1 /* 2 * Copyright 2019 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/firmware.h> 24 #include <linux/pci.h> 25 #include "amdgpu.h" 26 #include "amdgpu_atomfirmware.h" 27 #include "gmc_v10_0.h" 28 29 #include "hdp/hdp_5_0_0_offset.h" 30 #include "hdp/hdp_5_0_0_sh_mask.h" 31 #include "gc/gc_10_1_0_sh_mask.h" 32 #include "mmhub/mmhub_2_0_0_sh_mask.h" 33 #include "athub/athub_2_0_0_sh_mask.h" 34 #include "athub/athub_2_0_0_offset.h" 35 #include "dcn/dcn_2_0_0_offset.h" 36 #include "dcn/dcn_2_0_0_sh_mask.h" 37 #include "oss/osssys_5_0_0_offset.h" 38 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h" 39 #include "navi10_enum.h" 40 41 #include "soc15.h" 42 #include "soc15d.h" 43 #include "soc15_common.h" 44 45 #include "nbio_v2_3.h" 46 47 #include "gfxhub_v2_0.h" 48 #include "mmhub_v2_0.h" 49 #include "athub_v2_0.h" 50 /* XXX Move this macro to navi10 header file, which is like vid.h for VI.*/ 51 #define AMDGPU_NUM_OF_VMIDS 8 52 53 #if 0 54 static const struct soc15_reg_golden golden_settings_navi10_hdp[] = 55 { 56 /* TODO add golden setting for hdp */ 57 }; 58 #endif 59 60 static int 61 gmc_v10_0_vm_fault_interrupt_state(struct amdgpu_device *adev, 62 struct amdgpu_irq_src *src, unsigned type, 63 enum amdgpu_interrupt_state state) 64 { 65 struct amdgpu_vmhub *hub; 66 u32 tmp, reg, bits[AMDGPU_MAX_VMHUBS], i; 67 68 bits[AMDGPU_GFXHUB_0] = GCVM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 69 GCVM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 70 GCVM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 71 GCVM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 72 GCVM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 73 GCVM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 74 GCVM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK; 75 76 bits[AMDGPU_MMHUB_0] = MMVM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 77 MMVM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 78 MMVM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 79 MMVM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 80 MMVM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 81 MMVM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 82 MMVM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK; 83 84 switch (state) { 85 case AMDGPU_IRQ_STATE_DISABLE: 86 /* MM HUB */ 87 hub = &adev->vmhub[AMDGPU_MMHUB_0]; 88 for (i = 0; i < 16; i++) { 89 reg = hub->vm_context0_cntl + i; 90 tmp = RREG32(reg); 91 tmp &= ~bits[AMDGPU_MMHUB_0]; 92 WREG32(reg, tmp); 93 } 94 95 /* GFX HUB */ 96 hub = &adev->vmhub[AMDGPU_GFXHUB_0]; 97 for (i = 0; i < 16; i++) { 98 reg = hub->vm_context0_cntl + i; 99 tmp = RREG32(reg); 100 tmp &= ~bits[AMDGPU_GFXHUB_0]; 101 WREG32(reg, tmp); 102 } 103 break; 104 case AMDGPU_IRQ_STATE_ENABLE: 105 /* MM HUB */ 106 hub = &adev->vmhub[AMDGPU_MMHUB_0]; 107 for (i = 0; i < 16; i++) { 108 reg = hub->vm_context0_cntl + i; 109 tmp = RREG32(reg); 110 tmp |= bits[AMDGPU_MMHUB_0]; 111 WREG32(reg, tmp); 112 } 113 114 /* GFX HUB */ 115 hub = &adev->vmhub[AMDGPU_GFXHUB_0]; 116 for (i = 0; i < 16; i++) { 117 reg = hub->vm_context0_cntl + i; 118 tmp = RREG32(reg); 119 tmp |= bits[AMDGPU_GFXHUB_0]; 120 WREG32(reg, tmp); 121 } 122 break; 123 default: 124 break; 125 } 126 127 return 0; 128 } 129 130 static int gmc_v10_0_process_interrupt(struct amdgpu_device *adev, 131 struct amdgpu_irq_src *source, 132 struct amdgpu_iv_entry *entry) 133 { 134 struct amdgpu_vmhub *hub = &adev->vmhub[entry->vmid_src]; 135 uint32_t status = 0; 136 u64 addr; 137 138 addr = (u64)entry->src_data[0] << 12; 139 addr |= ((u64)entry->src_data[1] & 0xf) << 44; 140 141 if (!amdgpu_sriov_vf(adev)) { 142 /* 143 * Issue a dummy read to wait for the status register to 144 * be updated to avoid reading an incorrect value due to 145 * the new fast GRBM interface. 