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 "cnxk_ethdev.h"
   9 #include "cnxk_eventdev.h"
  10 #include "cnxk_worker.h"
  11
  12 #include "cn9k_ethdev.h"
  13 #include "cn9k_rx.h"
  14
  15 /* SSO Operations */
  16
  17 static __rte_always_inline uint8_t
  18 cn9k_sso_hws_new_event(struct cn9k_sso_hws *ws, const struct rte_event *ev)
  19 {
  20         const uint32_t tag = (uint32_t)ev->event;
  21         const uint8_t new_tt = ev->sched_type;
  22         const uint64_t event_ptr = ev->u64;
  23         const uint16_t grp = ev->queue_id;
  24
  25         rte_atomic_thread_fence(__ATOMIC_ACQ_REL);
  26         if (ws->xaq_lmt <= *ws->fc_mem)
  27                 return 0;
  28
  29         cnxk_sso_hws_add_work(event_ptr, tag, new_tt, ws->grps_base[grp]);
  30         return 1;
  31 }
  32
  33 static __rte_always_inline void
  34 cn9k_sso_hws_fwd_swtag(struct cn9k_sso_hws_state *vws,
  35                        const struct rte_event *ev)
  36 {
  37         const uint32_t tag = (uint32_t)ev->event;
  38         const uint8_t new_tt = ev->sched_type;
  39         const uint8_t cur_tt = CNXK_TT_FROM_TAG(plt_read64(vws->tag_op));
  40
  41         /* CNXK model
  42          * cur_tt/new_tt     SSO_TT_ORDERED SSO_TT_ATOMIC SSO_TT_UNTAGGED
  43          *
  44          * SSO_TT_ORDERED        norm           norm             untag
  45          * SSO_TT_ATOMIC         norm           norm               untag
  46          * SSO_TT_UNTAGGED       norm           norm             NOOP
  47          */
  48
  49         if (new_tt == SSO_TT_UNTAGGED) {
  50                 if (cur_tt != SSO_TT_UNTAGGED)
  51                         cnxk_sso_hws_swtag_untag(
  52                                 CN9K_SSOW_GET_BASE_ADDR(vws->getwrk_op) +
  53                                 SSOW_LF_GWS_OP_SWTAG_UNTAG);
  54         } else {
  55                 cnxk_sso_hws_swtag_norm(tag, new_tt, vws->swtag_norm_op);
  56         }
  57 }
  58
  59 static __rte_always_inline void
  60 cn9k_sso_hws_fwd_group(struct cn9k_sso_hws_state *ws,
  61                        const struct rte_event *ev, const uint16_t grp)
  62 {
  63         const uint32_t tag = (uint32_t)ev->event;
  64         const uint8_t new_tt = ev->sched_type;
  65
  66         plt_write64(ev->u64, CN9K_SSOW_GET_BASE_ADDR(ws->getwrk_op) +
  67                                      SSOW_LF_GWS_OP_UPD_WQP_GRP1);
  68         cnxk_sso_hws_swtag_desched(tag, new_tt, grp, ws->swtag_desched_op);
  69 }
  70
  71 static __rte_always_inline void
  72 cn9k_sso_hws_forward_event(struct cn9k_sso_hws *ws, const struct rte_event *ev)
  73 {
  74         const uint8_t grp = ev->queue_id;
  75
  76         /* Group hasn't changed, Use SWTAG to forward the event */
  77         if (CNXK_GRP_FROM_TAG(plt_read64(ws->tag_op)) == grp) {
  78                 cn9k_sso_hws_fwd_swtag((struct cn9k_sso_hws_state *)ws, ev);
  79                 ws->swtag_req = 1;
  80         } else {
  81                 /*
  82                  * Group has been changed for group based work pipelining,
  83                  * Use deschedule/add_work operation to transfer the event to
  84                  * new group/core
  85                  */
  86                 cn9k_sso_hws_fwd_group((struct cn9k_sso_hws_state *)ws, ev,
  87                                        grp);
  88         }
  89 }
  90
  91 /* Dual ws ops. */
  92
  93 static __rte_always_inline uint8_t
  94 cn9k_sso_hws_dual_new_event(struct cn9k_sso_hws_dual *dws,
  95                             const struct rte_event *ev)
  96 {
  97         const uint32_t tag = (uint32_t)ev->event;
  98         const uint8_t new_tt = ev->sched_type;
  99         const uint64_t event_ptr = ev->u64;
 100         const uint16_t grp = ev->queue_id;
 101
 102         rte_atomic_thread_fence(__ATOMIC_ACQ_REL);
 103         if (dws->xaq_lmt <= *dws->fc_mem)
 104                 return 0;
 105
 106         cnxk_sso_hws_add_work(event_ptr, tag, new_tt, dws->grps_base[grp]);
 107         return 1;
 108 }
 109
 110 static __rte_always_inline void
 111 cn9k_sso_hws_dual_forward_event(struct cn9k_sso_hws_dual *dws,
 112                                 struct cn9k_sso_hws_state *vws,
 113                                 const struct rte_event *ev)
 114 {
 115         const uint8_t grp = ev->queue_id;
 116
 117         /* Group hasn't changed, Use SWTAG to forward the event */
 118         if (CNXK_GRP_FROM_TAG(plt_read64(vws->tag_op)) == grp) {
