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 "hdp/hdp_5_0_0_offset.h" 31 #include "hdp/hdp_5_0_0_sh_mask.h" 32 #include "athub/athub_2_0_0_sh_mask.h" 33 #include "athub/athub_2_0_0_offset.h" 34 #include "dcn/dcn_2_0_0_offset.h" 35 #include "dcn/dcn_2_0_0_sh_mask.h" 36 #include "oss/osssys_5_0_0_offset.h" 37 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h" 38 #include "navi10_enum.h" 39 40 #include "soc15.h" 41 #include "soc15d.h" 42 #include "soc15_common.h" 43 44 #include "nbio_v2_3.h" 45 46 #include "gfxhub_v2_0.h" 47 #include "gfxhub_v2_1.h" 48 #include "mmhub_v2_0.h" 49 #include "athub_v2_0.h" 50 #include "athub_v2_1.h" 51 52 #if 0 53 static const struct soc15_reg_golden golden_settings_navi10_hdp[] = 54 { 55 /* TODO add golden setting for hdp */ 56 }; 57 #endif 58 59 static int gmc_v10_0_ecc_interrupt_state(struct amdgpu_device *adev, 60 struct amdgpu_irq_src *src, 61 unsigned type, 62 enum amdgpu_interrupt_state state) 63 { 64 return 0; 65 } 66 67 static int 68 gmc_v10_0_vm_fault_interrupt_state(struct amdgpu_device *adev, 69 struct amdgpu_irq_src *src, unsigned type, 70 enum amdgpu_interrupt_state state) 71 { 72 switch (state) { 73 case AMDGPU_IRQ_STATE_DISABLE: 74 /* MM HUB */ 75 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB_0, false); 76 /* GFX HUB */ 77 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB_0, false); 78 break; 79 case AMDGPU_IRQ_STATE_ENABLE: 80 /* MM HUB */ 81 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB_0, true); 82 /* GFX HUB */ 83 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB_0, true); 84 break; 85 default: 86 break; 87 } 88 89 return 0; 90 } 91 92 static int gmc_v10_0_process_interrupt(struct amdgpu_device *adev, 93 struct amdgpu_irq_src *source, 94 struct amdgpu_iv_entry *entry) 95 { 96 struct amdgpu_vmhub *hub = &adev->vmhub[entry->vmid_src]; 97 uint32_t status = 0; 98 u64 addr; 99 100 addr = (u64)entry->src_data[0] << 12; 101 addr |= ((u64)entry->src_data[1] & 0xf) << 44; 102 103 if (!amdgpu_sriov_vf(adev)) { 104 /* 105 * Issue a dummy read to wait for the status register to 106 * be updated to avoid reading an incorrect value due to 107 * the new fast GRBM interface. 108 */ 109 if (entry->vmid_src == AMDGPU_GFXHUB_0) 110 RREG32(hub->vm_l2_pro_fault_status); 111 112 status = RREG32(hub->vm_l2_pro_fault_status); 113 WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1); 114 } 115 116 if (printk_ratelimit()) { 117 struct amdgpu_task_info task_info; 118 119 memset(&task_info, 0, sizeof(struct amdgpu_task_info)); 120 amdgpu_vm_get_task_info(adev, entry->pasid, &task_info); 121 122 dev_err(adev->dev, 123 "[%s] page fault (src_id:%u ring:%u vmid:%u pasid:%u, " 124 "for process %s pid %d thread %s pid %d)\n", 125 entry->vmid_src ? "mmhub" : "gfxhub", 126 entry->src_id, entry->ring_id, entry->vmid, 127 entry->pasid, task_info.process_name, task_info.tgid, 128 task_info.task_name, task_info.pid); 129 dev_err(adev->dev, " in page starting at address 0x%016llx from client %d\n", 130 addr, entry->client_id); 131 if (!amdgpu_sriov_vf(adev)) 132 hub->vmhub_funcs->print_l2_protection_fault_status(adev, status); 133 } 134 135 return 0; 136 } 137 138 static const struct amdgpu_irq_src_funcs gmc_v10_0_irq_funcs = { 139 .set = gmc_v10_0_vm_fault_interrupt_state, 140 .process = gmc_v10_0_process_interrupt, 141 }; 142 143 static const struct amdgpu_irq_src_funcs gmc_v10_0_ecc_funcs = { 144 .set = gmc_v10_0_ecc_interrupt_state, 145 .process = amdgpu_umc_process_ecc_irq, 146 }; 147 148 static void gmc_v10_0_set_irq_funcs(struct amdgpu_device *adev) 149 { 150 adev->gmc.vm_fault.num_types = 1; 151 adev->gmc.vm_fault.funcs = &gmc_v10_0_irq_funcs; 152 153 if (!amdgpu_sriov_vf(adev)) { 154 adev->gmc.ecc_irq.num_types = 1; 155 adev->gmc.ecc_irq.funcs = &gmc_v10_0_ecc_funcs; 156 } 157 } 158 159 /** 160 * gmc_v10_0_use_invalidate_semaphore - judge whether to use semaphore 161 * 162 * @adev: amdgpu_device pointer 