drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c

   1 /*
   2  * Copyright 2014 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  */
  23
  24 #include <linux/slab.h>
  25 #include <linux/mutex.h>
  26 #include "kfd_device_queue_manager.h"
  27 #include "kfd_kernel_queue.h"
  28 #include "kfd_priv.h"
  29 #include "kfd_pm4_headers_vi.h"
  30 #include "kfd_pm4_opcodes.h"
  31
  32 static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes,
  33                                 unsigned int buffer_size_bytes)
  34 {
  35         unsigned int temp = *wptr + increment_bytes / sizeof(uint32_t);
  36
  37         WARN((temp * sizeof(uint32_t)) > buffer_size_bytes,
  38              "Runlist IB overflow");
  39         *wptr = temp;
  40 }
  41
  42 static unsigned int build_pm4_header(unsigned int opcode, size_t packet_size)
  43 {
  44         union PM4_MES_TYPE_3_HEADER header;
  45
  46         header.u32All = 0;
  47         header.opcode = opcode;
  48         header.count = packet_size/sizeof(uint32_t) - 2;
  49         header.type = PM4_TYPE_3;
  50
  51         return header.u32All;
  52 }
  53
  54 static void pm_calc_rlib_size(struct packet_manager *pm,
  55                                 unsigned int *rlib_size,
  56                                 bool *over_subscription)
  57 {
  58         unsigned int process_count, queue_count;
  59         unsigned int map_queue_size;
  60
  61         process_count = pm->dqm->processes_count;
  62         queue_count = pm->dqm->queue_count;
  63
  64         /* check if there is over subscription*/
  65         *over_subscription = false;
  66         if ((process_count > 1) || queue_count > get_queues_num(pm->dqm)) {
  67                 *over_subscription = true;
  68                 pr_debug("Over subscribed runlist\n");
  69         }
  70
  71         map_queue_size = sizeof(struct pm4_mes_map_queues);
  72         /* calculate run list ib allocation size */
  73         *rlib_size = process_count * sizeof(struct pm4_mes_map_process) +
  74                      queue_count * map_queue_size;
  75
  76         /*
  77          * Increase the allocation size in case we need a chained run list
  78          * when over subscription
  79          */
  80         if (*over_subscription)
  81                 *rlib_size += sizeof(struct pm4_mes_runlist);
  82
  83         pr_debug("runlist ib size %d\n", *rlib_size);
  84 }
  85
  86 static int pm_allocate_runlist_ib(struct packet_manager *pm,
  87                                 unsigned int **rl_buffer,
  88                                 uint64_t *rl_gpu_buffer,
  89                                 unsigned int *rl_buffer_size,
  90                                 bool *is_over_subscription)
  91 {
  92         int retval;
  93
  94         if (WARN_ON(pm->allocated))
  95                 return -EINVAL;
  96
  97         pm_calc_rlib_size(pm, rl_buffer_size, is_over_subscription);
  98
  99         retval = kfd_gtt_sa_allocate(pm->dqm->dev, *rl_buffer_size,
 100                                         &pm->ib_buffer_obj);
 101
 102         if (retval) {
 103                 pr_err("Failed to allocate runlist IB\n");
 104                 return retval;
 105         }
 106
 107         *(void **)rl_buffer = pm->ib_buffer_obj->cpu_ptr;
 108         *rl_gpu_buffer = pm->ib_buffer_obj->gpu_addr;
 109
 110         memset(*rl_buffer, 0, *rl_buffer_size);
 111         pm->allocated = true;
 112         return retval;
 113 }
 114
 115 static int pm_create_runlist(struct packet_manager *pm, uint32_t *buffer,
 116                         uint64_t ib, size_t ib_size_in_dwords, bool chain)
 117 {
 118         struct pm4_mes_runlist *packet;
 119
 120         if (WARN_ON(!ib))
 121                 return -EFAULT;
 122
 123         packet = (struct pm4_mes_runlist *)buffer;
 124
 125         memset(buffer, 0, sizeof(struct pm4_mes_runlist));
 126         packet->header.u32All = build_pm4_header(IT_RUN_LIST,
 127                                                 sizeof(struct pm4_mes_runlist));
 128
 129         packet->bitfields4.ib_size = ib_size_in_dwords;
 130         packet->bitfields4.chain = chain ? 1 : 0;
 131         packet->bitfields4.offload_polling = 0;
 132         packet->bitfields4.valid = 1;
 133         packet->ordinal2 = lower_32_bits(ib);
 134         packet->bitfields3.ib_base_hi = upper_32_bits(ib);
 135
 136         return 0;
 137 }
 138
 139 static int pm_create_map_process(struct packet_manager *pm, uint32_t *buffer,
