drivers/crypto/scheduler/scheduler_multicore.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2017 Intel Corporation. All rights reserved.
   5  *
   6  *   Redistribution and use in source and binary forms, with or without
   7  *   modification, are permitted provided that the following conditions
   8  *   are met:
   9  *
  10  *     * Redistributions of source code must retain the above copyright
  11  *       notice, this list of conditions and the following disclaimer.
  12  *     * Redistributions in binary form must reproduce the above copyright
  13  *       notice, this list of conditions and the following disclaimer in
  14  *       the documentation and/or other materials provided with the
  15  *       distribution.
  16  *     * Neither the name of Intel Corporation nor the names of its
  17  *       contributors may be used to endorse or promote products derived
  18  *       from this software without specific prior written permission.
  19  *
  20  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  21  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  22  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  23  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  24  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  25  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  26  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  27  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  28  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  29  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  30  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  31  */
  32 #include <unistd.h>
  33
  34 #include <rte_cryptodev.h>
  35 #include <rte_malloc.h>
  36
  37 #include "rte_cryptodev_scheduler_operations.h"
  38 #include "scheduler_pmd_private.h"
  39
  40 #define MC_SCHED_ENQ_RING_NAME_PREFIX   "MCS_ENQR_"
  41 #define MC_SCHED_DEQ_RING_NAME_PREFIX   "MCS_DEQR_"
  42
  43 #define MC_SCHED_BUFFER_SIZE 32
  44
  45 /** multi-core scheduler context */
  46 struct mc_scheduler_ctx {
  47         uint32_t num_workers;             /**< Number of workers polling */
  48         uint32_t stop_signal;
  49
  50         struct rte_ring *sched_enq_ring[RTE_CRYPTODEV_SCHEDULER_MAX_NB_WORKER_CORES];
  51         struct rte_ring *sched_deq_ring[RTE_CRYPTODEV_SCHEDULER_MAX_NB_WORKER_CORES];
  52 };
  53
  54 struct mc_scheduler_qp_ctx {
  55         struct scheduler_slave slaves[RTE_CRYPTODEV_SCHEDULER_MAX_NB_SLAVES];
  56         uint32_t nb_slaves;
  57
  58         uint32_t last_enq_worker_idx;
  59         uint32_t last_deq_worker_idx;
  60
  61         struct mc_scheduler_ctx *mc_private_ctx;
  62 };
  63
  64 static uint16_t
  65 schedule_enqueue(void *qp, struct rte_crypto_op **ops, uint16_t nb_ops)
  66 {
  67         struct mc_scheduler_qp_ctx *mc_qp_ctx =
  68                         ((struct scheduler_qp_ctx *)qp)->private_qp_ctx;
  69         struct mc_scheduler_ctx *mc_ctx = mc_qp_ctx->mc_private_ctx;
  70         uint32_t worker_idx = mc_qp_ctx->last_enq_worker_idx;
  71         uint16_t i, processed_ops = 0;
  72
  73         if (unlikely(nb_ops == 0))
  74                 return 0;
  75
  76         for (i = 0; i <  mc_ctx->num_workers && nb_ops != 0; i++) {
  77                 struct rte_ring *enq_ring = mc_ctx->sched_enq_ring[worker_idx];
  78                 uint16_t nb_queue_ops = rte_ring_enqueue_burst(enq_ring,
  79                         (void *)(&ops[processed_ops]), nb_ops, NULL);
  80
  81                 nb_ops -= nb_queue_ops;
  82                 processed_ops += nb_queue_ops;
  83
  84                 if (++worker_idx == mc_ctx->num_workers)
  85                         worker_idx = 0;
  86         }
  87         mc_qp_ctx->last_enq_worker_idx = worker_idx;
  88
  89         return processed_ops;
  90 }
  91
  92 static uint16_t
  93 schedule_enqueue_ordering(void *qp, struct rte_crypto_op **ops,
  94                 uint16_t nb_ops)
  95 {
  96         struct rte_ring *order_ring =
  97                         ((struct scheduler_qp_ctx *)qp)->order_ring;
  98         uint16_t nb_ops_to_enq = get_max_enqueue_order_count(order_ring,
  99                         nb_ops);
 100         uint16_t nb_ops_enqd = schedule_enqueue(qp, ops,
 101                         nb_ops_to_enq);
 102
 103         scheduler_order_insert(order_ring, ops, nb_ops_enqd);
 104
 105         return nb_ops_enqd;
 106 }
 107
 108
 109 static uint16_t
