drivers/gpu/drm/radeon/radeon_kfd.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 #include <linux/module.h>
  24 #include <linux/fdtable.h>
  25 #include <linux/uaccess.h>
  26 #include <drm/drmP.h>
  27 #include "radeon.h"
  28 #include "cikd.h"
  29 #include "cik_reg.h"
  30 #include "radeon_kfd.h"
  31 #include "radeon_ucode.h"
  32 #include <linux/firmware.h>
  33 #include "cik_structs.h"
  34
  35 #define CIK_PIPE_PER_MEC        (4)
  36
  37 static const uint32_t watchRegs[MAX_WATCH_ADDRESSES * ADDRESS_WATCH_REG_MAX] = {
  38         TCP_WATCH0_ADDR_H, TCP_WATCH0_ADDR_L, TCP_WATCH0_CNTL,
  39         TCP_WATCH1_ADDR_H, TCP_WATCH1_ADDR_L, TCP_WATCH1_CNTL,
  40         TCP_WATCH2_ADDR_H, TCP_WATCH2_ADDR_L, TCP_WATCH2_CNTL,
  41         TCP_WATCH3_ADDR_H, TCP_WATCH3_ADDR_L, TCP_WATCH3_CNTL
  42 };
  43
  44 struct kgd_mem {
  45         struct radeon_bo *bo;
  46         uint64_t gpu_addr;
  47         void *cpu_ptr;
  48 };
  49
  50
  51 static int alloc_gtt_mem(struct kgd_dev *kgd, size_t size,
  52                         void **mem_obj, uint64_t *gpu_addr,
  53                         void **cpu_ptr);
  54
  55 static void free_gtt_mem(struct kgd_dev *kgd, void *mem_obj);
  56
  57 static uint64_t get_vmem_size(struct kgd_dev *kgd);
  58 static uint64_t get_gpu_clock_counter(struct kgd_dev *kgd);
  59
  60 static uint32_t get_max_engine_clock_in_mhz(struct kgd_dev *kgd);
  61 static uint16_t get_fw_version(struct kgd_dev *kgd, enum kgd_engine_type type);
  62
  63 /*
  64  * Register access functions
  65  */
  66
  67 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
  68                 uint32_t sh_mem_config, uint32_t sh_mem_ape1_base,
  69                 uint32_t sh_mem_ape1_limit, uint32_t sh_mem_bases);
  70
  71 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
  72                                         unsigned int vmid);
  73
  74 static int kgd_init_pipeline(struct kgd_dev *kgd, uint32_t pipe_id,
  75                                 uint32_t hpd_size, uint64_t hpd_gpu_addr);
  76 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id);
  77 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
  78                         uint32_t queue_id, uint32_t __user *wptr);
  79 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd);
  80 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
  81                                 uint32_t pipe_id, uint32_t queue_id);
  82
  83 static int kgd_hqd_destroy(struct kgd_dev *kgd, uint32_t reset_type,
  84                                 unsigned int timeout, uint32_t pipe_id,
  85                                 uint32_t queue_id);
  86 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd);
  87 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
  88                                 unsigned int timeout);
  89 static int kgd_address_watch_disable(struct kgd_dev *kgd);
  90 static int kgd_address_watch_execute(struct kgd_dev *kgd,
  91                                         unsigned int watch_point_id,
  92                                         uint32_t cntl_val,
  93                                         uint32_t addr_hi,
  94                                         uint32_t addr_lo);
  95 static int kgd_wave_control_execute(struct kgd_dev *kgd,
  96                                         uint32_t gfx_index_val,
  97                                         uint32_t sq_cmd);
  98 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
  99                                         unsigned int watch_point_id,
 100                                         unsigned int reg_offset);
 101
 102 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd, uint8_t vmid);
 103 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
 104                                                         uint8_t vmid);
 105 static void write_vmid_invalidate_request(struct kgd_dev *kgd, uint8_t vmid);
 106
 107 static const struct kfd2kgd_calls kfd2kgd = {