 140                                 struct qcm_process_device *qpd)
 141 {
 142         struct pm4_mes_map_process *packet;
 143         struct queue *cur;
 144         uint32_t num_queues;
 145
 146         packet = (struct pm4_mes_map_process *)buffer;
 147
 148         memset(buffer, 0, sizeof(struct pm4_mes_map_process));
 149
 150         packet->header.u32All = build_pm4_header(IT_MAP_PROCESS,
 151                                         sizeof(struct pm4_mes_map_process));
 152         packet->bitfields2.diq_enable = (qpd->is_debug) ? 1 : 0;
 153         packet->bitfields2.process_quantum = 1;
 154         packet->bitfields2.pasid = qpd->pqm->process->pasid;
 155         packet->bitfields3.page_table_base = qpd->page_table_base;
 156         packet->bitfields10.gds_size = qpd->gds_size;
 157         packet->bitfields10.num_gws = qpd->num_gws;
 158         packet->bitfields10.num_oac = qpd->num_oac;
 159         num_queues = 0;
 160         list_for_each_entry(cur, &qpd->queues_list, list)
 161                 num_queues++;
 162         packet->bitfields10.num_queues = (qpd->is_debug) ? 0 : num_queues;
 163
 164         packet->sh_mem_config = qpd->sh_mem_config;
 165         packet->sh_mem_bases = qpd->sh_mem_bases;
 166         packet->sh_mem_ape1_base = qpd->sh_mem_ape1_base;
 167         packet->sh_mem_ape1_limit = qpd->sh_mem_ape1_limit;
 168
 169         packet->sh_hidden_private_base_vmid = qpd->sh_hidden_private_base;
 170
 171         packet->gds_addr_lo = lower_32_bits(qpd->gds_context_area);
 172         packet->gds_addr_hi = upper_32_bits(qpd->gds_context_area);
 173
 174         return 0;
 175 }
 176
 177 static int pm_create_map_queue(struct packet_manager *pm, uint32_t *buffer,
 178                 struct queue *q, bool is_static)
 179 {
 180         struct pm4_mes_map_queues *packet;
 181         bool use_static = is_static;
 182
 183         packet = (struct pm4_mes_map_queues *)buffer;
 184         memset(buffer, 0, sizeof(struct pm4_mes_map_queues));
 185
 186         packet->header.u32All = build_pm4_header(IT_MAP_QUEUES,
 187                                                 sizeof(struct pm4_mes_map_queues));
 188         packet->bitfields2.alloc_format =
 189                 alloc_format__mes_map_queues__one_per_pipe_vi;
 190         packet->bitfields2.num_queues = 1;
 191         packet->bitfields2.queue_sel =
 192                 queue_sel__mes_map_queues__map_to_hws_determined_queue_slots_vi;
 193
 194         packet->bitfields2.engine_sel =
 195                 engine_sel__mes_map_queues__compute_vi;
 196         packet->bitfields2.queue_type =
 197                 queue_type__mes_map_queues__normal_compute_vi;
 198
 199         switch (q->properties.type) {
 200         case KFD_QUEUE_TYPE_COMPUTE:
 201                 if (use_static)
 202                         packet->bitfields2.queue_type =
 203                 queue_type__mes_map_queues__normal_latency_static_queue_vi;
 204                 break;
 205         case KFD_QUEUE_TYPE_DIQ:
 206                 packet->bitfields2.queue_type =
 207                         queue_type__mes_map_queues__debug_interface_queue_vi;
 208                 break;
 209         case KFD_QUEUE_TYPE_SDMA:
 210                 packet->bitfields2.engine_sel =
 211                                 engine_sel__mes_map_queues__sdma0_vi;
 212                 use_static = false; /* no static queues under SDMA */
 213                 break;
 214         default:
 215                 WARN(1, "queue type %d", q->properties.type);
 216                 return -EINVAL;
 217         }
 218         packet->bitfields3.doorbell_offset =
 219                         q->properties.doorbell_off;
 220
 221         packet->mqd_addr_lo =
 222                         lower_32_bits(q->gart_mqd_addr);
 223
 224         packet->mqd_addr_hi =
 225                         upper_32_bits(q->gart_mqd_addr);
 226
 227         packet->wptr_addr_lo =
 228                         lower_32_bits((uint64_t)q->properties.write_ptr);
 229
 230         packet->wptr_addr_hi =
 231                         upper_32_bits((uint64_t)q->properties.write_ptr);
 232
 233         return 0;
 234 }
 235
 236 static int pm_create_runlist_ib(struct packet_manager *pm,
 237                                 struct list_head *queues,
 238                                 uint64_t *rl_gpu_addr,
 239                                 size_t *rl_size_bytes)
 240 {
 241         unsigned int alloc_size_bytes;
 242         unsigned int *rl_buffer, rl_wptr, i;
 243         int retval, proccesses_mapped;
