drivers/event/cnxk/cn9k_worker.h

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(C) 2021 Marvell.
   3  */
   4
   5 #ifndef __CN9K_WORKER_H__
   6 #define __CN9K_WORKER_H__
   7
   8 #include <rte_eventdev.h>
   9 #include <rte_vect.h>
  10
  11 #include "cnxk_ethdev.h"
  12 #include "cnxk_eventdev.h"
  13 #include "cnxk_worker.h"
  14 #include "cn9k_cryptodev_ops.h"
  15
  16 #include "cn9k_ethdev.h"
  17 #include "cn9k_rx.h"
  18 #include "cn9k_tx.h"
  19
  20 /* SSO Operations */
  21
  22 static __rte_always_inline uint8_t
  23 cn9k_sso_hws_new_event(struct cn9k_sso_hws *ws, const struct rte_event *ev)
  24 {
  25         const uint32_t tag = (uint32_t)ev->event;
  26         const uint8_t new_tt = ev->sched_type;
  27         const uint64_t event_ptr = ev->u64;
  28         const uint16_t grp = ev->queue_id;
  29
  30         rte_atomic_thread_fence(__ATOMIC_ACQ_REL);
  31         if (ws->xaq_lmt <= *ws->fc_mem)
  32                 return 0;
  33
  34         cnxk_sso_hws_add_work(event_ptr, tag, new_tt,
  35                               ws->grp_base + (grp << 12));
  36         return 1;
  37 }
  38
  39 static __rte_always_inline void
  40 cn9k_sso_hws_fwd_swtag(uint64_t base, const struct rte_event *ev)
  41 {
  42         const uint32_t tag = (uint32_t)ev->event;
  43         const uint8_t new_tt = ev->sched_type;
  44         const uint8_t cur_tt =
  45                 CNXK_TT_FROM_TAG(plt_read64(base + SSOW_LF_GWS_TAG));
  46
  47         /* CNXK model
  48          * cur_tt/new_tt     SSO_TT_ORDERED SSO_TT_ATOMIC SSO_TT_UNTAGGED
  49          *
  50          * SSO_TT_ORDERED        norm           norm             untag
  51          * SSO_TT_ATOMIC         norm           norm               untag
  52          * SSO_TT_UNTAGGED       norm           norm             NOOP
  53          */
  54
  55         if (new_tt == SSO_TT_UNTAGGED) {
  56                 if (cur_tt != SSO_TT_UNTAGGED)
  57                         cnxk_sso_hws_swtag_untag(base +
  58                                                  SSOW_LF_GWS_OP_SWTAG_UNTAG);
  59         } else {
  60                 cnxk_sso_hws_swtag_norm(tag, new_tt,
  61                                         base + SSOW_LF_GWS_OP_SWTAG_NORM);
  62         }
  63 }
  64
  65 static __rte_always_inline void
  66 cn9k_sso_hws_fwd_group(uint64_t base, const struct rte_event *ev,
  67                        const uint16_t grp)
  68 {
  69         const uint32_t tag = (uint32_t)ev->event;
  70         const uint8_t new_tt = ev->sched_type;
  71
  72         plt_write64(ev->u64, base + SSOW_LF_GWS_OP_UPD_WQP_GRP1);
  73         cnxk_sso_hws_swtag_desched(tag, new_tt, grp,
  74                                    base + SSOW_LF_GWS_OP_SWTAG_DESCHED);
  75 }
  76
  77 static __rte_always_inline void
  78 cn9k_sso_hws_forward_event(struct cn9k_sso_hws *ws, const struct rte_event *ev)
  79 {
  80         const uint8_t grp = ev->queue_id;
  81
  82         /* Group hasn't changed, Use SWTAG to forward the event */
  83         if (CNXK_GRP_FROM_TAG(plt_read64(ws->base + SSOW_LF_GWS_TAG)) == grp) {
  84                 cn9k_sso_hws_fwd_swtag(ws->base, ev);
  85                 ws->swtag_req = 1;
  86         } else {
  87                 /*
  88                  * Group has been changed for group based work pipelining,
  89                  * Use deschedule/add_work operation to transfer the event to
  90                  * new group/core
  91                  */
  92                 cn9k_sso_hws_fwd_group(ws->base, ev, grp);
  93         }
  94 }
  95
  96 /* Dual ws ops. */
  97
  98 static __rte_always_inline uint8_t
  99 cn9k_sso_hws_dual_new_event(struct cn9k_sso_hws_dual *dws,
 100                             const struct rte_event *ev)
 101 {
 102         const uint32_t tag = (uint32_t)ev->event;
 103         const uint8_t new_tt = ev->sched_type;
 104         const uint64_t event_ptr = ev->u64;
 105         const uint16_t grp = ev->queue_id;
 106
 107         rte_atomic_thread_fence(__ATOMIC_ACQ_REL);
 108         if (dws->xaq_lmt <= *dws->fc_mem)
 109                 return 0;
 110
 111         cnxk_sso_hws_add_work(event_ptr, tag, new_tt,
 112                               dws->grp_base + (grp << 12));
 113         return 1;
 114 }
 115
 116 static __rte_always_inline void
 117 cn9k_sso_hws_dual_forward_event(struct cn9k_sso_hws_dual *dws, uint64_t base,
 118                                 const struct rte_event *ev)
 119 {
 120         const uint8_t grp = ev->queue_id;
 121
 122         /* Group hasn't changed, Use SWTAG to forward the event */
 123         if (CNXK_GRP_FROM_TAG(plt_read64(base + SSOW_LF_GWS_TAG)) == grp) {
 124                 cn9k_sso_hws_fwd_swtag(base, ev);
 125                 dws->swtag_req = 1;
 126         } else {
 127                 /*
 128                  * Group has been changed for group based work pipelining,
 129                  * Use deschedule/add_work operation to transfer the event to
 130                  * new group/core
 131                  */
 132                 cn9k_sso_hws_fwd_group(base, ev, grp);
 133         }
 134 }
 135
 136 static __rte_always_inline void
 137 cn9k_wqe_to_mbuf(uint64_t wqe, const uint64_t mbuf, uint8_t port_id,
 138                  const uint32_t tag, const uint32_t flags,
 139                  const void *const lookup_mem)
 140 {
 141         const uint64_t mbuf_init = 0x100010000ULL | RTE_PKTMBUF_HEADROOM |
 142                                    (flags & NIX_RX_OFFLOAD_TSTAMP_F ? 8 : 0);
 143
 144         cn9k_nix_cqe_to_mbuf((struct nix_cqe_hdr_s *)wqe, tag,
 145                              (struct rte_mbuf *)mbuf, lookup_mem,
 146                              mbuf_init | ((uint64_t)port_id) << 48, flags);
 147 }
 148
 149 static __rte_always_inline uint16_t