 108         .init_gtt_mem_allocation = alloc_gtt_mem,
 109         .free_gtt_mem = free_gtt_mem,
 110         .get_vmem_size = get_vmem_size,
 111         .get_gpu_clock_counter = get_gpu_clock_counter,
 112         .get_max_engine_clock_in_mhz = get_max_engine_clock_in_mhz,
 113         .program_sh_mem_settings = kgd_program_sh_mem_settings,
 114         .set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
 115         .init_pipeline = kgd_init_pipeline,
 116         .init_interrupts = kgd_init_interrupts,
 117         .hqd_load = kgd_hqd_load,
 118         .hqd_sdma_load = kgd_hqd_sdma_load,
 119         .hqd_is_occupied = kgd_hqd_is_occupied,
 120         .hqd_sdma_is_occupied = kgd_hqd_sdma_is_occupied,
 121         .hqd_destroy = kgd_hqd_destroy,
 122         .hqd_sdma_destroy = kgd_hqd_sdma_destroy,
 123         .address_watch_disable = kgd_address_watch_disable,
 124         .address_watch_execute = kgd_address_watch_execute,
 125         .wave_control_execute = kgd_wave_control_execute,
 126         .address_watch_get_offset = kgd_address_watch_get_offset,
 127         .get_atc_vmid_pasid_mapping_pasid = get_atc_vmid_pasid_mapping_pasid,
 128         .get_atc_vmid_pasid_mapping_valid = get_atc_vmid_pasid_mapping_valid,
 129         .write_vmid_invalidate_request = write_vmid_invalidate_request,
 130         .get_fw_version = get_fw_version
 131 };
 132
 133 static const struct kgd2kfd_calls *kgd2kfd;
 134
 135 int radeon_kfd_init(void)
 136 {
 137         int ret;
 138
 139 #if defined(CONFIG_HSA_AMD_MODULE)
 140         int (*kgd2kfd_init_p)(unsigned, const struct kgd2kfd_calls**);
 141
 142         kgd2kfd_init_p = symbol_request(kgd2kfd_init);
 143
 144         if (kgd2kfd_init_p == NULL)
 145                 return -ENOENT;
 146
 147         ret = kgd2kfd_init_p(KFD_INTERFACE_VERSION, &kgd2kfd);
 148         if (ret) {
 149                 symbol_put(kgd2kfd_init);
 150                 kgd2kfd = NULL;
 151         }
 152
 153 #elif defined(CONFIG_HSA_AMD)
 154         ret = kgd2kfd_init(KFD_INTERFACE_VERSION, &kgd2kfd);
 155         if (ret)
 156                 kgd2kfd = NULL;
 157
 158 #else
 159         ret = -ENOENT;
 160 #endif
 161
 162         return ret;
 163 }
 164
 165 void radeon_kfd_fini(void)
 166 {
 167         if (kgd2kfd) {
 168                 kgd2kfd->exit();
 169                 symbol_put(kgd2kfd_init);
 170         }
 171 }
 172
 173 void radeon_kfd_device_probe(struct radeon_device *rdev)
 174 {
 175         if (kgd2kfd)
 176                 rdev->kfd = kgd2kfd->probe((struct kgd_dev *)rdev,
 177                         rdev->pdev, &kfd2kgd);
 178 }
 179
 180 void radeon_kfd_device_init(struct radeon_device *rdev)
 181 {
 182         if (rdev->kfd) {
 183                 struct kgd2kfd_shared_resources gpu_resources = {
 184                         .compute_vmid_bitmap = 0xFF00,
 185
 186                         .first_compute_pipe = 1,
 187                         .compute_pipe_count = 4 - 1,
 188                 };
 189
 190                 radeon_doorbell_get_kfd_info(rdev,
 191                                 &gpu_resources.doorbell_physical_address,
 192                                 &gpu_resources.doorbell_aperture_size,
 193                                 &gpu_resources.doorbell_start_offset);
 194
 195                 kgd2kfd->device_init(rdev->kfd, &gpu_resources);
 196         }
 197 }
 198
 199 void radeon_kfd_device_fini(struct radeon_device *rdev)
 200 {
 201         if (rdev->kfd) {
 202                 kgd2kfd->device_exit(rdev->kfd);
 203                 rdev->kfd = NULL;
 204         }
 205 }
 206
 207 void radeon_kfd_interrupt(struct radeon_device *rdev, const void *ih_ring_entry)
 208 {
 209         if (rdev->kfd)
 210                 kgd2kfd->interrupt(rdev->kfd, ih_ring_entry);
 211 }
 212
 213 void radeon_kfd_suspend(struct radeon_device *rdev)
 214 {
 215         if (rdev->kfd)
 216                 kgd2kfd->suspend(rdev->kfd);
 217 }
 218