146 */ 147 if (entry->vmid_src == AMDGPU_GFXHUB_0) 148 RREG32(hub->vm_l2_pro_fault_status); 149 150 status = RREG32(hub->vm_l2_pro_fault_status); 151 WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1); 152 } 153 154 if (printk_ratelimit()) { 155 struct amdgpu_task_info task_info; 156 157 memset(&task_info, 0, sizeof(struct amdgpu_task_info)); 158 amdgpu_vm_get_task_info(adev, entry->pasid, &task_info); 159 160 dev_err(adev->dev, 161 "[%s] page fault (src_id:%u ring:%u vmid:%u pasid:%u, " 162 "for process %s pid %d thread %s pid %d)\n", 163 entry->vmid_src ? "mmhub" : "gfxhub", 164 entry->src_id, entry->ring_id, entry->vmid, 165 entry->pasid, task_info.process_name, task_info.tgid, 166 task_info.task_name, task_info.pid); 167 dev_err(adev->dev, " in page starting at address 0x%016llx from client %d\n", 168 addr, entry->client_id); 169 if (!amdgpu_sriov_vf(adev)) { 170 dev_err(adev->dev, 171 "GCVM_L2_PROTECTION_FAULT_STATUS:0x%08X\n", 172 status); 173 dev_err(adev->dev, "\t MORE_FAULTS: 0x%lx\n", 174 REG_GET_FIELD(status, 175 GCVM_L2_PROTECTION_FAULT_STATUS, MORE_FAULTS)); 176 dev_err(adev->dev, "\t WALKER_ERROR: 0x%lx\n", 177 REG_GET_FIELD(status, 178 GCVM_L2_PROTECTION_FAULT_STATUS, WALKER_ERROR)); 179 dev_err(adev->dev, "\t PERMISSION_FAULTS: 0x%lx\n", 180 REG_GET_FIELD(status, 181 GCVM_L2_PROTECTION_FAULT_STATUS, PERMISSION_FAULTS)); 182 dev_err(adev->dev, "\t MAPPING_ERROR: 0x%lx\n", 183 REG_GET_FIELD(status, 184 GCVM_L2_PROTECTION_FAULT_STATUS, MAPPING_ERROR)); 185 dev_err(adev->dev, "\t RW: 0x%lx\n", 186 REG_GET_FIELD(status, 187 GCVM_L2_PROTECTION_FAULT_STATUS, RW)); 188 } 189 } 190 191 return 0; 192 } 193 194 static const struct amdgpu_irq_src_funcs gmc_v10_0_irq_funcs = { 195 .set = gmc_v10_0_vm_fault_interrupt_state, 196 .process = gmc_v10_0_process_interrupt, 197 }; 198 199 static void gmc_v10_0_set_irq_funcs(struct amdgpu_device *adev) 200 { 201 adev->gmc.vm_fault.num_types = 1; 202 adev->gmc.vm_fault.funcs = &gmc_v10_0_irq_funcs; 203 } 204 205 static uint32_t gmc_v10_0_get_invalidate_req(unsigned int vmid, 206 uint32_t flush_type) 207 { 208 u32 req = 0; 209 210 /* invalidate using legacy mode on vmid*/ 211 req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, 212 PER_VMID_INVALIDATE_REQ, 1 << vmid); 213 req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, FLUSH_TYPE, flush_type); 214 req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PTES, 1); 215 req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE0, 1); 216 req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE1, 1); 217 req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE2, 1); 218 req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, INVALIDATE_L1_PTES, 1); 219 req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, 220 CLEAR_PROTECTION_FAULT_STATUS_ADDR, 0); 221 222 return req; 223 } 224 225 /** 226 * gmc_v10_0_use_invalidate_semaphore - judge whether to use semaphore 227 * 228 * @adev: amdgpu_device pointer 229 * @vmhub: vmhub type 230 * 231 */ 232 static bool gmc_v10_0_use_invalidate_semaphore(struct amdgpu_device *adev, 233 uint32_t vmhub) 234 { 235 return ((vmhub == AMDGPU_MMHUB_0 || 236 vmhub == AMDGPU_MMHUB_1) && 237 (!amdgpu_sriov_vf(adev))); 238 } 239 240 static bool gmc_v10_0_get_atc_vmid_pasid_mapping_info( 241 struct amdgpu_device *adev, 242 uint8_t vmid, uint16_t *p_pasid) 243 { 244 uint32_t value; 245 246 value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) 247 + vmid); 248 *p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK; 249 250 return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK); 251 } 252 253 /* 254 * GART 255 * VMID 0 is the physical GPU addresses as used by the kernel. 