 150 cn9k_sso_hws_dual_get_work(uint64_t base, uint64_t pair_base,
 151                            struct rte_event *ev, const uint32_t flags,
 152                            const void *const lookup_mem,
 153                            struct cnxk_timesync_info *const tstamp)
 154 {
 155         const uint64_t set_gw = BIT_ULL(16) | 1;
 156         union {
 157                 __uint128_t get_work;
 158                 uint64_t u64[2];
 159         } gw;
 160         uint64_t tstamp_ptr;
 161         uint64_t mbuf;
 162
 163         if (flags & NIX_RX_OFFLOAD_PTYPE_F)
 164                 rte_prefetch_non_temporal(lookup_mem);
 165 #ifdef RTE_ARCH_ARM64
 166         asm volatile(PLT_CPU_FEATURE_PREAMBLE
 167                      "rty%=:                                    \n"
 168                      "          ldr %[tag], [%[tag_loc]]        \n"
 169                      "          ldr %[wqp], [%[wqp_loc]]        \n"
 170                      "          tbnz %[tag], 63, rty%=          \n"
 171                      "done%=:   str %[gw], [%[pong]]            \n"
 172                      "          dmb ld                          \n"
 173                      "          sub %[mbuf], %[wqp], #0x80      \n"
 174                      "          prfm pldl1keep, [%[mbuf]]       \n"
 175                      : [tag] "=&r"(gw.u64[0]), [wqp] "=&r"(gw.u64[1]),
 176                        [mbuf] "=&r"(mbuf)
 177                      : [tag_loc] "r"(base + SSOW_LF_GWS_TAG),
 178                        [wqp_loc] "r"(base + SSOW_LF_GWS_WQP), [gw] "r"(set_gw),
 179                        [pong] "r"(pair_base + SSOW_LF_GWS_OP_GET_WORK0));
 180 #else
 181         gw.u64[0] = plt_read64(base + SSOW_LF_GWS_TAG);
 182         while ((BIT_ULL(63)) & gw.u64[0])
 183                 gw.u64[0] = plt_read64(base + SSOW_LF_GWS_TAG);
 184         gw.u64[1] = plt_read64(base + SSOW_LF_GWS_WQP);
 185         plt_write64(set_gw, pair_base + SSOW_LF_GWS_OP_GET_WORK0);
 186         mbuf = (uint64_t)((char *)gw.u64[1] - sizeof(struct rte_mbuf));
 187 #endif
 188
 189         gw.u64[0] = (gw.u64[0] & (0x3ull << 32)) << 6 |
 190                     (gw.u64[0] & (0x3FFull << 36)) << 4 |
 191                     (gw.u64[0] & 0xffffffff);
 192
 193         if (CNXK_TT_FROM_EVENT(gw.u64[0]) != SSO_TT_EMPTY) {
 194                 if ((flags & CPT_RX_WQE_F) &&
 195                     (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
 196                      RTE_EVENT_TYPE_CRYPTODEV)) {
 197                         gw.u64[1] = cn9k_cpt_crypto_adapter_dequeue(gw.u64[1]);
 198                 } else if (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
 199                            RTE_EVENT_TYPE_ETHDEV) {
 200                         uint8_t port = CNXK_SUB_EVENT_FROM_TAG(gw.u64[0]);
 201
 202                         gw.u64[0] = CNXK_CLR_SUB_EVENT(gw.u64[0]);
 203                         cn9k_wqe_to_mbuf(gw.u64[1], mbuf, port,
 204                                          gw.u64[0] & 0xFFFFF, flags,
 205                                          lookup_mem);
 206                         /* Extracting tstamp, if PTP enabled*/
 207                         tstamp_ptr = *(uint64_t *)(((struct nix_wqe_hdr_s *)
 208                                                             gw.u64[1]) +
 209                                                    CNXK_SSO_WQE_SG_PTR);
 210                         cnxk_nix_mbuf_to_tstamp((struct rte_mbuf *)mbuf, tstamp,
 211                                                 flags & NIX_RX_OFFLOAD_TSTAMP_F,
 212                                                 flags & NIX_RX_MULTI_SEG_F,
 213                                                 (uint64_t *)tstamp_ptr);
 214                         gw.u64[1] = mbuf;
 215                 }
 216         }
 217
 218         ev->event = gw.u64[0];
 219         ev->u64 = gw.u64[1];
 220
 221         return !!gw.u64[1];
 222 }
 223
 224 static __rte_always_inline uint16_t
 225 cn9k_sso_hws_get_work(struct cn9k_sso_hws *ws, struct rte_event *ev,
 226                       const uint32_t flags, const void *const lookup_mem)
 227 {
 228         union {
 229                 __uint128_t get_work;
 230                 uint64_t u64[2];
 231         } gw;
 232         uint64_t tstamp_ptr;
 233         uint64_t mbuf;
 234
 235         plt_write64(BIT_ULL(16) | /* wait for work. */
 236                             1,    /* Use Mask set 0. */
 237                     ws->base + SSOW_LF_GWS_OP_GET_WORK0);
 238
 239         if (flags & NIX_RX_OFFLOAD_PTYPE_F)
 240                 rte_prefetch_non_temporal(lookup_mem);
 241 #ifdef RTE_ARCH_ARM64
 242         asm volatile(PLT_CPU_FEATURE_PREAMBLE
 243                      "          ldr %[tag], [%[tag_loc]]        \n"
 244                      "          ldr %[wqp], [%[wqp_loc]]        \n"
 245                      "          tbz %[tag], 63, done%=          \n"
 246                      "          sevl                            \n"
 247                      "rty%=:    wfe                             \n"
 248                      "          ldr %[tag], [%[tag_loc]]        \n"
 249                      "          ldr %[wqp], [%[wqp_loc]]        \n"
 250                      "          tbnz %[tag], 63, rty%=          \n"
 251                      "done%=:   dmb ld                          \n"
 252                      "          sub %[mbuf], %[wqp], #0x80      \n"
 253                      "          prfm pldl1keep, [%[mbuf]]       \n"
 254                      : [tag] "=&r"(gw.u64[0]), [wqp] "=&r"(gw.u64[1]),
 255                        [mbuf] "=&r"(mbuf)
 256                      : [tag_loc] "r"(ws->base + SSOW_LF_GWS_TAG),
 257                        [wqp_loc] "r"(ws->base + SSOW_LF_GWS_WQP));
 258 #else
