1 /* 2 * Copyright 2016-2024 Advanced Micro Devices, Inc. 3 * All Rights Reserved. 4 * 5 * Permission is hereby granted, free of charge, to any person obtaining a 6 * copy of this software and associated documentation files (the 7 * "Software"), to deal in the Software without restriction, including 8 * without limitation the rights to use, copy, modify, merge, publish, 9 * distribute, sub license, and/or sell copies of the Software, and to 10 * permit persons to whom the Software is furnished to do so, subject to 11 * the following conditions: 12 * 13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 15 * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL 16 * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, 17 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR 18 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE 19 * USE OR OTHER DEALINGS IN THE SOFTWARE. 20 * 21 * The above copyright notice and this permission notice (including the 22 * next paragraph) shall be included in all copies or substantial portions 23 * of the Software. 24 * 25 */ 26 27 #include <linux/firmware.h> 28 #include <linux/module.h> 29 #include <linux/dmi.h> 30 #include <linux/pci.h> 31 #include <linux/debugfs.h> 32 #include <drm/drm_drv.h> 33 34 #include "amdgpu.h" 35 #include "amdgpu_pm.h" 36 #include "amdgpu_vcn.h" 37 #include "soc15d.h" 38 39 /* Firmware Names */ 40 #define FIRMWARE_RAVEN "amdgpu/raven_vcn.bin" 41 #define FIRMWARE_PICASSO "amdgpu/picasso_vcn.bin" 42 #define FIRMWARE_RAVEN2 "amdgpu/raven2_vcn.bin" 43 #define FIRMWARE_ARCTURUS "amdgpu/arcturus_vcn.bin" 44 #define FIRMWARE_RENOIR "amdgpu/renoir_vcn.bin" 45 #define FIRMWARE_GREEN_SARDINE "amdgpu/green_sardine_vcn.bin" 46 #define FIRMWARE_NAVI10 "amdgpu/navi10_vcn.bin" 47 #define FIRMWARE_NAVI14 "amdgpu/navi14_vcn.bin" 48 #define FIRMWARE_NAVI12 "amdgpu/navi12_vcn.bin" 49 #define FIRMWARE_SIENNA_CICHLID "amdgpu/sienna_cichlid_vcn.bin" 50 #define FIRMWARE_NAVY_FLOUNDER "amdgpu/navy_flounder_vcn.bin" 51 #define FIRMWARE_VANGOGH "amdgpu/vangogh_vcn.bin" 52 #define FIRMWARE_DIMGREY_CAVEFISH "amdgpu/dimgrey_cavefish_vcn.bin" 53 #define FIRMWARE_ALDEBARAN "amdgpu/aldebaran_vcn.bin" 54 #define FIRMWARE_BEIGE_GOBY "amdgpu/beige_goby_vcn.bin" 55 #define FIRMWARE_YELLOW_CARP "amdgpu/yellow_carp_vcn.bin" 56 #define FIRMWARE_VCN_3_1_2 "amdgpu/vcn_3_1_2.bin" 57 #define FIRMWARE_VCN4_0_0 "amdgpu/vcn_4_0_0.bin" 58 #define FIRMWARE_VCN4_0_2 "amdgpu/vcn_4_0_2.bin" 59 #define FIRMWARE_VCN4_0_3 "amdgpu/vcn_4_0_3.bin" 60 #define FIRMWARE_VCN4_0_4 "amdgpu/vcn_4_0_4.bin" 61 #define FIRMWARE_VCN4_0_5 "amdgpu/vcn_4_0_5.bin" 62 #define FIRMWARE_VCN4_0_6 "amdgpu/vcn_4_0_6.bin" 63 #define FIRMWARE_VCN4_0_6_1 "amdgpu/vcn_4_0_6_1.bin" 64 #define FIRMWARE_VCN5_0_0 "amdgpu/vcn_5_0_0.bin" 65 #define FIRMWARE_VCN5_0_1 "amdgpu/vcn_5_0_1.bin" 66 67 MODULE_FIRMWARE(FIRMWARE_RAVEN); 68 MODULE_FIRMWARE(FIRMWARE_PICASSO); 69 MODULE_FIRMWARE(FIRMWARE_RAVEN2); 70 MODULE_FIRMWARE(FIRMWARE_ARCTURUS); 71 MODULE_FIRMWARE(FIRMWARE_RENOIR); 72 MODULE_FIRMWARE(FIRMWARE_GREEN_SARDINE); 73 MODULE_FIRMWARE(FIRMWARE_ALDEBARAN); 74 MODULE_FIRMWARE(FIRMWARE_NAVI10); 75 MODULE_FIRMWARE(FIRMWARE_NAVI14); 76 MODULE_FIRMWARE(FIRMWARE_NAVI12); 77 MODULE_FIRMWARE(FIRMWARE_SIENNA_CICHLID); 78 MODULE_FIRMWARE(FIRMWARE_NAVY_FLOUNDER); 79 MODULE_FIRMWARE(FIRMWARE_VANGOGH); 80 MODULE_FIRMWARE(FIRMWARE_DIMGREY_CAVEFISH); 81 MODULE_FIRMWARE(FIRMWARE_BEIGE_GOBY); 82 MODULE_FIRMWARE(FIRMWARE_YELLOW_CARP); 83 MODULE_FIRMWARE(FIRMWARE_VCN_3_1_2); 84 MODULE_FIRMWARE(FIRMWARE_VCN4_0_0); 85 MODULE_FIRMWARE(FIRMWARE_VCN4_0_2); 86 MODULE_FIRMWARE(FIRMWARE_VCN4_0_3); 87 MODULE_FIRMWARE(FIRMWARE_VCN4_0_4); 88 MODULE_FIRMWARE(FIRMWARE_VCN4_0_5); 89 MODULE_FIRMWARE(FIRMWARE_VCN4_0_6); 90 MODULE_FIRMWARE(FIRMWARE_VCN4_0_6_1); 91 MODULE_FIRMWARE(FIRMWARE_VCN5_0_0); 92 MODULE_FIRMWARE(FIRMWARE_VCN5_0_1); 93 94 static void amdgpu_vcn_idle_work_handler(struct work_struct *work); 95 static void amdgpu_vcn_reg_dump_fini(struct amdgpu_device *adev); 96 97 int amdgpu_vcn_early_init(struct amdgpu_device *adev, int i) 98 { 99 char ucode_prefix[25]; 100 int r; 101 102 adev->vcn.inst[i].adev = adev; 103 adev->vcn.inst[i].inst = i; 104 amdgpu_ucode_ip_version_decode(adev, UVD_HWIP, ucode_prefix, sizeof(ucode_prefix)); 105 106 if (i != 0 && adev->vcn.per_inst_fw) { 107 r = amdgpu_ucode_request(adev, &adev->vcn.inst[i].fw, 108 AMDGPU_UCODE_REQUIRED, 109 "amdgpu/%s_%d.bin", ucode_prefix, i); 110 if (r) 111 amdgpu_ucode_release(&adev->vcn.inst[i].fw); 112 } else { 113 if (!adev->vcn.inst[0].fw) { 114 r = amdgpu_ucode_request(adev, &adev->vcn.inst[0].fw, 115 AMDGPU_UCODE_REQUIRED, 116 "amdgpu/%s.bin", ucode_prefix); 117 if (r) 118 amdgpu_ucode_release(&adev->vcn.inst[0].fw); 119 } else { 120 r = 0; 121 } 122 adev->vcn.inst[i].fw = adev->vcn.inst[0].fw; 123 } 124 125 return r; 126 } 127 128 int amdgpu_vcn_sw_init(struct amdgpu_device *adev, int i) 129 { 130 unsigned long bo_size; 131 const struct common_firmware_header *hdr; 132 unsigned char fw_check; 133 unsigned int fw_shared_size, log_offset; 134 int r; 135 136 mutex_init(&adev->vcn.inst[i].vcn1_jpeg1_workaround); 137 mutex_init(&adev->vcn.inst[i].vcn_pg_lock); 138 mutex_init(&adev->vcn.inst[i].engine_reset_mutex); 139 atomic_set(&adev->vcn.inst[i].total_submission_cnt, 0); 140 INIT_DELAYED_WORK(&adev->vcn.inst[i].idle_work, amdgpu_vcn_idle_work_handler); 141 atomic_set(&adev->vcn.inst[i].dpg_enc_submission_cnt, 0); 142 if ((adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) && 143 (adev->pg_flags & AMD_PG_SUPPORT_VCN_DPG)) 144 adev->vcn.inst[i].indirect_sram = true; 145 146 /* 147 * Some Steam Deck's BIOS versions are incompatible with the 148 * indirect SRAM mode, leading to amdgpu being unable to get 149 * properly probed (and even potentially crashing the kernel). 150 * Hence, check for these versions here - notice this is 151 * restricted to Vangogh (Deck's APU). 