 219 int radeon_kfd_resume(struct radeon_device *rdev)
 220 {
 221         int r = 0;
 222
 223         if (rdev->kfd)
 224                 r = kgd2kfd->resume(rdev->kfd);
 225
 226         return r;
 227 }
 228
 229 static int alloc_gtt_mem(struct kgd_dev *kgd, size_t size,
 230                         void **mem_obj, uint64_t *gpu_addr,
 231                         void **cpu_ptr)
 232 {
 233         struct radeon_device *rdev = (struct radeon_device *)kgd;
 234         struct kgd_mem **mem = (struct kgd_mem **) mem_obj;
 235         int r;
 236
 237         BUG_ON(kgd == NULL);
 238         BUG_ON(gpu_addr == NULL);
 239         BUG_ON(cpu_ptr == NULL);
 240
 241         *mem = kmalloc(sizeof(struct kgd_mem), GFP_KERNEL);
 242         if ((*mem) == NULL)
 243                 return -ENOMEM;
 244
 245         r = radeon_bo_create(rdev, size, PAGE_SIZE, true, RADEON_GEM_DOMAIN_GTT,
 246                                 RADEON_GEM_GTT_WC, NULL, NULL, &(*mem)->bo);
 247         if (r) {
 248                 dev_err(rdev->dev,
 249                         "failed to allocate BO for amdkfd (%d)\n", r);
 250                 return r;
 251         }
 252
 253         /* map the buffer */
 254         r = radeon_bo_reserve((*mem)->bo, true);
 255         if (r) {
 256                 dev_err(rdev->dev, "(%d) failed to reserve bo for amdkfd\n", r);
 257                 goto allocate_mem_reserve_bo_failed;
 258         }
 259
 260         r = radeon_bo_pin((*mem)->bo, RADEON_GEM_DOMAIN_GTT,
 261                                 &(*mem)->gpu_addr);
 262         if (r) {
 263                 dev_err(rdev->dev, "(%d) failed to pin bo for amdkfd\n", r);
 264                 goto allocate_mem_pin_bo_failed;
 265         }
 266         *gpu_addr = (*mem)->gpu_addr;
 267
 268         r = radeon_bo_kmap((*mem)->bo, &(*mem)->cpu_ptr);
 269         if (r) {
 270                 dev_err(rdev->dev,
 271                         "(%d) failed to map bo to kernel for amdkfd\n", r);
 272                 goto allocate_mem_kmap_bo_failed;
 273         }
 274         *cpu_ptr = (*mem)->cpu_ptr;
 275
 276         radeon_bo_unreserve((*mem)->bo);
 277
 278         return 0;
 279
 280 allocate_mem_kmap_bo_failed:
 281         radeon_bo_unpin((*mem)->bo);
 282 allocate_mem_pin_bo_failed:
 283         radeon_bo_unreserve((*mem)->bo);
 284 allocate_mem_reserve_bo_failed:
 285         radeon_bo_unref(&(*mem)->bo);
 286
 287         return r;
 288 }
 289
 290 static void free_gtt_mem(struct kgd_dev *kgd, void *mem_obj)
 291 {
 292         struct kgd_mem *mem = (struct kgd_mem *) mem_obj;
 293
 294         BUG_ON(mem == NULL);
 295
 296         radeon_bo_reserve(mem->bo, true);
 297         radeon_bo_kunmap(mem->bo);
 298         radeon_bo_unpin(mem->bo);
 299         radeon_bo_unreserve(mem->bo);
 300         radeon_bo_unref(&(mem->bo));
 301         kfree(mem);
 302 }
 303
 304 static uint64_t get_vmem_size(struct kgd_dev *kgd)
 305 {
 306         struct radeon_device *rdev = (struct radeon_device *)kgd;
 307
 308         BUG_ON(kgd == NULL);
 309
 310         return rdev->mc.real_vram_size;
 311 }
 312
 313 static uint64_t get_gpu_clock_counter(struct kgd_dev *kgd)
 314 {
 315         struct radeon_device *rdev = (struct radeon_device *)kgd;
 316
 317         return rdev->asic->get_gpu_clock_counter(rdev);
 318 }
 319
 320 static uint32_t get_max_engine_clock_in_mhz(struct kgd_dev *kgd)
 321 {
 322         struct radeon_device *rdev = (struct radeon_device *)kgd;
 323
 324         /* The sclk is in quantas of 10kHz */
 325         return rdev->pm.dpm.dyn_state.max_clock_voltage_on_ac.sclk / 100;
 326 }
 327
 328 static inline struct radeon_device *get_radeon_device(struct kgd_dev *kgd)
 329 {
 330         return (struct radeon_device *)kgd;
 331 }
 332
 333 static void write_register(struct kgd_dev *kgd, uint32_t offset, uint32_t value)
 334 {
 335         struct radeon_device *rdev = get_radeon_device(kgd);
 336
 337         writel(value, (void __iomem *)(rdev->rmmio + offset));
