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 #include "umc_v8_7.h" 29 30 #include "athub/athub_2_0_0_sh_mask.h" 31 #include "athub/athub_2_0_0_offset.h" 32 #include "dcn/dcn_2_0_0_offset.h" 33 #include "dcn/dcn_2_0_0_sh_mask.h" 34 #include "oss/osssys_5_0_0_offset.h" 35 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h" 36 #include "navi10_enum.h" 37 38 #include "soc15.h" 39 #include "soc15d.h" 40 #include "soc15_common.h" 41 42 #include "nbio_v2_3.h" 43 44 #include "gfxhub_v2_0.h" 45 #include "gfxhub_v2_1.h" 46 #include "mmhub_v2_0.h" 47 #include "mmhub_v2_3.h" 48 #include "athub_v2_0.h" 49 #include "athub_v2_1.h" 50 51 #if 0 52 static const struct soc15_reg_golden golden_settings_navi10_hdp[] = 53 { 54 /* TODO add golden setting for hdp */ 55 }; 56 #endif 57 58 static int gmc_v10_0_ecc_interrupt_state(struct amdgpu_device *adev, 59 struct amdgpu_irq_src *src, 60 unsigned type, 61 enum amdgpu_interrupt_state state) 62 { 63 return 0; 64 } 65 66 static int 67 gmc_v10_0_vm_fault_interrupt_state(struct amdgpu_device *adev, 68 struct amdgpu_irq_src *src, unsigned type, 69 enum amdgpu_interrupt_state state) 70 { 71 switch (state) { 72 case AMDGPU_IRQ_STATE_DISABLE: 73 /* MM HUB */ 74 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB_0, false); 75 /* GFX HUB */ 76 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB_0, false); 77 break; 78 case AMDGPU_IRQ_STATE_ENABLE: 79 /* MM HUB */ 80 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB_0, true); 81 /* GFX HUB */ 82 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB_0, true); 83 break; 84 default: 85 break; 86 } 87 88 return 0; 89 } 90 91 static int gmc_v10_0_process_interrupt(struct amdgpu_device *adev, 92 struct amdgpu_irq_src *source, 93 struct amdgpu_iv_entry *entry) 94 { 95 bool retry_fault = !!(entry->src_data[1] & 0x80); 96 struct amdgpu_vmhub *hub = &adev->vmhub[entry->vmid_src]; 97 struct amdgpu_task_info task_info; 98 uint32_t status = 0; 99 u64 addr; 100 101 addr = (u64)entry->src_data[0] << 12; 102 addr |= ((u64)entry->src_data[1] & 0xf) << 44; 103 104 if (retry_fault) { 105 /* Returning 1 here also prevents sending the IV to the KFD */ 106 107 /* Process it onyl if it's the first fault for this address */ 108 if (entry->ih != &adev->irq.ih_soft && 109 amdgpu_gmc_filter_faults(adev, addr, entry->pasid, 110 entry->timestamp)) 111 return 1; 112 113 /* Delegate it to a different ring if the hardware hasn't 114 * already done it. 115 */ 116 if (entry->ih == &adev->irq.ih) { 117 amdgpu_irq_delegate(adev, entry, 8); 118 return 1; 119 } 120 121 /* Try to handle the recoverable page faults by filling page 122 * tables 123 */ 124 if (amdgpu_vm_handle_fault(adev, entry->pasid, addr)) 125 return 1; 126 } 127 128 if (!amdgpu_sriov_vf(adev)) { 129 /* 130 * Issue a dummy read to wait for the status register to 131 * be updated to avoid reading an incorrect value due to 132 * the new fast GRBM interface. 133 */ 134 if ((entry->vmid_src == AMDGPU_GFXHUB_0) && 135 (adev->asic_type < CHIP_SIENNA_CICHLID)) 136 RREG32(hub->vm_l2_pro_fault_status); 137 138 status = RREG32(hub->vm_l2_pro_fault_status); 139 WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1); 140 } 141 142 if (!printk_ratelimit()) 143 return 0; 144 145 memset(&task_info, 0, sizeof(struct amdgpu_task_info)); 146 amdgpu_vm_get_task_info(adev, entry->pasid, &task_info); 147 148 dev_err(adev->dev, 149 "[%s] page fault (src_id:%u ring:%u vmid:%u pasid:%u, " 150 "for process %s pid %d thread %s pid %d)\n", 151 entry->vmid_src ? "mmhub" : "gfxhub", 152 entry->src_id, entry->ring_id, entry->vmid, 153 entry->pasid, task_info.process_name, task_info.tgid, 154 task_info.task_name, task_info.pid); 155 dev_err(adev->dev, " in page starting at address 0x%016llx from client 0x%x (%s)\n", 156 addr, entry->client_id, 157 soc15_ih_clientid_name[entry->client_id]); 158 159 if (!amdgpu_sriov_vf(adev)) 160 hub->vmhub_funcs->print_l2_protection_fault_status(adev, 161 status); 162 163 return 0; 164 } 165 166 static const struct amdgpu_irq_src_funcs gmc_v10_0_irq_funcs = { 167 .set = gmc_v10_0_vm_fault_interrupt_state, 