256 * VMIDs 1-15 are used for userspace clients and are handled 257 * by the amdgpu vm/hsa code. 258 */ 259 260 static void gmc_v10_0_flush_vm_hub(struct amdgpu_device *adev, uint32_t vmid, 261 unsigned int vmhub, uint32_t flush_type) 262 { 263 bool use_semaphore = gmc_v10_0_use_invalidate_semaphore(adev, vmhub); 264 struct amdgpu_vmhub *hub = &adev->vmhub[vmhub]; 265 u32 tmp = gmc_v10_0_get_invalidate_req(vmid, flush_type); 266 /* Use register 17 for GART */ 267 const unsigned eng = 17; 268 unsigned int i; 269 270 spin_lock(&adev->gmc.invalidate_lock); 271 /* 272 * It may lose gpuvm invalidate acknowldege state across power-gating 273 * off cycle, add semaphore acquire before invalidation and semaphore 274 * release after invalidation to avoid entering power gated state 275 * to WA the Issue 276 */ 277 278 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 279 if (use_semaphore) { 280 for (i = 0; i < adev->usec_timeout; i++) { 281 /* a read return value of 1 means semaphore acuqire */ 282 tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_sem + eng); 283 if (tmp & 0x1) 284 break; 285 udelay(1); 286 } 287 288 if (i >= adev->usec_timeout) 289 DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n"); 290 } 291 292 WREG32_NO_KIQ(hub->vm_inv_eng0_req + eng, tmp); 293 294 /* 295 * Issue a dummy read to wait for the ACK register to be cleared 296 * to avoid a false ACK due to the new fast GRBM interface. 297 */ 298 if (vmhub == AMDGPU_GFXHUB_0) 299 RREG32_NO_KIQ(hub->vm_inv_eng0_req + eng); 300 301 /* Wait for ACK with a delay.*/ 302 for (i = 0; i < adev->usec_timeout; i++) { 303 tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_ack + eng); 304 tmp &= 1 << vmid; 305 if (tmp) 306 break; 307 308 udelay(1); 309 } 310 311 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 312 if (use_semaphore) 313 /* 314 * add semaphore release after invalidation, 315 * write with 0 means semaphore release 316 */ 317 WREG32_NO_KIQ(hub->vm_inv_eng0_sem + eng, 0); 318 319 spin_unlock(&adev->gmc.invalidate_lock); 320 321 if (i < adev->usec_timeout) 322 return; 323 324 DRM_ERROR("Timeout waiting for VM flush ACK!\n"); 325 } 326 327 /** 328 * gmc_v10_0_flush_gpu_tlb - gart tlb flush callback 329 * 330 * @adev: amdgpu_device pointer 331 * @vmid: vm instance to flush 332 * 333 * Flush the TLB for the requested page table. 334 */ 335 static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, 336 uint32_t vmhub, uint32_t flush_type) 337 { 338 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 339 struct dma_fence *fence; 340 struct amdgpu_job *job; 341 342 int r; 343 344 /* flush hdp cache */ 345 adev->nbio.funcs->hdp_flush(adev, NULL); 346 347 mutex_lock(&adev->mman.gtt_window_lock); 348 349 if (vmhub == AMDGPU_MMHUB_0) { 350 gmc_v10_0_flush_vm_hub(adev, vmid, AMDGPU_MMHUB_0, 0); 351 mutex_unlock(&adev->mman.gtt_window_lock); 352 return; 353 } 354 355 BUG_ON(vmhub != AMDGPU_GFXHUB_0); 356 357 if (!adev->mman.buffer_funcs_enabled || 358 !adev->ib_pool_ready || 359 adev->in_gpu_reset || 360 ring->sched.ready == false) { 361 gmc_v10_0_flush_vm_hub(adev, vmid, AMDGPU_GFXHUB_0, 0); 362 mutex_unlock(&adev->mman.gtt_window_lock); 363 return; 364 } 365 366 /* The SDMA on Navi has a bug which can theoretically result in memory 367 * corruption if an invalidation happens at the same time as an VA 368 * translation. Avoid this by doing the invalidation from the SDMA 369 * itself. 370 */ 371 r = amdgpu_job_alloc_with_ib(adev, 16 * 4, &job); 372 if (r) 373 goto error_alloc; 374 375 job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo); 376 job->vm_needs_flush = true; 377 job->ibs->ptr[job->ibs->length_dw++] = ring->funcs->nop; 378 amdgpu_ring_pad_ib(ring, &job->ibs[0]); 379 r = amdgpu_job_submit(job, &adev->mman.entity, 380 AMDGPU_FENCE_OWNER_UNDEFINED, &fence); 381 if (r) 382 goto error_submit; 383 384 mutex_unlock(&adev->mman.gtt_window_lock); 385 386 dma_fence_wait(fence, false); 387 dma_fence_put(fence); 388 389 return; 390 391 error_submit: 392 amdgpu_job_free(job); 393 394 error_alloc: 395 mutex_unlock(&adev->mman.gtt_window_lock); 396 DRM_ERROR("Error flushing GPU TLB using the SDMA (%d)!