 259         gw.u64[0] = plt_read64(ws->base + SSOW_LF_GWS_TAG);
 260         while ((BIT_ULL(63)) & gw.u64[0])
 261                 gw.u64[0] = plt_read64(ws->base + SSOW_LF_GWS_TAG);
 262
 263         gw.u64[1] = plt_read64(ws->base + SSOW_LF_GWS_WQP);
 264         mbuf = (uint64_t)((char *)gw.u64[1] - sizeof(struct rte_mbuf));
 265 #endif
 266
 267         gw.u64[0] = (gw.u64[0] & (0x3ull << 32)) << 6 |
 268                     (gw.u64[0] & (0x3FFull << 36)) << 4 |
 269                     (gw.u64[0] & 0xffffffff);
 270
 271         if (CNXK_TT_FROM_EVENT(gw.u64[0]) != SSO_TT_EMPTY) {
 272                 if ((flags & CPT_RX_WQE_F) &&
 273                     (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
 274                      RTE_EVENT_TYPE_CRYPTODEV)) {
 275                         gw.u64[1] = cn9k_cpt_crypto_adapter_dequeue(gw.u64[1]);
 276                 } else if (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
 277                            RTE_EVENT_TYPE_ETHDEV) {
 278                         uint8_t port = CNXK_SUB_EVENT_FROM_TAG(gw.u64[0]);
 279
 280                         gw.u64[0] = CNXK_CLR_SUB_EVENT(gw.u64[0]);
 281                         cn9k_wqe_to_mbuf(gw.u64[1], mbuf, port,
 282                                          gw.u64[0] & 0xFFFFF, flags,
 283                                          lookup_mem);
 284                         /* Extracting tstamp, if PTP enabled*/
 285                         tstamp_ptr = *(uint64_t *)(((struct nix_wqe_hdr_s *)
 286                                                             gw.u64[1]) +
 287                                                    CNXK_SSO_WQE_SG_PTR);
 288                         cnxk_nix_mbuf_to_tstamp((struct rte_mbuf *)mbuf,
 289                                                 ws->tstamp,
 290                                                 flags & NIX_RX_OFFLOAD_TSTAMP_F,
 291                                                 flags & NIX_RX_MULTI_SEG_F,
 292                                                 (uint64_t *)tstamp_ptr);
 293                         gw.u64[1] = mbuf;
 294                 }
 295         }
 296
 297         ev->event = gw.u64[0];
 298         ev->u64 = gw.u64[1];
 299
 300         return !!gw.u64[1];
 301 }
 302
 303 /* Used in cleaning up workslot. */
 304 static __rte_always_inline uint16_t
 305 cn9k_sso_hws_get_work_empty(uint64_t base, struct rte_event *ev)
 306 {
 307         union {
 308                 __uint128_t get_work;
 309                 uint64_t u64[2];
 310         } gw;
 311         uint64_t mbuf;
 312
 313 #ifdef RTE_ARCH_ARM64
 314         asm volatile(PLT_CPU_FEATURE_PREAMBLE
 315                      "          ldr %[tag], [%[tag_loc]]        \n"
 316                      "          ldr %[wqp], [%[wqp_loc]]        \n"
 317                      "          tbz %[tag], 63, done%=          \n"
 318                      "          sevl                            \n"
 319                      "rty%=:    wfe                             \n"
 320                      "          ldr %[tag], [%[tag_loc]]        \n"
 321                      "          ldr %[wqp], [%[wqp_loc]]        \n"
 322                      "          tbnz %[tag], 63, rty%=          \n"
 323                      "done%=:   dmb ld                          \n"
 324                      "          sub %[mbuf], %[wqp], #0x80      \n"
 325                      : [tag] "=&r"(gw.u64[0]), [wqp] "=&r"(gw.u64[1]),
 326                        [mbuf] "=&r"(mbuf)
 327                      : [tag_loc] "r"(base + SSOW_LF_GWS_TAG),
 328                        [wqp_loc] "r"(base + SSOW_LF_GWS_WQP));
 329 #else
 330         gw.u64[0] = plt_read64(base + SSOW_LF_GWS_TAG);
 331         while ((BIT_ULL(63)) & gw.u64[0])
 332                 gw.u64[0] = plt_read64(base + SSOW_LF_GWS_TAG);
 333
 334         gw.u64[1] = plt_read64(base + SSOW_LF_GWS_WQP);
 335         mbuf = (uint64_t)((char *)gw.u64[1] - sizeof(struct rte_mbuf));
 336 #endif
 337
 338         gw.u64[0] = (gw.u64[0] & (0x3ull << 32)) << 6 |
 339                     (gw.u64[0] & (0x3FFull << 36)) << 4 |
 340                     (gw.u64[0] & 0xffffffff);
 341
 342         if (CNXK_TT_FROM_EVENT(gw.u64[0]) != SSO_TT_EMPTY) {
 343                 if (CNXK_EVENT_TYPE_FROM_TAG(gw.u64[0]) ==
 344                     RTE_EVENT_TYPE_ETHDEV) {
 345                         uint8_t port = CNXK_SUB_EVENT_FROM_TAG(gw.u64[0]);
 346
 347                         gw.u64[0] = CNXK_CLR_SUB_EVENT(gw.u64[0]);
 348                         cn9k_wqe_to_mbuf(gw.u64[1], mbuf, port,
 349                                          gw.u64[0] & 0xFFFFF, 0, NULL);
 350                         gw.u64[1] = mbuf;
 351                 }
 352         }
 353
 354         ev->event = gw.u64[0];
 355         ev->u64 = gw.u64[1];
 356
 357         return !!gw.u64[1];
 358 }
 359
 360 /* CN9K Fastpath functions. */
 361 uint16_t __rte_hot cn9k_sso_hws_enq(void *port, const struct rte_event *ev);
 362 uint16_t __rte_hot cn9k_sso_hws_enq_burst(void *port,
 363                                           const struct rte_event ev[],
 364                                           uint16_t nb_events);
 365 uint16_t __rte_hot cn9k_sso_hws_enq_new_burst(void *port,
 366                                               const struct rte_event ev[],
 367                                               uint16_t nb_events);
 368 uint16_t __rte_hot cn9k_sso_hws_enq_fwd_burst(void *port,
 369                                               const struct rte_event ev[],
 370                                               uint16_t nb_events);
 371
 372 uint16_t __rte_hot cn9k_sso_hws_dual_enq(void *port,
 373                                          const struct rte_event *ev);
 374 uint16_t __rte_hot cn9k_sso_hws_dual_enq_burst(void *port,
