drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd.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 "amdgpu_amdkfd.h"
  24 #include "amd_shared.h"
  25 #include <drm/drmP.h>
  26 #include "amdgpu.h"
  27 #include "amdgpu_gfx.h"
  28 #include <linux/module.h>
  29
  30 const struct kgd2kfd_calls *kgd2kfd;
  31 bool (*kgd2kfd_init_p)(unsigned int, const struct kgd2kfd_calls**);
  32
  33 int amdgpu_amdkfd_init(void)
  34 {
  35         int ret;
  36
  37 #if defined(CONFIG_HSA_AMD_MODULE)
  38         int (*kgd2kfd_init_p)(unsigned int, const struct kgd2kfd_calls**);
  39
  40         kgd2kfd_init_p = symbol_request(kgd2kfd_init);
  41
  42         if (kgd2kfd_init_p == NULL)
  43                 return -ENOENT;
  44
  45         ret = kgd2kfd_init_p(KFD_INTERFACE_VERSION, &kgd2kfd);
  46         if (ret) {
  47                 symbol_put(kgd2kfd_init);
  48                 kgd2kfd = NULL;
  49         }
  50
  51 #elif defined(CONFIG_HSA_AMD)
  52         ret = kgd2kfd_init(KFD_INTERFACE_VERSION, &kgd2kfd);
  53         if (ret)
  54                 kgd2kfd = NULL;
  55
  56 #else
  57         ret = -ENOENT;
  58 #endif
  59
  60         return ret;
  61 }
  62
  63 void amdgpu_amdkfd_fini(void)
  64 {
  65         if (kgd2kfd) {
  66                 kgd2kfd->exit();
  67                 symbol_put(kgd2kfd_init);
  68         }
  69 }
  70
  71 void amdgpu_amdkfd_device_probe(struct amdgpu_device *adev)
  72 {
  73         const struct kfd2kgd_calls *kfd2kgd;
  74
  75         if (!kgd2kfd)
  76                 return;
  77
  78         switch (adev->asic_type) {
  79 #ifdef CONFIG_DRM_AMDGPU_CIK
  80         case CHIP_KAVERI:
  81                 kfd2kgd = amdgpu_amdkfd_gfx_7_get_functions();
  82                 break;
  83 #endif
  84         case CHIP_CARRIZO:
  85                 kfd2kgd = amdgpu_amdkfd_gfx_8_0_get_functions();
  86                 break;
  87         default:
  88                 dev_dbg(adev->dev, "kfd not supported on this ASIC\n");
  89                 return;
  90         }
  91
  92         adev->kfd = kgd2kfd->probe((struct kgd_dev *)adev,
  93                                    adev->pdev, kfd2kgd);
  94 }
  95
  96 /**
  97  * amdgpu_doorbell_get_kfd_info - Report doorbell configuration required to
  98  *                                setup amdkfd
  99  *
 100  * @adev: amdgpu_device pointer
 101  * @aperture_base: output returning doorbell aperture base physical address
 102  * @aperture_size: output returning doorbell aperture size in bytes
 103  * @start_offset: output returning # of doorbell bytes reserved for amdgpu.
 104  *
 105  * amdgpu and amdkfd share the doorbell aperture. amdgpu sets it up,
 106  * takes doorbells required for its own rings and reports the setup to amdkfd.
 107  * amdgpu reserved doorbells are at the start of the doorbell aperture.
 108  */
 109 static void amdgpu_doorbell_get_kfd_info(struct amdgpu_device *adev,
 110                                          phys_addr_t *aperture_base,
 111                                          size_t *aperture_size,
 112                                          size_t *start_offset)
 113 {
 114         /*
 115          * The first num_doorbells are used by amdgpu.
 116          * amdkfd takes whatever's left in the aperture.