\n", r); 397 } 398 399 /** 400 * gmc_v10_0_flush_gpu_tlb_pasid - tlb flush via pasid 401 * 402 * @adev: amdgpu_device pointer 403 * @pasid: pasid to be flush 404 * 405 * Flush the TLB for the requested pasid. 406 */ 407 static int gmc_v10_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, 408 uint16_t pasid, uint32_t flush_type, 409 bool all_hub) 410 { 411 int vmid, i; 412 signed long r; 413 uint32_t seq; 414 uint16_t queried_pasid; 415 bool ret; 416 struct amdgpu_ring *ring = &adev->gfx.kiq.ring; 417 struct amdgpu_kiq *kiq = &adev->gfx.kiq; 418 419 if (amdgpu_emu_mode == 0 && ring->sched.ready) { 420 spin_lock(&adev->gfx.kiq.ring_lock); 421 /* 2 dwords flush + 8 dwords fence */ 422 amdgpu_ring_alloc(ring, kiq->pmf->invalidate_tlbs_size + 8); 423 kiq->pmf->kiq_invalidate_tlbs(ring, 424 pasid, flush_type, all_hub); 425 amdgpu_fence_emit_polling(ring, &seq); 426 amdgpu_ring_commit(ring); 427 spin_unlock(&adev->gfx.kiq.ring_lock); 428 r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout); 429 if (r < 1) { 430 DRM_ERROR("wait for kiq fence error: %ld.\n", r); 431 return -ETIME; 432 } 433 434 return 0; 435 } 436 437 for (vmid = 1; vmid < 16; vmid++) { 438 439 ret = gmc_v10_0_get_atc_vmid_pasid_mapping_info(adev, vmid, 440 &queried_pasid); 441 if (ret && queried_pasid == pasid) { 442 if (all_hub) { 443 for (i = 0; i < adev->num_vmhubs; i++) 444 gmc_v10_0_flush_gpu_tlb(adev, vmid, 445 i, 0); 446 } else { 447 gmc_v10_0_flush_gpu_tlb(adev, vmid, 448 AMDGPU_GFXHUB_0, 0); 449 } 450 break; 451 } 452 } 453 454 return 0; 455 } 456 457 static uint64_t gmc_v10_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring, 458 unsigned vmid, uint64_t pd_addr) 459 { 460 bool use_semaphore = gmc_v10_0_use_invalidate_semaphore(ring->adev, ring->funcs->vmhub); 461 struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->funcs->vmhub]; 462 uint32_t req = gmc_v10_0_get_invalidate_req(vmid, 0); 463 unsigned eng = ring->vm_inv_eng; 464 465 /* 466 * It may lose gpuvm invalidate acknowldege state across power-gating 467 * off cycle, add semaphore acquire before invalidation and semaphore 468 * release after invalidation to avoid entering power gated state 469 * to WA the Issue 470 */ 471 472 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 473 if (use_semaphore) 474 /* a read return value of 1 means semaphore acuqire */ 475 amdgpu_ring_emit_reg_wait(ring, 476 hub->vm_inv_eng0_sem + eng, 0x1, 0x1); 477 478 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 + (2 * vmid), 479 lower_32_bits(pd_addr)); 480 481 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + (2 * vmid), 482 upper_32_bits(pd_addr)); 483 484 amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req + eng, 485 hub->vm_inv_eng0_ack + eng, 486 req, 1 << vmid); 487 488 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 489 if (use_semaphore) 490 /* 491 * add semaphore release after invalidation, 492 * write with 0 means semaphore release 493 */ 494 amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem + eng, 0); 495 496 return pd_addr; 497 } 498 499 static void gmc_v10_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid, 500 unsigned pasid) 501 { 502 struct amdgpu_device *adev = ring->adev; 503 uint32_t reg; 504 505 if (ring->funcs->vmhub == AMDGPU_GFXHUB_0) 506 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid; 507 else 508 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid; 509 510 amdgpu_ring_emit_wreg(ring, reg, pasid); 511 } 512 513 /* 514 * PTE format on NAVI 10: 515 * 63:59 reserved 516 * 58:57 reserved 517 * 56 F 518 * 55 L 519 * 54 reserved 520 * 53:52 SW 521 * 51 T 522 * 50:48 mtype 523 * 47:12 4k physical page base address 524 * 11:7 fragment 525 * 6 write 526 * 5 read 527 * 4 exe 528 * 3 Z 529 * 2 snooped 530 * 1 system 531 * 0 valid 532 * 533 * PDE format on NAVI 10: 534 * 63:59 block fragment size 535 * 58:55 reserved 536 * 54 P 537 * 53:48 reserved 538 * 47:6 physical base address of PD or PTE 539 * 5:3 reserved 540 * 2 C 541 * 1 system 