163 * @vmhub: vmhub type 164 * 165 */ 166 static bool gmc_v10_0_use_invalidate_semaphore(struct amdgpu_device *adev, 167 uint32_t vmhub) 168 { 169 return ((vmhub == AMDGPU_MMHUB_0 || 170 vmhub == AMDGPU_MMHUB_1) && 171 (!amdgpu_sriov_vf(adev))); 172 } 173 174 static bool gmc_v10_0_get_atc_vmid_pasid_mapping_info( 175 struct amdgpu_device *adev, 176 uint8_t vmid, uint16_t *p_pasid) 177 { 178 uint32_t value; 179 180 value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) 181 + vmid); 182 *p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK; 183 184 return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK); 185 } 186 187 /* 188 * GART 189 * VMID 0 is the physical GPU addresses as used by the kernel. 190 * VMIDs 1-15 are used for userspace clients and are handled 191 * by the amdgpu vm/hsa code. 192 */ 193 194 static void gmc_v10_0_flush_vm_hub(struct amdgpu_device *adev, uint32_t vmid, 195 unsigned int vmhub, uint32_t flush_type) 196 { 197 bool use_semaphore = gmc_v10_0_use_invalidate_semaphore(adev, vmhub); 198 struct amdgpu_vmhub *hub = &adev->vmhub[vmhub]; 199 u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type); 200 u32 tmp; 201 /* Use register 17 for GART */ 202 const unsigned eng = 17; 203 unsigned int i; 204 205 spin_lock(&adev->gmc.invalidate_lock); 206 /* 207 * It may lose gpuvm invalidate acknowldege state across power-gating 208 * off cycle, add semaphore acquire before invalidation and semaphore 209 * release after invalidation to avoid entering power gated state 210 * to WA the Issue 211 */ 212 213 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 214 if (use_semaphore) { 215 for (i = 0; i < adev->usec_timeout; i++) { 216 /* a read return value of 1 means semaphore acuqire */ 217 tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_sem + 218 hub->eng_distance * eng); 219 if (tmp & 0x1) 220 break; 221 udelay(1); 222 } 223 224 if (i >= adev->usec_timeout) 225 DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n"); 226 } 227 228 WREG32_NO_KIQ(hub->vm_inv_eng0_req + hub->eng_distance * eng, inv_req); 229 230 /* 231 * Issue a dummy read to wait for the ACK register to be cleared 232 * to avoid a false ACK due to the new fast GRBM interface. 233 */ 234 if (vmhub == AMDGPU_GFXHUB_0) 235 RREG32_NO_KIQ(hub->vm_inv_eng0_req + hub->eng_distance * eng); 236 237 /* Wait for ACK with a delay.*/ 238 for (i = 0; i < adev->usec_timeout; i++) { 239 tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_ack + 240 hub->eng_distance * eng); 241 tmp &= 1 << vmid; 242 if (tmp) 243 break; 244 245 udelay(1); 246 } 247 248 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 249 if (use_semaphore) 250 /* 251 * add semaphore release after invalidation, 252 * write with 0 means semaphore release 253 */ 254 WREG32_NO_KIQ(hub->vm_inv_eng0_sem + 255 hub->eng_distance * eng, 0); 256 257 spin_unlock(&adev->gmc.invalidate_lock); 258 259 if (i < adev->usec_timeout) 260 return; 261 262 DRM_ERROR("Timeout waiting for VM flush ACK!\n"); 263 } 264 265 /** 266 * gmc_v10_0_flush_gpu_tlb - gart tlb flush callback 267 * 268 * @adev: amdgpu_device pointer 269 * @vmid: vm instance to flush 270 * 271 * Flush the TLB for the requested page table. 272 */ 273 static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, 274 uint32_t vmhub, uint32_t flush_type) 275 { 276 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 277 struct dma_fence *fence; 278 struct amdgpu_job *job; 279 280 int r; 281 282 /* flush hdp cache */ 283 adev->nbio.funcs->hdp_flush(adev, NULL); 284 285 /* For SRIOV run time, driver shouldn't access the register through MMIO 286 * Directly use kiq to do the vm invalidation instead 287 */ 288 if (adev->gfx.kiq.ring.sched.ready && 289 (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev)) && 290 down_read_trylock(&adev->reset_sem)) { 291 struct amdgpu_vmhub *hub = &adev->vmhub[vmhub]; 292 const unsigned eng = 17; 293 u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type); 