168 .process = gmc_v10_0_process_interrupt, 169 }; 170 171 static const struct amdgpu_irq_src_funcs gmc_v10_0_ecc_funcs = { 172 .set = gmc_v10_0_ecc_interrupt_state, 173 .process = amdgpu_umc_process_ecc_irq, 174 }; 175 176 static void gmc_v10_0_set_irq_funcs(struct amdgpu_device *adev) 177 { 178 adev->gmc.vm_fault.num_types = 1; 179 adev->gmc.vm_fault.funcs = &gmc_v10_0_irq_funcs; 180 181 if (!amdgpu_sriov_vf(adev)) { 182 adev->gmc.ecc_irq.num_types = 1; 183 adev->gmc.ecc_irq.funcs = &gmc_v10_0_ecc_funcs; 184 } 185 } 186 187 /** 188 * gmc_v10_0_use_invalidate_semaphore - judge whether to use semaphore 189 * 190 * @adev: amdgpu_device pointer 191 * @vmhub: vmhub type 192 * 193 */ 194 static bool gmc_v10_0_use_invalidate_semaphore(struct amdgpu_device *adev, 195 uint32_t vmhub) 196 { 197 return ((vmhub == AMDGPU_MMHUB_0 || 198 vmhub == AMDGPU_MMHUB_1) && 199 (!amdgpu_sriov_vf(adev))); 200 } 201 202 static bool gmc_v10_0_get_atc_vmid_pasid_mapping_info( 203 struct amdgpu_device *adev, 204 uint8_t vmid, uint16_t *p_pasid) 205 { 206 uint32_t value; 207 208 value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) 209 + vmid); 210 *p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK; 211 212 return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK); 213 } 214 215 /* 216 * GART 217 * VMID 0 is the physical GPU addresses as used by the kernel. 218 * VMIDs 1-15 are used for userspace clients and are handled 219 * by the amdgpu vm/hsa code. 220 */ 221 222 static void gmc_v10_0_flush_vm_hub(struct amdgpu_device *adev, uint32_t vmid, 223 unsigned int vmhub, uint32_t flush_type) 224 { 225 bool use_semaphore = gmc_v10_0_use_invalidate_semaphore(adev, vmhub); 226 struct amdgpu_vmhub *hub = &adev->vmhub[vmhub]; 227 u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type); 228 u32 tmp; 229 /* Use register 17 for GART */ 230 const unsigned eng = 17; 231 unsigned int i; 232 unsigned char hub_ip = 0; 233 234 hub_ip = (vmhub == AMDGPU_GFXHUB_0) ? 235 GC_HWIP : MMHUB_HWIP; 236 237 spin_lock(&adev->gmc.invalidate_lock); 238 /* 239 * It may lose gpuvm invalidate acknowldege state across power-gating 240 * off cycle, add semaphore acquire before invalidation and semaphore 241 * release after invalidation to avoid entering power gated state 242 * to WA the Issue 243 */ 244 245 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 246 if (use_semaphore) { 247 for (i = 0; i < adev->usec_timeout; i++) { 248 /* a read return value of 1 means semaphore acuqire */ 249 tmp = RREG32_RLC_NO_KIQ(hub->vm_inv_eng0_sem + 250 hub->eng_distance * eng, hub_ip); 251 252 if (tmp & 0x1) 253 break; 254 udelay(1); 255 } 256 257 if (i >= adev->usec_timeout) 258 DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n"); 259 } 260 261 WREG32_RLC_NO_KIQ(hub->vm_inv_eng0_req + 262 hub->eng_distance * eng, 263 inv_req, hub_ip); 264 265 /* 266 * Issue a dummy read to wait for the ACK register to be cleared 267 * to avoid a false ACK due to the new fast GRBM interface. 268 */ 269 if ((vmhub == AMDGPU_GFXHUB_0) && 270 (adev->asic_type < CHIP_SIENNA_CICHLID)) 271 RREG32_RLC_NO_KIQ(hub->vm_inv_eng0_req + 272 hub->eng_distance * eng, hub_ip); 273 274 /* Wait for ACK with a delay.*/ 275 for (i = 0; i < adev->usec_timeout; i++) { 276 tmp = RREG32_RLC_NO_KIQ(hub->vm_inv_eng0_ack + 277 hub->eng_distance * eng, hub_ip); 278 279 tmp &= 1 << vmid; 280 if (tmp) 281 break; 282 283 udelay(1); 284 } 285 286 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 287 if (use_semaphore) 288 /* 289 * add semaphore release after invalidation, 290 * write with 0 means semaphore release 291 */ 292 WREG32_RLC_NO_KIQ(hub->vm_inv_eng0_sem + 293 hub->eng_distance * eng, 0, hub_ip); 294 295 spin_unlock(&adev->gmc.invalidate_lock); 296 297 if (i < adev->usec_timeout) 298 return; 299 300 DRM_ERROR("Timeout waiting for VM flush hub: %d!\n", vmhub); 301 } 302 303 /** 304 * gmc_v10_0_flush_gpu_tlb - gart tlb flush callback 305 * 306 * @adev: amdgpu_device pointer 307 * @vmid: vm instance to flush 308 * @vmhub: vmhub type 309 * @flush_type: the flush type 310 * 311 * Flush the TLB for the requested page table. 