 110 schedule_dequeue(void *qp, struct rte_crypto_op **ops, uint16_t nb_ops)
 111 {
 112         struct mc_scheduler_qp_ctx *mc_qp_ctx =
 113                         ((struct scheduler_qp_ctx *)qp)->private_qp_ctx;
 114         struct mc_scheduler_ctx *mc_ctx = mc_qp_ctx->mc_private_ctx;
 115         uint32_t worker_idx = mc_qp_ctx->last_deq_worker_idx;
 116         uint16_t i, processed_ops = 0;
 117
 118         for (i = 0; i < mc_ctx->num_workers && nb_ops != 0; i++) {
 119                 struct rte_ring *deq_ring = mc_ctx->sched_deq_ring[worker_idx];
 120                 uint16_t nb_deq_ops = rte_ring_dequeue_burst(deq_ring,
 121                         (void *)(&ops[processed_ops]), nb_ops, NULL);
 122
 123                 nb_ops -= nb_deq_ops;
 124                 processed_ops += nb_deq_ops;
 125                 if (++worker_idx == mc_ctx->num_workers)
 126                         worker_idx = 0;
 127         }
 128
 129         mc_qp_ctx->last_deq_worker_idx = worker_idx;
 130
 131         return processed_ops;
 132
 133 }
 134
 135 static uint16_t
 136 schedule_dequeue_ordering(void *qp, struct rte_crypto_op **ops,
 137                 uint16_t nb_ops)
 138 {
 139         struct rte_ring *order_ring =
 140                         ((struct scheduler_qp_ctx *)qp)->order_ring;
 141
 142         return scheduler_order_drain(order_ring, ops, nb_ops);
 143 }
 144
 145 static int
 146 slave_attach(__rte_unused struct rte_cryptodev *dev,
 147                 __rte_unused uint8_t slave_id)
 148 {
 149         return 0;
 150 }
 151
 152 static int
 153 slave_detach(__rte_unused struct rte_cryptodev *dev,
 154                 __rte_unused uint8_t slave_id)
 155 {
 156         return 0;
 157 }
 158
 159 static int
 160 mc_scheduler_worker(struct rte_cryptodev *dev)
 161 {
 162         struct scheduler_ctx *sched_ctx = dev->data->dev_private;
 163         struct mc_scheduler_ctx *mc_ctx = sched_ctx->private_ctx;
 164         struct rte_ring *enq_ring;
 165         struct rte_ring *deq_ring;
 166         uint32_t core_id = rte_lcore_id();
 167         int i, worker_idx = -1;
 168         struct scheduler_slave *slave;
 169         struct rte_crypto_op *enq_ops[MC_SCHED_BUFFER_SIZE];
 170         struct rte_crypto_op *deq_ops[MC_SCHED_BUFFER_SIZE];
 171         struct scheduler_session *sess0, *sess1, *sess2, *sess3;
 172         uint16_t processed_ops;
 173         uint16_t left_op = 0;
 174         uint16_t left_op_idx = 0;
 175         uint16_t inflight_ops = 0;
 176
 177         for (i = 0; i < (int)sched_ctx->nb_wc; i++) {
 178                 if (sched_ctx->wc_pool[i] == core_id) {
 179                         worker_idx = i;
 180                         break;
 181                 }
 182         }
 183         if (worker_idx == -1) {
 184                 CS_LOG_ERR("worker on core %u:cannot find worker index!\n", core_id);
 185                 return -1;
 186         }
 187
 188         slave = &sched_ctx->slaves[worker_idx];
 189         enq_ring = mc_ctx->sched_enq_ring[worker_idx];
 190         deq_ring = mc_ctx->sched_deq_ring[worker_idx];
 191
 192         while (!mc_ctx->stop_signal) {
 193                 if (left_op) {
 194                         processed_ops =
 195                                 rte_cryptodev_enqueue_burst(slave->dev_id,
 196                                                 slave->qp_id,
 197                                                 &enq_ops[left_op_idx], left_op);
 198
 199                         left_op -= processed_ops;
 200                         left_op_idx += processed_ops;
 201                 } else {
 202                         uint16_t nb_deq_ops = rte_ring_dequeue_burst(enq_ring,
 203                                 (void *)enq_ops, MC_SCHED_BUFFER_SIZE, NULL);
 204                         if (nb_deq_ops) {
 205                                 uint16_t i;
 206
 207                                 for (i = 0; i < nb_deq_ops && i < 4; i++)
 208                                         rte_prefetch0(enq_ops[i]->sym->session);
 209
 210                                 for (i = 0; (i < (nb_deq_ops - 8))
 211                                                         && (nb_deq_ops > 8); i += 4) {
 212                                         sess0 = (struct scheduler_session *)
 213                                                 enq_ops[i]->sym->session->_private;
 214                                         sess1 = (struct scheduler_session *)
