuffle from extracted CQE to mbuf. */
611         const vector unsigned char shuf_mask = (vector unsigned char){
612                  5,  4,           /* bswap16, pkt_len */
613                 -1, -1,           /* zero out 2nd half of pkt_len */
614                  5,  4,           /* bswap16, data_len */
615                 11, 10,           /* bswap16, vlan+tci */
616                 15, 14, 13, 12,   /* bswap32, rss */
617                  1,  2,  3, -1};  /* fdir.hi */
618         /* Mask to blend from the last Qword to the first DQword. */
619         /* Mask to blend from the last Qword to the first DQword. */
620         const vector unsigned char blend_mask = (vector unsigned char){
621                 -1,  0,  0,  0,
622                  0,  0,  0,  0,
623                 -1, -1, -1, -1,
624                 -1, -1, -1, -1};
625         const vector unsigned char crc_adj =
626                 (vector unsigned char)(vector unsigned short){
627                 rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
628                 rxq->crc_present * RTE_ETHER_CRC_LEN, 0, 0, 0, 0, 0};
629         const vector unsigned char flow_mark_adj =
630                 (vector unsigned char)(vector unsigned int){
631                 0, 0, 0, rxq->mark * (-1)};
632         const vector unsigned short cqe_sel_mask1 =
633                 (vector unsigned short){0, 0, 0, 0, 0xffff, 0xffff, 0, 0};
634         const vector unsigned short cqe_sel_mask2 =
635                 (vector unsigned short){0, 0, 0xffff, 0, 0, 0, 0, 0};
636
637         MLX5_ASSERT(rxq->sges_n == 0);
638         MLX5_ASSERT(rxq->cqe_n == rxq->elts_n);
639         cq = &(*rxq->cqes)[cq_idx];
640         rte_prefetch0(cq);
641         rte_prefetch0(cq + 1);
642         rte_prefetch0(cq + 2);
643         rte_prefetch0(cq + 3);
644         pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
645
646         repl_n = q_n - (rxq->rq_ci - rxq->rq_pi);
647         if (repl_n >= rxq->rq_repl_thresh)
648                 mlx5_rx_replenish_bulk_mbuf(rxq, repl_n);
649         /* See if there're unreturned mbufs from compressed CQE. */
650         rcvd_pkt = rxq->decompressed;
651         if (rcvd_pkt > 0) {
652                 rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
653                 rxq_copy_mbuf_v(rxq, pkts, rcvd_pkt);
654                 rxq->rq_pi += rcvd_pkt;
655                 rxq->decompressed -= rcvd_pkt;
656                 pkts += rcvd_pkt;
657         }
658         elts_idx = rxq->rq_pi & q_mask;
659         elts = &(*rxq->elts)[elts_idx];
660         /* Not to overflow pkts array. */
661         pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
662         /* Not to cross queue end. */
663         pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
664         pkts_n = RTE_MIN(pkts_n, q_n - cq_idx);
665         if (!pkts_n)
666                 return rcvd_pkt;
667         /* At this point, there shouldn't be any remaining packets. */
668         MLX5_ASSERT(rxq->decompressed == 0);
669
670         /*
671          * A. load first Qword (8bytes) in one loop.
672          * B. copy 4 mbuf pointers from elts ring to returing pkts.
673          * C. load remaining CQE data and extract necessary fields.
674          *    Final 16bytes cqes[] extracted from original 64bytes CQE has the
675          *    following structure:
676          *        struct {
677          *          uint8_t  pkt_info;
678          *          uint8_t  flow_tag[3];
679          *          uint16_t byte_cnt;
680          *          uint8_t  rsvd4;
681          *          uint8_t  op_own;
682          *          uint16_t hdr_type_etc;
683          *          uint16_t vlan_info;
684          *          uint32_t rx_has_res;
685          *        } c;
686          * D. fill in mbuf.
687          * E. get valid CQEs.
688          * F. find compressed CQE.