312 */ 313 static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, 314 uint32_t vmhub, uint32_t flush_type) 315 { 316 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 317 struct dma_fence *fence; 318 struct amdgpu_job *job; 319 320 int r; 321 322 /* flush hdp cache */ 323 adev->hdp.funcs->flush_hdp(adev, NULL); 324 325 /* For SRIOV run time, driver shouldn't access the register through MMIO 326 * Directly use kiq to do the vm invalidation instead 327 */ 328 if (adev->gfx.kiq.ring.sched.ready && 329 (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev)) && 330 down_read_trylock(&adev->reset_sem)) { 331 struct amdgpu_vmhub *hub = &adev->vmhub[vmhub]; 332 const unsigned eng = 17; 333 u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type); 334 u32 req = hub->vm_inv_eng0_req + hub->eng_distance * eng; 335 u32 ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng; 336 337 amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, inv_req, 338 1 << vmid); 339 340 up_read(&adev->reset_sem); 341 return; 342 } 343 344 mutex_lock(&adev->mman.gtt_window_lock); 345 346 if (vmhub == AMDGPU_MMHUB_0) { 347 gmc_v10_0_flush_vm_hub(adev, vmid, AMDGPU_MMHUB_0, 0); 348 mutex_unlock(&adev->mman.gtt_window_lock); 349 return; 350 } 351 352 BUG_ON(vmhub != AMDGPU_GFXHUB_0); 353 354 if (!adev->mman.buffer_funcs_enabled || 355 !adev->ib_pool_ready || 356 amdgpu_in_reset(adev) || 357 ring->sched.ready == false) { 358 gmc_v10_0_flush_vm_hub(adev, vmid, AMDGPU_GFXHUB_0, 0); 359 mutex_unlock(&adev->mman.gtt_window_lock); 360 return; 361 } 362 363 /* The SDMA on Navi has a bug which can theoretically result in memory 364 * corruption if an invalidation happens at the same time as an VA 365 * translation. Avoid this by doing the invalidation from the SDMA 366 * itself. 367 */ 368 r = amdgpu_job_alloc_with_ib(adev, 16 * 4, AMDGPU_IB_POOL_IMMEDIATE, 369 &job); 370 if (r) 371 goto error_alloc; 372 373 job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo); 374 job->vm_needs_flush = true; 375 job->ibs->ptr[job->ibs->length_dw++] = ring->funcs->nop; 376 amdgpu_ring_pad_ib(ring, &job->ibs[0]); 377 r = amdgpu_job_submit(job, &adev->mman.entity, 378 AMDGPU_FENCE_OWNER_UNDEFINED, &fence); 379 if (r) 380 goto error_submit; 381 382 mutex_unlock(&adev->mman.gtt_window_lock); 383 384 dma_fence_wait(fence, false); 385 dma_fence_put(fence); 386 387 return; 388 389 error_submit: 390 amdgpu_job_free(job); 391 392 error_alloc: 393 mutex_unlock(&adev->mman.gtt_window_lock); 394 DRM_ERROR("Error flushing GPU TLB using the SDMA (%d)!\n", r); 395 } 396 397 /** 398 * gmc_v10_0_flush_gpu_tlb_pasid - tlb flush via pasid 399 * 400 * @adev: amdgpu_device pointer 401 * @pasid: pasid to be flush 402 * @flush_type: the flush type 403 * @all_hub: Used with PACKET3_INVALIDATE_TLBS_ALL_HUB() 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 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT); 426 if (r) { 427 amdgpu_ring_undo(ring); 428 spin_unlock(&adev->gfx.kiq.ring_lock); 429 return -ETIME; 430 } 431 432 amdgpu_ring_commit(ring); 433 spin_unlock(&adev->gfx.kiq.ring_lock); 434 r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout); 435 if (r < 1) { 436 dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r); 437 return -ETIME; 438 } 439 440 return 0; 441 } 442 443 for (vmid = 1; vmid < AMDGPU_NUM_VMID; vmid++) { 444 445 ret = gmc_v10_0_get_atc_vmid_pasid_mapping_info(adev, vmid, 446 &queried_pasid); 447 if (ret && queried_pasid == pasid) { 448 if (all_hub) { 449 for (i = 0; i < adev->num_vmhubs; i++) 450 gmc_v10_0_flush_gpu_tlb(adev, vmid, 451 i, flush_type); 452 } else { 453 gmc_v10_0_flush_gpu_tlb(adev, vmid, 454 AMDGPU_GFXHUB_0, flush_type); 455 } 456 break; 457 } 458 } 459 460 return 0; 461 } 462 463 static uint64_t gmc_v10_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring, 464 unsigned vmid, uint64_t pd_addr) 465 { 466 bool use_semaphore = gmc_v10_0_use_invalidate_semaphore(ring->adev, ring->funcs->vmhub); 467 struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->funcs->vmhub]; 468 uint32_t