 215                                                 enq_ops[i+1]->sym->session->_private;
 216                                         sess2 = (struct scheduler_session *)
 217                                                 enq_ops[i+2]->sym->session->_private;
 218                                         sess3 = (struct scheduler_session *)
 219                                                 enq_ops[i+3]->sym->session->_private;
 220
 221                                         enq_ops[i]->sym->session =
 222                                                         sess0->sessions[worker_idx];
 223                                         enq_ops[i + 1]->sym->session =
 224                                                         sess1->sessions[worker_idx];
 225                                         enq_ops[i + 2]->sym->session =
 226                                                         sess2->sessions[worker_idx];
 227                                         enq_ops[i + 3]->sym->session =
 228                                                         sess3->sessions[worker_idx];
 229
 230                                         rte_prefetch0(enq_ops[i + 4]->sym->session);
 231                                         rte_prefetch0(enq_ops[i + 5]->sym->session);
 232                                         rte_prefetch0(enq_ops[i + 6]->sym->session);
 233                                         rte_prefetch0(enq_ops[i + 7]->sym->session);
 234                                 }
 235
 236                                 for (; i < nb_deq_ops; i++) {
 237                                         sess0 = (struct scheduler_session *)
 238                                                 enq_ops[i]->sym->session->_private;
 239                                         enq_ops[i]->sym->session =
 240                                                         sess0->sessions[worker_idx];
 241                                 }
 242
 243                                 processed_ops = rte_cryptodev_enqueue_burst(slave->dev_id,
 244                                                 slave->qp_id, enq_ops, nb_deq_ops);
 245
 246                                 if (unlikely(processed_ops < nb_deq_ops)) {
 247                                         left_op = nb_deq_ops - processed_ops;
 248                                         left_op_idx = processed_ops;
 249                                 }
 250
 251                                 inflight_ops += processed_ops;
 252                         }
 253                 }
 254
 255                 if (inflight_ops > 0) {
 256                         processed_ops = rte_cryptodev_dequeue_burst(slave->dev_id,
 257                                         slave->qp_id, deq_ops, MC_SCHED_BUFFER_SIZE);
 258                         if (processed_ops) {
 259                                 uint16_t nb_enq_ops = rte_ring_enqueue_burst(deq_ring,
 260                                         (void *)deq_ops, processed_ops, NULL);
 261                                 inflight_ops -= nb_enq_ops;
 262                         }
 263                 }
 264
 265                 rte_pause();
 266         }
 267
 268         return 0;
 269 }
 270
 271 static int
 272 scheduler_start(struct rte_cryptodev *dev)
 273 {
 274         struct scheduler_ctx *sched_ctx = dev->data->dev_private;
 275         struct mc_scheduler_ctx *mc_ctx = sched_ctx->private_ctx;
 276         uint16_t i;
 277
 278         mc_ctx->stop_signal = 0;
 279
 280         for (i = 0; i < sched_ctx->nb_wc; i++)
 281                 rte_eal_remote_launch(
 282                         (lcore_function_t *)mc_scheduler_worker, dev,
 283                                         sched_ctx->wc_pool[i]);
 284
 285         if (sched_ctx->reordering_enabled) {
 286                 dev->enqueue_burst = &schedule_enqueue_ordering;
 287                 dev->dequeue_burst = &schedule_dequeue_ordering;
 288         } else {
 289                 dev->enqueue_burst = &schedule_enqueue;
 290                 dev->dequeue_burst = &schedule_dequeue;
 291         }
 292
 293         for (i = 0; i < dev->data->nb_queue_pairs; i++) {
 294                 struct scheduler_qp_ctx *qp_ctx = dev->data->queue_pairs[i];
 295                 struct mc_scheduler_qp_ctx *mc_qp_ctx =
 296                                 qp_ctx->private_qp_ctx;
 297                 uint32_t j;
 298
 299                 memset(mc_qp_ctx->slaves, 0,