689          */
690         for (pos = 0;
691              pos < pkts_n;
692              pos += MLX5_VPMD_DESCS_PER_LOOP) {
693                 vector unsigned char cqes[MLX5_VPMD_DESCS_PER_LOOP];
694                 vector unsigned char cqe_tmp1, cqe_tmp2;
695                 vector unsigned char pkt_mb0, pkt_mb1, pkt_mb2, pkt_mb3;
696                 vector unsigned char op_own, op_own_tmp1, op_own_tmp2;
697                 vector unsigned char opcode, owner_mask, invalid_mask;
698                 vector unsigned char comp_mask;
699                 vector unsigned char mask;
700 #ifdef MLX5_PMD_SOFT_COUNTERS
701                 const vector unsigned char lower_half = {
702                         0, 1, 4, 5, 8, 9, 12, 13,
703                         16, 17, 20, 21, 24, 25, 28, 29};
704                 const vector unsigned char upper_half = {
705                         2, 3, 6, 7, 10, 11, 14, 15,
706                         18, 19, 22, 23, 26, 27, 30, 31};
707                 const vector unsigned long shmax = {64, 64};
708                 vector unsigned char byte_cnt;
709                 vector unsigned short left, right;
710                 vector unsigned long lshift;
711                 vector __attribute__((altivec(bool__)))
712                         unsigned long shmask;
713 #endif
714                 vector unsigned char mbp1, mbp2;
715                 vector unsigned char p =
716                         (vector unsigned char)(vector unsigned short){
717                                 0, 1, 2, 3, 0, 0, 0, 0};
718                 unsigned int p1, p2, p3;
719
720                 /* Prefetch next 4 CQEs. */
721                 if (pkts_n - pos >= 2 * MLX5_VPMD_DESCS_PER_LOOP) {
722                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP]);
723                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 1]);
724                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 2]);
725                         rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 3]);
726                 }
727
728                 /* A.0 do not cross the end of CQ. */
729                 mask = (vector unsigned char)(vector unsigned long){
730                         (pkts_n - pos) * sizeof(uint16_t) * 8, 0};
731
732                 {
733                         vector unsigned long lshift;
734                         vector __attribute__((altivec(bool__)))
735                                 unsigned long shmask;
736                         const vector unsigned long shmax = {64, 64};
737
738                         lshift = vec_splat((vector unsigned long)mask, 0);
739                         shmask = vec_cmpgt(shmax, lshift);
740                         mask = (vector unsigned char)
741                                 vec_sl((vector unsigned long)ones, lshift);
742                         mask = (vector unsigned char)
743                                 vec_sel((vector unsigned long)shmask,
744                                 (vector unsigned long)mask, shmask);
745                 }
746
747                 p = (vector unsigned char)
748                         vec_andc((vector unsigned long)p,
749                         (vector unsigned long)mask);
750
751                 /* A.1 load cqes. */
752                 p3 = (unsigned int)((vector unsigned short)p)[3];
753                 cqes[3] = (vector unsigned char)(vector unsigned long){
754                         *(__rte_aligned(8) unsigned long *)
755                         &cq[pos + p3].sop_drop_qpn, 0LL};
756                 rte_compiler_barrier();
757
758                 p2 = (unsigned int)((vector unsigned short)p)[2];
759                 cqes[2] = (vector unsigned char)(vector unsigned long){
760                         *(__rte_aligned(8) unsigned long *)
761                         &cq[pos + p2].sop_drop_qpn, 0LL};
762                 rte_compiler_barrier();
763
764                 /* B.1 load mbuf pointers. */
765                 mbp1 = (vector unsigned char)vec_vsx_ld(0,
766                         (signed int const *)&elts[pos]);
767                 mbp2 = (vector unsigned char)vec_vsx_ld(0,
768                         (signed int const *)&elts[pos + 2]);
769
770                 /* A.1 load a block having op_own. */
771                 p1 = (unsigned int)((vector unsigned short)p)[1];
772                 cqes[1] = (vector unsigned char)(vector unsigned long){
773                         *(__rte_aligned(8) unsigned long *)
774                         &cq[pos + p1].sop_drop_qpn, 0LL};
775                 rte_compiler_barrier();
776
777                 cqes[0] = (vector unsigned char)(vector unsigned long){