 119                 cn9k_sso_hws_fwd_swtag(vws, ev);
 120                 dws->swtag_req = 1;
 121         } else {
 122                 /*
 123                  * Group has been changed for group based work pipelining,
 124                  * Use deschedule/add_work operation to transfer the event to
 125                  * new group/core
 126                  */
 127                 cn9k_sso_hws_fwd_group(vws, ev, grp);
 128         }
 129 }
 130
 131 static __rte_always_inline uint16_t
 132 cn9k_sso_hws_dual_get_work(struct cn9k_sso_hws_state *ws,
 133                            struct cn9k_sso_hws_state *ws_pair,
 134                            struct rte_event *ev)
 135 {
 136         const uint64_t set_gw = BIT_ULL(16) | 1;
 137         union {
 138                 __uint128_t get_work;
 139                 uint64_t u64[2];
 140         } gw;
 141
 142 #ifdef RTE_ARCH_ARM64
 143         asm volatile(PLT_CPU_FEATURE_PREAMBLE
 144                      "rty%=:                                    \n"
 145                      "          ldr %[tag], [%[tag_loc]]        \n"
 146                      "          ldr %[wqp], [%[wqp_loc]]        \n"
 147                      "          tbnz %[tag], 63, rty%=          \n"
 148                      "done%=:   str %[gw], [%[pong]]            \n"
 149                      "          dmb ld                          \n"
 150                      : [tag] "=&r"(gw.u64[0]), [wqp] "=&r"(gw.u64[1])
 151                      : [tag_loc] "r"(ws->tag_op), [wqp_loc] "r"(ws->wqp_op),
 152                        [gw] "r"(set_gw), [pong] "r"(ws_pair->getwrk_op));
 153 #else
 154         gw.u64[0] = plt_read64(ws->tag_op);
 155         while ((BIT_ULL(63)) & gw.u64[0])
 156                 gw.u64[0] = plt_read64(ws->tag_op);
 157         gw.u64[1] = plt_read64(ws->wqp_op);
 158         plt_write64(set_gw, ws_pair->getwrk_op);
 159 #endif
 160
 161         gw.u64[0] = (gw.u64[0] & (0x3ull << 32)) << 6 |
 162                     (gw.u64[0] & (0x3FFull << 36)) << 4 |
 163                     (gw.u64[0] & 0xffffffff);
 164
 165         ev->event = gw.u64[0];
 166         ev->u64 = gw.u64[1];
 167
 168         return !!gw.u64[1];
 169 }
 170
 171 static __rte_always_inline uint16_t
 172 cn9k_sso_hws_get_work(struct cn9k_sso_hws *ws, struct rte_event *ev)
 173 {
 174         union {
 175                 __uint128_t get_work;
 176                 uint64_t u64[2];
 177         } gw;
 178
 179         plt_write64(BIT_ULL(16) | /* wait for work. */
 180                             1,    /* Use Mask set 0. */
 181                     ws->getwrk_op);
 182 #ifdef RTE_ARCH_ARM64
 183         asm volatile(PLT_CPU_FEATURE_PREAMBLE
 184                      "          ldr %[tag], [%[tag_loc]]        \n"
 185                      "          ldr %[wqp], [%[wqp_loc]]        \n"
 186                      "          tbz %[tag], 63, done%=          \n"
 187                      "          sevl                            \n"
 188                      "rty%=:    wfe                             \n"
 189                      "          ldr %[tag], [%[tag_loc]]        \n"
 190                      "          ldr %[wqp], [%[wqp_loc]]        \n"
 191                      "          tbnz %[tag], 63, rty%=          \n"
 192                      "done%=:   dmb ld                          \n"
 193                      : [tag] "=&r"(gw.u64[0]), [wqp] "=&r"(gw.u64[1])
 194                      : [tag_loc] "r"(ws->tag_op), [wqp_loc] "r"(ws->wqp_op));
 195 #else
 196         gw.u64[0] = plt_read64(ws->tag_op);
 197         while ((BIT_ULL(63)) & gw.u64[0])
 198                 gw.u64[0] = plt_read64(ws->tag_op);
 199
 200         gw.u64[1] = plt_read64(ws->wqp_op);
 201 #endif
 202
 203         gw.u64[0] = (gw.u64[0] & (0x3ull << 32)) << 6 |
 204                     (gw.u64[0] & (0x3FFull << 36)) << 4 |
 205                     (gw.u64[0] & 0xffffffff);
 206
 207         ev->event = gw.u64[0];
 208         ev->u64 = gw.u64[1];
 209
 210         return !!gw.u64[1];
 211 }
 212
 213 /* Used in cleaning up workslot. */
 214 static __rte_always_inline uint16_t
 215 cn9k_sso_hws_get_work_empty(struct cn9k_sso_hws_state *ws, struct rte_event *ev)
 216 {
 217         union {
 218                 __uint128_t get_work;
 219                 uint64_t u64[2];
 220         } gw;
 221
 222 #ifdef RTE_ARCH_ARM64
 223         asm volatile(PLT_CPU_FEATURE_PREAMBLE
 224                      "          ldr %[tag], [%[tag_loc]]        \n"