 338 }
 339
 340 static uint32_t read_register(struct kgd_dev *kgd, uint32_t offset)
 341 {
 342         struct radeon_device *rdev = get_radeon_device(kgd);
 343
 344         return readl((void __iomem *)(rdev->rmmio + offset));
 345 }
 346
 347 static void lock_srbm(struct kgd_dev *kgd, uint32_t mec, uint32_t pipe,
 348                         uint32_t queue, uint32_t vmid)
 349 {
 350         struct radeon_device *rdev = get_radeon_device(kgd);
 351         uint32_t value = PIPEID(pipe) | MEID(mec) | VMID(vmid) | QUEUEID(queue);
 352
 353         mutex_lock(&rdev->srbm_mutex);
 354         write_register(kgd, SRBM_GFX_CNTL, value);
 355 }
 356
 357 static void unlock_srbm(struct kgd_dev *kgd)
 358 {
 359         struct radeon_device *rdev = get_radeon_device(kgd);
 360
 361         write_register(kgd, SRBM_GFX_CNTL, 0);
 362         mutex_unlock(&rdev->srbm_mutex);
 363 }
 364
 365 static void acquire_queue(struct kgd_dev *kgd, uint32_t pipe_id,
 366                                 uint32_t queue_id)
 367 {
 368         uint32_t mec = (++pipe_id / CIK_PIPE_PER_MEC) + 1;
 369         uint32_t pipe = (pipe_id % CIK_PIPE_PER_MEC);
 370
 371         lock_srbm(kgd, mec, pipe, queue_id, 0);
 372 }
 373
 374 static void release_queue(struct kgd_dev *kgd)
 375 {
 376         unlock_srbm(kgd);
 377 }
 378
 379 static void kgd_program_sh_mem_settings(struct kgd_dev *kgd, uint32_t vmid,
 380                                         uint32_t sh_mem_config,
 381                                         uint32_t sh_mem_ape1_base,
 382                                         uint32_t sh_mem_ape1_limit,
 383                                         uint32_t sh_mem_bases)
 384 {
 385         lock_srbm(kgd, 0, 0, 0, vmid);
 386
 387         write_register(kgd, SH_MEM_CONFIG, sh_mem_config);
 388         write_register(kgd, SH_MEM_APE1_BASE, sh_mem_ape1_base);
 389         write_register(kgd, SH_MEM_APE1_LIMIT, sh_mem_ape1_limit);
 390         write_register(kgd, SH_MEM_BASES, sh_mem_bases);
 391
 392         unlock_srbm(kgd);
 393 }
 394
 395 static int kgd_set_pasid_vmid_mapping(struct kgd_dev *kgd, unsigned int pasid,
 396                                         unsigned int vmid)
 397 {
 398         /*
 399          * We have to assume that there is no outstanding mapping.
 400          * The ATC_VMID_PASID_MAPPING_UPDATE_STATUS bit could be 0
 401          * because a mapping is in progress or because a mapping finished and
 402          * the SW cleared it.
 403          * So the protocol is to always wait & clear.
 404          */
 405         uint32_t pasid_mapping = (pasid == 0) ? 0 : (uint32_t)pasid |
 406                                         ATC_VMID_PASID_MAPPING_VALID_MASK;
 407
 408         write_register(kgd, ATC_VMID0_PASID_MAPPING + vmid*sizeof(uint32_t),
 409                         pasid_mapping);
 410
 411         while (!(read_register(kgd, ATC_VMID_PASID_MAPPING_UPDATE_STATUS) &
 412                                                                 (1U << vmid)))
 413                 cpu_relax();
 414         write_register(kgd, ATC_VMID_PASID_MAPPING_UPDATE_STATUS, 1U << vmid);
 415
 416         /* Mapping vmid to pasid also for IH block */
 417         write_register(kgd, IH_VMID_0_LUT + vmid * sizeof(uint32_t),
 418                         pasid_mapping);
 419
 420         return 0;
 421 }
 422
 423 static int kgd_init_pipeline(struct kgd_dev *kgd, uint32_t pipe_id,
 424                                 uint32_t hpd_size, uint64_t hpd_gpu_addr)
 425 {
 426         uint32_t mec = (pipe_id / CIK_PIPE_PER_MEC) + 1;
 427         uint32_t pipe = (pipe_id % CIK_PIPE_PER_MEC);
 428
 429         lock_srbm(kgd, mec, pipe, 0, 0);
 430         write_register(kgd, CP_HPD_EOP_BASE_ADDR,
 431                         lower_32_bits(hpd_gpu_addr >> 8));
 432         write_register(kgd, CP_HPD_EOP_BASE_ADDR_HI,
 433                         upper_32_bits(hpd_gpu_addr >> 8));
 434         write_register(kgd, CP_HPD_EOP_VMID, 0);