294 u32 req = hub->vm_inv_eng0_req + hub->eng_distance * eng; 295 u32 ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng; 296 297 amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, inv_req, 298 1 << vmid); 299 300 up_read(&adev->reset_sem); 301 return; 302 } 303 304 mutex_lock(&adev->mman.gtt_window_lock); 305 306 if (vmhub == AMDGPU_MMHUB_0) { 307 gmc_v10_0_flush_vm_hub(adev, vmid, AMDGPU_MMHUB_0, 0); 308 mutex_unlock(&adev->mman.gtt_window_lock); 309 return; 310 } 311 312 BUG_ON(vmhub != AMDGPU_GFXHUB_0); 313 314 if (!adev->mman.buffer_funcs_enabled || 315 !adev->ib_pool_ready || 316 amdgpu_in_reset(adev) || 317 ring->sched.ready == false) { 318 gmc_v10_0_flush_vm_hub(adev, vmid, AMDGPU_GFXHUB_0, 0); 319 mutex_unlock(&adev->mman.gtt_window_lock); 320 return; 321 } 322 323 /* The SDMA on Navi has a bug which can theoretically result in memory 324 * corruption if an invalidation happens at the same time as an VA 325 * translation. Avoid this by doing the invalidation from the SDMA 326 * itself. 327 */ 328 r = amdgpu_job_alloc_with_ib(adev, 16 * 4, AMDGPU_IB_POOL_IMMEDIATE, 329 &job); 330 if (r) 331 goto error_alloc; 332 333 job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo); 334 job->vm_needs_flush = true; 335 job->ibs->ptr[job->ibs->length_dw++] = ring->funcs->nop; 336 amdgpu_ring_pad_ib(ring, &job->ibs[0]); 337 r = amdgpu_job_submit(job, &adev->mman.entity, 338 AMDGPU_FENCE_OWNER_UNDEFINED, &fence); 339 if (r) 340 goto error_submit; 341 342 mutex_unlock(&adev->mman.gtt_window_lock); 343 344 dma_fence_wait(fence, false); 345 dma_fence_put(fence); 346 347 return; 348 349 error_submit: 350 amdgpu_job_free(job); 351 352 error_alloc: 353 mutex_unlock(&adev->mman.gtt_window_lock); 354 DRM_ERROR("Error flushing GPU TLB using the SDMA (%d)!\n", r); 355 } 356 357 /** 358 * gmc_v10_0_flush_gpu_tlb_pasid - tlb flush via pasid 359 * 360 * @adev: amdgpu_device pointer 361 * @pasid: pasid to be flush 362 * 363 * Flush the TLB for the requested pasid. 364 */ 365 static int gmc_v10_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, 366 uint16_t pasid, uint32_t flush_type, 367 bool all_hub) 368 { 369 int vmid, i; 370 signed long r; 371 uint32_t seq; 372 uint16_t queried_pasid; 373 bool ret; 374 struct amdgpu_ring *ring = &adev->gfx.kiq.ring; 375 struct amdgpu_kiq *kiq = &adev->gfx.kiq; 376 377 if (amdgpu_emu_mode == 0 && ring->sched.ready) { 378 spin_lock(&adev->gfx.kiq.ring_lock); 379 /* 2 dwords flush + 8 dwords fence */ 380 amdgpu_ring_alloc(ring, kiq->pmf->invalidate_tlbs_size + 8); 381 kiq->pmf->kiq_invalidate_tlbs(ring, 382 pasid, flush_type, all_hub); 383 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT); 384 if (r) { 385 amdgpu_ring_undo(ring); 386 spin_unlock(&adev->gfx.kiq.ring_lock); 387 return -ETIME; 388 } 389 390 amdgpu_ring_commit(ring); 391 spin_unlock(&adev->gfx.kiq.ring_lock); 392 r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout); 393 if (r < 1) { 394 dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r); 395 return -ETIME; 396 } 397 398 return 0; 399 } 400 401 for (vmid = 1; vmid < 16; vmid++) { 402 403 ret = gmc_v10_0_get_atc_vmid_pasid_mapping_info(adev, vmid, 404 &queried_pasid); 405 if (ret && queried_pasid == pasid) { 406 if (all_hub) { 407 for (i = 0; i < adev->num_vmhubs; i++) 408 gmc_v10_0_flush_gpu_tlb(adev, vmid, 409 i, flush_type); 410 } else { 411 gmc_v10_0_flush_gpu_tlb(adev, vmid, 412 AMDGPU_GFXHUB_0, flush_type); 413 } 414 break; 415 } 416 } 417 418 return 0; 419 } 420 421 static uint64_t gmc_v10_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring, 422 unsigned vmid, uint64_t pd_addr) 423 { 424 bool use_semaphore = gmc_v10_0_use_invalidate_semaphore(ring->adev, ring->funcs->vmhub); 425 struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->funcs->vmhub]; 426 uint32_t req = hub->vmhub_funcs->get_invalidate_req(vmid, 0); 427 unsigned eng = ring->vm_inv_eng; 428 429 /* 430 * It may lose gpuvm invalidate acknowldege state across power-gating 431 * off cycle, add semaphore acquire before invalidation and semaphore 432 * release after invalidation to avoid entering power gated state 433 * to WA the Issue 434 */ 435 436 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 437 if (use_semaphore) 438 /* a read return value of 1 means semaphore acuqire */ 439 amdgpu_ring_emit_reg_wait(ring, 440 hub->vm_inv_eng0_sem + 441 hub->eng_distance * eng, 0x1, 0x1); 442 443 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 + 444 (hub->ctx_addr_distance * vmid), 445 lower_32_bits(pd_addr)); 446 447 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + 448 (hub->ctx_addr_distance * vmid), 449 upper_32_bits(pd_addr)); 450 451 amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req + 452 hub->eng_distance * eng, 453 hub->vm_inv_eng0_ack + 454 hub->eng_distance * eng, 455 req, 1 << vmid); 456 457 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 458 if (use_semaphore) 459 /* 460 * add semaphore release after invalidation, 461 * write with 0 means semaphore release 462 */ 463 amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem + 464 hub->eng_distance * eng, 0); 465 466 return pd_addr; 467 } 468 469 static void gmc_v10_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid, 470 unsigned pasid) 471 { 472 struct amdgpu_device *adev = ring->adev; 473 uint32_t reg; 474 475 if (ring->funcs->vmhub == AMDGPU_GFXHUB_0) 476 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid; 477 else 478 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid; 479 480 amdgpu_ring_emit_wreg(ring, reg, pasid); 481 } 482 483 /* 484 * PTE format on NAVI 10: 485 * 63:59 reserved 486 * 58:57 reserved 487 * 56 F 488 * 55 L 489 * 54 reserved 490 * 53:52 SW 491 * 51 T 492 * 50:48 mtype 493 * 47:12 4k physical page base address 494 * 11:7 fragment 495 * 6 write 496 * 5 read 497 * 4 exe 498 * 3 Z 499 * 2 snooped 500 * 1 system 501 * 0 valid 502 * 503 * PDE format on NAVI 10: 504 * 63:59 block fragment size 505 * 58:55 reserved 506 * 54 P 507 * 53:48 reserved 508 * 47:6 physical base address of PD or PTE 509 * 5:3 reserved 510 * 2 C 511 * 1 system 512 * 0 valid 513 */ 514 515 static uint64_t gmc_v10_0_map_mtype(struct amdgpu_device *adev, uint32_t flags) 516 { 517 switch (flags) { 518 case AMDGPU_VM_MTYPE_DEFAULT: 519 return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC); 520 case AMDGPU_VM_MTYPE_NC: 521 return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC); 522 case AMDGPU_VM_MTYPE_WC: 523 return AMDGPU_PTE_MTYPE_NV10(MTYPE_WC); 524 case AMDGPU_VM_MTYPE_CC: 525 return AMDGPU_PTE_MTYPE_NV10(MTYPE_CC); 526 case AMDGPU_VM_MTYPE_UC: 527 return AMDGPU_PTE_MTYPE_NV10(MTYPE_UC); 528 default: 529 return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC); 530 } 531 } 532 533 static void gmc_v10_0_get_vm_pde(struct amdgpu_device *adev, int level, 534 uint64_t *addr, uint64_t *flags) 535 { 536 if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM)) 537 *addr = adev->vm_manager.vram_base_offset + *addr - 538 adev->gmc.vram_start; 539 BUG_ON(*addr & 0xFFFF00000000003FULL); 540 541 if (!adev->gmc.translate_further) 542 return; 543 544 if (level == AMDGPU_VM_PDB1) { 545 /* Set the block fragment size */ 546 if (!(*flags & AMDGPU_PDE_PTE)) 547 *flags |= AMDGPU_PDE_BFS(0x9); 548 549 } else if (level == AMDGPU_VM_PDB0) { 550 if (*flags & AMDGPU_PDE_PTE) 551 *flags &= ~AMDGPU_PDE_PTE; 552 else 553 *flags |= AMDGPU_PTE_TF; 554 } 555 } 556 557 static void gmc_v10_0_get_vm_pte(struct amdgpu_device *adev, 558 struct amdgpu_bo_va_mapping *mapping, 559 uint64_t *flags) 560 { 561 *flags &= ~AMDGPU_PTE_EXECUTABLE; 562 *flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE; 563 564 *flags &= ~AMDGPU_PTE_MTYPE_NV10_MASK; 565 *flags |= (mapping->flags & AMDGPU_PTE_MTYPE_NV10_MASK); 566 567 if (mapping->flags & AMDGPU_PTE_PRT) { 568 *flags |= AMDGPU_PTE_PRT; 569 *flags |= AMDGPU_PTE_SNOOPED; 570 *flags |= AMDGPU_PTE_LOG; 571 *flags |= AMDGPU_PTE_SYSTEM; 572 *flags &= ~AMDGPU_PTE_VALID; 573 } 574 } 575 576 static unsigned gmc_v10_0_get_vbios_fb_size(struct amdgpu_device *adev) 577 { 578 u32 d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL); 579 unsigned size; 580 581 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) { 582 size = AMDGPU_VBIOS_VGA_ALLOCATION; 583 } else { 584 u32 viewport; 585 u32 pitch; 586 587 viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION); 588 pitch = RREG32_SOC15(DCE, 0, mmHUBPREQ0_DCSURF_SURFACE_PITCH); 589 size = (REG_GET_FIELD(viewport, 590 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) * 591 REG_GET_FIELD(pitch, HUBPREQ0_DCSURF_SURFACE_PITCH, PITCH) * 592 4); 593 } 594 595 return size; 596 } 597 598 static const struct amdgpu_gmc_funcs gmc_v10_0_gmc_funcs = { 599 .flush_gpu_tlb = gmc_v10_0_flush_gpu_tlb, 600 .flush_gpu_tlb_pasid = gmc_v10_0_flush_gpu_tlb_pasid, 601 .emit_flush_gpu_tlb = gmc_v10_0_emit_flush_gpu_tlb, 602 .emit_pasid_mapping = gmc_v10_0_emit_pasid_mapping, 603 .map_mtype = gmc_v10_0_map_mtype, 604 .get_vm_pde = gmc_v10_0_get_vm_pde, 605 .get_vm_pte = gmc_v10_0_get_vm_pte, 606 .get_vbios_fb_size = gmc_v10_0_get_vbios_fb_size, 607 }; 608 609 static void gmc_v10_0_set_gmc_funcs(struct amdgpu_device *adev) 610 { 611 if (adev->gmc.gmc_funcs == NULL) 612 adev->gmc.gmc_funcs = &gmc_v10_0_gmc_funcs; 613 } 614 615 static void gmc_v10_0_set_umc_funcs(struct amdgpu_device *adev) 616 { 617 switch (adev->asic_type) { 618 case CHIP_SIENNA_CICHLID: 619 adev->umc.max_ras_err_cnt_per_query = UMC_V8_7_TOTAL_CHANNEL_NUM; 620 adev->umc.channel_inst_num = UMC_V8_7_CHANNEL_INSTANCE_NUM; 621 adev->umc.umc_inst_num = UMC_V8_7_UMC_INSTANCE_NUM; 622 adev->umc.channel_offs = UMC_V8_7_PER_CHANNEL_OFFSET_SIENNA; 623 adev->umc.channel_idx_tbl = &umc_v8_7_channel_idx_tbl[0][0]; 624 adev->umc.funcs = &umc_v8_7_funcs; 625 break; 626 default: 627 break; 628 } 629 } 630 631 632 static void gmc_v10_0_set_mmhub_funcs(struct amdgpu_device *adev) 633 { 634 adev->mmhub.funcs = &mmhub_v2_0_funcs; 635 } 636 637 static int gmc_v10_0_early_init(void *handle) 638 { 639 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 640 641 gmc_v10_0_set_mmhub_funcs(adev); 642 gmc_v10_0_set_gmc_funcs(adev); 643 gmc_v10_0_set_irq_funcs(adev); 644 gmc_v10_0_set_umc_funcs(adev); 645 646 adev->gmc.shared_aperture_start = 0x2000000000000000ULL; 647 adev->gmc.shared_aperture_end = 648 adev->gmc.shared_aperture_start + (4ULL << 30) - 1; 649 adev->gmc.private_aperture_start = 0x1000000000000000ULL; 650 adev->gmc.private_aperture_end = 651 adev->gmc.private_aperture_start + (4ULL << 30) - 1; 652 653 return 0; 654 } 655 656 static int gmc_v10_0_late_init(void *handle) 657 { 658 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 659 int r; 660 661 amdgpu_bo_late_init(adev); 662 663 r = amdgpu_gmc_allocate_vm_inv_eng(adev); 664 if (r) 665 return r; 666 667 r = amdgpu_gmc_ras_late_init(adev); 668 if (r) 669 return r; 670 671 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0); 672 } 673 674 static void gmc_v10_0_vram_gtt_location(struct amdgpu_device *adev, 675 struct amdgpu_gmc *mc) 676 { 677 u64 base = 0; 678 679 if (adev->asic_type == CHIP_SIENNA_CICHLID || 680 adev->asic_type == CHIP_NAVY_FLOUNDER) 681 base = gfxhub_v2_1_get_fb_location(adev); 682 else 683 base = gfxhub_v2_0_get_fb_location(adev); 684 685 /* add the xgmi offset of the physical node */ 686 base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size; 687 688 amdgpu_gmc_vram_location(adev, &adev->gmc, base); 689 amdgpu_gmc_gart_location(adev, mc); 690 691 /* base offset of vram pages */ 692 if (adev->asic_type == CHIP_SIENNA_CICHLID || 693 adev->asic_type == CHIP_NAVY_FLOUNDER) 694 adev->vm_manager.vram_base_offset = gfxhub_v2_1_get_mc_fb_offset(adev); 695 else 696 adev->vm_manager.vram_base_offset = gfxhub_v2_0_get_mc_fb_offset(adev); 697 698 /* add the xgmi offset of the physical node */ 699 adev->vm_manager.vram_base_offset += 700 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size; 701 } 702 703 /** 704 * gmc_v10_0_mc_init - initialize the memory controller driver params 705 * 706 * @adev: amdgpu_device pointer 707 * 708 * Look up the amount of vram, vram width, and decide how to place 709 * vram and gart within the GPU's physical address space. 