152 */ 153 if (amdgpu_ip_version(adev, UVD_HWIP, 0) == IP_VERSION(3, 0, 2)) { 154 const char *bios_ver = dmi_get_system_info(DMI_BIOS_VERSION); 155 156 if (bios_ver && (!strncmp("F7A0113", bios_ver, 7) || 157 !strncmp("F7A0114", bios_ver, 7))) { 158 adev->vcn.inst[i].indirect_sram = false; 159 dev_info(adev->dev, 160 "Steam Deck quirk: indirect SRAM disabled on BIOS %s\n", bios_ver); 161 } 162 } 163 164 /* from vcn4 and above, only unified queue is used */ 165 adev->vcn.inst[i].using_unified_queue = 166 amdgpu_ip_version(adev, UVD_HWIP, 0) >= IP_VERSION(4, 0, 0); 167 168 hdr = (const struct common_firmware_header *)adev->vcn.inst[i].fw->data; 169 adev->vcn.inst[i].fw_version = le32_to_cpu(hdr->ucode_version); 170 adev->vcn.fw_version = le32_to_cpu(hdr->ucode_version); 171 172 /* Bit 20-23, it is encode major and non-zero for new naming convention. 173 * This field is part of version minor and DRM_DISABLED_FLAG in old naming 174 * convention. Since the l:wq!atest version minor is 0x5B and DRM_DISABLED_FLAG 175 * is zero in old naming convention, this field is always zero so far. 176 * These four bits are used to tell which naming convention is present. 177 */ 178 fw_check = (le32_to_cpu(hdr->ucode_version) >> 20) & 0xf; 179 if (fw_check) { 180 unsigned int dec_ver, enc_major, enc_minor, vep, fw_rev; 181 182 fw_rev = le32_to_cpu(hdr->ucode_version) & 0xfff; 183 enc_minor = (le32_to_cpu(hdr->ucode_version) >> 12) & 0xff; 184 enc_major = fw_check; 185 dec_ver = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xf; 186 vep = (le32_to_cpu(hdr->ucode_version) >> 28) & 0xf; 187 dev_info(adev->dev, 188 "[VCN instance %d] Found VCN firmware Version ENC: %u.%u DEC: %u VEP: %u Revision: %u\n", 189 i, enc_major, enc_minor, dec_ver, vep, fw_rev); 190 } else { 191 unsigned int version_major, version_minor, family_id; 192 193 family_id = le32_to_cpu(hdr->ucode_version) & 0xff; 194 version_major = (le32_to_cpu(hdr->ucode_version) >> 24) & 0xff; 195 version_minor = (le32_to_cpu(hdr->ucode_version) >> 8) & 0xff; 196 dev_info(adev->dev, "[VCN instance %d] Found VCN firmware Version: %u.%u Family ID: %u\n", 197 i, version_major, version_minor, family_id); 198 } 199 200 bo_size = AMDGPU_VCN_STACK_SIZE + AMDGPU_VCN_CONTEXT_SIZE; 201 if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) 202 bo_size += AMDGPU_GPU_PAGE_ALIGN(le32_to_cpu(hdr->ucode_size_bytes) + 8); 203 204 if (amdgpu_ip_version(adev, UVD_HWIP, 0) >= IP_VERSION(5, 0, 0)) { 205 fw_shared_size = AMDGPU_GPU_PAGE_ALIGN(sizeof(struct amdgpu_vcn5_fw_shared)); 206 log_offset = offsetof(struct amdgpu_vcn5_fw_shared, fw_log); 207 } else if (amdgpu_ip_version(adev, UVD_HWIP, 0) >= IP_VERSION(4, 0, 0)) { 208 fw_shared_size = AMDGPU_GPU_PAGE_ALIGN(sizeof(struct amdgpu_vcn4_fw_shared)); 209 log_offset = offsetof(struct amdgpu_vcn4_fw_shared, fw_log); 210 } else { 211 fw_shared_size = AMDGPU_GPU_PAGE_ALIGN(sizeof(struct amdgpu_fw_shared)); 212 log_offset = offsetof(struct amdgpu_fw_shared, fw_log); 213 } 214 215 bo_size += fw_shared_size; 216 217 if (amdgpu_vcnfw_log) 218 bo_size += AMDGPU_VCNFW_LOG_SIZE; 219 220 r = amdgpu_bo_create_kernel(adev, bo_size, PAGE_SIZE, 221 AMDGPU_GEM_DOMAIN_VRAM | 222 AMDGPU_GEM_DOMAIN_GTT, 223 &adev->vcn.inst[i].vcpu_bo, 224 &adev->vcn.inst[i].gpu_addr, 225 &adev->vcn.inst[i].cpu_addr); 226 if (r) { 227 dev_err(adev->dev, "(%d) failed to allocate vcn bo\n", r); 228 return r; 229 } 230 231 adev->vcn.inst[i].fw_shared.cpu_addr = adev->vcn.inst[i].cpu_addr + 232 bo_size - fw_shared_size; 233 adev->vcn.inst[i].fw_shared.gpu_addr = adev->vcn.inst[i].gpu_addr + 234 bo_size - fw_shared_size; 235 236 adev->vcn.inst[i].fw_shared.mem_size = fw_shared_size; 237 238 if (amdgpu_vcnfw_log) { 239 adev->vcn.inst[i].fw_shared.cpu_addr -= AMDGPU_VCNFW_LOG_SIZE; 240 adev->vcn.inst[i].fw_shared.gpu_addr -= AMDGPU_VCNFW_LOG_SIZE; 241 adev->vcn.inst[i].fw_shared.log_offset = log_offset; 242 } 243 244 if (adev->vcn.inst[i].indirect_sram) { 245 r = amdgpu_bo_create_kernel(adev, 64 * 2 * 4, PAGE_SIZE, 246 AMDGPU_GEM_DOMAIN_VRAM | 247 AMDGPU_GEM_DOMAIN_GTT, 248 &adev->vcn.inst[i].dpg_sram_bo, 249 &adev->vcn.inst[i].dpg_sram_gpu_addr, 250 &adev->vcn.inst[i].dpg_sram_cpu_addr); 251 if (r) { 252 dev_err(adev->dev, "VCN %d (%d) failed to allocate DPG bo\n", i, r); 253 return r; 254 } 255 } 256 257 return 0; 258 } 259 260 void amdgpu_vcn_sw_fini(struct amdgpu_device *adev, int i) 261 { 262 int j; 263 264 if (adev->vcn.harvest_config & (1 << i)) 265 return; 266 267 amdgpu_bo_free_kernel( 268 &adev->vcn.inst[i].dpg_sram_bo, 269 &adev->vcn.inst[i].dpg_sram_gpu_addr, 270 (void **)&adev->vcn.inst[i].dpg_sram_cpu_addr); 271 272 kvfree(adev->vcn.inst[i].saved_bo); 273 274 amdgpu_bo_free_kernel(&adev->vcn.inst[i].vcpu_bo, 275 &adev->vcn.inst[i].gpu_addr, 276 (void **)&adev->vcn.inst[i].cpu_addr); 277 278 amdgpu_ring_fini(&adev->vcn.inst[i].ring_dec); 279 280 for (j = 0; j < adev->vcn.inst[i].num_enc_rings; ++j) 281 amdgpu_ring_fini(&adev->vcn.inst[i].ring_enc[j]); 282 283 if (adev->vcn.per_inst_fw) { 284 amdgpu_ucode_release(&adev->vcn.inst[i].fw); 285 } else { 286 amdgpu_ucode_release(&adev->vcn.inst[0].fw); 287 adev->vcn.inst[i].fw = NULL; 288 } 289 290 if (adev->vcn.reg_list) 291 amdgpu_vcn_reg_dump_fini(adev); 292 293 mutex_destroy(&adev->vcn.inst[i].vcn_pg_lock); 294 mutex_destroy(&adev->vcn.inst[i].vcn1_jpeg1_workaround); 295 } 296 297 bool amdgpu_vcn_is_disabled_vcn(struct amdgpu_device *adev, enum vcn_ring_type type, uint32_t vcn_instance) 298 { 299 bool ret = false; 300 int vcn_config = adev->vcn.inst[vcn_instance].vcn_config; 301 302 if ((type == VCN_ENCODE_RING) && (vcn_config & VCN_BLOCK_ENCODE_DISABLE_MASK)) 303 ret = true; 304 else if ((type == VCN_DECODE_RING) && (vcn_config & VCN_BLOCK_DECODE_DISABLE_MASK)) 305 ret = true; 306 else if ((type == VCN_UNIFIED_RING) && (vcn_config & VCN_BLOCK_QUEUE_DISABLE_MASK)) 307 ret = true; 308 309 return ret; 310 } 311 312 static int amdgpu_vcn_save_vcpu_bo_inst(struct amdgpu_device *adev, int i) 313 { 314 unsigned int size; 315 void *ptr; 316 int idx; 317 318 if (adev->vcn.harvest_config & (1 << i)) 319 return 0; 320 if (adev->vcn.inst[i].vcpu_bo == NULL) 321 return 0; 322 323 size = amdgpu_bo_size(adev->vcn.inst[i].vcpu_bo); 324 ptr = adev->vcn.inst[i].cpu_addr; 325 326 adev->vcn.inst[i].saved_bo = kvmalloc(size, GFP_KERNEL); 327 if (!adev->vcn.inst[i].saved_bo) 328 return -ENOMEM; 329 330 if (drm_dev_enter(adev_to_drm(adev), &idx)) { 331 memcpy_fromio(adev->vcn.inst[i].saved_bo, ptr, size); 332 drm_dev_exit(idx); 333 } 334 335 return 0; 336 } 