 300                                 RTE_CRYPTODEV_SCHEDULER_MAX_NB_SLAVES *
 301                                 sizeof(struct scheduler_slave));
 302                 for (j = 0; j < sched_ctx->nb_slaves; j++) {
 303                         mc_qp_ctx->slaves[j].dev_id =
 304                                         sched_ctx->slaves[j].dev_id;
 305                         mc_qp_ctx->slaves[j].qp_id = i;
 306                 }
 307
 308                 mc_qp_ctx->nb_slaves = sched_ctx->nb_slaves;
 309
 310                 mc_qp_ctx->last_enq_worker_idx = 0;
 311                 mc_qp_ctx->last_deq_worker_idx = 0;
 312         }
 313
 314         return 0;
 315 }
 316
 317 static int
 318 scheduler_stop(struct rte_cryptodev *dev)
 319 {
 320         struct scheduler_ctx *sched_ctx = dev->data->dev_private;
 321         struct mc_scheduler_ctx *mc_ctx = sched_ctx->private_ctx;
 322
 323         mc_ctx->stop_signal = 1;
 324
 325         for (uint16_t i = 0; i < sched_ctx->nb_wc; i++)
 326                 rte_eal_wait_lcore(sched_ctx->wc_pool[i]);
 327
 328         return 0;
 329 }
 330
 331 static int
 332 scheduler_config_qp(struct rte_cryptodev *dev, uint16_t qp_id)
 333 {
 334         struct scheduler_qp_ctx *qp_ctx = dev->data->queue_pairs[qp_id];
 335         struct mc_scheduler_qp_ctx *mc_qp_ctx;
 336         struct scheduler_ctx *sched_ctx = dev->data->dev_private;
 337         struct mc_scheduler_ctx *mc_ctx = sched_ctx->private_ctx;
 338
 339         mc_qp_ctx = rte_zmalloc_socket(NULL, sizeof(*mc_qp_ctx), 0,
 340                         rte_socket_id());
 341         if (!mc_qp_ctx) {
 342                 CS_LOG_ERR("failed allocate memory for private queue pair");
 343                 return -ENOMEM;
 344         }
 345
 346         mc_qp_ctx->mc_private_ctx = mc_ctx;
 347         qp_ctx->private_qp_ctx = (void *)mc_qp_ctx;
 348
 349
 350         return 0;
 351 }
 352
 353 static int
 354 scheduler_create_private_ctx(struct rte_cryptodev *dev)
 355 {
 356         struct scheduler_ctx *sched_ctx = dev->data->dev_private;
 357         struct mc_scheduler_ctx *mc_ctx;
 358
 359         if (sched_ctx->private_ctx)
 360                 rte_free(sched_ctx->private_ctx);
 361
 362         mc_ctx = rte_zmalloc_socket(NULL, sizeof(struct mc_scheduler_ctx), 0,
 363                         rte_socket_id());
 364         if (!mc_ctx) {
 365                 CS_LOG_ERR("failed allocate memory");
 366                 return -ENOMEM;
 367         }
 368
 369         mc_ctx->num_workers = sched_ctx->nb_wc;
 370         for (uint16_t i = 0; i < sched_ctx->nb_wc; i++) {
 371                 char r_name[16];
 372
 373                 snprintf(r_name, sizeof(r_name), MC_SCHED_ENQ_RING_NAME_PREFIX "%u", i);
 374                 mc_ctx->sched_enq_ring[i] = rte_ring_create(r_name, PER_SLAVE_BUFF_SIZE,
 375                                         rte_socket_id(), RING_F_SC_DEQ | RING_F_SP_ENQ);
 376                 if (!mc_ctx->sched_enq_ring[i]) {
 377                         CS_LOG_ERR("Cannot create ring for worker %u", i);
 378                         return -1;
 379                 }
 380                 snprintf(r_name, sizeof(r_name), MC_SCHED_DEQ_RING_NAME_PREFIX "%u", i);
 381                 mc_ctx->sched_deq_ring[i] = rte_ring_create(r_name, PER_SLAVE_BUFF_SIZE,
 382                                         rte_socket_id(), RING_F_SC_DEQ | RING_F_SP_ENQ);
 383                 if (!mc_ctx->sched_deq_ring[i]) {
 384                         CS_LOG_ERR("Cannot create ring for worker %u", i);
 385                         return -1;
 386                 }
 387         }
 388
 389         sched_ctx->private_ctx = (void *)mc_ctx;
 390
 391         return 0;
 392 }
 393
 394 struct rte_cryptodev_scheduler_ops scheduler_mc_ops = {
 395         slave_attach,
 396         slave_detach,
 397         scheduler_start,
 398         scheduler_stop,
 399         scheduler_config_qp,
 400         scheduler_create_private_ctx,
 401         NULL,   /* option_set */
 402         NULL    /* option_get */
 403 };
 404
 405 struct rte_cryptodev_scheduler mc_scheduler = {
 406                 .name = "multicore-scheduler",
 407                 .description = "scheduler which will run burst across multiple cpu cores",
 408                 .mode = CDEV_SCHED_MODE_MULTICORE,
 409                 .ops = &scheduler_mc_ops
 410 };
 411
 412 struct rte_cryptodev_scheduler *multicore_scheduler = &mc_scheduler;