 375                                                const struct rte_event ev[],
 376                                                uint16_t nb_events);
 377 uint16_t __rte_hot cn9k_sso_hws_dual_enq_new_burst(void *port,
 378                                                    const struct rte_event ev[],
 379                                                    uint16_t nb_events);
 380 uint16_t __rte_hot cn9k_sso_hws_dual_enq_fwd_burst(void *port,
 381                                                    const struct rte_event ev[],
 382                                                    uint16_t nb_events);
 383 uint16_t __rte_hot cn9k_sso_hws_ca_enq(void *port, struct rte_event ev[],
 384                                        uint16_t nb_events);
 385 uint16_t __rte_hot cn9k_sso_hws_dual_ca_enq(void *port, struct rte_event ev[],
 386                                             uint16_t nb_events);
 387
 388 #define R(name, flags)                                                         \
 389         uint16_t __rte_hot cn9k_sso_hws_deq_##name(                            \
 390                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 391         uint16_t __rte_hot cn9k_sso_hws_deq_burst_##name(                      \
 392                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 393                 uint64_t timeout_ticks);                                       \
 394         uint16_t __rte_hot cn9k_sso_hws_deq_tmo_##name(                        \
 395                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 396         uint16_t __rte_hot cn9k_sso_hws_deq_tmo_burst_##name(                  \
 397                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 398                 uint64_t timeout_ticks);                                       \
 399         uint16_t __rte_hot cn9k_sso_hws_deq_ca_##name(                         \
 400                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 401         uint16_t __rte_hot cn9k_sso_hws_deq_ca_burst_##name(                   \
 402                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 403                 uint64_t timeout_ticks);                                       \
 404         uint16_t __rte_hot cn9k_sso_hws_deq_tmo_ca_##name(                     \
 405                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 406         uint16_t __rte_hot cn9k_sso_hws_deq_tmo_ca_burst_##name(               \
 407                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 408                 uint64_t timeout_ticks);                                       \
 409         uint16_t __rte_hot cn9k_sso_hws_deq_seg_##name(                        \
 410                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 411         uint16_t __rte_hot cn9k_sso_hws_deq_seg_burst_##name(                  \
 412                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 413                 uint64_t timeout_ticks);                                       \
 414         uint16_t __rte_hot cn9k_sso_hws_deq_tmo_seg_##name(                    \
 415                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 416         uint16_t __rte_hot cn9k_sso_hws_deq_tmo_seg_burst_##name(              \
 417                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 418                 uint64_t timeout_ticks);                                       \
 419         uint16_t __rte_hot cn9k_sso_hws_deq_ca_seg_##name(                     \
 420                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 421         uint16_t __rte_hot cn9k_sso_hws_deq_ca_seg_burst_##name(               \
 422                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 423                 uint64_t timeout_ticks);                                       \
 424         uint16_t __rte_hot cn9k_sso_hws_deq_tmo_ca_seg_##name(                 \
 425                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 426         uint16_t __rte_hot cn9k_sso_hws_deq_tmo_ca_seg_burst_##name(           \
 427                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 428                 uint64_t timeout_ticks);
 429
 430 NIX_RX_FASTPATH_MODES
 431 #undef R
 432
 433 #define SSO_DEQ(fn, flags)                                                     \
 434         uint16_t __rte_hot fn(void *port, struct rte_event *ev,                \
 435                               uint64_t timeout_ticks)                          \
 436         {                                                                      \
 437                 struct cn9k_sso_hws *ws = port;                                \
 438                 RTE_SET_USED(timeout_ticks);                                   \
 439                 if (ws->swtag_req) {                                           \
 440                         ws->swtag_req = 0;                                     \
 441                         cnxk_sso_hws_swtag_wait(ws->base + SSOW_LF_GWS_TAG);   \
 442                         return 1;                                              \
 443                 }                                                              \
 444                 return cn9k_sso_hws_get_work(ws, ev, flags, ws->lookup_mem);   \
 445         }
 446
 447 #define SSO_DEQ_SEG(fn, flags)    SSO_DEQ(fn, flags | NIX_RX_MULTI_SEG_F)
 448 #define SSO_DEQ_CA(fn, flags)     SSO_DEQ(fn, flags | CPT_RX_WQE_F)
 449 #define SSO_DEQ_CA_SEG(fn, flags) SSO_DEQ_SEG(fn, flags | CPT_RX_WQE_F)
 450
 451 #define SSO_DEQ_TMO(fn, flags)                                                 \
 452         uint16_t __rte_hot fn(void *port, struct rte_event *ev,                \
 453                               uint64_t timeout_ticks)                          \