 435         write_register(kgd, CP_HPD_EOP_CONTROL, hpd_size);
 436         unlock_srbm(kgd);
 437
 438         return 0;
 439 }
 440
 441 static int kgd_init_interrupts(struct kgd_dev *kgd, uint32_t pipe_id)
 442 {
 443         uint32_t mec;
 444         uint32_t pipe;
 445
 446         mec = (pipe_id / CIK_PIPE_PER_MEC) + 1;
 447         pipe = (pipe_id % CIK_PIPE_PER_MEC);
 448
 449         lock_srbm(kgd, mec, pipe, 0, 0);
 450
 451         write_register(kgd, CPC_INT_CNTL,
 452                         TIME_STAMP_INT_ENABLE | OPCODE_ERROR_INT_ENABLE);
 453
 454         unlock_srbm(kgd);
 455
 456         return 0;
 457 }
 458
 459 static inline uint32_t get_sdma_base_addr(struct cik_sdma_rlc_registers *m)
 460 {
 461         uint32_t retval;
 462
 463         retval = m->sdma_engine_id * SDMA1_REGISTER_OFFSET +
 464                         m->sdma_queue_id * KFD_CIK_SDMA_QUEUE_OFFSET;
 465
 466         pr_debug("kfd: sdma base address: 0x%x\n", retval);
 467
 468         return retval;
 469 }
 470
 471 static inline struct cik_mqd *get_mqd(void *mqd)
 472 {
 473         return (struct cik_mqd *)mqd;
 474 }
 475
 476 static inline struct cik_sdma_rlc_registers *get_sdma_mqd(void *mqd)
 477 {
 478         return (struct cik_sdma_rlc_registers *)mqd;
 479 }
 480
 481 static int kgd_hqd_load(struct kgd_dev *kgd, void *mqd, uint32_t pipe_id,
 482                         uint32_t queue_id, uint32_t __user *wptr)
 483 {
 484         uint32_t wptr_shadow, is_wptr_shadow_valid;
 485         struct cik_mqd *m;
 486
 487         m = get_mqd(mqd);
 488
 489         is_wptr_shadow_valid = !get_user(wptr_shadow, wptr);
 490
 491         acquire_queue(kgd, pipe_id, queue_id);
 492         write_register(kgd, CP_MQD_BASE_ADDR, m->cp_mqd_base_addr_lo);
 493         write_register(kgd, CP_MQD_BASE_ADDR_HI, m->cp_mqd_base_addr_hi);
 494         write_register(kgd, CP_MQD_CONTROL, m->cp_mqd_control);
 495
 496         write_register(kgd, CP_HQD_PQ_BASE, m->cp_hqd_pq_base_lo);
 497         write_register(kgd, CP_HQD_PQ_BASE_HI, m->cp_hqd_pq_base_hi);
 498         write_register(kgd, CP_HQD_PQ_CONTROL, m->cp_hqd_pq_control);
 499
 500         write_register(kgd, CP_HQD_IB_CONTROL, m->cp_hqd_ib_control);
 501         write_register(kgd, CP_HQD_IB_BASE_ADDR, m->cp_hqd_ib_base_addr_lo);
 502         write_register(kgd, CP_HQD_IB_BASE_ADDR_HI, m->cp_hqd_ib_base_addr_hi);
 503
 504         write_register(kgd, CP_HQD_IB_RPTR, m->cp_hqd_ib_rptr);
 505
 506         write_register(kgd, CP_HQD_PERSISTENT_STATE,
 507                         m->cp_hqd_persistent_state);
 508         write_register(kgd, CP_HQD_SEMA_CMD, m->cp_hqd_sema_cmd);
 509         write_register(kgd, CP_HQD_MSG_TYPE, m->cp_hqd_msg_type);
 510
 511         write_register(kgd, CP_HQD_ATOMIC0_PREOP_LO,
 512                         m->cp_hqd_atomic0_preop_lo);
 513
 514         write_register(kgd, CP_HQD_ATOMIC0_PREOP_HI,
 515                         m->cp_hqd_atomic0_preop_hi);
 516
 517         write_register(kgd, CP_HQD_ATOMIC1_PREOP_LO,
 518                         m->cp_hqd_atomic1_preop_lo);
 519
 520         write_register(kgd, CP_HQD_ATOMIC1_PREOP_HI,
 521                         m->cp_hqd_atomic1_preop_hi);
 522
 523         write_register(kgd, CP_HQD_PQ_RPTR_REPORT_ADDR,
 524                         m->cp_hqd_pq_rptr_report_addr_lo);
 525
 526         write_register(kgd, CP_HQD_PQ_RPTR_REPORT_ADDR_HI,
 527                         m->cp_hqd_pq_rptr_report_addr_hi);
 528
 529         write_register(kgd, CP_HQD_PQ_RPTR, m->cp_hqd_pq_rptr);
 530
 531         write_register(kgd, CP_HQD_PQ_WPTR_POLL_ADDR,
 532                         m->cp_hqd_pq_wptr_poll_addr_lo);
 533
 534         write_register(kgd, CP_HQD_PQ_WPTR_POLL_ADDR_HI,
 535                         m->cp_hqd_pq_wptr_poll_addr_hi);
 536
 537         write_register(kgd, CP_HQD_PQ_DOORBELL_CONTROL,
 538                         m->cp_hqd_pq_doorbell_control);
 539