337 338 int amdgpu_vcn_save_vcpu_bo(struct amdgpu_device *adev) 339 { 340 int ret, i; 341 342 for (i = 0; i < adev->vcn.num_vcn_inst; ++i) { 343 ret = amdgpu_vcn_save_vcpu_bo_inst(adev, i); 344 if (ret) 345 return ret; 346 } 347 348 return 0; 349 } 350 351 int amdgpu_vcn_suspend(struct amdgpu_device *adev, int i) 352 { 353 bool in_ras_intr = amdgpu_ras_intr_triggered(); 354 355 if (adev->vcn.harvest_config & (1 << i)) 356 return 0; 357 358 /* err_event_athub and dpc recovery will corrupt VCPU buffer, so we need to 359 * restore fw data and clear buffer in amdgpu_vcn_resume() */ 360 if (in_ras_intr || adev->pcie_reset_ctx.in_link_reset) 361 return 0; 362 363 return amdgpu_vcn_save_vcpu_bo_inst(adev, i); 364 } 365 366 int amdgpu_vcn_resume(struct amdgpu_device *adev, int i) 367 { 368 unsigned int size; 369 void *ptr; 370 int idx; 371 372 if (adev->vcn.harvest_config & (1 << i)) 373 return 0; 374 if (adev->vcn.inst[i].vcpu_bo == NULL) 375 return -EINVAL; 376 377 size = amdgpu_bo_size(adev->vcn.inst[i].vcpu_bo); 378 ptr = adev->vcn.inst[i].cpu_addr; 379 380 if (adev->vcn.inst[i].saved_bo != NULL) { 381 if (drm_dev_enter(adev_to_drm(adev), &idx)) { 382 memcpy_toio(ptr, adev->vcn.inst[i].saved_bo, size); 383 drm_dev_exit(idx); 384 } 385 kvfree(adev->vcn.inst[i].saved_bo); 386 adev->vcn.inst[i].saved_bo = NULL; 387 } else { 388 const struct common_firmware_header *hdr; 389 unsigned int offset; 390 391 hdr = (const struct common_firmware_header *)adev->vcn.inst[i].fw->data; 392 if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) { 393 offset = le32_to_cpu(hdr->ucode_array_offset_bytes); 394 if (drm_dev_enter(adev_to_drm(adev), &idx)) { 395 memcpy_toio(adev->vcn.inst[i].cpu_addr, 396 adev->vcn.inst[i].fw->data + offset, 397 le32_to_cpu(hdr->ucode_size_bytes)); 398 drm_dev_exit(idx); 399 } 400 size -= le32_to_cpu(hdr->ucode_size_bytes); 401 ptr += le32_to_cpu(hdr->ucode_size_bytes); 402 } 403 memset_io(ptr, 0, size); 404 } 405 406 return 0; 407 } 408 409 void amdgpu_vcn_get_profile(struct amdgpu_device *adev) 410 { 411 int r; 412 413 mutex_lock(&adev->vcn.workload_profile_mutex); 414 415 if (adev->vcn.workload_profile_active) { 416 mutex_unlock(&adev->vcn.workload_profile_mutex); 417 return; 418 } 419 r = amdgpu_dpm_switch_power_profile(adev, PP_SMC_POWER_PROFILE_VIDEO, 420 true); 421 if (r) 422 dev_warn(adev->dev, 423 "(%d) failed to enable video power profile mode\n", r); 424 else 425 adev->vcn.workload_profile_active = true; 426 mutex_unlock(&adev->vcn.workload_profile_mutex); 427 } 428 429 void amdgpu_vcn_put_profile(struct amdgpu_device *adev) 430 { 431 bool pg = true; 432 int r, i; 433 434 mutex_lock(&adev->vcn.workload_profile_mutex); 435 for (i = 0; i < adev->vcn.num_vcn_inst; i++) { 436 if (adev->vcn.inst[i].cur_state != AMD_PG_STATE_GATE) { 437 pg = false; 438 break; 439 } 440 } 441 442 if (pg) { 443 r = amdgpu_dpm_switch_power_profile( 444 adev, PP_SMC_POWER_PROFILE_VIDEO, false); 445 if (r) 446 dev_warn( 447 adev->dev, 448 "(%d) failed to disable video power profile mode\n", 449 r); 450 else 451 adev->vcn.workload_profile_active = false; 452 } 453 454 mutex_unlock(&adev->vcn.workload_profile_mutex); 455 } 456 457 static void amdgpu_vcn_idle_work_handler(struct work_struct *work) 458 { 459 struct amdgpu_vcn_inst *vcn_inst = 460 container_of(work, struct amdgpu_vcn_inst, idle_work.work); 461 struct amdgpu_device *adev = vcn_inst->adev; 462 unsigned int fences = 0, fence[AMDGPU_MAX_VCN_INSTANCES] = {0}; 463 unsigned int i = vcn_inst->inst, j; 464 465 if (adev->vcn.harvest_config & (1 << i)) 466 return; 467 468 for (j = 0; j < adev->vcn.inst[i].num_enc_rings; ++j) 469 fence[i] += amdgpu_fence_count_emitted(&vcn_inst->ring_enc[j]); 470 471 /* Only set DPG pause for VCN3 or below, VCN4 and above will be handled by FW */ 472 if (adev->pg_flags & AMD_PG_SUPPORT_VCN_DPG && 473 !adev->vcn.inst[i].using_unified_queue) { 474 struct dpg_pause_state new_state; 475 476 if (fence[i] || 477 unlikely(atomic_read(&vcn_inst->dpg_enc_submission_cnt))) 478 new_state.fw_based = VCN_DPG_STATE__PAUSE; 479 else 480 new_state.fw_based = VCN_DPG_STATE__UNPAUSE; 481 482 adev->vcn.inst[i].pause_dpg_mode(vcn_inst, &new_state); 483 } 484 485 fence[i] += amdgpu_fence_count_emitted(&vcn_inst->ring_dec); 486 fences += fence[i]; 487 488 if (!fences && !atomic_read(&vcn_inst->total_submission_cnt)) { 489 mutex_lock(&vcn_inst->vcn_pg_lock); 490 vcn_inst->set_pg_state(vcn_inst, AMD_PG_STATE_GATE); 491 mutex_unlock(&vcn_inst->vcn_pg_lock); 492 amdgpu_vcn_put_profile(adev); 493 494 } else { 495 schedule_delayed_work(&vcn_inst->idle_work, VCN_IDLE_TIMEOUT); 496 } 497 } 498 499 void amdgpu_vcn_ring_begin_use(struct amdgpu_ring *ring) 500 { 501 struct amdgpu_device *adev = ring->adev; 502 struct amdgpu_vcn_inst *vcn_inst = &adev->vcn.inst[ring->me]; 503 504 atomic_inc(&vcn_inst->total_submission_cnt); 505 506 cancel_delayed_work_sync(&vcn_inst->idle_work); 507 508 mutex_lock(&vcn_inst->vcn_pg_lock); 509 vcn_inst->set_pg_state(vcn_inst, AMD_PG_STATE_UNGATE); 510 511 /* Only set DPG pause for VCN3 or below, VCN4 and above will be handled by FW */ 512 if (adev->pg_flags & AMD_PG_SUPPORT_VCN_DPG && 513 !vcn_inst->using_unified_queue) { 514 struct dpg_pause_state new_state; 515 516 if (ring->funcs->type == AMDGPU_RING_TYPE_VCN_ENC) { 517 atomic_inc(&vcn_inst->dpg_enc_submission_cnt); 518 new_state.fw_based = VCN_DPG_STATE__PAUSE; 519 } else { 520 unsigned int fences = 0; 521 unsigned int i; 522 523 for (i = 0; i < vcn_inst->num_enc_rings; ++i) 524 fences += amdgpu_fence_count_emitted(&vcn_inst->ring_enc[i]); 525 526 if (fences || atomic_read(&vcn_inst->dpg_enc_submission_cnt)) 527 new_state.fw_based = VCN_DPG_STATE__PAUSE; 528 else 529 new_state.fw_based = VCN_DPG_STATE__UNPAUSE; 530 } 531 532 vcn_inst->pause_dpg_mode(vcn_inst, &new_state); 533 } 534 mutex_unlock(&vcn_inst->vcn_pg_lock); 535 amdgpu_vcn_get_profile(adev); 536 } 537 538 void amdgpu_vcn_ring_end_use(struct amdgpu_ring *ring) 539 { 540 struct amdgpu_device *adev = ring->adev; 541 542 /* Only set DPG pause for VCN3 or below, VCN4 and above will be handled by FW */ 543 if (ring->adev->pg_flags & AMD_PG_SUPPORT_VCN_DPG && 544 ring->funcs->type == AMDGPU_RING_TYPE_VCN_ENC && 545 !adev->vcn.inst[ring->me].using_unified_queue) 546 atomic_dec(&ring->adev->vcn.inst[ring->me].dpg_enc_submission_cnt); 547 548 atomic_dec(&ring->adev->vcn.inst[ring->me].total_submission_cnt); 549 550 schedule_delayed_work(&ring->adev->vcn.inst[ring->me].idle_work, 551 VCN_IDLE_TIMEOUT); 552 } 553 554 int amdgpu_vcn_dec_ring_test_ring(struct amdgpu_ring *ring) 555 { 556 struct amdgpu_device *adev = ring->adev; 557 uint32_t tmp = 0; 558 unsigned int i; 559 int r; 560 561 /* VCN in SRIOV does not support direct register read/write */ 562 if (amdgpu_sriov_vf(adev)) 563 return 0; 564 565 WREG32(adev->vcn.inst[ring->me].external.scratch9, 0xCAFEDEAD); 566 r = amdgpu_ring_alloc(ring, 3); 567 if (r) 568 return r; 569 amdgpu_ring_write(ring, PACKET0(adev->vcn.inst[ring->me].internal.scratch9, 0)); 570 amdgpu_ring_write(ring, 0xDEADBEEF); 571 amdgpu_ring_commit(ring); 572 for (i = 0; i < adev->usec_timeout; i++) { 573 tmp = RREG32(adev->vcn.inst[ring->me].external.scratch9); 574 if (tmp == 0xDEADBEEF) 575 break; 576 udelay(1); 577 } 578 579 if (i >= adev->usec_timeout) 580 r = -ETIMEDOUT; 581 582 return r; 583 } 584 585 int amdgpu_vcn_dec_sw_ring_test_ring(struct amdgpu_ring *ring) 586 { 587 struct amdgpu_device *adev = ring->adev; 588 uint32_t rptr; 589 unsigned int i; 590 int r; 591 592 if (amdgpu_sriov_vf(adev)) 593 return 0; 594 595 r = amdgpu_ring_alloc(ring, 16); 596 if (r) 597 return r; 598 599 rptr = amdgpu_ring_get_rptr(ring); 600 601 amdgpu_ring_write(ring, VCN_DEC_SW_CMD_END); 602 amdgpu_ring_commit(ring); 603 604 for (i = 0; i < adev->usec_timeout; i++) { 605 if (amdgpu_ring_get_rptr(ring) != rptr) 606 break; 607 udelay(1); 608 } 609 610 if (i >= adev->usec_timeout) 611 r = -ETIMEDOUT; 612 613 return r; 614 } 615 616 static int amdgpu_vcn_dec_send_msg(struct amdgpu_ring *ring, 617 struct amdgpu_ib *ib_msg, 618 struct dma_fence **fence) 619 { 620 u64 addr = AMDGPU_GPU_PAGE_ALIGN(ib_msg->gpu_addr); 621 struct amdgpu_device *adev = ring->adev; 622 struct dma_fence *f = NULL; 623 struct amdgpu_job *job; 624 struct amdgpu_ib *ib; 625 int i, r; 626 627 r = amdgpu_job_alloc_with_ib(ring->adev, NULL, NULL, 628 64, AMDGPU_IB_POOL_DIRECT, 629 &job, AMDGPU_KERNEL_JOB_ID_VCN_RING_TEST); 630 if (r) 631 goto err; 632 633 ib = &job->ibs[0]; 634 ib->ptr[0] = PACKET0(adev->vcn.inst[ring->me].internal.data0, 0); 635 ib->ptr[1] = addr; 636 ib->ptr[2] = PACKET0(adev->vcn.inst[ring->me].internal.data1, 0); 637 ib->ptr[3] = addr >> 32; 638 ib->ptr[4] = PACKET0(adev->vcn.inst[ring->me].internal.cmd, 0); 639 ib->ptr[5] = 0; 640 for (i = 6; i < 16; i += 2) { 641 ib->ptr[i] = PACKET0(adev->vcn.inst[ring->me].internal.nop, 0); 642 ib->ptr[i+1] = 0; 643 } 644 ib->length_dw = 16; 645 646 r = amdgpu_job_submit_direct(job, ring, &f); 647 if (r) 648 goto err_free; 649 650 amdgpu_ib_free(ib_msg, f); 651 652 if (fence) 653 *fence = dma_fence_get(f); 654 dma_fence_put(f); 655 656 return 0; 657 658 err_free: 659 amdgpu_job_free(job); 660 err: 661 amdgpu_ib_free(ib_msg, f); 662 return r; 663 } 664 665 static int amdgpu_vcn_dec_get_create_msg(struct amdgpu_ring *ring, uint32_t handle, 666 struct amdgpu_ib *ib) 667 { 668 struct amdgpu_device *adev = ring->adev; 669 uint32_t *msg; 670 int r, i; 671 672 memset(ib, 0, sizeof(*ib)); 673 r = amdgpu_ib_get(adev, NULL, AMDGPU_GPU_PAGE_SIZE * 2, 674 AMDGPU_IB_POOL_DIRECT, 675 ib); 676 if (r) 677 return r; 678 679 msg = (uint32_t *)AMDGPU_GPU_PAGE_ALIGN((unsigned long)ib->ptr); 680 msg[0] = cpu_to_le32(0x00000028); 681 msg[1] = cpu_to_le32(0x00000038); 682 msg[2] = cpu_to_le32(0x00000001); 683 msg[3] = cpu_to_le32(0x00000000); 684 msg[4] = cpu_to_le32(handle); 685 msg[5] = cpu_to_le32(0x00000000); 686 msg[6] = cpu_to_le32(0x00000001); 687 msg[7] = cpu_to_le32(0x00000028); 688 msg[8] = cpu_to_le32(0x00000010); 689 msg[9] = cpu_to_le32(0x00000000); 690 msg[10] = cpu_to_le32(0x00000007); 691 msg[11] = cpu_to_le32(0x00000000); 692 msg[12] = cpu_to_le32(0x00000780); 693 msg[13] = cpu_to_le32(0x00000440); 694 for (i = 14; i < 1024; ++i) 695 msg[i] = cpu_to_le32(0x0); 696 697 return 0; 698 } 699 700 static int amdgpu_vcn_dec_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle, 701 struct amdgpu_ib *ib) 702 { 703 struct amdgpu_device *adev = ring->adev; 704 uint32_t *msg; 705 int r, i; 706 707 memset(ib, 0, sizeof(*ib)); 708 r = amdgpu_ib_get(adev, NULL, AMDGPU_GPU_PAGE_SIZE * 2, 709 AMDGPU_IB_POOL_DIRECT, 710 ib); 711 if (r) 712 return r; 713 714 msg = (uint32_t *)AMDGPU_GPU_PAGE_ALIGN((unsigned long)ib->ptr); 715 msg[0] = cpu_to_le32(0x00000028); 716 msg[1] = cpu_to_le32(0x00000018); 717 msg[2] = cpu_to_le32(0x00000000); 718 msg[3] = cpu_to_le32(0x00000002); 719 msg[4] = cpu_to_le32(handle); 720 msg[5] = cpu_to_le32(0x00000000); 721 for (i = 6; i < 1024; ++i) 722 msg[i] = cpu_to_le32(0x0); 723 724 return 0; 725 } 726 727 int amdgpu_vcn_dec_ring_test_ib(struct amdgpu_ring *ring, long timeout) 728 { 729 struct dma_fence *fence = NULL; 730 struct amdgpu_ib ib; 731 long r; 732 733 r = amdgpu_vcn_dec_get_create_msg(ring, 1, &ib); 734 if (r) 735 goto error; 736 737 r = amdgpu_vcn_dec_send_msg(ring, &ib, NULL); 738 if (r) 739 goto error; 740 r = amdgpu_vcn_dec_get_destroy_msg(ring, 1, &ib); 741 if (r) 742 goto error; 743 744 r = amdgpu_vcn_dec_send_msg(ring, &ib, &fence); 745 if (r) 746 goto error; 747 748 r = dma_fence_wait_timeout(fence, false, timeout); 749 if (r == 0) 750 r = -ETIMEDOUT; 751 else if (r > 0) 752 r = 0; 753 754 dma_fence_put(fence); 755 error: 756 return r; 757 } 758 759 static uint32_t *amdgpu_vcn_unified_ring_ib_header(struct amdgpu_ib *ib, 760 uint32_t ib_pack_in_dw, bool enc) 761 { 762 uint32_t *ib_checksum; 763 764 ib->ptr[ib->length_dw++] = 0x00000010; /* single queue checksum */ 765 ib->ptr[ib->length_dw++] = 0x30000002; 766 ib_checksum = &ib->ptr[ib->length_dw++]; 767 ib->ptr[ib->length_dw++] = ib_pack_in_dw; 768 769 ib->ptr[ib->length_dw++] = 0x00000010; /* engine info */ 770 ib->ptr[ib->length_dw++] = 0x30000001; 771 ib->ptr[ib->length_dw++] = enc ? 