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 26 #include <drm/drm_cache.h> 27 28 #include "amdgpu.h" 29 #include "amdgpu_atomfirmware.h" 30 #include "gmc_v10_0.h" 31 #include "umc_v8_7.h" 32 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 "gfxhub_v2_1.h" 49 #include "mmhub_v2_0.h" 50 #include "mmhub_v2_3.h" 51 #include "athub_v2_0.h" 52 #include "athub_v2_1.h" 53 54 static int gmc_v10_0_ecc_interrupt_state(struct amdgpu_device *adev, 55 struct amdgpu_irq_src *src, 56 unsigned int type, 57 enum amdgpu_interrupt_state state) 58 { 59 return 0; 60 } 61 62 static int 63 gmc_v10_0_vm_fault_interrupt_state(struct amdgpu_device *adev, 64 struct amdgpu_irq_src *src, unsigned int type, 65 enum amdgpu_interrupt_state state) 66 { 67 switch (state) { 68 case AMDGPU_IRQ_STATE_DISABLE: 69 /* MM HUB */ 70 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB0(0), false); 71 /* GFX HUB */ 72 /* This works because this interrupt is only 73 * enabled at init/resume and disabled in 74 * fini/suspend, so the overall state doesn't 75 * change over the course of suspend/resume. 76 */ 77 if (!adev->in_s0ix) 78 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB(0), false); 79 break; 80 case AMDGPU_IRQ_STATE_ENABLE: 81 /* MM HUB */ 82 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB0(0), true); 83 /* GFX HUB */ 84 /* This works because this interrupt is only 85 * enabled at init/resume and disabled in 86 * fini/suspend, so the overall state doesn't 87 * change over the course of suspend/resume. 88 */ 89 if (!adev->in_s0ix) 90 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB(0), true); 91 break; 92 default: 93 break; 94 } 95 96 return 0; 97 } 98 99 static int gmc_v10_0_process_interrupt(struct amdgpu_device *adev, 100 struct amdgpu_irq_src *source, 101 struct amdgpu_iv_entry *entry) 102 { 103 uint32_t vmhub_index = entry->client_id == SOC15_IH_CLIENTID_VMC ? 104 AMDGPU_MMHUB0(0) : AMDGPU_GFXHUB(0); 105 struct amdgpu_vmhub *hub = &adev->vmhub[vmhub_index]; 106 bool retry_fault = !!(entry->src_data[1] & 0x80); 107 bool write_fault = !!(entry->src_data[1] & 0x20); 108 struct amdgpu_task_info *task_info; 109 uint32_t status = 0; 110 u64 addr; 111 112 addr = (u64)entry->src_data[0] << 12; 113 addr |= ((u64)entry->src_data[1] & 0xf) << 44; 114 115 if (retry_fault) { 116 /* Returning 1 here also prevents sending the IV to the KFD */ 117 118 /* Process it onyl if it's the first fault for this address */ 119 if (entry->ih != &adev->irq.ih_soft && 120 amdgpu_gmc_filter_faults(adev, entry->ih, addr, entry->pasid, 121 entry->timestamp)) 122 return 1; 123 124 /* Delegate it to a different ring if the hardware hasn't 125 * already done it. 126 */ 127 if (entry->ih == &adev->irq.ih) { 128 amdgpu_irq_delegate(adev, entry, 8); 129 return 1; 130 } 131 132 /* Try to handle the recoverable page faults by filling page 133 * tables 134 */ 135 if (amdgpu_vm_handle_fault(adev, entry->pasid, 0, 0, addr, 136 entry->timestamp, write_fault)) 137 return 1; 138 } 139 140 if (!amdgpu_sriov_vf(adev)) { 141 /* 142 * Issue a dummy read to wait for the status register to 143 * be updated to avoid reading an incorrect value due to 144 * the new fast GRBM interface. 