778                         *(__rte_aligned(8) unsigned long *)
779                         &cq[pos].sop_drop_qpn, 0LL};
780                 rte_compiler_barrier();
781
782                 /* B.2 copy mbuf pointers. */
783                 *(vector unsigned char *)&pkts[pos] = mbp1;
784                 *(vector unsigned char *)&pkts[pos + 2] = mbp2;
785                 rte_cio_rmb();
786
787                 /* C.1 load remaining CQE data and extract necessary fields. */
788                 cqe_tmp2 = *(vector unsigned char *)
789                         &cq[pos + p3].pkt_info;
790                 cqe_tmp1 = *(vector unsigned char *)
791                         &cq[pos + p2].pkt_info;
792                 cqes[3] = vec_sel(cqes[3], cqe_tmp2, blend_mask);
793                 cqes[2] = vec_sel(cqes[2], cqe_tmp1, blend_mask);
794                 cqe_tmp2 = (vector unsigned char)vec_vsx_ld(0,
795                         (signed int const *)&cq[pos + p3].csum);
796                 cqe_tmp1 = (vector unsigned char)vec_vsx_ld(0,
797                         (signed int const *)&cq[pos + p2].csum);
798                 cqes[3] = (vector unsigned char)
799                         vec_sel((vector unsigned short)cqes[3],
800                         (vector unsigned short)cqe_tmp2, cqe_sel_mask1);
801                 cqes[2] = (vector unsigned char)
802                         vec_sel((vector unsigned short)cqes[2],
803                         (vector unsigned short)cqe_tmp1, cqe_sel_mask1);
804                 cqe_tmp2 = (vector unsigned char)(vector unsigned long){
805                         *(__rte_aligned(8) unsigned long *)
806                         &cq[pos + p3].rsvd3[9], 0LL};
807                 cqe_tmp1 = (vector unsigned char)(vector unsigned long){
808                         *(__rte_aligned(8) unsigned long *)
809                         &cq[pos + p2].rsvd3[9], 0LL};
810                 cqes[3] = (vector unsigned char)
811                         vec_sel((vector unsigned short)cqes[3],
812                         (vector unsigned short)cqe_tmp2,
813                         (vector unsigned short)cqe_sel_mask2);
814                 cqes[2] = (vector unsigned char)
815                         vec_sel((vector unsigned short)cqes[2],
816                         (vector unsigned short)cqe_tmp1,
817                         (vector unsigned short)cqe_sel_mask2);
818
819                 /* C.2 generate final structure for mbuf with swapping bytes. */
820                 pkt_mb3 = vec_perm(cqes[3], zero, shuf_mask);
821                 pkt_mb2 = vec_perm(cqes[2], zero, shuf_mask);
822
823                 /* C.3 adjust CRC length. */
824                 pkt_mb3 = (vector unsigned char)
825                         ((vector unsigned short)pkt_mb3 -
826                         (vector unsigned short)crc_adj);
827                 pkt_mb2 = (vector unsigned char)
828                         ((vector unsigned short)pkt_mb2 -
829                         (vector unsigned short)crc_adj);
830
831                 /* C.4 adjust flow mark. */
832                 pkt_mb3 = (vector unsigned char)
833                         ((vector unsigned int)pkt_mb3 +
834                         (vector unsigned int)flow_mark_adj);
835                 pkt_mb2 = (vector unsigned char)
836                         ((vector unsigned int)pkt_mb2 +
837                         (vector unsigned int)flow_mark_adj);
838
839                 /* D.1 fill in mbuf - rx_descriptor_fields1. */
840                 *(vector unsigned char *)
841                         &pkts[pos + 3]->pkt_len = pkt_mb3;
842                 *(vector unsigned char *)
843                         &pkts[pos + 2]->pkt_len = pkt_mb2;
844
845                 /* E.1 extract op_own field. */
846                 op_own_tmp2 = (vector unsigned char)
847                         vec_mergeh((vector unsigned int)cqes[2],
848                         (vector unsigned int)cqes[3]);
849
850                 /* C.1 load remaining CQE data and extract necessary fields. */
851                 cqe_tmp2 = *(vector unsigned char *)
852                         &cq[pos + p1].pkt_info;
853                 cqe_tmp1 = *(vector unsigned char *)
854                         &cq[pos].pkt_info;
855                 cqes[1] = vec_sel(cqes[1], cqe_tmp2, blend_mask);
856                 cqes[0] = vec_sel(cqes[0], cqe_tmp2, blend_mask);
857                 cqe_tmp2 = (vector unsigned char)vec_vsx_ld(0,
858                         (signed int const *)&cq[pos + p1].csum);
859                 cqe_tmp1 = (vector unsigned char)vec_vsx_ld(0,