 117          */
 118         if (adev->doorbell.size > adev->doorbell.num_doorbells * sizeof(u32)) {
 119                 *aperture_base = adev->doorbell.base;
 120                 *aperture_size = adev->doorbell.size;
 121                 *start_offset = adev->doorbell.num_doorbells * sizeof(u32);
 122         } else {
 123                 *aperture_base = 0;
 124                 *aperture_size = 0;
 125                 *start_offset = 0;
 126         }
 127 }
 128
 129 void amdgpu_amdkfd_device_init(struct amdgpu_device *adev)
 130 {
 131         int i;
 132         int last_valid_bit;
 133         if (adev->kfd) {
 134                 struct kgd2kfd_shared_resources gpu_resources = {
 135                         .compute_vmid_bitmap = 0xFF00,
 136                         .num_pipe_per_mec = adev->gfx.mec.num_pipe_per_mec,
 137                         .num_queue_per_pipe = adev->gfx.mec.num_queue_per_pipe
 138                 };
 139
 140                 /* this is going to have a few of the MSBs set that we need to
 141                  * clear */
 142                 bitmap_complement(gpu_resources.queue_bitmap,
 143                                   adev->gfx.mec.queue_bitmap,
 144                                   KGD_MAX_QUEUES);
 145
 146                 /* remove the KIQ bit as well */
 147                 if (adev->gfx.kiq.ring.ready)
 148                         clear_bit(amdgpu_gfx_queue_to_bit(adev,
 149                                                           adev->gfx.kiq.ring.me - 1,
 150                                                           adev->gfx.kiq.ring.pipe,
 151                                                           adev->gfx.kiq.ring.queue),
 152                                   gpu_resources.queue_bitmap);
 153
 154                 /* According to linux/bitmap.h we shouldn't use bitmap_clear if
 155                  * nbits is not compile time constant */
 156                 last_valid_bit = 1 /* only first MEC can have compute queues */
 157                                 * adev->gfx.mec.num_pipe_per_mec
 158                                 * adev->gfx.mec.num_queue_per_pipe;
 159                 for (i = last_valid_bit; i < KGD_MAX_QUEUES; ++i)
 160                         clear_bit(i, gpu_resources.queue_bitmap);
 161
 162                 amdgpu_doorbell_get_kfd_info(adev,
 163                                 &gpu_resources.doorbell_physical_address,
 164                                 &gpu_resources.doorbell_aperture_size,
 165                                 &gpu_resources.doorbell_start_offset);
 166
 167                 kgd2kfd->device_init(adev->kfd, &gpu_resources);
 168         }
 169 }
 170
 171 void amdgpu_amdkfd_device_fini(struct amdgpu_device *adev)
 172 {
 173         if (adev->kfd) {
 174                 kgd2kfd->device_exit(adev->kfd);
 175                 adev->kfd = NULL;
 176         }
 177 }
 178
 179 void amdgpu_amdkfd_interrupt(struct amdgpu_device *adev,
 180                 const void *ih_ring_entry)
 181 {
 182         if (adev->kfd)
 183                 kgd2kfd->interrupt(adev->kfd, ih_ring_entry);
 184 }
 185
 186 void amdgpu_amdkfd_suspend(struct amdgpu_device *adev)
 187 {
 188         if (adev->kfd)
 189                 kgd2kfd->suspend(adev->kfd);
 190 }
 191
 192 int amdgpu_amdkfd_resume(struct amdgpu_device *adev)
 193 {
 194         int r = 0;
 195
 196         if (adev->kfd)
 197                 r = kgd2kfd->resume(adev->kfd);
 198
 199         return r;
 200 }
 201
 202 int alloc_gtt_mem(struct kgd_dev *kgd, size_t size,
 203                         void **mem_obj, uint64_t *gpu_addr,
 204                         void **cpu_ptr)
 205 {
 206         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 207         struct kgd_mem **mem = (struct kgd_mem **) mem_obj;
 208         int r;
 209
 210         BUG_ON(kgd == NULL);
 211         BUG_ON(gpu_addr == NULL);
 212         BUG_ON(cpu_ptr == NULL);
 213
 214         *mem = kmalloc(sizeof(struct kgd_mem), GFP_KERNEL);
 215         if ((*mem) == NULL)
 216                 return -ENOMEM;
 217
 218         r = amdgpu_bo_create(adev, size, PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_GTT,
 219                              AMDGPU_GEM_CREATE_CPU_GTT_USWC, NULL, NULL, 0,
 220                              &(*mem)->bo);
 221         if (r) {
 222                 dev_err(adev->dev,
 223                         "failed to allocate BO for amdkfd (%d)\n", r);
 224                 return r;
 225         }
 226
 227         /* map the buffer */
 228         r = amdgpu_bo_reserve((*mem)->bo, true);
 229         if (r) {
 230                 dev_err(adev->dev, "(%d) failed to reserve bo for amdkfd\n", r);
 231                 goto allocate_mem_reserve_bo_failed;
 232         }
 233
 234         r = amdgpu_bo_pin((*mem)->bo, AMDGPU_GEM_DOMAIN_GTT,
 235                                 &(*mem)->gpu_addr);
 236         if (r) {