145 */ 146 if ((entry->vmid_src == AMDGPU_GFXHUB(0)) && 147 (amdgpu_ip_version(adev, GC_HWIP, 0) < 148 IP_VERSION(10, 3, 0))) 149 RREG32(hub->vm_l2_pro_fault_status); 150 151 status = RREG32(hub->vm_l2_pro_fault_status); 152 WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1); 153 154 amdgpu_vm_update_fault_cache(adev, entry->pasid, addr, status, 155 entry->vmid_src ? AMDGPU_MMHUB0(0) : AMDGPU_GFXHUB(0)); 156 } 157 158 if (!printk_ratelimit()) 159 return 0; 160 161 dev_err(adev->dev, 162 "[%s] page fault (src_id:%u ring:%u vmid:%u pasid:%u)\n", 163 entry->vmid_src ? "mmhub" : "gfxhub", 164 entry->src_id, entry->ring_id, entry->vmid, entry->pasid); 165 task_info = amdgpu_vm_get_task_info_pasid(adev, entry->pasid); 166 if (task_info) { 167 dev_err(adev->dev, 168 " in process %s pid %d thread %s pid %d\n", 169 task_info->process_name, task_info->tgid, 170 task_info->task_name, task_info->pid); 171 amdgpu_vm_put_task_info(task_info); 172 } 173 174 dev_err(adev->dev, " in page starting at address 0x%016llx from client 0x%x (%s)\n", 175 addr, entry->client_id, 176 soc15_ih_clientid_name[entry->client_id]); 177 178 /* Only print L2 fault status if the status register could be read and 179 * contains useful information 180 */ 181 if (status != 0) 182 hub->vmhub_funcs->print_l2_protection_fault_status(adev, 183 status); 184 185 return 0; 186 } 187 188 static const struct amdgpu_irq_src_funcs gmc_v10_0_irq_funcs = { 189 .set = gmc_v10_0_vm_fault_interrupt_state, 190 .process = gmc_v10_0_process_interrupt, 191 }; 192 193 static const struct amdgpu_irq_src_funcs gmc_v10_0_ecc_funcs = { 194 .set = gmc_v10_0_ecc_interrupt_state, 195 .process = amdgpu_umc_process_ecc_irq, 196 }; 197 198 static void gmc_v10_0_set_irq_funcs(struct amdgpu_device *adev) 199 { 200 adev->gmc.vm_fault.num_types = 1; 201 adev->gmc.vm_fault.funcs = &gmc_v10_0_irq_funcs; 202 203 if (!amdgpu_sriov_vf(adev)) { 204 adev->gmc.ecc_irq.num_types = 1; 205 adev->gmc.ecc_irq.funcs = &gmc_v10_0_ecc_funcs; 206 } 207 } 208 209 /** 210 * gmc_v10_0_use_invalidate_semaphore - judge whether to use semaphore 211 * 212 * @adev: amdgpu_device pointer 213 * @vmhub: vmhub type 214 * 215 */ 216 static bool gmc_v10_0_use_invalidate_semaphore(struct amdgpu_device *adev, 217 uint32_t vmhub) 218 { 219 return ((vmhub == AMDGPU_MMHUB0(0)) && 220 (!amdgpu_sriov_vf(adev))); 221 } 222 223 static bool gmc_v10_0_get_atc_vmid_pasid_mapping_info( 224 struct amdgpu_device *adev, 225 uint8_t vmid, uint16_t *p_pasid) 226 { 227 uint32_t value; 228 229 value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING) 230 + vmid); 231 *p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK; 232 233 return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK); 234 } 235 236 /* 237 * GART 238 * VMID 0 is the physical GPU addresses as used by the kernel. 239 * VMIDs 1-15 are used for userspace clients and are handled 240 * by the amdgpu vm/hsa code. 241 */ 242 243 /** 244 * gmc_v10_0_flush_gpu_tlb - gart tlb flush callback 245 * 246 * @adev: amdgpu_device pointer 247 * @vmid: vm instance to flush 248 * @vmhub: vmhub type 249 * @flush_type: the flush type 250 * 251 * Flush the TLB for the requested page table. 252 */ 253 static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, 254 uint32_t vmhub, uint32_t flush_type) 255 { 256 bool use_semaphore = gmc_v10_0_use_invalidate_semaphore(adev, vmhub); 257 struct amdgpu_vmhub *hub = &adev->vmhub[vmhub]; 258 u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type); 259 /* Use register 17 for GART */ 260 const unsigned int eng = 17; 261 unsigned char hub_ip = 0; 262 u32 sem, req, ack; 263 unsigned int i; 264 u32 tmp; 265 266 sem = hub->vm_inv_eng0_sem + hub->eng_distance * eng; 267 req = hub->vm_inv_eng0_req + hub->eng_distance * eng; 268 ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng; 269 270 /* flush hdp cache */ 271 amdgpu_device_flush_hdp(adev, NULL); 272 273 /* This is necessary for SRIOV as well as for GFXOFF to function 274 * properly under bare metal 275 */ 276 if (adev->gfx.kiq[0].ring.sched.ready && !adev->enable_mes && 277 (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev))) { 278 amdgpu_gmc_fw_reg_write_reg_wait(adev, req, ack, inv_req, 