 454         {                                                                      \
 455                 struct cn9k_sso_hws *ws = port;                                \
 456                 uint16_t ret = 1;                                              \
 457                 uint64_t iter;                                                 \
 458                 if (ws->swtag_req) {                                           \
 459                         ws->swtag_req = 0;                                     \
 460                         cnxk_sso_hws_swtag_wait(ws->base + SSOW_LF_GWS_TAG);   \
 461                         return ret;                                            \
 462                 }                                                              \
 463                 ret = cn9k_sso_hws_get_work(ws, ev, flags, ws->lookup_mem);    \
 464                 for (iter = 1; iter < timeout_ticks && (ret == 0); iter++)     \
 465                         ret = cn9k_sso_hws_get_work(ws, ev, flags,             \
 466                                                     ws->lookup_mem);           \
 467                 return ret;                                                    \
 468         }
 469
 470 #define SSO_DEQ_TMO_SEG(fn, flags)    SSO_DEQ_TMO(fn, flags | NIX_RX_MULTI_SEG_F)
 471 #define SSO_DEQ_TMO_CA(fn, flags)     SSO_DEQ_TMO(fn, flags | CPT_RX_WQE_F)
 472 #define SSO_DEQ_TMO_CA_SEG(fn, flags) SSO_DEQ_TMO_SEG(fn, flags | CPT_RX_WQE_F)
 473
 474 #define R(name, flags)                                                         \
 475         uint16_t __rte_hot cn9k_sso_hws_dual_deq_##name(                       \
 476                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 477         uint16_t __rte_hot cn9k_sso_hws_dual_deq_burst_##name(                 \
 478                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 479                 uint64_t timeout_ticks);                                       \
 480         uint16_t __rte_hot cn9k_sso_hws_dual_deq_tmo_##name(                   \
 481                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 482         uint16_t __rte_hot cn9k_sso_hws_dual_deq_tmo_burst_##name(             \
 483                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 484                 uint64_t timeout_ticks);                                       \
 485         uint16_t __rte_hot cn9k_sso_hws_dual_deq_ca_##name(                    \
 486                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 487         uint16_t __rte_hot cn9k_sso_hws_dual_deq_ca_burst_##name(              \
 488                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 489                 uint64_t timeout_ticks);                                       \
 490         uint16_t __rte_hot cn9k_sso_hws_dual_deq_tmo_ca_##name(                \
 491                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 492         uint16_t __rte_hot cn9k_sso_hws_dual_deq_tmo_ca_burst_##name(          \
 493                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 494                 uint64_t timeout_ticks);                                       \
 495         uint16_t __rte_hot cn9k_sso_hws_dual_deq_seg_##name(                   \
 496                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 497         uint16_t __rte_hot cn9k_sso_hws_dual_deq_seg_burst_##name(             \
 498                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 499                 uint64_t timeout_ticks);                                       \
 500         uint16_t __rte_hot cn9k_sso_hws_dual_deq_tmo_seg_##name(               \
 501                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 502         uint16_t __rte_hot cn9k_sso_hws_dual_deq_tmo_seg_burst_##name(         \
 503                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 504                 uint64_t timeout_ticks);                                       \
 505         uint16_t __rte_hot cn9k_sso_hws_dual_deq_ca_seg_##name(                \
 506                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 507         uint16_t __rte_hot cn9k_sso_hws_dual_deq_ca_seg_burst_##name(          \
 508                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 509                 uint64_t timeout_ticks);                                       \
 510         uint16_t __rte_hot cn9k_sso_hws_dual_deq_tmo_ca_seg_##name(            \
 511                 void *port, struct rte_event *ev, uint64_t timeout_ticks);     \
 512         uint16_t __rte_hot cn9k_sso_hws_dual_deq_tmo_ca_seg_burst_##name(      \
 513                 void *port, struct rte_event ev[], uint16_t nb_events,         \
 514                 uint64_t timeout_ticks);
 515
 516 NIX_RX_FASTPATH_MODES
 517 #undef R
 518
 519 #define SSO_DUAL_DEQ(fn, flags)                                                \
 520         uint16_t __rte_hot fn(void *port, struct rte_event *ev,                \
 521                               uint64_t timeout_ticks)                          \
 522         {                                                                      \
 523                 struct cn9k_sso_hws_dual *dws = port;                          \
 524                 uint16_t gw;                                                   \
 525                 RTE_SET_USED(timeout_ticks);                                   \
 526                 if (dws->swtag_req) {                                          \
 527                         dws->swtag_req = 0;                                    \
 528                         cnxk_sso_hws_swtag_wait(dws->base[!dws->vws] +         \
 529                                                 SSOW_LF_GWS_TAG);              \
 530                         return 1;                                              \