 225                      "          ldr %[wqp], [%[wqp_loc]]        \n"
 226                      "          tbz %[tag], 63, done%=          \n"
 227                      "          sevl                            \n"
 228                      "rty%=:    wfe                             \n"
 229                      "          ldr %[tag], [%[tag_loc]]        \n"
 230                      "          ldr %[wqp], [%[wqp_loc]]        \n"
 231                      "          tbnz %[tag], 63, rty%=          \n"
 232                      "done%=:   dmb ld                          \n"
 233                      : [tag] "=&r"(gw.u64[0]), [wqp] "=&r"(gw.u64[1])
 234                      : [tag_loc] "r"(ws->tag_op), [wqp_loc] "r"(ws->wqp_op));
 235 #else
 236         gw.u64[0] = plt_read64(ws->tag_op);
 237         while ((BIT_ULL(63)) & gw.u64[0])
 238                 gw.u64[0] = plt_read64(ws->tag_op);
 239
 240         gw.u64[1] = plt_read64(ws->wqp_op);
 241 #endif
 242
 243         gw.u64[0] = (gw.u64[0] & (0x3ull << 32)) << 6 |
 244                     (gw.u64[0] & (0x3FFull << 36)) << 4 |
 245                     (gw.u64[0] & 0xffffffff);
 246
 247         ev->event = gw.u64[0];
 248         ev->u64 = gw.u64[1];
 249
 250         return !!gw.u64[1];
 251 }
 252
 253 /* CN9K Fastpath functions. */
 254 uint16_t __rte_hot cn9k_sso_hws_enq(void *port, const struct rte_event *ev);
 255 uint16_t __rte_hot cn9k_sso_hws_enq_burst(void *port,
 256                                           const struct rte_event ev[],
 257                                           uint16_t nb_events);
 258 uint16_t __rte_hot cn9k_sso_hws_enq_new_burst(void *port,
 259                                               const struct rte_event ev[],
 260                                               uint16_t nb_events);
 261 uint16_t __rte_hot cn9k_sso_hws_enq_fwd_burst(void *port,
 262                                               const struct rte_event ev[],
 263                                               uint16_t nb_events);
 264
 265 uint16_t __rte_hot cn9k_sso_hws_dual_enq(void *port,
 266                                          const struct rte_event *ev);
 267 uint16_t __rte_hot cn9k_sso_hws_dual_enq_burst(void *port,
 268                                                const struct rte_event ev[],
 269                                                uint16_t nb_events);
 270 uint16_t __rte_hot cn9k_sso_hws_dual_enq_new_burst(void *port,
 271                                                    const struct rte_event ev[],
 272                                                    uint16_t nb_events);
 273 uint16_t __rte_hot cn9k_sso_hws_dual_enq_fwd_burst(void *port,
 274                                                    const struct rte_event ev[],
 275                                                    uint16_t nb_events);
 276
 277 uint16_t __rte_hot cn9k_sso_hws_deq(void *port, struct rte_event *ev,
 278                                     uint64_t timeout_ticks);
 279 uint16_t __rte_hot cn9k_sso_hws_deq_burst(void *port, struct rte_event ev[],
 280                                           uint16_t nb_events,
 281                                           uint64_t timeout_ticks);
 282 uint16_t __rte_hot cn9k_sso_hws_tmo_deq(void *port, struct rte_event *ev,
 283                                         uint64_t timeout_ticks);
 284 uint16_t __rte_hot cn9k_sso_hws_tmo_deq_burst(void *port, struct rte_event ev[],
 285                                               uint16_t nb_events,
 286                                               uint64_t timeout_ticks);
 287
 288 uint16_t __rte_hot cn9k_sso_hws_dual_deq(void *port, struct rte_event *ev,
 289                                          uint64_t timeout_ticks);
 290 uint16_t __rte_hot cn9k_sso_hws_dual_deq_burst(void *port,
 291                                                struct rte_event ev[],
 292                                                uint16_t nb_events,
 293                                                uint64_t timeout_ticks);
 294 uint16_t __rte_hot cn9k_sso_hws_dual_tmo_deq(void *port, struct rte_event *ev,
 295                                              uint64_t timeout_ticks);
 296 uint16_t __rte_hot cn9k_sso_hws_dual_tmo_deq_burst(void *port,
 297                                                    struct rte_event ev[],
 298                                                    uint16_t nb_events,
 299                                                    uint64_t timeout_ticks);
 300
 301 #endif