279 1 << vmid, GET_INST(GC, 0)); 280 return; 281 } 282 283 /* This path is needed before KIQ/MES/GFXOFF are set up */ 284 hub_ip = (vmhub == AMDGPU_GFXHUB(0)) ? GC_HWIP : MMHUB_HWIP; 285 286 spin_lock(&adev->gmc.invalidate_lock); 287 /* 288 * It may lose gpuvm invalidate acknowldege state across power-gating 289 * off cycle, add semaphore acquire before invalidation and semaphore 290 * release after invalidation to avoid entering power gated state 291 * to WA the Issue 292 */ 293 294 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 295 if (use_semaphore) { 296 for (i = 0; i < adev->usec_timeout; i++) { 297 /* a read return value of 1 means semaphore acuqire */ 298 tmp = RREG32_RLC_NO_KIQ(sem, hub_ip); 299 if (tmp & 0x1) 300 break; 301 udelay(1); 302 } 303 304 if (i >= adev->usec_timeout) 305 DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n"); 306 } 307 308 WREG32_RLC_NO_KIQ(req, inv_req, hub_ip); 309 310 /* 311 * Issue a dummy read to wait for the ACK register to be cleared 312 * to avoid a false ACK due to the new fast GRBM interface. 313 */ 314 if ((vmhub == AMDGPU_GFXHUB(0)) && 315 (amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(10, 3, 0))) 316 RREG32_RLC_NO_KIQ(req, hub_ip); 317 318 /* Wait for ACK with a delay.*/ 319 for (i = 0; i < adev->usec_timeout; i++) { 320 tmp = RREG32_RLC_NO_KIQ(ack, hub_ip); 321 tmp &= 1 << vmid; 322 if (tmp) 323 break; 324 325 udelay(1); 326 } 327 328 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 329 if (use_semaphore) 330 WREG32_RLC_NO_KIQ(sem, 0, hub_ip); 331 332 spin_unlock(&adev->gmc.invalidate_lock); 333 334 if (i >= adev->usec_timeout) 335 dev_err(adev->dev, "Timeout waiting for VM flush hub: %d!\n", 336 vmhub); 337 } 338 339 /** 340 * gmc_v10_0_flush_gpu_tlb_pasid - tlb flush via pasid 341 * 342 * @adev: amdgpu_device pointer 343 * @pasid: pasid to be flush 344 * @flush_type: the flush type 345 * @all_hub: Used with PACKET3_INVALIDATE_TLBS_ALL_HUB() 346 * @inst: is used to select which instance of KIQ to use for the invalidation 347 * 348 * Flush the TLB for the requested pasid. 349 */ 350 static void gmc_v10_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, 351 uint16_t pasid, uint32_t flush_type, 352 bool all_hub, uint32_t inst) 353 { 354 uint16_t queried; 355 int vmid, i; 356 357 for (vmid = 1; vmid < AMDGPU_NUM_VMID; vmid++) { 358 bool valid; 359 360 valid = gmc_v10_0_get_atc_vmid_pasid_mapping_info(adev, vmid, 361 &queried); 362 if (!valid || queried != pasid) 363 continue; 364 365 if (all_hub) { 366 for_each_set_bit(i, adev->vmhubs_mask, 367 AMDGPU_MAX_VMHUBS) 368 gmc_v10_0_flush_gpu_tlb(adev, vmid, i, 369 flush_type); 370 } else { 371 gmc_v10_0_flush_gpu_tlb(adev, vmid, AMDGPU_GFXHUB(0), 372 flush_type); 373 } 374 } 375 } 376 377 static uint64_t gmc_v10_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring, 378 unsigned int vmid, uint64_t pd_addr) 379 { 380 bool use_semaphore = gmc_v10_0_use_invalidate_semaphore(ring->adev, ring->vm_hub); 381 struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->vm_hub]; 382 uint32_t req = hub->vmhub_funcs->get_invalidate_req(vmid, 0); 383 unsigned int eng = ring->vm_inv_eng; 384 385 /* 386 * It may lose gpuvm invalidate acknowldege state across power-gating 387 * off cycle, add semaphore acquire before invalidation and semaphore 388 * release after invalidation to avoid entering power gated state 389 * to WA the Issue 390 */ 391 392 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 393 if (use_semaphore) 394 /* a read return value of 1 means semaphore acuqire */ 395 amdgpu_ring_emit_reg_wait(ring, 396 hub->vm_inv_eng0_sem + 397 hub->eng_distance * eng, 0x1, 0x1); 398 399 