req = hub->vmhub_funcs->get_invalidate_req(vmid, 0); 469 unsigned eng = ring->vm_inv_eng; 470 471 /* 472 * It may lose gpuvm invalidate acknowldege state across power-gating 473 * off cycle, add semaphore acquire before invalidation and semaphore 474 * release after invalidation to avoid entering power gated state 475 * to WA the Issue 476 */ 477 478 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 479 if (use_semaphore) 480 /* a read return value of 1 means semaphore acuqire */ 481 amdgpu_ring_emit_reg_wait(ring, 482 hub->vm_inv_eng0_sem + 483 hub->eng_distance * eng, 0x1, 0x1); 484 485 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 + 486 (hub->ctx_addr_distance * vmid), 487 lower_32_bits(pd_addr)); 488 489 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + 490 (hub->ctx_addr_distance * vmid), 491 upper_32_bits(pd_addr)); 492 493 amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req + 494 hub->eng_distance * eng, 495 hub->vm_inv_eng0_ack + 496 hub->eng_distance * eng, 497 req, 1 << vmid); 498 499 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 500 if (use_semaphore) 501 /* 502 * add semaphore release after invalidation, 503 * write with 0 means semaphore release 504 */ 505 amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem + 506 hub->eng_distance * eng, 0); 507 508 return pd_addr; 509 } 510 511 static void gmc_v10_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid, 512 unsigned pasid) 513 { 514 struct amdgpu_device *adev = ring->adev; 515 uint32_t reg; 516 517 if (ring->funcs->vmhub == AMDGPU_GFXHUB_0) 518 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid; 519 else 520 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid; 521 522 amdgpu_ring_emit_wreg(ring, reg, pasid); 523 } 524 525 /* 526 * PTE format on NAVI 10: 527 * 63:59 reserved 528 * 58 reserved and for sienna_cichlid is used for MALL noalloc 529 * 57 reserved 530 * 56 F 531 * 55 L 532 * 54 reserved 533 * 53:52 SW 534 * 51 T 535 * 50:48 mtype 536 * 47:12 4k physical page base address 537 * 11:7 fragment 538 * 6 write 539 * 5 read 540 * 4 exe 541 * 3 Z 542 * 2 snooped 543 * 1 system 544 * 0 valid 545 * 546 * PDE format on NAVI 10: 547 * 63:59 block fragment size 548 * 58:55 reserved 549 * 54 P 550 * 53:48 reserved 551 * 47:6 physical base address of PD or PTE 552 * 5:3 reserved 553 * 2 C 554 * 1 system 555 * 0 valid 556 */ 557 558 static uint64_t gmc_v10_0_map_mtype(struct amdgpu_device *adev, uint32_t flags) 559 { 560 switch (flags) { 561 case AMDGPU_VM_MTYPE_DEFAULT: 562 return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC); 563 case AMDGPU_VM_MTYPE_NC: 564 return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC); 565 case AMDGPU_VM_MTYPE_WC: 566 return AMDGPU_PTE_MTYPE_NV10(MTYPE_WC); 567 case AMDGPU_VM_MTYPE_CC: 568 return AMDGPU_PTE_MTYPE_NV10(MTYPE_CC); 569 case AMDGPU_VM_MTYPE_UC: 570 return AMDGPU_PTE_MTYPE_NV10(MTYPE_UC); 571 default: 572 return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC); 573 } 574 } 575 576 static void gmc_v10_0_get_vm_pde(struct amdgpu_device *adev, int level, 577 uint64_t *addr, uint64_t *flags) 578 { 579 if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM)) 580 *addr = amdgpu_gmc_vram_mc2pa(adev, *addr); 581 BUG_ON(*addr & 0xFFFF00000000003FULL); 582 583 if (!adev->gmc.translate_further) 584 return; 585 586 if (level == AMDGPU_VM_PDB1) { 587 /* Set the block fragment size */ 588 if (!(*flags & AMDGPU_PDE_PTE)) 589 *flags |= AMDGPU_PDE_BFS(0x9); 590 591 } else if (level == AMDGPU_VM_PDB0) { 592 if (*flags & AMDGPU_PDE_PTE) 593 *flags &= ~AMDGPU_PDE_PTE; 594 else 595 *flags |= AMDGPU_PTE_TF; 596 } 597 } 598 599 static void gmc_v10_0_get_vm_pte(struct amdgpu_device *adev, 600 struct amdgpu_bo_va_mapping *mapping, 601 uint64_t *flags) 602 { 603 *flags &= ~AMDGPU_PTE_EXECUTABLE; 604 *flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE; 605 606 *flags &= ~AMDGPU_PTE_MTYPE_NV10_MASK; 607 *flags |= (mapping->flags & AMDGPU_PTE_MTYPE_NV10_MASK); 608 609 if (mapping->flags & AMDGPU_PTE_PRT) { 610 *flags |= AMDGPU_PTE_PRT; 611 *flags |= AMDGPU_PTE_SNOOPED; 612 *flags |= AMDGPU_PTE_LOG; 613 *flags |= AMDGPU_PTE_SYSTEM; 614 *flags &= ~AMDGPU_PTE_VALID; 615 } 616 } 617 618 static unsigned gmc_v10_0_get_vbios_fb_size(struct amdgpu_device *adev) 619 { 620 u32 d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL); 621 unsigned size; 622 623 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) { 624 size = AMDGPU_VBIOS_VGA_ALLOCATION; 625 } else { 626 u32 viewport; 627 u32 pitch; 628 629 viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION); 630 pitch = RREG32_SOC15(DCE, 0, mmHUBPREQ0_DCSURF_SURFACE_PITCH); 631 size = (REG_GET_FIELD(viewport, 632 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) * 633 REG_GET_FIELD(pitch, HUBPREQ0_DCSURF_SURFACE_PITCH, PITCH) * 634 4); 635 } 636 637 return size; 638 } 639 640 static const struct amdgpu_gmc_funcs gmc_v10_0_gmc_funcs = { 641 .flush_gpu_tlb = gmc_v10_0_flush_gpu_tlb, 642 .flush_gpu_tlb_pasid = gmc_v10_0_flush_gpu_tlb_pasid, 643 .emit_flush_gpu_tlb = gmc_v10_0_emit_flush_gpu_tlb, 644 .emit_pasid_mapping = gmc_v10_0_emit_pasid_mapping, 645 .map_mtype = gmc_v10_0_map_mtype, 646 .get_vm_pde = gmc_v10_0_get_vm_pde, 647 .get_vm_pte = gmc_v10_0_get_vm_pte, 648 .get_vbios_fb_size = gmc_v10_0_get_vbios_fb_size, 649 }; 650 651 static void gmc_v10_0_set_gmc_funcs(struct amdgpu_device *adev) 652 { 653 if (adev->gmc.gmc_funcs == NULL) 654 adev->gmc.gmc_funcs = &gmc_v10_0_gmc_funcs; 655 } 656 657 static void gmc_v10_0_set_umc_funcs(struct amdgpu_device *adev) 658 { 659 switch (adev->asic_type) { 660 case CHIP_SIENNA_CICHLID: 661 adev->umc.max_ras_err_cnt_per_query = UMC_V8_7_TOTAL_CHANNEL_NUM; 662 adev->umc.channel_inst_num = UMC_V8_7_CHANNEL_INSTANCE_NUM; 663 adev->umc.umc_inst_num = UMC_V8_7_UMC_INSTANCE_NUM; 664 adev->umc.channel_offs = UMC_V8_7_PER_CHANNEL_OFFSET_SIENNA; 665 adev->umc.channel_idx_tbl = &umc_v8_7_channel_idx_tbl[0][0]; 666 adev->umc.ras_funcs = &umc_v8_7_ras_funcs; 667 break; 668 default: 669 break; 670 } 671 } 672 673 674 static void gmc_v10_0_set_mmhub_funcs(struct amdgpu_device *adev) 675 { 676 switch (adev->asic_type) { 677 case CHIP_VANGOGH: 678 case CHIP_YELLOW_CARP: 679 adev->mmhub.funcs = &mmhub_v2_3_funcs; 680 break; 681 default: 682 adev->mmhub.funcs = &mmhub_v2_0_funcs; 683 break; 684 } 685 } 686 687 static void gmc_v10_0_set_gfxhub_funcs(struct amdgpu_device *adev) 688 { 689 switch (adev->asic_type) { 690 case CHIP_SIENNA_CICHLID: 691 case CHIP_NAVY_FLOUNDER: 692 case CHIP_VANGOGH: 693 case CHIP_DIMGREY_CAVEFISH: 694 case CHIP_BEIGE_GOBY: 695 case CHIP_YELLOW_CARP: 696 adev->gfxhub.funcs = &gfxhub_v2_1_funcs; 697 break; 698 default: 699 adev->gfxhub.funcs = &gfxhub_v2_0_funcs; 700 break; 701 } 702 } 703 704 705 static int gmc_v10_0_early_init(void *handle) 706 { 707 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 708 709 gmc_v10_0_set_mmhub_funcs(adev); 710 gmc_v10_0_set_gfxhub_funcs(adev); 711 gmc_v10_0_set_gmc_funcs(adev); 712 gmc_v10_0_set_irq_funcs(adev); 713 gmc_v10_0_set_umc_funcs(adev); 714 715 adev->gmc.shared_aperture_start = 0x2000000000000000ULL; 716 adev->gmc.shared_aperture_end = 717 adev->gmc.shared_aperture_start + (4ULL << 30) - 1; 718 adev->gmc.private_aperture_start = 0x1000000000000000ULL; 719 adev->gmc.private_aperture_end = 720 adev->gmc.private_aperture_start + (4ULL << 30) - 1; 721 722 return 0; 723 } 724 725 static int gmc_v10_0_late_init(void *handle) 726 { 727 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 728 int r; 729 730 r = amdgpu_gmc_allocate_vm_inv_eng(adev); 731 if (r) 732 return r; 733 734 r = amdgpu_gmc_ras_late_init(adev); 735 if (r) 736 return r; 737 738 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0); 739 } 740 741 static void gmc_v10_0_vram_gtt_location(struct amdgpu_device *adev, 742 struct amdgpu_gmc *mc) 743 { 744 u64 base = 0; 745 746 base = adev->gfxhub.funcs->get_fb_location(adev); 747 748 /* add the xgmi offset of the physical node */ 749 base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size; 750 751 amdgpu_gmc_vram_location(adev, &adev->gmc, base); 752 amdgpu_gmc_gart_location(adev, mc); 753 amdgpu_gmc_agp_location(adev, mc); 754 755 /* base offset of vram pages */ 756 adev->vm_manager.vram_base_offset = adev->gfxhub.funcs->get_mc_fb_offset(adev); 757 758 /* add the xgmi offset of the physical node */ 759 adev->vm_manager.vram_base_offset += 760 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size; 761 } 762 763 /** 764 * gmc_v10_0_mc_init - initialize the memory controller driver params 765 * 766 * @adev: amdgpu_device pointer 767 * 768 * Look up the amount of vram, vram width, and decide how to place 769 * vram and gart within the GPU's physical address space. 