 237                 dev_err(adev->dev, "(%d) failed to pin bo for amdkfd\n", r);
 238                 goto allocate_mem_pin_bo_failed;
 239         }
 240         *gpu_addr = (*mem)->gpu_addr;
 241
 242         r = amdgpu_bo_kmap((*mem)->bo, &(*mem)->cpu_ptr);
 243         if (r) {
 244                 dev_err(adev->dev,
 245                         "(%d) failed to map bo to kernel for amdkfd\n", r);
 246                 goto allocate_mem_kmap_bo_failed;
 247         }
 248         *cpu_ptr = (*mem)->cpu_ptr;
 249
 250         amdgpu_bo_unreserve((*mem)->bo);
 251
 252         return 0;
 253
 254 allocate_mem_kmap_bo_failed:
 255         amdgpu_bo_unpin((*mem)->bo);
 256 allocate_mem_pin_bo_failed:
 257         amdgpu_bo_unreserve((*mem)->bo);
 258 allocate_mem_reserve_bo_failed:
 259         amdgpu_bo_unref(&(*mem)->bo);
 260
 261         return r;
 262 }
 263
 264 void free_gtt_mem(struct kgd_dev *kgd, void *mem_obj)
 265 {
 266         struct kgd_mem *mem = (struct kgd_mem *) mem_obj;
 267
 268         BUG_ON(mem == NULL);
 269
 270         amdgpu_bo_reserve(mem->bo, true);
 271         amdgpu_bo_kunmap(mem->bo);
 272         amdgpu_bo_unpin(mem->bo);
 273         amdgpu_bo_unreserve(mem->bo);
 274         amdgpu_bo_unref(&(mem->bo));
 275         kfree(mem);
 276 }
 277
 278 uint64_t get_vmem_size(struct kgd_dev *kgd)
 279 {
 280         struct amdgpu_device *adev =
 281                 (struct amdgpu_device *)kgd;
 282
 283         BUG_ON(kgd == NULL);
 284
 285         return adev->mc.real_vram_size;
 286 }
 287
 288 void get_local_mem_info(struct kgd_dev *kgd,
 289                         struct kfd_local_mem_info *mem_info)
 290 {
 291         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 292         uint64_t address_mask = adev->dev->dma_mask ? ~*adev->dev->dma_mask :
 293                                              ~((1ULL << 32) - 1);
 294         resource_size_t aper_limit = adev->mc.aper_base + adev->mc.aper_size;
 295
 296         memset(mem_info, 0, sizeof(*mem_info));
 297         if (!(adev->mc.aper_base & address_mask || aper_limit & address_mask)) {
 298                 mem_info->local_mem_size_public = adev->mc.visible_vram_size;
 299                 mem_info->local_mem_size_private = adev->mc.real_vram_size -
 300                                 adev->mc.visible_vram_size;
 301         } else {
 302                 mem_info->local_mem_size_public = 0;
 303                 mem_info->local_mem_size_private = adev->mc.real_vram_size;
 304         }
 305         mem_info->vram_width = adev->mc.vram_width;
 306
 307         pr_debug("Address base: 0x%llx limit 0x%llx public 0x%llx private 0x%llx\n",
 308                         adev->mc.aper_base, aper_limit,
 309                         mem_info->local_mem_size_public,
 310                         mem_info->local_mem_size_private);
 311
 312         if (amdgpu_sriov_vf(adev))
 313                 mem_info->mem_clk_max = adev->clock.default_mclk / 100;
 314         else
 315                 mem_info->mem_clk_max = amdgpu_dpm_get_mclk(adev, false) / 100;
 316 }
 317
 318 uint64_t get_gpu_clock_counter(struct kgd_dev *kgd)
 319 {
 320         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 321
 322         if (adev->gfx.funcs->get_gpu_clock_counter)
 323                 return adev->gfx.funcs->get_gpu_clock_counter(adev);
 324         return 0;
 325 }
 326
 327 uint32_t get_max_engine_clock_in_mhz(struct kgd_dev *kgd)
 328 {
 329         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 330
 331         /* the sclk is in quantas of 10kHz */
 332         if (amdgpu_sriov_vf(adev))
 333                 return adev->clock.default_sclk / 100;
 334
 335         return amdgpu_dpm_get_sclk(adev, false) / 100;
 336 }
 337
 338 void get_cu_info(struct kgd_dev *kgd, struct kfd_cu_info *cu_info)
 339 {
 340         struct amdgpu_device *adev = (struct amdgpu_device *)kgd;
 341         struct amdgpu_cu_info acu_info = adev->gfx.cu_info;
 342
 343         memset(cu_info, 0, sizeof(*cu_info));
 344         if (sizeof(cu_info->cu_bitmap) != sizeof(acu_info.bitmap))
 345                 return;
 346
 347         cu_info->cu_active_number = acu_info.number;
 348         cu_info->cu_ao_mask = acu_info.ao_cu_mask;
 349         memcpy(&cu_info->cu_bitmap[0], &acu_info.bitmap[0],
 350                sizeof(acu_info.bitmap));
 351         cu_info->num_shader_engines = adev->gfx.config.max_shader_engines;
 352         cu_info->num_shader_arrays_per_engine = adev->gfx.config.max_sh_per_se;
 353         cu_info->num_cu_per_sh = adev->gfx.config.max_cu_per_sh;
 354         cu_info->simd_per_cu = acu_info.simd_per_cu;
 355         cu_info->max_waves_per_simd = acu_info.max_waves_per_simd;
 356         cu_info->wave_front_size = acu_info.wave_front_size;
 357         cu_info->max_scratch_slots_per_cu = acu_info.max_scratch_slots_per_cu;
 358         cu_info->lds_size = acu_info.lds_size;
 359 }