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 + 400 (hub->ctx_addr_distance * vmid), 401 lower_32_bits(pd_addr)); 402 403 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + 404 (hub->ctx_addr_distance * vmid), 405 upper_32_bits(pd_addr)); 406 407 amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req + 408 hub->eng_distance * eng, 409 hub->vm_inv_eng0_ack + 410 hub->eng_distance * eng, 411 req, 1 << vmid); 412 413 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 414 if (use_semaphore) 415 /* 416 * add semaphore release after invalidation, 417 * write with 0 means semaphore release 418 */ 419 amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem + 420 hub->eng_distance * eng, 0); 421 422 return pd_addr; 423 } 424 425 static void gmc_v10_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned int vmid, 426 unsigned int pasid) 427 { 428 struct amdgpu_device *adev = ring->adev; 429 uint32_t reg; 430 431 if (ring->vm_hub == AMDGPU_GFXHUB(0)) 432 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid; 433 else 434 reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid; 435 436 amdgpu_ring_emit_wreg(ring, reg, pasid); 437 } 438 439 /* 440 * PTE format on NAVI 10: 441 * 63:59 reserved 442 * 58 reserved and for sienna_cichlid is used for MALL noalloc 443 * 57 reserved 444 * 56 F 445 * 55 L 446 * 54 reserved 447 * 53:52 SW 448 * 51 T 449 * 50:48 mtype 450 * 47:12 4k physical page base address 451 * 11:7 fragment 452 * 6 write 453 * 5 read 454 * 4 exe 455 * 3 Z 456 * 2 snooped 457 * 1 system 458 * 0 valid 459 * 460 * PDE format on NAVI 10: 461 * 63:59 block fragment size 462 * 58:55 reserved 463 * 54 P 464 * 53:48 reserved 465 * 47:6 physical base address of PD or PTE 466 * 5:3 reserved 467 * 2 C 468 * 1 system 469 * 0 valid 470 */ 471 472 static uint64_t gmc_v10_0_map_mtype(struct amdgpu_device *adev, uint32_t flags) 473 { 474 switch (flags) { 475 case AMDGPU_VM_MTYPE_DEFAULT: 476 return AMDGPU_PTE_MTYPE_NV10(0ULL, MTYPE_NC); 477 case AMDGPU_VM_MTYPE_NC: 478 return AMDGPU_PTE_MTYPE_NV10(0ULL, MTYPE_NC); 479 case AMDGPU_VM_MTYPE_WC: 480 return AMDGPU_PTE_MTYPE_NV10(0ULL, MTYPE_WC); 481 case AMDGPU_VM_MTYPE_CC: 482 return AMDGPU_PTE_MTYPE_NV10(0ULL, MTYPE_CC); 483 case AMDGPU_VM_MTYPE_UC: 484 return AMDGPU_PTE_MTYPE_NV10(0ULL, MTYPE_UC); 485 default: 486 return AMDGPU_PTE_MTYPE_NV10(0ULL, MTYPE_NC); 487 } 488 } 489 490 static void gmc_v10_0_get_vm_pde(struct amdgpu_device *adev, int level, 491 uint64_t *addr, uint64_t *flags) 492 { 493 if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM)) 494 *addr = amdgpu_gmc_vram_mc2pa(adev, *addr); 495 BUG_ON(*addr & 0xFFFF00000000003FULL); 496 497 if (!adev->gmc.translate_further) 498 return; 499 500 if (level == AMDGPU_VM_PDB1) { 501 /* Set the block fragment size */ 502 if (!(*flags & AMDGPU_PDE_PTE)) 503 *flags |= AMDGPU_PDE_BFS(0x9); 504 505 } else if (level == AMDGPU_VM_PDB0) { 506 if (*flags & AMDGPU_PDE_PTE) 507 *flags &= ~AMDGPU_PDE_PTE; 508 else 509 *flags |= AMDGPU_PTE_TF; 510 } 511 } 512 513 static void gmc_v10_0_get_vm_pte(struct amdgpu_device *adev, 514 struct amdgpu_bo_va_mapping *mapping, 515 uint64_t *flags) 516 { 517 struct amdgpu_bo *bo = mapping->bo_va->base.bo; 518 519 *flags &= ~AMDGPU_PTE_EXECUTABLE; 520 *flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE; 521 522 *flags &= ~AMDGPU_PTE_MTYPE_NV10_MASK; 523 *flags |= (mapping->flags & AMDGPU_PTE_MTYPE_NV10_MASK); 524 525 *flags &= ~AMDGPU_PTE_NOALLOC; 526 *flags |= (mapping->flags & AMDGPU_PTE_NOALLOC); 527 528 if (mapping->flags & AMDGPU_PTE_PRT) { 529 *flags |= AMDGPU_PTE_PRT; 530 *flags |= AMDGPU_PTE_SNOOPED; 531 *flags |= AMDGPU_PTE_LOG; 532 *flags |= AMDGPU_PTE_SYSTEM; 533 *flags &= ~AMDGPU_PTE_VALID; 534 } 535 536 if (bo && bo->flags & (AMDGPU_GEM_CREATE_COHERENT | 537 AMDGPU_GEM_CREATE_EXT_COHERENT | 538 AMDGPU_GEM_CREATE_UNCACHED)) 539 *flags = AMDGPU_PTE_MTYPE_NV10(*flags, MTYPE_UC); 540 } 541 542 static unsigned int gmc_v10_0_get_vbios_fb_size(struct amdgpu_device *adev) 543 { 544 u32 d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL); 545 unsigned int size; 546 547 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) { 548 size = AMDGPU_VBIOS_VGA_ALLOCATION; 549 } else { 550 u32 viewport; 551 u32 pitch; 552 553 viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION); 554 pitch = RREG32_SOC15(DCE, 0, mmHUBPREQ0_DCSURF_SURFACE_PITCH); 555 size = (REG_GET_FIELD(viewport, 556 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) * 557 REG_GET_FIELD(pitch, HUBPREQ0_DCSURF_SURFACE_PITCH, PITCH) * 558 4); 559 } 560 561 return size; 562 } 563 564 static const struct amdgpu_gmc_funcs gmc_v10_0_gmc_funcs = { 565 .flush_gpu_tlb = gmc_v10_0_flush_gpu_tlb, 566 .flush_gpu_tlb_pasid = gmc_v10_0_flush_gpu_tlb_pasid, 567 .emit_flush_gpu_tlb = gmc_v10_0_emit_flush_gpu_tlb, 568 .emit_pasid_mapping = gmc_v10_0_emit_pasid_mapping, 569 .map_mtype = gmc_v10_0_map_mtype, 570 .get_vm_pde = gmc_v10_0_get_vm_pde, 571 .get_vm_pte = gmc_v10_0_get_vm_pte, 572 .get_vbios_fb_size = gmc_v10_0_get_vbios_fb_size, 573 }; 574 575 static void gmc_v10_0_set_gmc_funcs(struct amdgpu_device *adev) 576 { 577 if (adev->gmc.gmc_funcs == NULL) 578 adev->gmc.gmc_funcs = &gmc_v10_0_gmc_funcs; 579 } 580 581 static void gmc_v10_0_set_umc_funcs(struct amdgpu_device *adev) 582 { 583 switch (amdgpu_ip_version(adev, UMC_HWIP, 0)) { 584 case IP_VERSION(8, 7, 0): 585 adev->umc.max_ras_err_cnt_per_query = UMC_V8_7_TOTAL_CHANNEL_NUM; 586 adev->umc.channel_inst_num = UMC_V8_7_CHANNEL_INSTANCE_NUM; 587 adev->umc.umc_inst_num = UMC_V8_7_UMC_INSTANCE_NUM; 588 adev->umc.channel_offs = UMC_V8_7_PER_CHANNEL_OFFSET_SIENNA; 589 adev->umc.retire_unit = 1; 590 adev->umc.channel_idx_tbl = &umc_v8_7_channel_idx_tbl[0][0]; 591 adev->umc.ras = &umc_v8_7_ras; 592 break; 593 default: 594 break; 595 } 596 } 597 598 static void gmc_v10_0_set_mmhub_funcs(struct amdgpu_device *adev) 599 { 600 switch (amdgpu_ip_version(adev, MMHUB_HWIP, 0)) { 601 case IP_VERSION(2, 3, 0): 602 case IP_VERSION(2, 4, 0): 603 case IP_VERSION(2, 4, 1): 604 adev->mmhub.funcs = &mmhub_v2_3_funcs; 605 break; 606 default: 607 adev->mmhub.funcs = &mmhub_v2_0_funcs; 608 break; 609 } 610 } 611 612 static void gmc_v10_0_set_gfxhub_funcs(struct amdgpu_device *adev) 613 { 614 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 615 case IP_VERSION(10, 3, 0): 616 case IP_VERSION(10, 3, 2): 617 case IP_VERSION(10, 3, 1): 618 case IP_VERSION(10, 3, 4): 619 case IP_VERSION(10, 3, 5): 620 case IP_VERSION(10, 3, 6): 621 case IP_VERSION(10, 3, 3): 622 case IP_VERSION(10, 3, 7): 623 adev->gfxhub.funcs = &gfxhub_v2_1_funcs; 624 break; 625 default: 626 adev->gfxhub.funcs = &gfxhub_v2_0_funcs; 627 break; 628 } 629 } 630 631 632 static int gmc_v10_0_early_init(struct amdgpu_ip_block *ip_block) 633 { 634 struct amdgpu_device *adev = ip_block->adev; 635 636 gmc_v10_0_set_mmhub_funcs(adev); 637 gmc_v10_0_set_gfxhub_funcs(adev); 638 gmc_v10_0_set_gmc_funcs(adev); 639 gmc_v10_0_set_irq_funcs(adev); 640 gmc_v10_0_set_umc_funcs(adev); 641 642 adev->gmc.shared_aperture_start = 0x2000000000000000ULL; 643 adev->gmc.shared_aperture_end = 644 adev->gmc.shared_aperture_start + (4ULL << 30) - 1; 645 adev->gmc.private_aperture_start = 0x1000000000000000ULL; 646 adev->gmc.private_aperture_end = 647 adev->gmc.private_aperture_start + (4ULL << 30) - 1; 648 adev->gmc.noretry_flags = AMDGPU_VM_NORETRY_FLAGS_TF; 649 650 return 0; 