 540         write_register(kgd, CP_HQD_VMID, m->cp_hqd_vmid);
 541
 542         write_register(kgd, CP_HQD_QUANTUM, m->cp_hqd_quantum);
 543
 544         write_register(kgd, CP_HQD_PIPE_PRIORITY, m->cp_hqd_pipe_priority);
 545         write_register(kgd, CP_HQD_QUEUE_PRIORITY, m->cp_hqd_queue_priority);
 546
 547         write_register(kgd, CP_HQD_IQ_RPTR, m->cp_hqd_iq_rptr);
 548
 549         if (is_wptr_shadow_valid)
 550                 write_register(kgd, CP_HQD_PQ_WPTR, wptr_shadow);
 551
 552         write_register(kgd, CP_HQD_ACTIVE, m->cp_hqd_active);
 553         release_queue(kgd);
 554
 555         return 0;
 556 }
 557
 558 static int kgd_hqd_sdma_load(struct kgd_dev *kgd, void *mqd)
 559 {
 560         struct cik_sdma_rlc_registers *m;
 561         uint32_t sdma_base_addr;
 562
 563         m = get_sdma_mqd(mqd);
 564         sdma_base_addr = get_sdma_base_addr(m);
 565
 566         write_register(kgd,
 567                         sdma_base_addr + SDMA0_RLC0_VIRTUAL_ADDR,
 568                         m->sdma_rlc_virtual_addr);
 569
 570         write_register(kgd,
 571                         sdma_base_addr + SDMA0_RLC0_RB_BASE,
 572                         m->sdma_rlc_rb_base);
 573
 574         write_register(kgd,
 575                         sdma_base_addr + SDMA0_RLC0_RB_BASE_HI,
 576                         m->sdma_rlc_rb_base_hi);
 577
 578         write_register(kgd,
 579                         sdma_base_addr + SDMA0_RLC0_RB_RPTR_ADDR_LO,
 580                         m->sdma_rlc_rb_rptr_addr_lo);
 581
 582         write_register(kgd,
 583                         sdma_base_addr + SDMA0_RLC0_RB_RPTR_ADDR_HI,
 584                         m->sdma_rlc_rb_rptr_addr_hi);
 585
 586         write_register(kgd,
 587                         sdma_base_addr + SDMA0_RLC0_DOORBELL,
 588                         m->sdma_rlc_doorbell);
 589
 590         write_register(kgd,
 591                         sdma_base_addr + SDMA0_RLC0_RB_CNTL,
 592                         m->sdma_rlc_rb_cntl);
 593
 594         return 0;
 595 }
 596
 597 static bool kgd_hqd_is_occupied(struct kgd_dev *kgd, uint64_t queue_address,
 598                                 uint32_t pipe_id, uint32_t queue_id)
 599 {
 600         uint32_t act;
 601         bool retval = false;
 602         uint32_t low, high;
 603
 604         acquire_queue(kgd, pipe_id, queue_id);
 605         act = read_register(kgd, CP_HQD_ACTIVE);
 606         if (act) {
 607                 low = lower_32_bits(queue_address >> 8);
 608                 high = upper_32_bits(queue_address >> 8);
 609
 610                 if (low == read_register(kgd, CP_HQD_PQ_BASE) &&
 611                                 high == read_register(kgd, CP_HQD_PQ_BASE_HI))
 612                         retval = true;
 613         }
 614         release_queue(kgd);
 615         return retval;
 616 }
 617
 618 static bool kgd_hqd_sdma_is_occupied(struct kgd_dev *kgd, void *mqd)
 619 {
 620         struct cik_sdma_rlc_registers *m;
 621         uint32_t sdma_base_addr;
 622         uint32_t sdma_rlc_rb_cntl;
 623
 624         m = get_sdma_mqd(mqd);
 625         sdma_base_addr = get_sdma_base_addr(m);
 626
 627         sdma_rlc_rb_cntl = read_register(kgd,
 628                                         sdma_base_addr + SDMA0_RLC0_RB_CNTL);
 629
 630         if (sdma_rlc_rb_cntl & SDMA_RB_ENABLE)
 631                 return true;
 632
 633         return false;
 634 }
 635
 636 static int kgd_hqd_destroy(struct kgd_dev *kgd, uint32_t reset_type,
 637                                 unsigned int timeout, uint32_t pipe_id,
 638                                 uint32_t queue_id)
 639 {
 640         uint32_t temp;
 641
 642         acquire_queue(kgd, pipe_id, queue_id);
 643         write_register(kgd, CP_HQD_PQ_DOORBELL_CONTROL, 0);
 644
 645         write_register(kgd, CP_HQD_DEQUEUE_REQUEST, reset_type);
 646
 647         while (true) {
 648                 temp = read_register(kgd, CP_HQD_ACTIVE);
 649                 if (temp & 0x1)