0x2 : 0x3; 772 ib->ptr[ib->length_dw++] = ib_pack_in_dw * sizeof(uint32_t); 773 774 return ib_checksum; 775 } 776 777 static void amdgpu_vcn_unified_ring_ib_checksum(uint32_t **ib_checksum, 778 uint32_t ib_pack_in_dw) 779 { 780 uint32_t i; 781 uint32_t checksum = 0; 782 783 for (i = 0; i < ib_pack_in_dw; i++) 784 checksum += *(*ib_checksum + 2 + i); 785 786 **ib_checksum = checksum; 787 } 788 789 static int amdgpu_vcn_dec_sw_send_msg(struct amdgpu_ring *ring, 790 struct amdgpu_ib *ib_msg, 791 struct dma_fence **fence) 792 { 793 struct amdgpu_vcn_decode_buffer *decode_buffer = NULL; 794 unsigned int ib_size_dw = 64; 795 struct amdgpu_device *adev = ring->adev; 796 struct dma_fence *f = NULL; 797 struct amdgpu_job *job; 798 struct amdgpu_ib *ib; 799 uint64_t addr = AMDGPU_GPU_PAGE_ALIGN(ib_msg->gpu_addr); 800 uint32_t *ib_checksum; 801 uint32_t ib_pack_in_dw; 802 int i, r; 803 804 if (adev->vcn.inst[ring->me].using_unified_queue) 805 ib_size_dw += 8; 806 807 r = amdgpu_job_alloc_with_ib(ring->adev, NULL, NULL, 808 ib_size_dw * 4, AMDGPU_IB_POOL_DIRECT, 809 &job, AMDGPU_KERNEL_JOB_ID_VCN_RING_TEST); 810 if (r) 811 goto err; 812 813 ib = &job->ibs[0]; 814 ib->length_dw = 0; 815 816 /* single queue headers */ 817 if (adev->vcn.inst[ring->me].using_unified_queue) { 818 ib_pack_in_dw = sizeof(struct amdgpu_vcn_decode_buffer) / sizeof(uint32_t) 819 + 4 + 2; /* engine info + decoding ib in dw */ 820 ib_checksum = amdgpu_vcn_unified_ring_ib_header(ib, ib_pack_in_dw, false); 821 } 822 823 ib->ptr[ib->length_dw++] = sizeof(struct amdgpu_vcn_decode_buffer) + 8; 824 ib->ptr[ib->length_dw++] = cpu_to_le32(AMDGPU_VCN_IB_FLAG_DECODE_BUFFER); 825 decode_buffer = (struct amdgpu_vcn_decode_buffer *)&(ib->ptr[ib->length_dw]); 826 ib->length_dw += sizeof(struct amdgpu_vcn_decode_buffer) / 4; 827 memset(decode_buffer, 0, sizeof(struct amdgpu_vcn_decode_buffer)); 828 829 decode_buffer->valid_buf_flag |= cpu_to_le32(AMDGPU_VCN_CMD_FLAG_MSG_BUFFER); 830 decode_buffer->msg_buffer_address_hi = cpu_to_le32(addr >> 32); 831 decode_buffer->msg_buffer_address_lo = cpu_to_le32(addr); 832 833 for (i = ib->length_dw; i < ib_size_dw; ++i) 834 ib->ptr[i] = 0x0; 835 836 if (adev->vcn.inst[ring->me].using_unified_queue) 837 amdgpu_vcn_unified_ring_ib_checksum(&ib_checksum, ib_pack_in_dw); 838 839 r = amdgpu_job_submit_direct(job, ring, &f); 840 if (r) 841 goto err_free; 842 843 amdgpu_ib_free(ib_msg, f); 844 845 if (fence) 846 *fence = dma_fence_get(f); 847 dma_fence_put(f); 848 849 return 0; 850 851 err_free: 852 amdgpu_job_free(job); 853 err: 854 amdgpu_ib_free(ib_msg, f); 855 return r; 856 } 857 858 int amdgpu_vcn_dec_sw_ring_test_ib(struct amdgpu_ring *ring, long timeout) 859 { 860 struct dma_fence *fence = NULL; 861 struct amdgpu_ib ib; 862 long r; 863 864 r = amdgpu_vcn_dec_get_create_msg(ring, 1, &ib); 865 if (r) 866 goto error; 867 868 r = amdgpu_vcn_dec_sw_send_msg(ring, &ib, NULL); 869 if (r) 870 goto error; 871 r = amdgpu_vcn_dec_get_destroy_msg(ring, 1, &ib); 872 if (r) 873 goto error; 874 875 r = amdgpu_vcn_dec_sw_send_msg(ring, &ib, &fence); 876 if (r) 877 goto error; 878 879 r = dma_fence_wait_timeout(fence, false, timeout); 880 if (r == 0) 881 r = -ETIMEDOUT; 882 else if (r > 0) 883 r = 0; 884 885 dma_fence_put(fence); 886 error: 887 return r; 888 } 889 890 int amdgpu_vcn_enc_ring_test_ring(struct amdgpu_ring *ring) 891 { 892 struct amdgpu_device *adev = ring->adev; 893 uint32_t rptr; 894 unsigned int i; 895 int r; 896 897 if (amdgpu_sriov_vf(adev)) 898 return 0; 899 900 r = amdgpu_ring_alloc(ring, 16); 901 if (r) 902 return r; 903 904 rptr = amdgpu_ring_get_rptr(ring); 905 906 amdgpu_ring_write(ring, VCN_ENC_CMD_END); 907 amdgpu_ring_commit(ring); 908 909 for (i = 0; i < adev->usec_timeout; i++) { 910 if (amdgpu_ring_get_rptr(ring) != rptr) 911 break; 912 udelay(1); 913 } 914 915 if (i >= adev->usec_timeout) 916 r = -ETIMEDOUT; 917 918 return r; 919 } 920 921 static int amdgpu_vcn_enc_get_create_msg(struct amdgpu_ring *ring, uint32_t handle, 922 struct amdgpu_ib *ib_msg, 923 struct dma_fence **fence) 924 { 925 unsigned int ib_size_dw = 16; 926 struct amdgpu_device *adev = ring->adev; 927 struct amdgpu_job *job; 928 struct amdgpu_ib *ib; 929 struct dma_fence *f = NULL; 930 uint32_t *ib_checksum = NULL; 931 uint64_t addr; 932 int i, r; 933 934 if (adev->vcn.inst[ring->me].using_unified_queue) 935 ib_size_dw += 8; 936 937 r = amdgpu_job_alloc_with_ib(ring->adev, NULL, NULL, 938 ib_size_dw * 4, AMDGPU_IB_POOL_DIRECT, 939 &job, AMDGPU_KERNEL_JOB_ID_VCN_RING_TEST); 940 if (r) 941 return r; 942 943 ib = &job->ibs[0]; 944 addr = AMDGPU_GPU_PAGE_ALIGN(ib_msg->gpu_addr); 945 946 ib->length_dw = 0; 947 948 if (adev->vcn.inst[ring->me].using_unified_queue) 949 ib_checksum = amdgpu_vcn_unified_ring_ib_header(ib, 0x11, true); 950 951 ib->ptr[ib->length_dw++] = 0x00000018; 952 ib->ptr[ib->length_dw++] = 0x00000001; /* session info */ 953 ib->ptr[ib->length_dw++] = handle; 954 ib->ptr[ib->length_dw++] = upper_32_bits(addr); 955 ib->ptr[ib->length_dw++] = addr; 956 ib->ptr[ib->length_dw++] = 0x00000000; 957 958 ib->ptr[ib->length_dw++] = 0x00000014; 959 ib->ptr[ib->length_dw++] = 0x00000002; /* task info */ 960 ib->ptr[ib->length_dw++] = 0x0000001c; 961 ib->ptr[ib->length_dw++] = 0x00000000; 962 ib->ptr[ib->length_dw++] = 0x00000000; 963 964 ib->ptr[ib->length_dw++] = 0x00000008; 965 ib->ptr[ib->length_dw++] = 0x08000001; /* op initialize */ 966 967 for (i = ib->length_dw; i < ib_size_dw; ++i) 968 ib->ptr[i] = 0x0; 969 970 if (adev->vcn.inst[ring->me].using_unified_queue) 971 amdgpu_vcn_unified_ring_ib_checksum(&ib_checksum, 0x11); 972 973 r = amdgpu_job_submit_direct(job, ring, &f); 974 if (r) 975 goto err; 976 977 if (fence) 978 *fence = dma_fence_get(f); 979 dma_fence_put(f); 980 981 return 0; 982 983 err: 984 amdgpu_job_free(job); 985 return r; 986 } 987 988 static int amdgpu_vcn_enc_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle, 989 struct amdgpu_ib *ib_msg, 990 struct dma_fence **fence) 991 { 992 unsigned int ib_size_dw = 16; 993 struct amdgpu_device *adev = ring->adev; 994 struct amdgpu_job *job; 995 struct amdgpu_ib *ib; 996 struct dma_fence *f = NULL; 997 uint32_t *ib_checksum = NULL; 998 uint64_t addr; 999 int i, r; 1000 1001 if (adev->vcn.inst[ring->me].using_unified_queue) 1002 ib_size_dw += 8; 1003 1004 r = amdgpu_job_alloc_with_ib(ring->adev, NULL, NULL, 1005 ib_size_dw * 4, AMDGPU_IB_POOL_DIRECT, 1006 &job, AMDGPU_KERNEL_JOB_ID_VCN_RING_TEST); 1007 if (r) 