 531                 }                                                              \
 532                 gw = cn9k_sso_hws_dual_get_work(                               \
 533                         dws->base[dws->vws], dws->base[!dws->vws], ev, flags,  \
 534                         dws->lookup_mem, dws->tstamp);                         \
 535                 dws->vws = !dws->vws;                                          \
 536                 return gw;                                                     \
 537         }
 538
 539 #define SSO_DUAL_DEQ_SEG(fn, flags) SSO_DUAL_DEQ(fn, flags | NIX_RX_MULTI_SEG_F)
 540 #define SSO_DUAL_DEQ_CA(fn, flags)  SSO_DUAL_DEQ(fn, flags | CPT_RX_WQE_F)
 541 #define SSO_DUAL_DEQ_CA_SEG(fn, flags)                                         \
 542         SSO_DUAL_DEQ_SEG(fn, flags | CPT_RX_WQE_F)
 543
 544 #define SSO_DUAL_DEQ_TMO(fn, flags)                                            \
 545         uint16_t __rte_hot fn(void *port, struct rte_event *ev,                \
 546                               uint64_t timeout_ticks)                          \
 547         {                                                                      \
 548                 struct cn9k_sso_hws_dual *dws = port;                          \
 549                 uint16_t ret = 1;                                              \
 550                 uint64_t iter;                                                 \
 551                 if (dws->swtag_req) {                                          \
 552                         dws->swtag_req = 0;                                    \
 553                         cnxk_sso_hws_swtag_wait(dws->base[!dws->vws] +         \
 554                                                 SSOW_LF_GWS_TAG);              \
 555                         return ret;                                            \
 556                 }                                                              \
 557                 ret = cn9k_sso_hws_dual_get_work(                              \
 558                         dws->base[dws->vws], dws->base[!dws->vws], ev, flags,  \
 559                         dws->lookup_mem, dws->tstamp);                         \
 560                 dws->vws = !dws->vws;                                          \
 561                 for (iter = 1; iter < timeout_ticks && (ret == 0); iter++) {   \
 562                         ret = cn9k_sso_hws_dual_get_work(                      \
 563                                 dws->base[dws->vws], dws->base[!dws->vws], ev, \
 564                                 flags, dws->lookup_mem, dws->tstamp);          \
 565                         dws->vws = !dws->vws;                                  \
 566                 }                                                              \
 567                 return ret;                                                    \
 568         }
 569
 570 #define SSO_DUAL_DEQ_TMO_SEG(fn, flags)                                        \
 571         SSO_DUAL_DEQ_TMO(fn, flags | NIX_RX_MULTI_SEG_F)
 572 #define SSO_DUAL_DEQ_TMO_CA(fn, flags)                                         \
 573         SSO_DUAL_DEQ_TMO(fn, flags | CPT_RX_WQE_F)
 574 #define SSO_DUAL_DEQ_TMO_CA_SEG(fn, flags)                                     \
 575         SSO_DUAL_DEQ_TMO_SEG(fn, flags | CPT_RX_WQE_F)
 576
 577 #define SSO_CMN_DEQ_BURST(fnb, fn, flags)                                      \
 578         uint16_t __rte_hot fnb(void *port, struct rte_event ev[],              \
 579                                uint16_t nb_events, uint64_t timeout_ticks)     \
 580         {                                                                      \
 581                 RTE_SET_USED(nb_events);                                       \
 582                 return fn(port, ev, timeout_ticks);                            \
 583         }
 584
 585 #define SSO_CMN_DEQ_SEG_BURST(fnb, fn, flags)                                  \
 586         uint16_t __rte_hot fnb(void *port, struct rte_event ev[],              \
 587                                uint16_t nb_events, uint64_t timeout_ticks)     \
 588         {                                                                      \
 589                 RTE_SET_USED(nb_events);                                       \
 590                 return fn(port, ev, timeout_ticks);                            \
 591         }
 592
 593 static __rte_always_inline void
 594 cn9k_sso_txq_fc_wait(const struct cn9k_eth_txq *txq)
 595 {
 596         while (!((txq->nb_sqb_bufs_adj -
 597                   __atomic_load_n(txq->fc_mem, __ATOMIC_RELAXED))
 598                  << (txq)->sqes_per_sqb_log2))
 599                 ;
 600 }
 601
 602 static __rte_always_inline const struct cn9k_eth_txq *
 603 cn9k_sso_hws_xtract_meta(struct rte_mbuf *m,
 604                          const uint64_t txq_data[][RTE_MAX_QUEUES_PER_PORT])
 605 {
 606         return (const struct cn9k_eth_txq *)
 607                 txq_data[m->port][rte_event_eth_tx_adapter_txq_get(m)];
 608 }
 609
 610 static __rte_always_inline void
 611 cn9k_sso_hws_prepare_pkt(const struct cn9k_eth_txq *txq, struct rte_mbuf *m,
 612                          uint64_t *cmd, const uint32_t flags)
 613 {
 614         roc_lmt_mov(cmd, txq->cmd, cn9k_nix_tx_ext_subs(flags));
 615         cn9k_nix_xmit_prepare(m, cmd, flags, txq->lso_tun_fmt);
 616 }
 617
 618 #if defined(RTE_ARCH_ARM64)
 619
 620 static __rte_always_inline void
 621 cn9k_sso_hws_xmit_sec_one(const struct cn9k_eth_txq *txq, uint64_t base,
 622                           struct rte_mbuf *m, uint64_t *cmd,
 623                           uint32_t flags)
 624 {
 625         struct cn9k_outb_priv_data *outb_priv;