770 * Returns 0 for success. 771 */ 772 static int gmc_v10_0_mc_init(struct amdgpu_device *adev) 773 { 774 int r; 775 776 /* size in MB on si */ 777 adev->gmc.mc_vram_size = 778 adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL; 779 adev->gmc.real_vram_size = adev->gmc.mc_vram_size; 780 781 if (!(adev->flags & AMD_IS_APU)) { 782 r = amdgpu_device_resize_fb_bar(adev); 783 if (r) 784 return r; 785 } 786 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0); 787 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0); 788 789 #ifdef CONFIG_X86_64 790 if (adev->flags & AMD_IS_APU) { 791 adev->gmc.aper_base = adev->gfxhub.funcs->get_mc_fb_offset(adev); 792 adev->gmc.aper_size = adev->gmc.real_vram_size; 793 } 794 #endif 795 796 /* In case the PCI BAR is larger than the actual amount of vram */ 797 adev->gmc.visible_vram_size = adev->gmc.aper_size; 798 if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size) 799 adev->gmc.visible_vram_size = adev->gmc.real_vram_size; 800 801 /* set the gart size */ 802 if (amdgpu_gart_size == -1) { 803 switch (adev->asic_type) { 804 case CHIP_NAVI10: 805 case CHIP_NAVI14: 806 case CHIP_NAVI12: 807 case CHIP_SIENNA_CICHLID: 808 case CHIP_NAVY_FLOUNDER: 809 case CHIP_VANGOGH: 810 case CHIP_DIMGREY_CAVEFISH: 811 case CHIP_BEIGE_GOBY: 812 case CHIP_YELLOW_CARP: 813 case CHIP_CYAN_SKILLFISH: 814 default: 815 adev->gmc.gart_size = 512ULL << 20; 816 break; 817 } 818 } else 819 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20; 820 821 gmc_v10_0_vram_gtt_location(adev, &adev->gmc); 822 823 return 0; 824 } 825 826 static int gmc_v10_0_gart_init(struct amdgpu_device *adev) 827 { 828 int r; 829 830 if (adev->gart.bo) { 831 WARN(1, "NAVI10 PCIE GART already initialized\n"); 832 return 0; 833 } 834 835 /* Initialize common gart structure */ 836 r = amdgpu_gart_init(adev); 837 if (r) 838 return r; 839 840 adev->gart.table_size = adev->gart.num_gpu_pages * 8; 841 adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_NV10(MTYPE_UC) | 842 AMDGPU_PTE_EXECUTABLE; 843 844 return amdgpu_gart_table_vram_alloc(adev); 845 } 846 847 static int gmc_v10_0_sw_init(void *handle) 848 { 849 int r, vram_width = 0, vram_type = 0, vram_vendor = 0; 850 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 851 852 adev->gfxhub.funcs->init(adev); 853 854 adev->mmhub.funcs->init(adev); 855 856 spin_lock_init(&adev->gmc.invalidate_lock); 857 858 if ((adev->flags & AMD_IS_APU) && amdgpu_emu_mode == 1) { 859 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_DDR4; 860 adev->gmc.vram_width = 64; 861 } else if (amdgpu_emu_mode == 1) { 862 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_GDDR6; 863 adev->gmc.vram_width = 1 * 128; /* numchan * chansize */ 864 } else { 865 r = amdgpu_atomfirmware_get_vram_info(adev, 866 &vram_width, &vram_type, &vram_vendor); 867 adev->gmc.vram_width = vram_width; 868 869 adev->gmc.vram_type = vram_type; 870 adev->gmc.vram_vendor = vram_vendor; 871 } 872 873 switch (adev->asic_type) { 874 case CHIP_NAVI10: 875 case CHIP_NAVI14: 876 case CHIP_NAVI12: 877 case CHIP_SIENNA_CICHLID: 878 case CHIP_NAVY_FLOUNDER: 879 case CHIP_VANGOGH: 880 case CHIP_DIMGREY_CAVEFISH: 881 case CHIP_BEIGE_GOBY: 882 case CHIP_YELLOW_CARP: 883 case CHIP_CYAN_SKILLFISH: 884 adev->num_vmhubs = 2; 885 /* 886 * To fulfill 4-level page support, 887 * vm size is 256TB (48bit), maximum size of Navi10/Navi14/Navi12, 888 * block size 512 (9bit) 889 */ 890 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 891 break; 892 default: 893 break; 894 } 895 896 /* This interrupt is VMC page fault.