651 } 652 653 static int gmc_v10_0_late_init(struct amdgpu_ip_block *ip_block) 654 { 655 struct amdgpu_device *adev = ip_block->adev; 656 int r; 657 658 r = amdgpu_gmc_allocate_vm_inv_eng(adev); 659 if (r) 660 return r; 661 662 r = amdgpu_gmc_ras_late_init(adev); 663 if (r) 664 return r; 665 666 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0); 667 } 668 669 static void gmc_v10_0_vram_gtt_location(struct amdgpu_device *adev, 670 struct amdgpu_gmc *mc) 671 { 672 u64 base = 0; 673 674 base = adev->gfxhub.funcs->get_fb_location(adev); 675 676 /* add the xgmi offset of the physical node */ 677 base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size; 678 679 amdgpu_gmc_set_agp_default(adev, mc); 680 amdgpu_gmc_vram_location(adev, &adev->gmc, base); 681 amdgpu_gmc_gart_location(adev, mc, AMDGPU_GART_PLACEMENT_BEST_FIT); 682 if (!amdgpu_sriov_vf(adev) && (amdgpu_agp == 1)) 683 amdgpu_gmc_agp_location(adev, mc); 684 685 /* base offset of vram pages */ 686 adev->vm_manager.vram_base_offset = adev->gfxhub.funcs->get_mc_fb_offset(adev); 687 688 /* add the xgmi offset of the physical node */ 689 adev->vm_manager.vram_base_offset += 690 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size; 691 } 692 693 /** 694 * gmc_v10_0_mc_init - initialize the memory controller driver params 695 * 696 * @adev: amdgpu_device pointer 697 * 698 * Look up the amount of vram, vram width, and decide how to place 699 * vram and gart within the GPU's physical address space. 700 * Returns 0 for success. 701 */ 702 static int gmc_v10_0_mc_init(struct amdgpu_device *adev) 703 { 704 int r; 705 706 /* size in MB on si */ 707 adev->gmc.mc_vram_size = 708 adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL; 709 adev->gmc.real_vram_size = adev->gmc.mc_vram_size; 710 711 if (!(adev->flags & AMD_IS_APU)) { 712 r = amdgpu_device_resize_fb_bar(adev); 713 if (r) 714 return r; 715 } 716 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0); 717 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0); 718 719 #ifdef CONFIG_X86_64 720 if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)) { 721 adev->gmc.aper_base = adev->gfxhub.funcs->get_mc_fb_offset(adev); 722 adev->gmc.aper_size = adev->gmc.real_vram_size; 723 } 724 #endif 725 726 adev->gmc.visible_vram_size = adev->gmc.aper_size; 727 728 /* set the gart size */ 729 if (amdgpu_gart_size == -1) { 730 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 731 default: 732 adev->gmc.gart_size = 512ULL << 20; 733 break; 734 case IP_VERSION(10, 3, 1): /* DCE SG support */ 735 case IP_VERSION(10, 3, 3): /* DCE SG support */ 736 case IP_VERSION(10, 3, 6): /* DCE SG support */ 737 case IP_VERSION(10, 3, 7): /* DCE SG support */ 738 adev->gmc.gart_size = 1024ULL << 20; 739 break; 740 } 741 } else { 742 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20; 743 } 744 745 gmc_v10_0_vram_gtt_location(adev, &adev->gmc); 746 747 return 0; 748 } 749 750 static int gmc_v10_0_gart_init(struct amdgpu_device *adev) 751 { 752 int r; 753 754 if (adev->gart.bo) { 755 WARN(1, "NAVI10 PCIE GART already initialized\n"); 756 return 0; 757 } 758 759 /* Initialize common gart structure */ 760 r = amdgpu_gart_init(adev); 761 if (r) 762 return r; 763 764 adev->gart.table_size = adev->gart.num_gpu_pages * 8; 765 adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_NV10(0ULL, MTYPE_UC) | 766 AMDGPU_PTE_EXECUTABLE; 767 768 return amdgpu_gart_table_vram_alloc(adev); 769 } 770 771 static int gmc_v10_0_sw_init(struct amdgpu_ip_block *ip_block) 772 { 773 int r, vram_width = 0, vram_type = 0, vram_vendor = 0; 774 struct amdgpu_device *adev = ip_block->adev; 775 776 adev->gfxhub.funcs->init(adev); 777 778 adev->mmhub.funcs->init(adev); 