 650                         break;
 651                 if (timeout == 0) {
 652                         pr_err("kfd: cp queue preemption time out (%dms)\n",
 653                                 temp);
 654                         release_queue(kgd);
 655                         return -ETIME;
 656                 }
 657                 msleep(20);
 658                 timeout -= 20;
 659         }
 660
 661         release_queue(kgd);
 662         return 0;
 663 }
 664
 665 static int kgd_hqd_sdma_destroy(struct kgd_dev *kgd, void *mqd,
 666                                 unsigned int timeout)
 667 {
 668         struct cik_sdma_rlc_registers *m;
 669         uint32_t sdma_base_addr;
 670         uint32_t temp;
 671
 672         m = get_sdma_mqd(mqd);
 673         sdma_base_addr = get_sdma_base_addr(m);
 674
 675         temp = read_register(kgd, sdma_base_addr + SDMA0_RLC0_RB_CNTL);
 676         temp = temp & ~SDMA_RB_ENABLE;
 677         write_register(kgd, sdma_base_addr + SDMA0_RLC0_RB_CNTL, temp);
 678
 679         while (true) {
 680                 temp = read_register(kgd, sdma_base_addr +
 681                                                 SDMA0_RLC0_CONTEXT_STATUS);
 682                 if (temp & SDMA_RLC_IDLE)
 683                         break;
 684                 if (timeout == 0)
 685                         return -ETIME;
 686                 msleep(20);
 687                 timeout -= 20;
 688         }
 689
 690         write_register(kgd, sdma_base_addr + SDMA0_RLC0_DOORBELL, 0);
 691         write_register(kgd, sdma_base_addr + SDMA0_RLC0_RB_RPTR, 0);
 692         write_register(kgd, sdma_base_addr + SDMA0_RLC0_RB_WPTR, 0);
 693         write_register(kgd, sdma_base_addr + SDMA0_RLC0_RB_BASE, 0);
 694
 695         return 0;
 696 }
 697
 698 static int kgd_address_watch_disable(struct kgd_dev *kgd)
 699 {
 700         union TCP_WATCH_CNTL_BITS cntl;
 701         unsigned int i;
 702
 703         cntl.u32All = 0;
 704
 705         cntl.bitfields.valid = 0;
 706         cntl.bitfields.mask = ADDRESS_WATCH_REG_CNTL_DEFAULT_MASK;
 707         cntl.bitfields.atc = 1;
 708
 709         /* Turning off this address until we set all the registers */
 710         for (i = 0; i < MAX_WATCH_ADDRESSES; i++)
 711                 write_register(kgd,
 712                                 watchRegs[i * ADDRESS_WATCH_REG_MAX +
 713                                         ADDRESS_WATCH_REG_CNTL],
 714                                 cntl.u32All);
 715
 716         return 0;
 717 }
 718
 719 static int kgd_address_watch_execute(struct kgd_dev *kgd,
 720                                         unsigned int watch_point_id,
 721                                         uint32_t cntl_val,
 722                                         uint32_t addr_hi,
 723                                         uint32_t addr_lo)
 724 {
 725         union TCP_WATCH_CNTL_BITS cntl;
 726
 727         cntl.u32All = cntl_val;
 728
 729         /* Turning off this watch point until we set all the registers */
 730         cntl.bitfields.valid = 0;
 731         write_register(kgd,
 732                         watchRegs[watch_point_id * ADDRESS_WATCH_REG_MAX +
 733                                 ADDRESS_WATCH_REG_CNTL],
 734                         cntl.u32All);
 735
 736         write_register(kgd,
 737                         watchRegs[watch_point_id * ADDRESS_WATCH_REG_MAX +
 738                                 ADDRESS_WATCH_REG_ADDR_HI],
 739                         addr_hi);
 740
 741         write_register(kgd,
 742                         watchRegs[watch_point_id * ADDRESS_WATCH_REG_MAX +
 743                                 ADDRESS_WATCH_REG_ADDR_LO],
 744                         addr_lo);
 745
 746         /* Enable the watch point */
 747         cntl.bitfields.valid = 1;
 748
 749         write_register(kgd,
 750                         watchRegs[watch_point_id * ADDRESS_WATCH_REG_MAX +