*/ 897 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC, 898 VMC_1_0__SRCID__VM_FAULT, 899 &adev->gmc.vm_fault); 900 901 if (r) 902 return r; 903 904 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2, 905 UTCL2_1_0__SRCID__FAULT, 906 &adev->gmc.vm_fault); 907 if (r) 908 return r; 909 910 if (!amdgpu_sriov_vf(adev)) { 911 /* interrupt sent to DF. */ 912 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0, 913 &adev->gmc.ecc_irq); 914 if (r) 915 return r; 916 } 917 918 /* 919 * Set the internal MC address mask This is the max address of the GPU's 920 * internal address space. 921 */ 922 adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */ 923 924 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44)); 925 if (r) { 926 printk(KERN_WARNING "amdgpu: No suitable DMA available.\n"); 927 return r; 928 } 929 930 if (adev->gmc.xgmi.supported) { 931 r = adev->gfxhub.funcs->get_xgmi_info(adev); 932 if (r) 933 return r; 934 } 935 936 r = gmc_v10_0_mc_init(adev); 937 if (r) 938 return r; 939 940 amdgpu_gmc_get_vbios_allocations(adev); 941 amdgpu_gmc_get_reserved_allocation(adev); 942 943 /* Memory manager */ 944 r = amdgpu_bo_init(adev); 945 if (r) 946 return r; 947 948 r = gmc_v10_0_gart_init(adev); 949 if (r) 950 return r; 951 952 /* 953 * number of VMs 954 * VMID 0 is reserved for System 955 * amdgpu graphics/compute will use VMIDs 1-7 956 * amdkfd will use VMIDs 8-15 957 */ 958 adev->vm_manager.first_kfd_vmid = 8; 959 960 amdgpu_vm_manager_init(adev); 961 962 return 0; 963 } 964 965 /** 966 * gmc_v10_0_gart_fini - vm fini callback 967 * 968 * @adev: amdgpu_device pointer 969 * 970 * Tears down the driver GART/VM setup (CIK). 971 */ 972 static void gmc_v10_0_gart_fini(struct amdgpu_device *adev) 973 { 974 amdgpu_gart_table_vram_free(adev); 975 } 976 977 static int gmc_v10_0_sw_fini(void *handle) 978 { 979 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 980 981 amdgpu_vm_manager_fini(adev); 982 gmc_v10_0_gart_fini(adev); 983 amdgpu_gem_force_release(adev); 984 amdgpu_bo_fini(adev); 985 986 return 0; 987 } 988 989 static void gmc_v10_0_init_golden_registers(struct amdgpu_device *adev) 990 { 991 switch (adev->asic_type) { 992 case CHIP_NAVI10: 993 case CHIP_NAVI14: 994 case CHIP_NAVI12: 995 case CHIP_SIENNA_CICHLID: 996 case CHIP_NAVY_FLOUNDER: 997 case CHIP_VANGOGH: 998 case CHIP_DIMGREY_CAVEFISH: 999 case CHIP_BEIGE_GOBY: 1000 case CHIP_YELLOW_CARP: 1001 case CHIP_CYAN_SKILLFISH: 1002 break; 1003 default: 1004 break; 1005 } 1006 } 1007 1008 /** 1009 * gmc_v10_0_gart_enable - gart enable 1010 * 1011 * @adev: amdgpu_device pointer 1012 */ 1013 static int gmc_v10_0_gart_enable(struct amdgpu_device *adev) 1014 { 1015 int r; 1016 bool value; 1017 1018 if (adev->gart.bo == NULL) { 1019 dev_err(adev->dev, "No VRAM object for PCIE GART.\n"); 1020 return -EINVAL; 1021 } 1022 1023 r = amdgpu_gart_table_vram_pin(adev); 1024 if (r) 1025 return r; 1026 1027 r = adev->gfxhub.funcs->gart_enable(adev); 1028 if (r) 1029 return r; 1030 1031 r = adev->mmhub.funcs->gart_enable(adev); 1032 if (r) 1033 return r; 1034 1035 adev->hdp.funcs->init_registers(adev); 1036 1037 /* Flush HDP after it is initialized */ 1038 adev->hdp.funcs->flush_hdp(adev, NULL); 1039 1040 value = (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) ? 1041 false : true; 1042 1043 adev->gfxhub.funcs->set_fault_enable_default(adev, value); 1044 adev->mmhub.funcs->set_fault_enable_default(adev, value); 1045 gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_MMHUB_0, 0); 1046 gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_GFXHUB_0, 0); 1047 1048 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n", 1049 (unsigned)(adev->gmc.gart_size >> 20), 1050 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo)); 1051 1052 adev->gart.ready = true; 1053 1054 return 0; 1055 } 1056 1057 static int gmc_v10_0_hw_init(void *handle) 1058 { 1059 int r; 1060 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1061 1062 /* The sequence of these two function calls matters.