779 780 spin_lock_init(&adev->gmc.invalidate_lock); 781 782 if ((adev->flags & AMD_IS_APU) && amdgpu_emu_mode == 1) { 783 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_DDR4; 784 adev->gmc.vram_width = 64; 785 } else if (amdgpu_emu_mode == 1) { 786 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_GDDR6; 787 adev->gmc.vram_width = 1 * 128; /* numchan * chansize */ 788 } else { 789 r = amdgpu_atomfirmware_get_vram_info(adev, 790 &vram_width, &vram_type, &vram_vendor); 791 adev->gmc.vram_width = vram_width; 792 793 adev->gmc.vram_type = vram_type; 794 adev->gmc.vram_vendor = vram_vendor; 795 } 796 797 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 798 case IP_VERSION(10, 3, 0): 799 adev->gmc.mall_size = 128 * 1024 * 1024; 800 break; 801 case IP_VERSION(10, 3, 2): 802 adev->gmc.mall_size = 96 * 1024 * 1024; 803 break; 804 case IP_VERSION(10, 3, 4): 805 adev->gmc.mall_size = 32 * 1024 * 1024; 806 break; 807 case IP_VERSION(10, 3, 5): 808 adev->gmc.mall_size = 16 * 1024 * 1024; 809 break; 810 default: 811 adev->gmc.mall_size = 0; 812 break; 813 } 814 815 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 816 case IP_VERSION(10, 1, 10): 817 case IP_VERSION(10, 1, 1): 818 case IP_VERSION(10, 1, 2): 819 case IP_VERSION(10, 1, 3): 820 case IP_VERSION(10, 1, 4): 821 case IP_VERSION(10, 3, 0): 822 case IP_VERSION(10, 3, 2): 823 case IP_VERSION(10, 3, 1): 824 case IP_VERSION(10, 3, 4): 825 case IP_VERSION(10, 3, 5): 826 case IP_VERSION(10, 3, 6): 827 case IP_VERSION(10, 3, 3): 828 case IP_VERSION(10, 3, 7): 829 set_bit(AMDGPU_GFXHUB(0), adev->vmhubs_mask); 830 set_bit(AMDGPU_MMHUB0(0), adev->vmhubs_mask); 831 /* 832 * To fulfill 4-level page support, 833 * vm size is 256TB (48bit), maximum size of Navi10/Navi14/Navi12, 834 * block size 512 (9bit) 835 */ 836 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 837 break; 838 default: 839 break; 840 } 841 842 /* This interrupt is VMC page fault.*/ 843 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC, 844 VMC_1_0__SRCID__VM_FAULT, 845 &adev->gmc.vm_fault); 846 847 if (r) 848 return r; 849 850 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2, 851 UTCL2_1_0__SRCID__FAULT, 852 &adev->gmc.vm_fault); 853 if (r) 854 return r; 855 856 if (!amdgpu_sriov_vf(adev)) { 857 /* interrupt sent to DF. */ 858 r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0, 859 &adev->gmc.ecc_irq); 860 if (r) 861 return r; 862 } 863 864 /* 865 * Set the internal MC address mask This is the max address of the GPU's 866 * internal address space. 867 */ 868 adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */ 869 870 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44)); 871 if (r) { 872 dev_warn(adev->dev, "amdgpu: No suitable DMA available.\n"); 873 return r; 874 } 875 876 adev->need_swiotlb = drm_need_swiotlb(44); 877 878 r = gmc_v10_0_mc_init(adev); 879 if (r) 880 return r; 881 882 amdgpu_gmc_get_vbios_allocations(adev); 883 884 /* Memory manager */ 885 r = amdgpu_bo_init(adev); 886 if (r) 887 return r; 888 889 r = gmc_v10_0_gart_init(adev); 890 if (r) 891 return r; 892 893 /* 894 * number of VMs 895 * VMID 0 is reserved for System 896 * amdgpu graphics/compute will use VMIDs 1-7 897 * amdkfd will use VMIDs 8-15 898 */ 899 adev->vm_manager.first_kfd_vmid = 8; 900 901 amdgpu_vm_manager_init(adev); 902 903 r = amdgpu_gmc_ras_sw_init(adev); 904 if (r) 905 return r; 906 907 return 0; 908 } 909 910 /** 911 * gmc_v10_0_gart_fini - vm fini callback 912 * 913 * @adev: amdgpu_device pointer 914 * 915 * Tears down the driver GART/VM setup (CIK). 