1008 return r; 1009 1010 ib = &job->ibs[0]; 1011 addr = AMDGPU_GPU_PAGE_ALIGN(ib_msg->gpu_addr); 1012 1013 ib->length_dw = 0; 1014 1015 if (adev->vcn.inst[ring->me].using_unified_queue) 1016 ib_checksum = amdgpu_vcn_unified_ring_ib_header(ib, 0x11, true); 1017 1018 ib->ptr[ib->length_dw++] = 0x00000018; 1019 ib->ptr[ib->length_dw++] = 0x00000001; 1020 ib->ptr[ib->length_dw++] = handle; 1021 ib->ptr[ib->length_dw++] = upper_32_bits(addr); 1022 ib->ptr[ib->length_dw++] = addr; 1023 ib->ptr[ib->length_dw++] = 0x00000000; 1024 1025 ib->ptr[ib->length_dw++] = 0x00000014; 1026 ib->ptr[ib->length_dw++] = 0x00000002; 1027 ib->ptr[ib->length_dw++] = 0x0000001c; 1028 ib->ptr[ib->length_dw++] = 0x00000000; 1029 ib->ptr[ib->length_dw++] = 0x00000000; 1030 1031 ib->ptr[ib->length_dw++] = 0x00000008; 1032 ib->ptr[ib->length_dw++] = 0x08000002; /* op close session */ 1033 1034 for (i = ib->length_dw; i < ib_size_dw; ++i) 1035 ib->ptr[i] = 0x0; 1036 1037 if (adev->vcn.inst[ring->me].using_unified_queue) 1038 amdgpu_vcn_unified_ring_ib_checksum(&ib_checksum, 0x11); 1039 1040 r = amdgpu_job_submit_direct(job, ring, &f); 1041 if (r) 1042 goto err; 1043 1044 if (fence) 1045 *fence = dma_fence_get(f); 1046 dma_fence_put(f); 1047 1048 return 0; 1049 1050 err: 1051 amdgpu_job_free(job); 1052 return r; 1053 } 1054 1055 int amdgpu_vcn_enc_ring_test_ib(struct amdgpu_ring *ring, long timeout) 1056 { 1057 struct amdgpu_device *adev = ring->adev; 1058 struct dma_fence *fence = NULL; 1059 struct amdgpu_ib ib; 1060 long r; 1061 1062 memset(&ib, 0, sizeof(ib)); 1063 r = amdgpu_ib_get(adev, NULL, (128 << 10) + AMDGPU_GPU_PAGE_SIZE, 1064 AMDGPU_IB_POOL_DIRECT, 1065 &ib); 1066 if (r) 1067 return r; 1068 1069 r = amdgpu_vcn_enc_get_create_msg(ring, 1, &ib, NULL); 1070 if (r) 1071 goto error; 1072 1073 r = amdgpu_vcn_enc_get_destroy_msg(ring, 1, &ib, &fence); 1074 if (r) 1075 goto error; 1076 1077 r = dma_fence_wait_timeout(fence, false, timeout); 1078 if (r == 0) 1079 r = -ETIMEDOUT; 1080 else if (r > 0) 1081 r = 0; 1082 1083 error: 1084 amdgpu_ib_free(&ib, fence); 1085 dma_fence_put(fence); 1086 1087 return r; 1088 } 1089 1090 int amdgpu_vcn_unified_ring_test_ib(struct amdgpu_ring *ring, long timeout) 1091 { 1092 struct amdgpu_device *adev = ring->adev; 1093 long r; 1094 1095 if ((amdgpu_ip_version(adev, UVD_HWIP, 0) != IP_VERSION(4, 0, 3)) && 1096 (amdgpu_ip_version(adev, UVD_HWIP, 0) != IP_VERSION(5, 0, 1))) { 1097 r = amdgpu_vcn_enc_ring_test_ib(ring, timeout); 1098 if (r) 1099 goto error; 1100 } 1101 1102 r = amdgpu_vcn_dec_sw_ring_test_ib(ring, timeout); 1103 1104 error: 1105 return r; 1106 } 1107 1108 enum amdgpu_ring_priority_level amdgpu_vcn_get_enc_ring_prio(int ring) 1109 { 1110 switch (ring) { 1111 case 0: 1112 return AMDGPU_RING_PRIO_0; 1113 case 1: 1114 return AMDGPU_RING_PRIO_1; 1115 case 2: 1116 return AMDGPU_RING_PRIO_2; 1117 default: 1118 return AMDGPU_RING_PRIO_0; 1119 } 1120 } 1121 1122 void amdgpu_vcn_setup_ucode(struct amdgpu_device *adev, int i) 1123 { 1124 unsigned int idx; 1125 1126 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) { 1127 const struct common_firmware_header *hdr; 1128 1129 if (adev->vcn.harvest_config & (1 << i)) 1130 return; 1131 1132 if ((amdgpu_ip_version(adev, UVD_HWIP, 0) == IP_VERSION(4, 0, 3) || 1133 amdgpu_ip_version(adev, UVD_HWIP, 0) == IP_VERSION(5, 0, 1)) 1134 && (i > 0)) 1135 return; 1136 1137 hdr = (const struct common_firmware_header *)adev->vcn.inst[i].fw->data; 1138 /* currently only support 2 FW instances */ 1139 if (i >= 2) { 1140 dev_info(adev->dev, "More then 2 VCN FW instances!\n"); 1141 return; 1142 } 1143 idx = AMDGPU_UCODE_ID_VCN + i; 1144 adev->firmware.ucode[idx].ucode_id = idx; 1145 adev->firmware.ucode[idx].fw = adev->vcn.inst[i].fw; 1146 adev->firmware.fw_size += 1147 ALIGN(le32_to_cpu(hdr->ucode_size_bytes), PAGE_SIZE); 1148 } 1149 } 1150 1151 /* 1152 * debugfs for mapping vcn firmware log buffer. 1153 */ 1154 #if defined(CONFIG_DEBUG_FS) 1155 static ssize_t amdgpu_debugfs_vcn_fwlog_read(struct file *f, char __user *buf, 1156 size_t size, loff_t *pos) 1157 { 1158 struct amdgpu_vcn_inst *vcn; 1159 void *log_buf; 1160 struct amdgpu_vcn_fwlog *plog; 1161 unsigned int read_pos, write_pos, available, i, read_bytes = 0; 1162 unsigned int read_num[2] = {0}; 1163 1164 vcn = file_inode(f)->i_private; 1165 if (!vcn) 1166 return -ENODEV; 1167 1168 if (!vcn->fw_shared.cpu_addr || !amdgpu_vcnfw_log) 1169 return -EFAULT; 1170 1171 log_buf = vcn->fw_shared.cpu_addr + vcn->fw_shared.mem_size; 1172 1173 plog = (struct amdgpu_vcn_fwlog *)log_buf; 1174 read_pos = plog->rptr; 1175 write_pos = plog->wptr; 1176 1177 if (read_pos > AMDGPU_VCNFW_LOG_SIZE || write_pos > AMDGPU_VCNFW_LOG_SIZE) 1178 return -EFAULT; 1179 1180 if (!size || (read_pos == write_pos)) 1181 return 0; 1182 1183 if (write_pos > read_pos) { 1184 available = write_pos - read_pos; 1185 read_num[0] = min_t(size_t, size, available); 1186 } else { 1187 read_num[0] = AMDGPU_VCNFW_LOG_SIZE - read_pos; 1188 available = read_num[0] + write_pos - plog->header_size; 1189 if (size > available) 1190 read_num[1] = write_pos - plog->header_size; 1191 else if (size > read_num[0]) 1192 read_num[1] = size - read_num[0]; 1193 else 1194 read_num[0] = size; 1195 } 1196 1197 for (i = 0; i < 2; i++) { 1198 if (read_num[i]) { 1199 if (read_pos == AMDGPU_VCNFW_LOG_SIZE) 1200 read_pos = plog->header_size; 1201 if (read_num[i] == copy_to_user((buf + read_bytes), 1202 (log_buf + read_pos), read_num[i])) 1203 return -EFAULT; 1204 1205 read_bytes += read_num[i]; 1206 read_pos += read_num[i]; 1207 } 1208 } 1209 1210 plog->rptr = read_pos; 1211 *pos += read_bytes; 1212 return read_bytes; 1213 } 1214 1215 static const struct file_operations amdgpu_debugfs_vcnfwlog_fops = { 1216 .owner = THIS_MODULE, 1217 .read = amdgpu_debugfs_vcn_fwlog_read, 1218 .llseek = default_llseek 1219 }; 1220 #endif 1221 1222 void amdgpu_debugfs_vcn_fwlog_init(struct amdgpu_device *adev, uint8_t i, 1223 struct amdgpu_vcn_inst *vcn) 1224 { 1225 #if defined(CONFIG_DEBUG_FS) 1226 struct drm_minor *minor = adev_to_drm(adev)->primary; 1227 struct dentry *root = minor->debugfs_root; 1228 char name[32]; 1229 1230 sprintf(name, "amdgpu_vcn_%d_fwlog", i); 1231 debugfs_create_file_size(name, S_IFREG | 0444, root, vcn, 1232 &amdgpu_debugfs_vcnfwlog_fops, 1233 AMDGPU_VCNFW_LOG_SIZE); 1234 #endif 1235 } 1236 1237 void amdgpu_vcn_fwlog_init(struct amdgpu_vcn_inst *vcn) 1238 { 1239 #if defined(CONFIG_DEBUG_FS) 1240 uint32_t *flag = vcn->fw_shared.cpu_addr; 1241 void *fw_log_cpu_addr = vcn->fw_shared.cpu_addr + vcn->fw_shared.mem_size; 1242 uint64_t fw_log_gpu_addr = vcn->fw_shared.gpu_addr + vcn->fw_shared.mem_size; 1243 struct amdgpu_vcn_fwlog *log_buf = fw_log_cpu_addr; 1244 struct amdgpu_fw_shared_fw_logging *fw_log = vcn->fw_shared.cpu_addr 1245 + vcn->fw_shared.log_offset; 1246 *flag |= cpu_to_le32(AMDGPU_VCN_FW_LOGGING_FLAG); 1247 fw_log->is_enabled = 1; 1248 fw_log->addr_lo = cpu_to_le32(fw_log_gpu_addr & 0xFFFFFFFF); 1249 fw_log->addr_hi = cpu_to_le32(fw_log_gpu_addr >> 32); 1250 fw_log->size = cpu_to_le32(AMDGPU_VCNFW_LOG_SIZE); 1251 1252 log_buf->header_size = sizeof(struct amdgpu_vcn_fwlog); 1253 log_buf->buffer_size = AMDGPU_VCNFW_LOG_SIZE; 1254 log_buf->rptr = log_buf->header_size; 1255 log_buf->wptr = log_buf->header_size; 1256 log_buf->wrapped = 0; 1257 #endif 1258 } 1259 1260 int amdgpu_vcn_process_poison_irq(struct amdgpu_device *adev, 1261 struct amdgpu_irq_src *source, 1262 struct amdgpu_iv_entry *entry) 1263 { 1264 struct ras_common_if *ras_if = adev->vcn.ras_if; 1265 struct ras_dispatch_if ih_data = { 1266 .entry = entry, 1267 }; 1268 1269 if (!ras_if) 1270 return 0; 1271 1272 if (!amdgpu_sriov_vf(adev)) { 1273 ih_data.head = *ras_if; 1274 amdgpu_ras_interrupt_dispatch(adev, &ih_data); 1275 } else { 1276 if (adev->virt.ops && adev->virt.ops->ras_poison_handler) 1277 adev->virt.ops->ras_poison_handler(adev, ras_if->block); 1278 else 1279 dev_warn(adev->dev, 1280 "No ras_poison_handler interface in SRIOV for VCN!\n"); 1281 } 1282 1283 return 0; 1284 } 1285 1286 int amdgpu_vcn_ras_late_init(struct amdgpu_device *adev, struct ras_common_if *ras_block) 1287 { 1288 int r, i; 1289 1290 r = amdgpu_ras_block_late_init(adev, ras_block); 1291 if (r) 1292 return r; 1293 1294 if (amdgpu_ras_is_supported(adev, ras_block->block)) { 1295 for (i = 0; i < adev->vcn.num_vcn_inst; i++) { 1296 if (adev->vcn.harvest_config & (1 << i) || 1297 !adev->vcn.inst[i].ras_poison_irq.funcs) 1298 continue; 1299 1300 r = amdgpu_irq_get(adev, &adev->vcn.inst[i].ras_poison_irq, 0); 1301 if (r) 1302 goto late_fini; 1303 } 1304 } 1305 return 0; 1306 1307 late_fini: 1308 amdgpu_ras_block_late_fini(adev, ras_block); 1309 return r; 1310 } 1311 1312 int amdgpu_vcn_ras_sw_init(struct amdgpu_device *adev) 1313 { 1314 int err; 1315 struct amdgpu_vcn_ras *ras; 1316 1317 if (!adev->vcn.ras) 1318 return 0; 1319 1320 ras = adev->vcn.ras; 1321 err = amdgpu_ras_register_ras_block(adev, &ras->ras_block); 1322 if (err) { 1323 dev_err(adev->dev, "Failed to register vcn ras block!\n"); 1324 return err; 1325 } 1326 1327 strcpy(ras->ras_block.ras_comm.name, "vcn"); 1328 ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__VCN; 1329 ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__POISON; 1330 adev->vcn.ras_if = &ras->ras_block.ras_comm; 1331 1332 if (!ras->ras_block.ras_late_init) 1333 ras->ras_block.ras_late_init = amdgpu_vcn_ras_late_init; 1334 1335 return 0; 1336 } 1337 1338 int amdgpu_vcn_psp_update_sram(struct amdgpu_device *adev, int inst_idx, 1339 enum AMDGPU_UCODE_ID ucode_id) 1340 { 1341 struct amdgpu_firmware_info ucode = { 1342 .ucode_id = (ucode_id ? ucode_id : 1343 (inst_idx ? AMDGPU_UCODE_ID_VCN1_RAM : 1344 AMDGPU_UCODE_ID_VCN0_RAM)), 1345 .mc_addr = adev->vcn.inst[inst_idx].dpg_sram_gpu_addr, 1346 .ucode_size = ((uintptr_t)adev->vcn.inst[inst_idx].dpg_sram_curr_addr - 1347 (uintptr_t)adev->vcn.inst[inst_idx].dpg_sram_cpu_addr), 1348 }; 1349 1350 return psp_execute_ip_fw_load(&adev->psp, &ucode); 1351 } 1352 1353 static ssize_t amdgpu_get_vcn_reset_mask(struct device *dev, 1354 struct device_attribute *attr, 1355 char *buf) 1356 { 1357 struct drm_device *ddev = dev_get_drvdata(dev); 1358 struct amdgpu_device *adev = drm_to_adev(ddev); 1359 1360 if (!adev) 1361 return -ENODEV; 1362 1363 return amdgpu_show_reset_mask(buf, adev->vcn.supported_reset); 1364 } 1365 1366 static DEVICE_ATTR(vcn_reset_mask, 0444, 1367 amdgpu_get_vcn_reset_mask, NULL); 1368 1369 int amdgpu_vcn_sysfs_reset_mask_init(struct amdgpu_device *adev) 1370 { 1371 int r = 0; 1372 1373 if (adev->vcn.num_vcn_inst) { 1374 r = device_create_file(adev->dev, &dev_attr_vcn_reset_mask); 1375 if (r) 1376 return r; 1377 } 1378 1379 return r; 1380 } 1381 1382 void amdgpu_vcn_sysfs_reset_mask_fini(struct amdgpu_device *adev) 1383 { 1384 if (adev->dev->kobj.sd) { 1385 if (adev->vcn.num_vcn_inst) 1386 device_remove_file(adev->dev, &dev_attr_vcn_reset_mask); 1387 } 1388 } 1389 1390 /* 1391 * debugfs to enable/disable vcn job submission to specific core or 1392 * instance. It is created only if the queue type is unified. 1393 */ 1394 #if defined(CONFIG_DEBUG_FS) 1395 static int amdgpu_debugfs_vcn_sched_mask_set(void *data, u64 val) 1396 { 1397 struct amdgpu_device *adev = (struct amdgpu_device *)data; 1398 u32 i; 1399 u64 mask; 1400 struct amdgpu_ring *ring; 1401 1402 if (!adev) 1403 return -ENODEV; 1404 1405 mask = (1ULL << adev->vcn.num_vcn_inst) - 1; 1406 if ((val & mask) == 0) 1407 return -EINVAL; 1408 for (i = 0; i < adev->vcn.num_vcn_inst; ++i) { 1409 ring = &adev->vcn.inst[i].ring_enc[0]; 1410 if (val & (1ULL << i)) 1411 ring->sched.ready = true; 1412 else 1413 ring->sched.ready = false; 1414 } 1415 /* publish sched.ready flag update effective immediately across smp */ 1416 smp_rmb(); 1417 return 0; 1418 } 1419 1420 static int amdgpu_debugfs_vcn_sched_mask_get(void *data, u64 *val) 1421 { 1422 struct amdgpu_device *adev = (struct amdgpu_device *)data; 1423 u32 i; 1424 u64 mask = 0; 1425 struct amdgpu_ring *ring; 1426 1427 if (!adev) 1428 return -ENODEV; 1429 for (i = 0; i < adev->vcn.num_vcn_inst; ++i) { 1430 ring = &adev->vcn.inst[i].ring_enc[0]; 1431 if (ring->sched.ready) 1432 mask |= 1ULL << i; 1433 } 1434 *val = mask; 1435 return 0; 1436 } 1437 1438 DEFINE_DEBUGFS_ATTRIBUTE(amdgpu_debugfs_vcn_sched_mask_fops, 1439 amdgpu_debugfs_vcn_sched_mask_get, 1440 amdgpu_debugfs_vcn_sched_mask_set, "%llx\n"); 1441 #endif 1442 1443 void