542 * 0 valid 543 */ 544 545 static uint64_t gmc_v10_0_map_mtype(struct amdgpu_device *adev, uint32_t flags) 546 { 547 switch (flags) { 548 case AMDGPU_VM_MTYPE_DEFAULT: 549 return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC); 550 case AMDGPU_VM_MTYPE_NC: 551 return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC); 552 case AMDGPU_VM_MTYPE_WC: 553 return AMDGPU_PTE_MTYPE_NV10(MTYPE_WC); 554 case AMDGPU_VM_MTYPE_CC: 555 return AMDGPU_PTE_MTYPE_NV10(MTYPE_CC); 556 case AMDGPU_VM_MTYPE_UC: 557 return AMDGPU_PTE_MTYPE_NV10(MTYPE_UC); 558 default: 559 return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC); 560 } 561 } 562 563 static void gmc_v10_0_get_vm_pde(struct amdgpu_device *adev, int level, 564 uint64_t *addr, uint64_t *flags) 565 { 566 if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM)) 567 *addr = adev->vm_manager.vram_base_offset + *addr - 568 adev->gmc.vram_start; 569 BUG_ON(*addr & 0xFFFF00000000003FULL); 570 571 if (!adev->gmc.translate_further) 572 return; 573 574 if (level == AMDGPU_VM_PDB1) { 575 /* Set the block fragment size */ 576 if (!(*flags & AMDGPU_PDE_PTE)) 577 *flags |= AMDGPU_PDE_BFS(0x9); 578 579 } else if (level == AMDGPU_VM_PDB0) { 580 if (*flags & AMDGPU_PDE_PTE) 581 *flags &= ~AMDGPU_PDE_PTE; 582 else 583 *flags |= AMDGPU_PTE_TF; 584 } 585 } 586 587 static void gmc_v10_0_get_vm_pte(struct amdgpu_device *adev, 588 struct amdgpu_bo_va_mapping *mapping, 589 uint64_t *flags) 590 { 591 *flags &= ~AMDGPU_PTE_EXECUTABLE; 592 *flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE; 593 594 *flags &= ~AMDGPU_PTE_MTYPE_NV10_MASK; 595 *flags |= (mapping->flags & AMDGPU_PTE_MTYPE_NV10_MASK); 596 597 if (mapping->flags & AMDGPU_PTE_PRT) { 598 *flags |= AMDGPU_PTE_PRT; 599 *flags |= AMDGPU_PTE_SNOOPED; 600 *flags |= AMDGPU_PTE_LOG; 601 *flags |= AMDGPU_PTE_SYSTEM; 602 *flags &= ~AMDGPU_PTE_VALID; 603 } 604 } 605 606 static const struct amdgpu_gmc_funcs gmc_v10_0_gmc_funcs = { 607 .flush_gpu_tlb = gmc_v10_0_flush_gpu_tlb, 608 .flush_gpu_tlb_pasid = gmc_v10_0_flush_gpu_tlb_pasid, 609 .emit_flush_gpu_tlb = gmc_v10_0_emit_flush_gpu_tlb, 610 .emit_pasid_mapping = gmc_v10_0_emit_pasid_mapping, 611 .map_mtype = gmc_v10_0_map_mtype, 612 .get_vm_pde = gmc_v10_0_get_vm_pde, 613 .get_vm_pte = gmc_v10_0_get_vm_pte 614 }; 615 616 static void gmc_v10_0_set_gmc_funcs(struct amdgpu_device *adev) 617 { 618 if (adev->gmc.gmc_funcs == NULL) 619 adev->gmc.gmc_funcs = &gmc_v10_0_gmc_funcs; 620 } 621 622 static int gmc_v10_0_early_init(void *handle) 623 { 624 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 625 626 gmc_v10_0_set_gmc_funcs(adev); 627 gmc_v10_0_set_irq_funcs(adev); 628 629 adev->gmc.shared_aperture_start = 0x2000000000000000ULL; 630 adev->gmc.shared_aperture_end = 631 adev->gmc.shared_aperture_start + (4ULL << 30) - 1; 632 adev->gmc.private_aperture_start = 0x1000000000000000ULL; 633 adev->gmc.private_aperture_end = 634 adev->gmc.private_aperture_start + (4ULL << 30) - 1; 635 636 return 0; 637 } 638 639 static int gmc_v10_0_late_init(void *handle) 640 { 641 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 642 int r; 643 644 /* 645 * Can't free the stolen VGA memory when it might be used for memory 646 * training again. 647 */ 648 if (!adev->fw_vram_usage.mem_train_support) 649 amdgpu_bo_late_init(adev); 650 651 r = amdgpu_gmc_allocate_vm_inv_eng(adev); 652 if (r) 653 return r; 654 655 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0); 656 } 657 658 static void gmc_v10_0_vram_gtt_location(struct amdgpu_device *adev, 659 struct amdgpu_gmc *mc) 660 { 661 u64 base = 0; 662 663 base = gfxhub_v2_0_get_fb_location(adev); 664 665 amdgpu_gmc_vram_location(adev, &adev->gmc, base); 666 amdgpu_gmc_gart_location(adev, mc); 667 668 /* base offset of vram pages */ 669 adev->vm_manager.vram_base_offset = gfxhub_v2_0_get_mc_fb_offset(adev); 670 } 671 672 /** 673 * gmc_v10_0_mc_init - initialize the memory controller driver params 674 * 675 * @adev: amdgpu_device pointer 676 * 677 * Look up the amount of vram, vram width, and decide how to place 678 * vram and gart within the GPU's physical address space. 