710 * Returns 0 for success. 711 */ 712 static int gmc_v10_0_mc_init(struct amdgpu_device *adev) 713 { 714 int r; 715 716 /* size in MB on si */ 717 adev->gmc.mc_vram_size = 718 adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL; 719 adev->gmc.real_vram_size = adev->gmc.mc_vram_size; 720 721 if (!(adev->flags & AMD_IS_APU)) { 722 r = amdgpu_device_resize_fb_bar(adev); 723 if (r) 724 return r; 725 } 726 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0); 727 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0); 728 729 /* In case the PCI BAR is larger than the actual amount of vram */ 730 adev->gmc.visible_vram_size = adev->gmc.aper_size; 731 if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size) 732 adev->gmc.visible_vram_size = adev->gmc.real_vram_size; 733 734 /* set the gart size */ 735 if (amdgpu_gart_size == -1) { 736 switch (adev->asic_type) { 737 case CHIP_NAVI10: 738 case CHIP_NAVI14: 739 case CHIP_NAVI12: 740 case CHIP_SIENNA_CICHLID: 741 case CHIP_NAVY_FLOUNDER: 742 default: 743 adev->gmc.gart_size = 512ULL << 20; 744 break; 745 } 746 } else 747 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20; 748 749 gmc_v10_0_vram_gtt_location(adev, &adev->gmc); 750 751 return 0; 752 } 753 754 static int gmc_v10_0_gart_init(struct amdgpu_device *adev) 755 { 756 int r; 757 758 if (adev->gart.bo) { 759 WARN(1, "NAVI10 PCIE GART already initialized\n"); 760 return 0; 761 } 762 763 /* Initialize common gart structure */ 764 r = amdgpu_gart_init(adev); 765 if (r) 766 return r; 767 768 adev->gart.table_size = adev->gart.num_gpu_pages * 8; 769 adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_NV10(MTYPE_UC) | 770 AMDGPU_PTE_EXECUTABLE; 771 772 return amdgpu_gart_table_vram_alloc(adev); 773 } 774 775 static int gmc_v10_0_sw_init(void *handle) 776 { 777 int r, vram_width = 0, vram_type = 0, vram_vendor = 0; 778 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 779 780 if (adev->asic_type == CHIP_SIENNA_CICHLID || 781 adev->asic_type == CHIP_NAVY_FLOUNDER) 782 gfxhub_v2_1_init(adev); 783 else 784 gfxhub_v2_0_init(adev); 785 786 adev->mmhub.funcs->init(adev); 787 788 spin_lock_init(&adev->gmc.invalidate_lock); 789 790 if (adev->asic_type == CHIP_SIENNA_CICHLID && amdgpu_emu_mode == 1) { 791 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_GDDR6; 792 adev->gmc.vram_width = 1 * 128; /* numchan * chansize */ 793 } else { 794 r = amdgpu_atomfirmware_get_vram_info(adev, 795 &vram_width, &vram_type, &vram_vendor); 796 adev->gmc.vram_width = vram_width; 797 798 adev->gmc.vram_type = vram_type; 799 adev->gmc.vram_vendor = vram_vendor; 800 } 801 802 switch (adev->asic_type) { 803 case CHIP_NAVI10: 804 case CHIP_NAVI14: 805 case CHIP_NAVI12: 806 case CHIP_SIENNA_CICHLID: 807 case CHIP_NAVY_FLOUNDER: 808 adev->num_vmhubs = 2; 809 /* 810 * To fulfill 4-level page support, 811 * vm size is 256TB (48bit), maximum size of Navi10/Navi14/Navi12, 812 * block size 512 (9bit) 813 */ 814 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 815 break; 816 default: 817 break; 818 } 819 820 /* This interrupt is VMC page fault.*/ 821 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC, 822 VMC_1_0__SRCID__VM_FAULT, 823 &adev->gmc.vm_fault); 824 825 if (r) 826 return r; 827 828 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2, 829 UTCL2_1_0__SRCID__FAULT, 830 &adev->gmc.vm_fault); 831 if (r) 832 return r; 833 834 if (!amdgpu_sriov_vf(adev)) { 835 /* interrupt sent to DF. */ 836 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0, 837 &adev->gmc.ecc_irq); 838 if (r) 839 return r; 840 } 841 842 /* 843 * Set the internal MC address mask This is the max address of the GPU's 844 * internal address space. 