 244         struct device_process_node *cur;
 245         struct qcm_process_device *qpd;
 246         struct queue *q;
 247         struct kernel_queue *kq;
 248         bool is_over_subscription;
 249
 250         rl_wptr = retval = proccesses_mapped = 0;
 251
 252         retval = pm_allocate_runlist_ib(pm, &rl_buffer, rl_gpu_addr,
 253                                 &alloc_size_bytes, &is_over_subscription);
 254         if (retval)
 255                 return retval;
 256
 257         *rl_size_bytes = alloc_size_bytes;
 258
 259         pr_debug("Building runlist ib process count: %d queues count %d\n",
 260                 pm->dqm->processes_count, pm->dqm->queue_count);
 261
 262         /* build the run list ib packet */
 263         list_for_each_entry(cur, queues, list) {
 264                 qpd = cur->qpd;
 265                 /* build map process packet */
 266                 if (proccesses_mapped >= pm->dqm->processes_count) {
 267                         pr_debug("Not enough space left in runlist IB\n");
 268                         pm_release_ib(pm);
 269                         return -ENOMEM;
 270                 }
 271
 272                 retval = pm_create_map_process(pm, &rl_buffer[rl_wptr], qpd);
 273                 if (retval)
 274                         return retval;
 275
 276                 proccesses_mapped++;
 277                 inc_wptr(&rl_wptr, sizeof(struct pm4_mes_map_process),
 278                                 alloc_size_bytes);
 279
 280                 list_for_each_entry(kq, &qpd->priv_queue_list, list) {
 281                         if (!kq->queue->properties.is_active)
 282                                 continue;
 283
 284                         pr_debug("static_queue, mapping kernel q %d, is debug status %d\n",
 285                                 kq->queue->queue, qpd->is_debug);
 286
 287                         retval = pm_create_map_queue(pm,
 288                                                 &rl_buffer[rl_wptr],
 289                                                 kq->queue,
 290                                                 qpd->is_debug);
 291                         if (retval)
 292                                 return retval;
 293
 294                         inc_wptr(&rl_wptr,
 295                                 sizeof(struct pm4_mes_map_queues),
 296                                 alloc_size_bytes);
 297                 }
 298
 299                 list_for_each_entry(q, &qpd->queues_list, list) {
 300                         if (!q->properties.is_active)
 301                                 continue;
 302
 303                         pr_debug("static_queue, mapping user queue %d, is debug status %d\n",
 304                                 q->queue, qpd->is_debug);
 305
 306                         retval = pm_create_map_queue(pm,
 307                                                 &rl_buffer[rl_wptr],
 308                                                 q,
 309                                                 qpd->is_debug);
 310
 311                         if (retval)
 312                                 return retval;
 313
 314                         inc_wptr(&rl_wptr,
 315                                 sizeof(struct pm4_mes_map_queues),
 316                                 alloc_size_bytes);
 317                 }
 318         }
 319
 320         pr_debug("Finished map process and queues to runlist\n");
 321
 322         if (is_over_subscription)
 323                 retval = pm_create_runlist(pm, &rl_buffer[rl_wptr],
 324                                         *rl_gpu_addr,
 325                                         alloc_size_bytes / sizeof(uint32_t),
 326                                         true);
 327
 328         for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++)
 329                 pr_debug("0x%2X ", rl_buffer[i]);
 330         pr_debug("\n");
 331
 332         return retval;
 333 }
 334
 335 int pm_init(struct packet_manager *pm, struct device_queue_manager *dqm)
 336 {
 337         pm->dqm = dqm;
 338         mutex_init(&pm->lock);
 339         pm->priv_queue = kernel_queue_init(dqm->dev, KFD_QUEUE_TYPE_HIQ);
 340         if (!pm->priv_queue) {
 341                 mutex_destroy(&pm->lock);
 342                 return -ENOMEM;
 343         }
 344         pm->allocated = false;
 345
 346         return 0;
 347 }
 348
 349 void pm_uninit(struct packet_manager *pm)
 350 {
 351         mutex_destroy(&pm->lock);
 352         kernel_queue_uninit(pm->priv_queue);