 751                                 ADDRESS_WATCH_REG_CNTL],
 752                         cntl.u32All);
 753
 754         return 0;
 755 }
 756
 757 static int kgd_wave_control_execute(struct kgd_dev *kgd,
 758                                         uint32_t gfx_index_val,
 759                                         uint32_t sq_cmd)
 760 {
 761         struct radeon_device *rdev = get_radeon_device(kgd);
 762         uint32_t data;
 763
 764         mutex_lock(&rdev->grbm_idx_mutex);
 765
 766         write_register(kgd, GRBM_GFX_INDEX, gfx_index_val);
 767         write_register(kgd, SQ_CMD, sq_cmd);
 768
 769         /*  Restore the GRBM_GFX_INDEX register  */
 770
 771         data = INSTANCE_BROADCAST_WRITES | SH_BROADCAST_WRITES |
 772                 SE_BROADCAST_WRITES;
 773
 774         write_register(kgd, GRBM_GFX_INDEX, data);
 775
 776         mutex_unlock(&rdev->grbm_idx_mutex);
 777
 778         return 0;
 779 }
 780
 781 static uint32_t kgd_address_watch_get_offset(struct kgd_dev *kgd,
 782                                         unsigned int watch_point_id,
 783                                         unsigned int reg_offset)
 784 {
 785         return watchRegs[watch_point_id * ADDRESS_WATCH_REG_MAX + reg_offset];
 786 }
 787
 788 static bool get_atc_vmid_pasid_mapping_valid(struct kgd_dev *kgd, uint8_t vmid)
 789 {
 790         uint32_t reg;
 791         struct radeon_device *rdev = (struct radeon_device *) kgd;
 792
 793         reg = RREG32(ATC_VMID0_PASID_MAPPING + vmid*4);
 794         return reg & ATC_VMID_PASID_MAPPING_VALID_MASK;
 795 }
 796
 797 static uint16_t get_atc_vmid_pasid_mapping_pasid(struct kgd_dev *kgd,
 798                                                         uint8_t vmid)
 799 {
 800         uint32_t reg;
 801         struct radeon_device *rdev = (struct radeon_device *) kgd;
 802
 803         reg = RREG32(ATC_VMID0_PASID_MAPPING + vmid*4);
 804         return reg & ATC_VMID_PASID_MAPPING_PASID_MASK;
 805 }
 806
 807 static void write_vmid_invalidate_request(struct kgd_dev *kgd, uint8_t vmid)
 808 {
 809         struct radeon_device *rdev = (struct radeon_device *) kgd;
 810
 811         return WREG32(VM_INVALIDATE_REQUEST, 1 << vmid);
 812 }
 813
 814 static uint16_t get_fw_version(struct kgd_dev *kgd, enum kgd_engine_type type)
 815 {
 816         struct radeon_device *rdev = (struct radeon_device *) kgd;
 817         const union radeon_firmware_header *hdr;
 818
 819         BUG_ON(kgd == NULL || rdev->mec_fw == NULL);
 820
 821         switch (type) {
 822         case KGD_ENGINE_PFP:
 823                 hdr = (const union radeon_firmware_header *) rdev->pfp_fw->data;
 824                 break;
 825
 826         case KGD_ENGINE_ME:
 827                 hdr = (const union radeon_firmware_header *) rdev->me_fw->data;
 828                 break;
 829
 830         case KGD_ENGINE_CE:
 831                 hdr = (const union radeon_firmware_header *) rdev->ce_fw->data;
 832                 break;
 833
 834         case KGD_ENGINE_MEC1:
 835                 hdr = (const union radeon_firmware_header *) rdev->mec_fw->data;
 836                 break;
 837
 838         case KGD_ENGINE_MEC2:
 839                 hdr = (const union radeon_firmware_header *)
 840                                                         rdev->mec2_fw->data;
 841                 break;
 842
 843         case KGD_ENGINE_RLC:
 844                 hdr = (const union radeon_firmware_header *) rdev->rlc_fw->data;
 845                 break;
 846
 847         case KGD_ENGINE_SDMA1:
 848         case KGD_ENGINE_SDMA2:
 849                 hdr = (const union radeon_firmware_header *)
 850                                                         rdev->sdma_fw->data;
 851                 break;
 852
 853         default:
 854                 return 0;
 855         }
 856
 857         if (hdr == NULL)
 858                 return 0;
 859
 860         /* Only 12 bit in use*/
 861         return hdr->common.ucode_version;
 862 }