845 */ 846 adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */ 847 848 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44)); 849 if (r) { 850 printk(KERN_WARNING "amdgpu: No suitable DMA available.\n"); 851 return r; 852 } 853 854 if (adev->gmc.xgmi.supported) { 855 r = gfxhub_v2_1_get_xgmi_info(adev); 856 if (r) 857 return r; 858 } 859 860 r = gmc_v10_0_mc_init(adev); 861 if (r) 862 return r; 863 864 amdgpu_gmc_get_vbios_allocations(adev); 865 866 /* Memory manager */ 867 r = amdgpu_bo_init(adev); 868 if (r) 869 return r; 870 871 r = gmc_v10_0_gart_init(adev); 872 if (r) 873 return r; 874 875 /* 876 * number of VMs 877 * VMID 0 is reserved for System 878 * amdgpu graphics/compute will use VMIDs 1-7 879 * amdkfd will use VMIDs 8-15 880 */ 881 adev->vm_manager.first_kfd_vmid = 8; 882 883 amdgpu_vm_manager_init(adev); 884 885 return 0; 886 } 887 888 /** 889 * gmc_v8_0_gart_fini - vm fini callback 890 * 891 * @adev: amdgpu_device pointer 892 * 893 * Tears down the driver GART/VM setup (CIK). 894 */ 895 static void gmc_v10_0_gart_fini(struct amdgpu_device *adev) 896 { 897 amdgpu_gart_table_vram_free(adev); 898 amdgpu_gart_fini(adev); 899 } 900 901 static int gmc_v10_0_sw_fini(void *handle) 902 { 903 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 904 905 amdgpu_vm_manager_fini(adev); 906 gmc_v10_0_gart_fini(adev); 907 amdgpu_gem_force_release(adev); 908 amdgpu_bo_fini(adev); 909 910 return 0; 911 } 912 913 static void gmc_v10_0_init_golden_registers(struct amdgpu_device *adev) 914 { 915 switch (adev->asic_type) { 916 case CHIP_NAVI10: 917 case CHIP_NAVI14: 918 case CHIP_NAVI12: 919 case CHIP_SIENNA_CICHLID: 920 case CHIP_NAVY_FLOUNDER: 921 break; 922 default: 923 break; 924 } 925 } 926 927 /** 928 * gmc_v10_0_gart_enable - gart enable 929 * 930 * @adev: amdgpu_device pointer 931 */ 932 static int gmc_v10_0_gart_enable(struct amdgpu_device *adev) 933 { 934 int r; 935 bool value; 936 u32 tmp; 937 938 if (adev->gart.bo == NULL) { 939 dev_err(adev->dev, "No VRAM object for PCIE GART.\n"); 940 return -EINVAL; 941 } 942 943 r = amdgpu_gart_table_vram_pin(adev); 944 if (r) 945 return r; 946 947 if (adev->asic_type == CHIP_SIENNA_CICHLID || 948 adev->asic_type == CHIP_NAVY_FLOUNDER) 949 r = gfxhub_v2_1_gart_enable(adev); 950 else 951 r = gfxhub_v2_0_gart_enable(adev); 952 if (r) 953 return r; 954 955 r = adev->mmhub.funcs->gart_enable(adev); 956 if (r) 957 return r; 958 959 tmp = RREG32_SOC15(HDP, 0, mmHDP_MISC_CNTL); 960 tmp |= HDP_MISC_CNTL__FLUSH_INVALIDATE_CACHE_MASK; 961 WREG32_SOC15(HDP, 0, mmHDP_MISC_CNTL, tmp); 962 963 tmp = RREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL); 964 WREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL, tmp); 965 966 /* Flush HDP after it is initialized */ 967 adev->nbio.funcs->hdp_flush(adev, NULL); 968 969 value = (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) ? 970 false : true; 971 972 if (adev->asic_type == CHIP_SIENNA_CICHLID || 973 adev->asic_type == CHIP_NAVY_FLOUNDER) 974 gfxhub_v2_1_set_fault_enable_default(adev, value); 975 else 976 gfxhub_v2_0_set_fault_enable_default(adev, value); 977 adev->mmhub.funcs->set_fault_enable_default(adev, value); 978 gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_MMHUB_0, 0); 979 gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_GFXHUB_0, 0); 980 981 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n", 982 (unsigned)(adev->gmc.gart_size >> 20), 983 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo)); 984 985 adev->gart.ready = true; 986 987 return 0; 988 } 989 990 static int gmc_v10_0_hw_init(void *handle) 991 { 992 int r; 993 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 994 995 /* The sequence of these two function calls matters.