*/ 1063 gmc_v10_0_init_golden_registers(adev); 1064 1065 /* 1066 * harvestable groups in gc_utcl2 need to be programmed before any GFX block 1067 * register setup within GMC, or else system hang when harvesting SA. 1068 */ 1069 if (adev->gfxhub.funcs && adev->gfxhub.funcs->utcl2_harvest) 1070 adev->gfxhub.funcs->utcl2_harvest(adev); 1071 1072 r = gmc_v10_0_gart_enable(adev); 1073 if (r) 1074 return r; 1075 1076 if (adev->umc.funcs && adev->umc.funcs->init_registers) 1077 adev->umc.funcs->init_registers(adev); 1078 1079 return 0; 1080 } 1081 1082 /** 1083 * gmc_v10_0_gart_disable - gart disable 1084 * 1085 * @adev: amdgpu_device pointer 1086 * 1087 * This disables all VM page table. 1088 */ 1089 static void gmc_v10_0_gart_disable(struct amdgpu_device *adev) 1090 { 1091 adev->gfxhub.funcs->gart_disable(adev); 1092 adev->mmhub.funcs->gart_disable(adev); 1093 amdgpu_gart_table_vram_unpin(adev); 1094 } 1095 1096 static int gmc_v10_0_hw_fini(void *handle) 1097 { 1098 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1099 1100 if (amdgpu_sriov_vf(adev)) { 1101 /* full access mode, so don't touch any GMC register */ 1102 DRM_DEBUG("For SRIOV client, shouldn't do anything.\n"); 1103 return 0; 1104 } 1105 1106 amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0); 1107 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0); 1108 gmc_v10_0_gart_disable(adev); 1109 1110 return 0; 1111 } 1112 1113 static int gmc_v10_0_suspend(void *handle) 1114 { 1115 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1116 1117 gmc_v10_0_hw_fini(adev); 1118 1119 return 0; 1120 } 1121 1122 static int gmc_v10_0_resume(void *handle) 1123 { 1124 int r; 1125 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1126 1127 r = gmc_v10_0_hw_init(adev); 1128 if (r) 1129 return r; 1130 1131 amdgpu_vmid_reset_all(adev); 1132 1133 return 0; 1134 } 1135 1136 static bool gmc_v10_0_is_idle(void *handle) 1137 { 1138 /* MC is always ready in GMC v10.*/ 1139 return true; 1140 } 1141 1142 static int gmc_v10_0_wait_for_idle(void *handle) 1143 { 1144 /* There is no need to wait for MC idle in GMC v10.*/ 1145 return 0; 1146 } 1147 1148 static int gmc_v10_0_soft_reset(void *handle) 1149 { 1150 return 0; 1151 } 1152 1153 static int gmc_v10_0_set_clockgating_state(void *handle, 1154 enum amd_clockgating_state state) 1155 { 1156 int r; 1157 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1158 1159 r = adev->mmhub.funcs->set_clockgating(adev, state); 1160 if (r) 1161 return r; 1162 1163 if (adev->asic_type >= CHIP_SIENNA_CICHLID && 1164 adev->asic_type <= CHIP_YELLOW_CARP) 1165 return athub_v2_1_set_clockgating(adev, state); 1166 else 1167 return athub_v2_0_set_clockgating(adev, state); 1168 } 1169 1170 static void gmc_v10_0_get_clockgating_state(void *handle, u32 *flags) 1171 { 1172 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1173 1174 adev->mmhub.funcs->get_clockgating(adev, flags); 1175 1176 if (adev->asic_type >= CHIP_SIENNA_CICHLID && 1177 adev->asic_type <= CHIP_YELLOW_CARP) 1178 athub_v2_1_get_clockgating(adev, flags); 1179 else 1180 athub_v2_0_get_clockgating(adev, flags); 1181 } 1182 1183 static int gmc_v10_0_set_powergating_state(void *handle, 1184 enum amd_powergating_state state) 1185 { 1186 return 0; 1187 } 1188 1189 const struct amd_ip_funcs gmc_v10_0_ip_funcs = { 1190 .name = "gmc_v10_0", 1191 .early_init = gmc_v10_0_early_init, 1192 .late_init = gmc_v10_0_late_init, 1193 .sw_init = gmc_v10_0_sw_init, 1194 .sw_fini = gmc_v10_0_sw_fini, 1195 .hw_init = gmc_v10_0_hw_init, 1196 .hw_fini = gmc_v10_0_hw_fini, 1197 .suspend = gmc_v10_0_suspend, 1198 .resume = gmc_v10_0_resume, 1199 .is_idle = gmc_v10_0_is_idle, 1200 .wait_for_idle = gmc_v10_0_wait_for_idle, 1201 .soft_reset = gmc_v10_0_soft_reset, 1202 .set_clockgating_state = gmc_v10_0_set_clockgating_state, 1203 .set_powergating_state = gmc_v10_0_set_powergating_state, 1204 .get_clockgating_state = gmc_v10_0_get_clockgating_state, 1205 }; 1206 1207 const struct amdgpu_ip_block_version gmc_v10_0_ip_block = 1208 { 1209 .type = AMD_IP_BLOCK_TYPE_GMC, 1210 .major = 10, 1211 .minor = 0, 1212 .rev = 0, 1213 .funcs = &gmc_v10_0_ip_funcs, 1214 }; 1215