 626         rte_iova_t io_addr = txq->cpt_io_addr;
 627         uint64_t *lmt_addr = txq->lmt_addr;
 628         struct cn9k_sec_sess_priv mdata;
 629         struct nix_send_hdr_s *send_hdr;
 630         uint64_t sa_base = txq->sa_base;
 631         uint32_t pkt_len, dlen_adj, rlen;
 632         uint64x2_t cmd01, cmd23;
 633         uint64_t lmt_status, sa;
 634         union nix_send_sg_s *sg;
 635         uintptr_t dptr, nixtx;
 636         uint64_t ucode_cmd[4];
 637         uint64_t esn, *iv;
 638         uint8_t l2_len;
 639
 640         mdata.u64 = *rte_security_dynfield(m);
 641         send_hdr = (struct nix_send_hdr_s *)cmd;
 642         if (flags & NIX_TX_NEED_EXT_HDR)
 643                 sg = (union nix_send_sg_s *)&cmd[4];
 644         else
 645                 sg = (union nix_send_sg_s *)&cmd[2];
 646
 647         if (flags & NIX_TX_NEED_SEND_HDR_W1)
 648                 l2_len = cmd[1] & 0xFF;
 649         else
 650                 l2_len = m->l2_len;
 651
 652         /* Retrieve DPTR */
 653         dptr = *(uint64_t *)(sg + 1);
 654         pkt_len = send_hdr->w0.total;
 655
 656         /* Calculate rlen */
 657         rlen = pkt_len - l2_len;
 658         rlen = (rlen + mdata.roundup_len) + (mdata.roundup_byte - 1);
 659         rlen &= ~(uint64_t)(mdata.roundup_byte - 1);
 660         rlen += mdata.partial_len;
 661         dlen_adj = rlen - pkt_len + l2_len;
 662
 663         /* Update send descriptors. Security is single segment only */
 664         send_hdr->w0.total = pkt_len + dlen_adj;
 665         sg->seg1_size = pkt_len + dlen_adj;
 666
 667         /* Get area where NIX descriptor needs to be stored */
 668         nixtx = dptr + pkt_len + dlen_adj;
 669         nixtx += BIT_ULL(7);
 670         nixtx = (nixtx - 1) & ~(BIT_ULL(7) - 1);
 671
 672         roc_lmt_mov((void *)(nixtx + 16), cmd, cn9k_nix_tx_ext_subs(flags));
 673
 674         /* Load opcode and cptr already prepared at pkt metadata set */
 675         pkt_len -= l2_len;
 676         pkt_len += sizeof(struct roc_onf_ipsec_outb_hdr) +
 677                     ROC_ONF_IPSEC_OUTB_MAX_L2_INFO_SZ;
 678         sa_base &= ~(ROC_NIX_INL_SA_BASE_ALIGN - 1);
 679
 680         sa = (uintptr_t)roc_nix_inl_onf_ipsec_outb_sa(sa_base, mdata.sa_idx);
 681         ucode_cmd[3] = (ROC_CPT_DFLT_ENG_GRP_SE_IE << 61 | sa);
 682         ucode_cmd[0] = (ROC_IE_ONF_MAJOR_OP_PROCESS_OUTBOUND_IPSEC << 48 |
 683                         0x40UL << 48 | pkt_len);
 684
 685         /* CPT Word 0 and Word 1 */
 686         cmd01 = vdupq_n_u64((nixtx + 16) | (cn9k_nix_tx_ext_subs(flags) + 1));
 687         /* CPT_RES_S is 16B above NIXTX */
 688         cmd01 = vsetq_lane_u8(nixtx & BIT_ULL(7), cmd01, 8);
 689
 690         /* CPT word 2 and 3 */
 691         cmd23 = vdupq_n_u64(0);
 692         cmd23 = vsetq_lane_u64((((uint64_t)RTE_EVENT_TYPE_CPU << 28) |
 693                                 CNXK_ETHDEV_SEC_OUTB_EV_SUB << 20), cmd23, 0);
 694         cmd23 = vsetq_lane_u64((uintptr_t)m | 1, cmd23, 1);
 695
 696         dptr += l2_len - ROC_ONF_IPSEC_OUTB_MAX_L2_INFO_SZ -
 697                 sizeof(struct roc_onf_ipsec_outb_hdr);
 698         ucode_cmd[1] = dptr;
 699         ucode_cmd[2] = dptr;
 700
 701         /* Update IV to zero and l2 sz */
 702         *(uint16_t *)(dptr + sizeof(struct roc_onf_ipsec_outb_hdr)) =
 703                 rte_cpu_to_be_16(ROC_ONF_IPSEC_OUTB_MAX_L2_INFO_SZ);
 704         iv = (uint64_t *)(dptr + 8);
 705         iv[0] = 0;
 706         iv[1] = 0;
 707
 708         /* Head wait if needed */
 709         if (base)
 710                 roc_sso_hws_head_wait(base + SSOW_LF_GWS_TAG);
 711
 712         /* ESN */
 713         outb_priv = roc_nix_inl_onf_ipsec_outb_sa_sw_rsvd((void *)sa);
 714         esn = outb_priv->esn;
 715         outb_priv->esn = esn + 1;
 716
 717         ucode_cmd[0] |= (esn >> 32) << 16;
 718         esn = rte_cpu_to_be_32(esn & (BIT_ULL(32) - 1));
 719
 720         /* Update ESN and IPID and IV */
 721         *(uint64_t *)dptr = esn << 32 | esn;
 722
 723         rte_io_wmb();
 724         cn9k_sso_txq_fc_wait(txq);
 725
 726         /* Write CPT instruction to lmt line */
 727         vst1q_u64(lmt_addr, cmd01);
 728         vst1q_u64(lmt_addr + 2, cmd23);
 729
 730         roc_lmt_mov_seg(lmt_addr + 4, ucode_cmd, 2);
 731
 732         if (roc_lmt_submit_ldeor(io_addr) == 0) {
 733                 do {
 734                         vst1q_u64(lmt_addr, cmd01);
 735                         vst1q_u64(lmt_addr + 2, cmd23);
 736                         roc_lmt_mov_seg(lmt_addr + 4, ucode_cmd, 2);
 737
 738                         lmt_status = roc_lmt_submit_ldeor(io_addr);
 739                 } while (lmt_status == 0);
 740         }
 741 }
 742 #else
 743
 744 static inline void
 745 cn9k_sso_hws_xmit_sec_one(const struct cn9k_eth_txq *txq, uint64_t base,
 746                           struct rte_mbuf *m, uint64_t *cmd,
 747                           uint32_t flags)
 748 {
 749         RTE_SET_USED(txq);
 750         RTE_SET_USED(base);
 751         RTE_SET_USED(m);
 752         RTE_SET_USED(cmd);
 753         RTE_SET_USED(flags);
 754 }
 755 #endif
 756
 757 static __rte_always_inline uint16_t
 758 cn9k_sso_hws_event_tx(uint64_t base, struct rte_event *ev, uint64_t *cmd,
 759                       const uint64_t txq_data[][RTE_MAX_QUEUES_PER_PORT],
 760                       const uint32_t flags)
 761 {
 762         struct rte_mbuf *m = ev->mbuf;
 763         const struct cn9k_eth_txq *txq;
 764         uint16_t ref_cnt = m->refcnt;
 765
 766         /* Perform header writes before barrier for TSO */
 767         cn9k_nix_xmit_prepare_tso(m, flags);
 768         /* Lets commit any changes in the packet here in case when
 769          * fast free is set as no further changes will be made to mbuf.