amdgpu_debugfs_vcn_sched_mask_init(struct amdgpu_device *adev) 1444 { 1445 #if defined(CONFIG_DEBUG_FS) 1446 struct drm_minor *minor = adev_to_drm(adev)->primary; 1447 struct dentry *root = minor->debugfs_root; 1448 char name[32]; 1449 1450 if (adev->vcn.num_vcn_inst <= 1 || !adev->vcn.inst[0].using_unified_queue) 1451 return; 1452 sprintf(name, "amdgpu_vcn_sched_mask"); 1453 debugfs_create_file(name, 0600, root, adev, 1454 &amdgpu_debugfs_vcn_sched_mask_fops); 1455 #endif 1456 } 1457 1458 /** 1459 * vcn_set_powergating_state - set VCN block powergating state 1460 * 1461 * @ip_block: amdgpu_ip_block pointer 1462 * @state: power gating state 1463 * 1464 * Set VCN block powergating state 1465 */ 1466 int vcn_set_powergating_state(struct amdgpu_ip_block *ip_block, 1467 enum amd_powergating_state state) 1468 { 1469 struct amdgpu_device *adev = ip_block->adev; 1470 int ret = 0, i; 1471 1472 for (i = 0; i < adev->vcn.num_vcn_inst; ++i) { 1473 struct amdgpu_vcn_inst *vinst = &adev->vcn.inst[i]; 1474 1475 ret |= vinst->set_pg_state(vinst, state); 1476 } 1477 1478 return ret; 1479 } 1480 1481 /** 1482 * amdgpu_vcn_reset_engine - Reset a specific VCN engine 1483 * @adev: Pointer to the AMDGPU device 1484 * @instance_id: VCN engine instance to reset 1485 * 1486 * Returns: 0 on success, or a negative error code on failure. 1487 */ 1488 static int amdgpu_vcn_reset_engine(struct amdgpu_device *adev, 1489 uint32_t instance_id) 1490 { 1491 struct amdgpu_vcn_inst *vinst = &adev->vcn.inst[instance_id]; 1492 int r, i; 1493 1494 mutex_lock(&vinst->engine_reset_mutex); 1495 /* Stop the scheduler's work queue for the dec and enc rings if they are running. 1496 * This ensures that no new tasks are submitted to the queues while 1497 * the reset is in progress. 1498 */ 1499 drm_sched_wqueue_stop(&vinst->ring_dec.sched); 1500 for (i = 0; i < vinst->num_enc_rings; i++) 1501 drm_sched_wqueue_stop(&vinst->ring_enc[i].sched); 1502 1503 /* Perform the VCN reset for the specified instance */ 1504 r = vinst->reset(vinst); 1505 if (r) 1506 goto unlock; 1507 r = amdgpu_ring_test_ring(&vinst->ring_dec); 1508 if (r) 1509 goto unlock; 1510 for (i = 0; i < vinst->num_enc_rings; i++) { 1511 r = amdgpu_ring_test_ring(&vinst->ring_enc[i]); 1512 if (r) 1513 goto unlock; 1514 } 1515 amdgpu_fence_driver_force_completion(&vinst->ring_dec); 1516 for (i = 0; i < vinst->num_enc_rings; i++) 1517 amdgpu_fence_driver_force_completion(&vinst->ring_enc[i]); 1518 1519 /* Restart the scheduler's work queue for the dec and enc rings 1520 * if they were stopped by this function. This allows new tasks 1521 * to be submitted to the queues after the reset is complete. 1522 */ 1523 drm_sched_wqueue_start(&vinst->ring_dec.sched); 1524 for (i = 0; i < vinst->num_enc_rings; i++) 1525 drm_sched_wqueue_start(&vinst->ring_enc[i].sched); 1526 1527 unlock: 1528 mutex_unlock(&vinst->engine_reset_mutex); 1529 1530 return r; 1531 } 1532 1533 /** 1534 * amdgpu_vcn_ring_reset - Reset a VCN ring 1535 * @ring: ring to reset 1536 * @vmid: vmid of guilty job 1537 * @timedout_fence: fence of timed out job 1538 * 1539 * This helper is for VCN blocks without unified queues because 1540 * resetting the engine resets all queues in that case. With 1541 * unified queues we have one queue per engine. 1542 * Returns: 0 on success, or a negative error code on failure. 1543 */ 1544 int amdgpu_vcn_ring_reset(struct amdgpu_ring *ring, 1545 unsigned int vmid, 1546 struct amdgpu_fence *timedout_fence) 1547 { 1548 struct amdgpu_device *adev = ring->adev; 1549 1550 if (adev->vcn.inst[ring->me].using_unified_queue) 1551 return -EINVAL; 1552 1553 return amdgpu_vcn_reset_engine(adev, ring->me); 1554 } 1555 1556 int amdgpu_vcn_reg_dump_init(struct amdgpu_device *adev, 1557 const struct amdgpu_hwip_reg_entry *reg, u32 count) 1558 { 1559 adev->vcn.ip_dump = kcalloc(adev->vcn.num_vcn_inst * count, 1560 sizeof(uint32_t), GFP_KERNEL); 1561 if (!adev->vcn.ip_dump) 1562 return -ENOMEM; 1563 adev->vcn.reg_list = reg; 1564 adev->vcn.reg_count = count; 1565 1566 return 0; 1567 } 1568 1569 static void amdgpu_vcn_reg_dump_fini(struct amdgpu_device *adev) 1570 { 1571 kfree(adev->vcn.ip_dump); 1572 adev->vcn.ip_dump = NULL; 1573 adev->vcn.reg_list = NULL; 1574 adev->vcn.reg_count = 0; 1575 } 1576 1577 void amdgpu_vcn_dump_ip_state(struct amdgpu_ip_block *ip_block) 1578 { 1579 struct amdgpu_device *adev = ip_block->adev; 1580 int i, j; 1581 bool is_powered; 1582 u32 inst_off; 1583 1584 if (!adev->vcn.ip_dump) 1585 return; 1586 1587 for (i = 0; i < adev->vcn.num_vcn_inst; i++) { 1588 if (adev->vcn.harvest_config & (1 << i)) 1589 continue; 1590 1591 inst_off = i * adev->vcn.reg_count; 1592 /* mmUVD_POWER_STATUS is always readable and is the first in reg_list */ 1593 adev->vcn.ip_dump[inst_off] = 1594 RREG32(SOC15_REG_ENTRY_OFFSET_INST(adev->vcn.reg_list[0], i)); 1595 is_powered = (adev->vcn.ip_dump[inst_off] & 1596 UVD_POWER_STATUS__UVD_POWER_STATUS_TILES_OFF) != 1597 UVD_POWER_STATUS__UVD_POWER_STATUS_TILES_OFF; 1598 1599 if (is_powered) 1600 for (j = 1; j < adev->vcn.reg_count; j++) 1601 adev->vcn.ip_dump[inst_off + j] = 1602 RREG32(SOC15_REG_ENTRY_OFFSET_INST(adev->vcn.reg_list[j], i)); 1603 } 1604 } 1605 1606 void amdgpu_vcn_print_ip_state(struct amdgpu_ip_block *ip_block, struct drm_printer *p) 1607 { 1608 struct amdgpu_device *adev = ip_block->adev; 1609 int i, j; 1610 bool is_powered; 1611 u32 inst_off; 1612 1613 if (!adev->vcn.ip_dump) 1614 return; 1615 1616 drm_printf(p, "num_instances:%d\n", adev->vcn.num_vcn_inst); 1617 for (i = 0; i < adev->vcn.num_vcn_inst; i++) { 1618 if (adev->vcn.harvest_config & (1 << i)) { 1619 drm_printf(p, "\nHarvested Instance:VCN%d Skipping dump\n", i); 1620 continue; 1621 } 1622 1623 inst_off = i * adev->vcn.reg_count; 1624 is_powered = (adev->vcn.ip_dump[inst_off] & 1625 UVD_POWER_STATUS__UVD_POWER_STATUS_TILES_OFF) != 1626 UVD_POWER_STATUS__UVD_POWER_STATUS_TILES_OFF; 1627 1628 if (is_powered) { 1629 drm_printf(p, "\nActive Instance:VCN%d\n", i); 1630 for (j = 0; j < adev->vcn.reg_count; j++) 1631 drm_printf(p, "%-50s \t 0x%08x\n", adev->vcn.reg_list[j].reg_name, 1632 adev->vcn.ip_dump[inst_off + j]); 1633 } else { 1634 drm_printf(p, "\nInactive Instance:VCN%d\n", i); 1635 } 1636 } 1637 } 1638