679 * Returns 0 for success. 680 */ 681 static int gmc_v10_0_mc_init(struct amdgpu_device *adev) 682 { 683 /* Could aper size report 0 ? */ 684 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0); 685 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0); 686 687 /* size in MB on si */ 688 adev->gmc.mc_vram_size = 689 adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL; 690 adev->gmc.real_vram_size = adev->gmc.mc_vram_size; 691 adev->gmc.visible_vram_size = adev->gmc.aper_size; 692 693 /* In case the PCI BAR is larger than the actual amount of vram */ 694 if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size) 695 adev->gmc.visible_vram_size = adev->gmc.real_vram_size; 696 697 /* set the gart size */ 698 if (amdgpu_gart_size == -1) { 699 switch (adev->asic_type) { 700 case CHIP_NAVI10: 701 case CHIP_NAVI14: 702 case CHIP_NAVI12: 703 default: 704 adev->gmc.gart_size = 512ULL << 20; 705 break; 706 } 707 } else 708 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20; 709 710 gmc_v10_0_vram_gtt_location(adev, &adev->gmc); 711 712 return 0; 713 } 714 715 static int gmc_v10_0_gart_init(struct amdgpu_device *adev) 716 { 717 int r; 718 719 if (adev->gart.bo) { 720 WARN(1, "NAVI10 PCIE GART already initialized\n"); 721 return 0; 722 } 723 724 /* Initialize common gart structure */ 725 r = amdgpu_gart_init(adev); 726 if (r) 727 return r; 728 729 adev->gart.table_size = adev->gart.num_gpu_pages * 8; 730 adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_NV10(MTYPE_UC) | 731 AMDGPU_PTE_EXECUTABLE; 732 733 return amdgpu_gart_table_vram_alloc(adev); 734 } 735 736 static unsigned gmc_v10_0_get_vbios_fb_size(struct amdgpu_device *adev) 737 { 738 u32 d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL); 739 unsigned size; 740 741 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) { 742 size = 9 * 1024 * 1024; /* reserve 8MB for vga emulator and 1 MB for FB */ 743 } else { 744 u32 viewport; 745 u32 pitch; 746 747 viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION); 748 pitch = RREG32_SOC15(DCE, 0, mmHUBPREQ0_DCSURF_SURFACE_PITCH); 749 size = (REG_GET_FIELD(viewport, 750 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) * 751 REG_GET_FIELD(pitch, HUBPREQ0_DCSURF_SURFACE_PITCH, PITCH) * 752 4); 753 } 754 /* return 0 if the pre-OS buffer uses up most of vram */ 755 if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024)) { 756 DRM_ERROR("Warning: pre-OS buffer uses most of vram, \ 757 be aware of gart table overwrite\n"); 758 return 0; 759 } 760 761 return size; 762 } 763 764 765 766 static int gmc_v10_0_sw_init(void *handle) 767 { 768 int r, vram_width = 0, vram_type = 0, vram_vendor = 0; 769 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 770 771 gfxhub_v2_0_init(adev); 772 mmhub_v2_0_init(adev); 773 774 spin_lock_init(&adev->gmc.invalidate_lock); 775 776 r = amdgpu_atomfirmware_get_vram_info(adev, 777 &vram_width, &vram_type, &vram_vendor); 778 if (!amdgpu_emu_mode) 779 adev->gmc.vram_width = vram_width; 780 else 781 adev->gmc.vram_width = 1 * 128; /* numchan * chansize */ 782 783 adev->gmc.vram_type = vram_type; 784 adev->gmc.vram_vendor = vram_vendor; 785 switch (adev->asic_type) { 786 case CHIP_NAVI10: 787 case CHIP_NAVI14: 788 case CHIP_NAVI12: 789 adev->num_vmhubs = 2; 790 /* 791 * To fulfill 4-level page support, 792 * vm size is 256TB (48bit), maximum size of Navi10/Navi14/Navi12, 793 * block size 512 (9bit) 794 */ 795 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 796 break; 797 default: 798 break; 799 } 800 801 /* This interrupt is VMC page fault.