 353 }
 354
 355 int pm_send_set_resources(struct packet_manager *pm,
 356                                 struct scheduling_resources *res)
 357 {
 358         struct pm4_mes_set_resources *packet;
 359         int retval = 0;
 360
 361         mutex_lock(&pm->lock);
 362         pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
 363                                         sizeof(*packet) / sizeof(uint32_t),
 364                                         (unsigned int **)&packet);
 365         if (!packet) {
 366                 pr_err("Failed to allocate buffer on kernel queue\n");
 367                 retval = -ENOMEM;
 368                 goto out;
 369         }
 370
 371         memset(packet, 0, sizeof(struct pm4_mes_set_resources));
 372         packet->header.u32All = build_pm4_header(IT_SET_RESOURCES,
 373                                         sizeof(struct pm4_mes_set_resources));
 374
 375         packet->bitfields2.queue_type =
 376                         queue_type__mes_set_resources__hsa_interface_queue_hiq;
 377         packet->bitfields2.vmid_mask = res->vmid_mask;
 378         packet->bitfields2.unmap_latency = KFD_UNMAP_LATENCY;
 379         packet->bitfields7.oac_mask = res->oac_mask;
 380         packet->bitfields8.gds_heap_base = res->gds_heap_base;
 381         packet->bitfields8.gds_heap_size = res->gds_heap_size;
 382
 383         packet->gws_mask_lo = lower_32_bits(res->gws_mask);
 384         packet->gws_mask_hi = upper_32_bits(res->gws_mask);
 385
 386         packet->queue_mask_lo = lower_32_bits(res->queue_mask);
 387         packet->queue_mask_hi = upper_32_bits(res->queue_mask);
 388
 389         pm->priv_queue->ops.submit_packet(pm->priv_queue);
 390
 391 out:
 392         mutex_unlock(&pm->lock);
 393
 394         return retval;
 395 }
 396
 397 int pm_send_runlist(struct packet_manager *pm, struct list_head *dqm_queues)
 398 {
 399         uint64_t rl_gpu_ib_addr;
 400         uint32_t *rl_buffer;
 401         size_t rl_ib_size, packet_size_dwords;
 402         int retval;
 403
 404         retval = pm_create_runlist_ib(pm, dqm_queues, &rl_gpu_ib_addr,
 405                                         &rl_ib_size);
 406         if (retval)
 407                 goto fail_create_runlist_ib;
 408
 409         pr_debug("runlist IB address: 0x%llX\n", rl_gpu_ib_addr);
 410
 411         packet_size_dwords = sizeof(struct pm4_mes_runlist) / sizeof(uint32_t);
 412         mutex_lock(&pm->lock);
 413
 414         retval = pm->priv_queue->ops.acquire_packet_buffer(pm->priv_queue,
 415                                         packet_size_dwords, &rl_buffer);
 416         if (retval)
 417                 goto fail_acquire_packet_buffer;
 418
 419         retval = pm_create_runlist(pm, rl_buffer, rl_gpu_ib_addr,
 420                                         rl_ib_size / sizeof(uint32_t), false);
 421         if (retval)
 422                 goto fail_create_runlist;
 423
 424         pm->priv_queue->ops.submit_packet(pm->priv_queue);
 425
 426         mutex_unlock(&pm->lock);
 427
 428         return retval;
 429
 430 fail_create_runlist:
 431         pm->priv_queue->ops.rollback_packet(pm->priv_queue);
 432 fail_acquire_packet_buffer:
 433         mutex_unlock(&pm->lock);
 434 fail_create_runlist_ib:
 435         pm_release_ib(pm);
 436         return retval;
 437 }
 438
 439 int pm_send_query_status(struct packet_manager *pm, uint64_t fence_address,
 440                         uint32_t fence_value)
 441 {
 442         int retval;
 443         struct pm4_mes_query_status *packet;
 444
 445         if (WARN_ON(!fence_address))
 446                 return -EFAULT;
 447
 448         mutex_lock(&pm->lock);
 449         retval = pm->priv_queue->ops.acquire_packet_buffer(
 450                         pm->priv_queue,
 451                         sizeof(struct pm4_mes_query_status) / sizeof(uint32_t),
 452                         (unsigned int **)&packet);
 453         if (retval)
 454                 goto fail_acquire_packet_buffer;
 455
 456         packet->header.u32All = build_pm4_header(IT_QUERY_STATUS,
 457                                         sizeof(struct pm4_mes_query_status));
 458
 459         packet->bitfields2.context_id = 0;
 460         packet->bitfields2.interrupt_sel =
 461                         interrupt_sel__mes_query_status__completion_status;
 462         packet->bitfields2.command =
 463                         command__mes_query_status__fence_only_after_write_ack;
 464
 465         packet->addr_hi = upper_32_bits((uint64_t)fence_address);