 770          * In case of fast free is not set, both cn9k_nix_prepare_mseg()
 771          * and cn9k_nix_xmit_prepare() has a barrier after refcnt update.
 772          */
 773         if (!(flags & NIX_TX_OFFLOAD_MBUF_NOFF_F) &&
 774             !(flags & NIX_TX_OFFLOAD_SECURITY_F))
 775                 rte_io_wmb();
 776         txq = cn9k_sso_hws_xtract_meta(m, txq_data);
 777         cn9k_sso_hws_prepare_pkt(txq, m, cmd, flags);
 778
 779         if (flags & NIX_TX_OFFLOAD_SECURITY_F) {
 780                 uint64_t ol_flags = m->ol_flags;
 781
 782                 if (ol_flags & RTE_MBUF_F_TX_SEC_OFFLOAD) {
 783                         uintptr_t ssow_base = base;
 784
 785                         if (ev->sched_type)
 786                                 ssow_base = 0;
 787
 788                         cn9k_sso_hws_xmit_sec_one(txq, ssow_base, m, cmd,
 789                                                   flags);
 790                         goto done;
 791                 }
 792
 793                 if (!(flags & NIX_TX_OFFLOAD_MBUF_NOFF_F))
 794                         rte_io_wmb();
 795         }
 796
 797         if (flags & NIX_TX_MULTI_SEG_F) {
 798                 const uint16_t segdw = cn9k_nix_prepare_mseg(m, cmd, flags);
 799                 if (!CNXK_TT_FROM_EVENT(ev->event)) {
 800                         cn9k_nix_xmit_mseg_prep_lmt(cmd, txq->lmt_addr, segdw);
 801                         roc_sso_hws_head_wait(base + SSOW_LF_GWS_TAG);
 802                         cn9k_sso_txq_fc_wait(txq);
 803                         if (cn9k_nix_xmit_submit_lmt(txq->io_addr) == 0)
 804                                 cn9k_nix_xmit_mseg_one(cmd, txq->lmt_addr,
 805                                                        txq->io_addr, segdw);
 806                 } else {
 807                         cn9k_nix_xmit_mseg_one(cmd, txq->lmt_addr, txq->io_addr,
 808                                                segdw);
 809                 }
 810         } else {
 811                 if (!CNXK_TT_FROM_EVENT(ev->event)) {
 812                         cn9k_nix_xmit_prep_lmt(cmd, txq->lmt_addr, flags);
 813                         roc_sso_hws_head_wait(base + SSOW_LF_GWS_TAG);
 814                         cn9k_sso_txq_fc_wait(txq);
 815                         if (cn9k_nix_xmit_submit_lmt(txq->io_addr) == 0)
 816                                 cn9k_nix_xmit_one(cmd, txq->lmt_addr,
 817                                                   txq->io_addr, flags);
 818                 } else {
 819                         cn9k_nix_xmit_one(cmd, txq->lmt_addr, txq->io_addr,
 820                                           flags);
 821                 }
 822         }
 823
 824 done:
 825         if (flags & NIX_TX_OFFLOAD_MBUF_NOFF_F) {
 826                 if (ref_cnt > 1)
 827                         return 1;
 828         }
 829
 830         cnxk_sso_hws_swtag_flush(base + SSOW_LF_GWS_TAG,
 831                                  base + SSOW_LF_GWS_OP_SWTAG_FLUSH);
 832
 833         return 1;
 834 }
 835
 836 #define T(name, f6, f5, f4, f3, f2, f1, f0, sz, flags)                         \
 837         uint16_t __rte_hot cn9k_sso_hws_tx_adptr_enq_##name(                   \
 838                 void *port, struct rte_event ev[], uint16_t nb_events);        \
 839         uint16_t __rte_hot cn9k_sso_hws_tx_adptr_enq_seg_##name(               \
 840                 void *port, struct rte_event ev[], uint16_t nb_events);        \
 841         uint16_t __rte_hot cn9k_sso_hws_dual_tx_adptr_enq_##name(              \
 842                 void *port, struct rte_event ev[], uint16_t nb_events);        \
 843         uint16_t __rte_hot cn9k_sso_hws_dual_tx_adptr_enq_seg_##name(          \
 844                 void *port, struct rte_event ev[], uint16_t nb_events);
 845
 846 NIX_TX_FASTPATH_MODES
 847 #undef T
 848
 849 #endif