*/ 802 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC, 803 VMC_1_0__SRCID__VM_FAULT, 804 &adev->gmc.vm_fault); 805 806 if (r) 807 return r; 808 809 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2, 810 UTCL2_1_0__SRCID__FAULT, 811 &adev->gmc.vm_fault); 812 if (r) 813 return r; 814 815 /* 816 * Set the internal MC address mask This is the max address of the GPU's 817 * internal address space. 818 */ 819 adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */ 820 821 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44)); 822 if (r) { 823 printk(KERN_WARNING "amdgpu: No suitable DMA available.\n"); 824 return r; 825 } 826 827 r = gmc_v10_0_mc_init(adev); 828 if (r) 829 return r; 830 831 adev->gmc.stolen_size = gmc_v10_0_get_vbios_fb_size(adev); 832 833 /* 834 * In dual GPUs scenario, stolen_size is assigned to zero on the 835 * secondary GPU, since there is no pre-OS console using that memory. 836 * Then the bottom region of VRAM was allocated as GTT, unfortunately a 837 * small region of bottom VRAM was encroached by UMC firmware during 838 * GDDR6 BIST training, this cause page fault. 839 * The page fault can be fixed by forcing stolen_size to 3MB, then the 840 * bottom region of VRAM was allocated as stolen memory, GTT corruption 841 * avoid. 842 */ 843 adev->gmc.stolen_size = max(adev->gmc.stolen_size, 844 AMDGPU_STOLEN_BIST_TRAINING_DEFAULT_SIZE); 845 846 /* Memory manager */ 847 r = amdgpu_bo_init(adev); 848 if (r) 849 return r; 850 851 r = gmc_v10_0_gart_init(adev); 852 if (r) 853 return r; 854 855 /* 856 * number of VMs 857 * VMID 0 is reserved for System 858 * amdgpu graphics/compute will use VMIDs 1-7 859 * amdkfd will use VMIDs 8-15 860 */ 861 adev->vm_manager.id_mgr[AMDGPU_GFXHUB_0].num_ids = AMDGPU_NUM_OF_VMIDS; 862 adev->vm_manager.id_mgr[AMDGPU_MMHUB_0].num_ids = AMDGPU_NUM_OF_VMIDS; 863 864 amdgpu_vm_manager_init(adev); 865 866 return 0; 867 } 868 869 /** 870 * gmc_v8_0_gart_fini - vm fini callback 871 * 872 * @adev: amdgpu_device pointer 873 * 874 * Tears down the driver GART/VM setup (CIK). 875 */ 876 static void gmc_v10_0_gart_fini(struct amdgpu_device *adev) 877 { 878 amdgpu_gart_table_vram_free(adev); 879 amdgpu_gart_fini(adev); 880 } 881 882 static int gmc_v10_0_sw_fini(void *handle) 883 { 884 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 885 void *stolen_vga_buf; 886 887 /* 888 * Free the stolen memory if it wasn't already freed in late_init 889 * because of memory training. 890 */ 891 amdgpu_bo_free_kernel(&adev->stolen_vga_memory, NULL, &stolen_vga_buf); 892 893 amdgpu_vm_manager_fini(adev); 894 gmc_v10_0_gart_fini(adev); 895 amdgpu_gem_force_release(adev); 896 amdgpu_bo_fini(adev); 897 898 return 0; 899 } 900 901 static void gmc_v10_0_init_golden_registers(struct amdgpu_device *adev) 902 { 903 switch (adev->asic_type) { 904 case CHIP_NAVI10: 905 case CHIP_NAVI14: 906 case CHIP_NAVI12: 907 break; 908 default: 909 break; 910 } 911 } 912 913 /** 914 * gmc_v10_0_gart_enable - gart enable 915 * 916 * @adev: amdgpu_device pointer 917 */ 918 static int gmc_v10_0_gart_enable(struct amdgpu_device *adev) 919 { 920 int r; 921 bool value; 922 u32 tmp; 923 924 if (adev->gart.bo == NULL) { 925 dev_err(adev->dev, "No VRAM object for PCIE GART.\n"); 926 return -EINVAL; 927 } 928 929 r = amdgpu_gart_table_vram_pin(adev); 930 if (r) 931 return r; 932 933 r = gfxhub_v2_0_gart_enable(adev); 934 if (r) 935 return r; 936 937 r = mmhub_v2_0_gart_enable(adev); 938 if (r) 939 return r; 940 941 tmp = RREG32_SOC15(HDP, 0, mmHDP_MISC_CNTL); 942 tmp |= HDP_MISC_CNTL__FLUSH_INVALIDATE_CACHE_MASK; 943 WREG32_SOC15(HDP, 0, mmHDP_MISC_CNTL, tmp); 944 945 tmp = RREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL); 946 WREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL, tmp); 947 948 /* Flush HDP after it is initialized */ 949 adev->nbio.funcs->hdp_flush(adev, NULL); 950 951 value = (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) ? 