860                         (signed int const *)&cq[pos].csum);
861                 cqes[1] = (vector unsigned char)
862                         vec_sel((vector unsigned short)cqes[1],
863                         (vector unsigned short)cqe_tmp2, cqe_sel_mask1);
864                 cqes[0] = (vector unsigned char)
865                         vec_sel((vector unsigned short)cqes[0],
866                         (vector unsigned short)cqe_tmp1, cqe_sel_mask1);
867                 cqe_tmp2 = (vector unsigned char)(vector unsigned long){
868                         *(__rte_aligned(8) unsigned long *)
869                         &cq[pos + p1].rsvd3[9], 0LL};
870                 cqe_tmp1 = (vector unsigned char)(vector unsigned long){
871                         *(__rte_aligned(8) unsigned long *)
872                         &cq[pos].rsvd3[9], 0LL};
873                 cqes[1] = (vector unsigned char)
874                         vec_sel((vector unsigned short)cqes[1],
875                         (vector unsigned short)cqe_tmp2, cqe_sel_mask2);
876                 cqes[0] = (vector unsigned char)
877                         vec_sel((vector unsigned short)cqes[0],
878                         (vector unsigned short)cqe_tmp1, cqe_sel_mask2);
879
880                 /* C.2 generate final structure for mbuf with swapping bytes. */
881                 pkt_mb1 = vec_perm(cqes[1], zero, shuf_mask);
882                 pkt_mb0 = vec_perm(cqes[0], zero, shuf_mask);
883
884                 /* C.3 adjust CRC length. */
885                 pkt_mb1 = (vector unsigned char)
886                         ((vector unsigned short)pkt_mb1 -
887                         (vector unsigned short)crc_adj);
888                 pkt_mb0 = (vector unsigned char)
889                         ((vector unsigned short)pkt_mb0 -
890                         (vector unsigned short)crc_adj);
891
892                 /* C.4 adjust flow mark. */
893                 pkt_mb1 = (vector unsigned char)
894                         ((vector unsigned int)pkt_mb1 +
895                         (vector unsigned int)flow_mark_adj);
896                 pkt_mb0 = (vector unsigned char)
897                         ((vector unsigned int)pkt_mb0 +
898                         (vector unsigned int)flow_mark_adj);
899
900                 /* E.1 extract op_own byte. */
901                 op_own_tmp1 = (vector unsigned char)
902                         vec_mergeh((vector unsigned int)cqes[0],
903                         (vector unsigned int)cqes[1]);
904                 op_own = (vector unsigned char)
905                         vec_mergel((vector unsigned long)op_own_tmp1,
906                         (vector unsigned long)op_own_tmp2);
907
908                 /* D.1 fill in mbuf - rx_descriptor_fields1. */
909                 *(vector unsigned char *)
910                         &pkts[pos + 1]->pkt_len = pkt_mb1;
911                 *(vector unsigned char *)
912                         &pkts[pos]->pkt_len = pkt_mb0;
913
914                 /* E.2 flip owner bit to mark CQEs from last round. */
915                 owner_mask = (vector unsigned char)
916                         vec_and((vector unsigned long)op_own,
917                         (vector unsigned long)owner_check);
918                 if (ownership)
919                         owner_mask = (vector unsigned char)
920                                 vec_xor((vector unsigned long)owner_mask,
921                                 (vector unsigned long)owner_check);
922                 owner_mask = (vector unsigned char)
923                         vec_cmpeq((vector unsigned int)owner_mask,
924                         (vector unsigned int)owner_check);
925                 owner_mask = (vector unsigned char)
926                         vec_packs((vector unsigned int)owner_mask,
927                         (vector unsigned int)zero);
928
929                 /* E.3 get mask for invalidated CQEs. */
930                 opcode = (vector unsigned char)
931                         vec_and((vector unsigned long)op_own,
932                         (vector unsigned long)opcode_check);
933                 invalid_mask = (vector unsigned char)
934                         vec_cmpeq((vector unsigned int)opcode_check,
935                         (vector unsigned int)opcode);
936                 invalid_mask = (vector unsigned char)
937                         vec_packs((vector unsigned int)invalid_mask,
938                         (vector unsigned int)zero);
939
940                 /* E.4 mask out beyond boundary. */
941                 invalid_mask = (vector unsigned char)