*/ 996 gmc_v10_0_init_golden_registers(adev); 997 998 r = gmc_v10_0_gart_enable(adev); 999 if (r) 1000 return r; 1001 1002 if (adev->umc.funcs && adev->umc.funcs->init_registers) 1003 adev->umc.funcs->init_registers(adev); 1004 1005 return 0; 1006 } 1007 1008 /** 1009 * gmc_v10_0_gart_disable - gart disable 1010 * 1011 * @adev: amdgpu_device pointer 1012 * 1013 * This disables all VM page table. 1014 */ 1015 static void gmc_v10_0_gart_disable(struct amdgpu_device *adev) 1016 { 1017 if (adev->asic_type == CHIP_SIENNA_CICHLID || 1018 adev->asic_type == CHIP_NAVY_FLOUNDER) 1019 gfxhub_v2_1_gart_disable(adev); 1020 else 1021 gfxhub_v2_0_gart_disable(adev); 1022 adev->mmhub.funcs->gart_disable(adev); 1023 amdgpu_gart_table_vram_unpin(adev); 1024 } 1025 1026 static int gmc_v10_0_hw_fini(void *handle) 1027 { 1028 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1029 1030 if (amdgpu_sriov_vf(adev)) { 1031 /* full access mode, so don't touch any GMC register */ 1032 DRM_DEBUG("For SRIOV client, shouldn't do anything.\n"); 1033 return 0; 1034 } 1035 1036 amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0); 1037 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0); 1038 gmc_v10_0_gart_disable(adev); 1039 1040 return 0; 1041 } 1042 1043 static int gmc_v10_0_suspend(void *handle) 1044 { 1045 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1046 1047 gmc_v10_0_hw_fini(adev); 1048 1049 return 0; 1050 } 1051 1052 static int gmc_v10_0_resume(void *handle) 1053 { 1054 int r; 1055 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1056 1057 r = gmc_v10_0_hw_init(adev); 1058 if (r) 1059 return r; 1060 1061 amdgpu_vmid_reset_all(adev); 1062 1063 return 0; 1064 } 1065 1066 static bool gmc_v10_0_is_idle(void *handle) 1067 { 1068 /* MC is always ready in GMC v10.*/ 1069 return true; 1070 } 1071 1072 static int gmc_v10_0_wait_for_idle(void *handle) 1073 { 1074 /* There is no need to wait for MC idle in GMC v10.*/ 1075 return 0; 1076 } 1077 1078 static int gmc_v10_0_soft_reset(void *handle) 1079 { 1080 return 0; 1081 } 1082 1083 static int gmc_v10_0_set_clockgating_state(void *handle, 1084 enum amd_clockgating_state state) 1085 { 1086 int r; 1087 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1088 1089 r = adev->mmhub.funcs->set_clockgating(adev, state); 1090 if (r) 1091 return r; 1092 1093 if (adev->asic_type == CHIP_SIENNA_CICHLID || 1094 adev->asic_type == CHIP_NAVY_FLOUNDER) 1095 return athub_v2_1_set_clockgating(adev, state); 1096 else 1097 return athub_v2_0_set_clockgating(adev, state); 1098 } 1099 1100 static void gmc_v10_0_get_clockgating_state(void *handle, u32 *flags) 1101 { 1102 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1103 1104 adev->mmhub.funcs->get_clockgating(adev, flags); 1105 1106 if (adev->asic_type == CHIP_SIENNA_CICHLID || 1107 adev->asic_type == CHIP_NAVY_FLOUNDER) 1108 athub_v2_1_get_clockgating(adev, flags); 1109 else 1110 athub_v2_0_get_clockgating(adev, flags); 1111 } 1112 1113 static int gmc_v10_0_set_powergating_state(void *handle, 1114 enum amd_powergating_state state) 1115 { 1116 return 0; 1117 } 1118 1119 const struct amd_ip_funcs gmc_v10_0_ip_funcs = { 1120 .name = "gmc_v10_0", 1121 .early_init = gmc_v10_0_early_init, 1122 .late_init = gmc_v10_0_late_init, 1123 .sw_init = gmc_v10_0_sw_init, 1124 .sw_fini = gmc_v10_0_sw_fini, 1125 .hw_init = gmc_v10_0_hw_init, 1126 .hw_fini = gmc_v10_0_hw_fini, 1127 .suspend = gmc_v10_0_suspend, 1128 .resume = gmc_v10_0_resume, 1129 .is_idle = gmc_v10_0_is_idle, 1130 .wait_for_idle = gmc_v10_0_wait_for_idle, 1131 .soft_reset = gmc_v10_0_soft_reset, 1132 .set_clockgating_state = gmc_v10_0_set_clockgating_state, 1133 .set_powergating_state = gmc_v10_0_set_powergating_state, 1134 .get_clockgating_state = gmc_v10_0_get_clockgating_state, 1135 }; 1136 1137 const struct amdgpu_ip_block_version gmc_v10_0_ip_block = 1138 { 1139 .type = AMD_IP_BLOCK_TYPE_GMC, 1140 .major = 10, 1141 .minor = 0, 1142 .rev = 0, 1143 .funcs = &gmc_v10_0_ip_funcs, 1144 }; 1145