952 false : true; 953 954 gfxhub_v2_0_set_fault_enable_default(adev, value); 955 mmhub_v2_0_set_fault_enable_default(adev, value); 956 gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_MMHUB_0, 0); 957 gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_GFXHUB_0, 0); 958 959 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n", 960 (unsigned)(adev->gmc.gart_size >> 20), 961 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo)); 962 963 adev->gart.ready = true; 964 965 return 0; 966 } 967 968 static int gmc_v10_0_hw_init(void *handle) 969 { 970 int r; 971 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 972 973 /* The sequence of these two function calls matters.*/ 974 gmc_v10_0_init_golden_registers(adev); 975 976 r = gmc_v10_0_gart_enable(adev); 977 if (r) 978 return r; 979 980 return 0; 981 } 982 983 /** 984 * gmc_v10_0_gart_disable - gart disable 985 * 986 * @adev: amdgpu_device pointer 987 * 988 * This disables all VM page table. 989 */ 990 static void gmc_v10_0_gart_disable(struct amdgpu_device *adev) 991 { 992 gfxhub_v2_0_gart_disable(adev); 993 mmhub_v2_0_gart_disable(adev); 994 amdgpu_gart_table_vram_unpin(adev); 995 } 996 997 static int gmc_v10_0_hw_fini(void *handle) 998 { 999 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1000 1001 if (amdgpu_sriov_vf(adev)) { 1002 /* full access mode, so don't touch any GMC register */ 1003 DRM_DEBUG("For SRIOV client, shouldn't do anything.\n"); 1004 return 0; 1005 } 1006 1007 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0); 1008 gmc_v10_0_gart_disable(adev); 1009 1010 return 0; 1011 } 1012 1013 static int gmc_v10_0_suspend(void *handle) 1014 { 1015 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1016 1017 gmc_v10_0_hw_fini(adev); 1018 1019 return 0; 1020 } 1021 1022 static int gmc_v10_0_resume(void *handle) 1023 { 1024 int r; 1025 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1026 1027 r = gmc_v10_0_hw_init(adev); 1028 if (r) 1029 return r; 1030 1031 amdgpu_vmid_reset_all(adev); 1032 1033 return 0; 1034 } 1035 1036 static bool gmc_v10_0_is_idle(void *handle) 1037 { 1038 /* MC is always ready in GMC v10.*/ 1039 return true; 1040 } 1041 1042 static int gmc_v10_0_wait_for_idle(void *handle) 1043 { 1044 /* There is no need to wait for MC idle in GMC v10.*/ 1045 return 0; 1046 } 1047 1048 static int gmc_v10_0_soft_reset(void *handle) 1049 { 1050 return 0; 1051 } 1052 1053 static int gmc_v10_0_set_clockgating_state(void *handle, 1054 enum amd_clockgating_state state) 1055 { 1056 int r; 1057 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1058 1059 r = mmhub_v2_0_set_clockgating(adev, state); 1060 if (r) 1061 return r; 1062 1063 return athub_v2_0_set_clockgating(adev, state); 1064 } 1065 1066 static void gmc_v10_0_get_clockgating_state(void *handle, u32 *flags) 1067 { 1068 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1069 1070 mmhub_v2_0_get_clockgating(adev, flags); 1071 1072 athub_v2_0_get_clockgating(adev, flags); 1073 } 1074 1075 static int gmc_v10_0_set_powergating_state(void *handle, 1076 enum amd_powergating_state state) 1077 { 1078 return 0; 1079 } 1080 1081 const struct amd_ip_funcs gmc_v10_0_ip_funcs = { 1082 .name = "gmc_v10_0", 1083 .early_init = gmc_v10_0_early_init, 1084 .late_init = gmc_v10_0_late_init, 1085 .sw_init = gmc_v10_0_sw_init, 1086 .sw_fini = gmc_v10_0_sw_fini, 1087 .hw_init = gmc_v10_0_hw_init, 1088 .hw_fini = gmc_v10_0_hw_fini, 1089 .suspend = gmc_v10_0_suspend, 1090 .resume = gmc_v10_0_resume, 1091 .is_idle = gmc_v10_0_is_idle, 1092 .wait_for_idle = gmc_v10_0_wait_for_idle, 1093 .soft_reset = gmc_v10_0_soft_reset, 1094 .set_clockgating_state = gmc_v10_0_set_clockgating_state, 1095 .set_powergating_state = gmc_v10_0_set_powergating_state, 1096 .get_clockgating_state = gmc_v10_0_get_clockgating_state, 1097 }; 1098 1099 const struct amdgpu_ip_block_version gmc_v10_0_ip_block = 1100 { 1101 .type = AMD_IP_BLOCK_TYPE_GMC, 1102 .major = 10, 1103 .minor = 0, 1104 .rev = 0, 1105 .funcs = &gmc_v10_0_ip_funcs, 1106 }; 1107