942                         vec_or((vector unsigned long)invalid_mask,
943                         (vector unsigned long)mask);
944
945                 /* E.5 merge invalid_mask with invalid owner. */
946                 invalid_mask = (vector unsigned char)
947                         vec_or((vector unsigned long)invalid_mask,
948                         (vector unsigned long)owner_mask);
949
950                 /* F.1 find compressed CQE format. */
951                 comp_mask = (vector unsigned char)
952                         vec_and((vector unsigned long)op_own,
953                         (vector unsigned long)format_check);
954                 comp_mask = (vector unsigned char)
955                         vec_cmpeq((vector unsigned int)comp_mask,
956                         (vector unsigned int)format_check);
957                 comp_mask = (vector unsigned char)
958                         vec_packs((vector unsigned int)comp_mask,
959                         (vector unsigned int)zero);
960
961                 /* F.2 mask out invalid entries. */
962                 comp_mask = (vector unsigned char)
963                         vec_andc((vector unsigned long)comp_mask,
964                         (vector unsigned long)invalid_mask);
965                 comp_idx = ((vector unsigned long)comp_mask)[0];
966
967                 /* F.3 get the first compressed CQE. */
968                 comp_idx = comp_idx ? __builtin_ctzll(comp_idx) /
969                         (sizeof(uint16_t) * 8) : MLX5_VPMD_DESCS_PER_LOOP;
970
971                 /* E.6 mask out entries after the compressed CQE. */
972                 mask = (vector unsigned char)(vector unsigned long){
973                         (comp_idx * sizeof(uint16_t) * 8), 0};
974                 lshift = vec_splat((vector unsigned long)mask, 0);
975                 shmask = vec_cmpgt(shmax, lshift);
976                 mask = (vector unsigned char)
977                         vec_sl((vector unsigned long)ones, lshift);
978                 mask = (vector unsigned char)
979                         vec_sel((vector unsigned long)shmask,
980                         (vector unsigned long)mask, shmask);
981                 invalid_mask = (vector unsigned char)
982                         vec_or((vector unsigned long)invalid_mask,
983                         (vector unsigned long)mask);
984
985                 /* E.7 count non-compressed valid CQEs. */
986                 n = ((vector unsigned long)invalid_mask)[0];
987                 n = n ? __builtin_ctzll(n) / (sizeof(uint16_t) * 8) :
988                         MLX5_VPMD_DESCS_PER_LOOP;
989                 nocmp_n += n;
990
991                 /* D.2 get the final invalid mask. */
992                 mask = (vector unsigned char)(vector unsigned long){
993                         (n * sizeof(uint16_t) * 8), 0};
994                 lshift = vec_splat((vector unsigned long)mask, 0);
995                 shmask = vec_cmpgt(shmax, lshift);
996                 mask = (vector unsigned char)
997                         vec_sl((vector unsigned long)ones, lshift);
998                 mask = (vector unsigned char)
999                         vec_sel((vector unsigned long)shmask,
1000                         (vector unsigned long)mask, shmask);
1001                 invalid_mask = (vector unsigned char)
1002                         vec_or((vector unsigned long)invalid_mask,
1003                         (vector unsigned long)mask);
1004
1005                 /* D.3 check error in opcode. */
1006                 opcode = (vector unsigned char)
1007                         vec_cmpeq((vector unsigned int)resp_err_check,
1008                         (vector unsigned int)opcode);
1009                 opcode = (vector unsigned char)
1010                         vec_packs((vector unsigned int)opcode,
1011                         (vector unsigned int)zero);
1012                 opcode = (vector unsigned char)
1013                         vec_andc((vector unsigned long)opcode,
1014                         (vector unsigned long)invalid_mask);
1015
1016                 /* D.4 mark if any error is set */
1017                 *err |= ((vector unsigned long)opcode)[0];
1018
1019                 /* D.5 fill in mbuf - rearm_data and packet_type. */
1020                 rxq_cq_to_ptype_oflags_v(rxq, cqes, opcode, &pkts[pos]);
1021                 if (rxq->hw_timestamp) {
1022                         pkts[pos]->timestamp =
1023                                 rte_be_to_cpu_64(cq[pos].timestamp);
1024                         pkts[pos + 1]->timestamp =
1025                                 rte_be_to_cpu_64(cq[pos + p1].timestamp);