 466         packet->addr_lo = lower_32_bits((uint64_t)fence_address);
 467         packet->data_hi = upper_32_bits((uint64_t)fence_value);
 468         packet->data_lo = lower_32_bits((uint64_t)fence_value);
 469
 470         pm->priv_queue->ops.submit_packet(pm->priv_queue);
 471
 472 fail_acquire_packet_buffer:
 473         mutex_unlock(&pm->lock);
 474         return retval;
 475 }
 476
 477 int pm_send_unmap_queue(struct packet_manager *pm, enum kfd_queue_type type,
 478                         enum kfd_preempt_type_filter mode,
 479                         uint32_t filter_param, bool reset,
 480                         unsigned int sdma_engine)
 481 {
 482         int retval;
 483         uint32_t *buffer;
 484         struct pm4_mes_unmap_queues *packet;
 485
 486         mutex_lock(&pm->lock);
 487         retval = pm->priv_queue->ops.acquire_packet_buffer(
 488                         pm->priv_queue,
 489                         sizeof(struct pm4_mes_unmap_queues) / sizeof(uint32_t),
 490                         &buffer);
 491         if (retval)
 492                 goto err_acquire_packet_buffer;
 493
 494         packet = (struct pm4_mes_unmap_queues *)buffer;
 495         memset(buffer, 0, sizeof(struct pm4_mes_unmap_queues));
 496         pr_debug("static_queue: unmapping queues: mode is %d , reset is %d , type is %d\n",
 497                 mode, reset, type);
 498         packet->header.u32All = build_pm4_header(IT_UNMAP_QUEUES,
 499                                         sizeof(struct pm4_mes_unmap_queues));
 500         switch (type) {
 501         case KFD_QUEUE_TYPE_COMPUTE:
 502         case KFD_QUEUE_TYPE_DIQ:
 503                 packet->bitfields2.engine_sel =
 504                         engine_sel__mes_unmap_queues__compute;
 505                 break;
 506         case KFD_QUEUE_TYPE_SDMA:
 507                 packet->bitfields2.engine_sel =
 508                         engine_sel__mes_unmap_queues__sdma0 + sdma_engine;
 509                 break;
 510         default:
 511                 WARN(1, "queue type %d", type);
 512                 retval = -EINVAL;
 513                 goto err_invalid;
 514         }
 515
 516         if (reset)
 517                 packet->bitfields2.action =
 518                                 action__mes_unmap_queues__reset_queues;
 519         else
 520                 packet->bitfields2.action =
 521                                 action__mes_unmap_queues__preempt_queues;
 522
 523         switch (mode) {
 524         case KFD_PREEMPT_TYPE_FILTER_SINGLE_QUEUE:
 525                 packet->bitfields2.queue_sel =
 526                                 queue_sel__mes_unmap_queues__perform_request_on_specified_queues;
 527                 packet->bitfields2.num_queues = 1;
 528                 packet->bitfields3b.doorbell_offset0 = filter_param;
 529                 break;
 530         case KFD_PREEMPT_TYPE_FILTER_BY_PASID:
 531                 packet->bitfields2.queue_sel =
 532                                 queue_sel__mes_unmap_queues__perform_request_on_pasid_queues;
 533                 packet->bitfields3a.pasid = filter_param;
 534                 break;
 535         case KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES:
 536                 packet->bitfields2.queue_sel =
 537                                 queue_sel__mes_unmap_queues__unmap_all_queues;
 538                 break;
 539         case KFD_PREEMPT_TYPE_FILTER_DYNAMIC_QUEUES:
 540                 /* in this case, we do not preempt static queues */
 541                 packet->bitfields2.queue_sel =
 542                                 queue_sel__mes_unmap_queues__unmap_all_non_static_queues;
 543                 break;
 544         default:
 545                 WARN(1, "filter %d", mode);
 546                 retval = -EINVAL;
 547                 goto err_invalid;
 548         }
 549
 550         pm->priv_queue->ops.submit_packet(pm->priv_queue);
 551
 552         mutex_unlock(&pm->lock);
 553         return 0;
 554
 555 err_invalid:
 556         pm->priv_queue->ops.rollback_packet(pm->priv_queue);
 557 err_acquire_packet_buffer:
 558         mutex_unlock(&pm->lock);
 559         return retval;
 560 }
 561
 562 void pm_release_ib(struct packet_manager *pm)
 563 {
 564         mutex_lock(&pm->lock);
 565         if (pm->allocated) {
 566                 kfd_gtt_sa_free(pm->dqm->dev, pm->ib_buffer_obj);
 567                 pm->allocated = false;
 568         }
 569         mutex_unlock(&pm->lock);
 570 }