1026                         pkts[pos + 2]->timestamp =
1027                                 rte_be_to_cpu_64(cq[pos + p2].timestamp);
1028                         pkts[pos + 3]->timestamp =
1029                                 rte_be_to_cpu_64(cq[pos + p3].timestamp);
1030                 }
1031                 if (!!rxq->flow_meta_mask) {
1032                         uint64_t flag = rxq->flow_meta_mask;
1033                         int32_t offs = rxq->flow_meta_offset;
1034                         uint32_t metadata;
1035
1036                         /* This code is subject for futher optimization. */
1037                         metadata = cq[pos].flow_table_metadata;
1038                         *RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *) =
1039                                                                 metadata;
1040                         pkts[pos]->ol_flags |= metadata ? flag : 0ULL;
1041                         metadata = cq[pos + 1].flow_table_metadata;
1042                         *RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *) =
1043                                                                 metadata;
1044                         pkts[pos + 1]->ol_flags |= metadata ? flag : 0ULL;
1045                         metadata = cq[pos + 2].flow_table_metadata;
1046                         *RTE_MBUF_DYNFIELD(pkts[pos + 2], offs, uint32_t *) =
1047                                                                 metadata;
1048                         pkts[pos + 2]->ol_flags |= metadata ? flag : 0ULL;
1049                         metadata = cq[pos + 3].flow_table_metadata;
1050                         *RTE_MBUF_DYNFIELD(pkts[pos + 3], offs, uint32_t *) =
1051                                                                 metadata;
1052                         pkts[pos + 3]->ol_flags |= metadata ? flag : 0ULL;
1053                 }
1054 #ifdef MLX5_PMD_SOFT_COUNTERS
1055                 /* Add up received bytes count. */
1056                 byte_cnt = vec_perm(op_own, zero, len_shuf_mask);
1057                 byte_cnt = (vector unsigned char)
1058                         vec_andc((vector unsigned long)byte_cnt,
1059                         (vector unsigned long)invalid_mask);
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                 left = vec_perm((vector unsigned short)byte_cnt,
1066                         (vector unsigned short)zero, lower_half);
1067                 right = vec_perm((vector unsigned short)byte_cnt,
1068                         (vector unsigned short)zero, upper_half);
1069                 byte_cnt = (vector unsigned char)vec_add(left, right);
1070                 rcvd_byte += ((vector unsigned long)byte_cnt)[0];
1071 #endif
1072
1073                 /*
1074                  * Break the loop unless more valid CQE is expected, or if
1075                  * there's a compressed CQE.
1076                  */
1077                 if (n != MLX5_VPMD_DESCS_PER_LOOP)
1078                         break;
1079         }
1080         /* If no new CQE seen, return without updating cq_db. */
1081         if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP))
1082                 return rcvd_pkt;
1083         /* Update the consumer indexes for non-compressed CQEs. */
1084         MLX5_ASSERT(nocmp_n <= pkts_n);
1085         rxq->cq_ci += nocmp_n;
1086         rxq->rq_pi += nocmp_n;
1087         rcvd_pkt += nocmp_n;
1088 #ifdef MLX5_PMD_SOFT_COUNTERS
1089         rxq->stats.ipackets += nocmp_n;
1090         rxq->stats.ibytes += rcvd_byte;
1091 #endif
1092         /* Decompress the last CQE if compressed. */
1093         if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP && comp_idx == n) {
1094                 MLX5_ASSERT(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
1095                 rxq->decompressed =
1096                         rxq_cq_decompress_v(rxq, &cq[nocmp_n], &elts[nocmp_n]);
1097                 /* Return more packets if needed. */
1098                 if (nocmp_n < pkts_n) {
1099                         uint16_t n = rxq->decompressed;
1100
1101                         n = RTE_MIN(n, pkts_n - nocmp_n);
1102                         rxq_copy_mbuf_v(rxq, &pkts[nocmp_n], n);
1103                         rxq->rq_pi += n;
1104                         rcvd_pkt += n;
1105                         rxq->decompressed -= n;
1106                 }
1107         }
1108         rte_compiler_barrier();
1109         *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1110         return rcvd_pkt;
1111 }
1112
1113 #endif /* RTE_PMD_MLX5_RXTX_VEC_ALTIVEC_H_ */