916 */ 917 static void gmc_v10_0_gart_fini(struct amdgpu_device *adev) 918 { 919 amdgpu_gart_table_vram_free(adev); 920 } 921 922 static int gmc_v10_0_sw_fini(struct amdgpu_ip_block *ip_block) 923 { 924 struct amdgpu_device *adev = ip_block->adev; 925 926 amdgpu_vm_manager_fini(adev); 927 gmc_v10_0_gart_fini(adev); 928 amdgpu_gem_force_release(adev); 929 amdgpu_bo_fini(adev); 930 931 return 0; 932 } 933 934 static void gmc_v10_0_init_golden_registers(struct amdgpu_device *adev) 935 { 936 } 937 938 /** 939 * gmc_v10_0_gart_enable - gart enable 940 * 941 * @adev: amdgpu_device pointer 942 */ 943 static int gmc_v10_0_gart_enable(struct amdgpu_device *adev) 944 { 945 int r; 946 bool value; 947 948 if (adev->gart.bo == NULL) { 949 dev_err(adev->dev, "No VRAM object for PCIE GART.\n"); 950 return -EINVAL; 951 } 952 953 amdgpu_gtt_mgr_recover(&adev->mman.gtt_mgr); 954 955 if (!adev->in_s0ix) { 956 r = adev->gfxhub.funcs->gart_enable(adev); 957 if (r) 958 return r; 959 } 960 961 r = adev->mmhub.funcs->gart_enable(adev); 962 if (r) 963 return r; 964 965 adev->hdp.funcs->init_registers(adev); 966 967 /* Flush HDP after it is initialized */ 968 amdgpu_device_flush_hdp(adev, NULL); 969 970 value = (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) ? 971 false : true; 972 973 if (!adev->in_s0ix) 974 adev->gfxhub.funcs->set_fault_enable_default(adev, value); 975 adev->mmhub.funcs->set_fault_enable_default(adev, value); 976 gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_MMHUB0(0), 0); 977 if (!adev->in_s0ix) 978 gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_GFXHUB(0), 0); 979 980 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n", 981 (unsigned int)(adev->gmc.gart_size >> 20), 982 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo)); 983 984 return 0; 985 } 986 987 static int gmc_v10_0_hw_init(struct amdgpu_ip_block *ip_block) 988 { 989 struct amdgpu_device *adev = ip_block->adev; 990 int r; 991 992 adev->gmc.flush_pasid_uses_kiq = !amdgpu_emu_mode; 993 994 /* The sequence of these two function calls matters.*/ 995 gmc_v10_0_init_golden_registers(adev); 996 997 /* 998 * harvestable groups in gc_utcl2 need to be programmed before any GFX block 999 * register setup within GMC, or else system hang when harvesting SA. 1000 */ 1001 if (!adev->in_s0ix && adev->gfxhub.funcs && adev->gfxhub.funcs->utcl2_harvest) 1002 adev->gfxhub.funcs->utcl2_harvest(adev); 1003 1004 r = gmc_v10_0_gart_enable(adev); 1005 if (r) 1006 return r; 1007 1008 if (amdgpu_emu_mode == 1) { 1009 r = amdgpu_gmc_vram_checking(adev); 1010 if (r) 1011 return r; 1012 } 1013 1014 if (adev->umc.funcs && adev->umc.funcs->init_registers) 1015 adev->umc.funcs->init_registers(adev); 1016 1017 return 0; 1018 } 1019 1020 /** 1021 * gmc_v10_0_gart_disable - gart disable 1022 * 1023 * @adev: amdgpu_device pointer 1024 * 1025 * This disables all VM page table. 1026 */ 1027 static void gmc_v10_0_gart_disable(struct amdgpu_device *adev) 1028 { 1029 if (!adev->in_s0ix) 1030 adev->gfxhub.funcs->gart_disable(adev); 1031 adev->mmhub.funcs->gart_disable(adev); 1032 } 1033 1034 static int gmc_v10_0_hw_fini(struct amdgpu_ip_block *ip_block) 1035 { 1036 struct amdgpu_device *adev = ip_block->adev; 1037 1038 gmc_v10_0_gart_disable(adev); 1039 1040 if (amdgpu_sriov_vf(adev)) { 1041 /* full access mode, so don't touch any GMC register */ 1042 DRM_DEBUG("For SRIOV client, shouldn't do anything.\n"); 1043 return 0; 1044 } 1045 1046 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0); 1047 1048 if (adev->gmc.ecc_irq.funcs && 1049 amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__UMC)) 1050 amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0); 1051 1052 return 0; 1053 } 1054 1055 static int gmc_v10_0_suspend(struct amdgpu_ip_block *ip_block) 1056 { 1057 gmc_v10_0_hw_fini(ip_block); 1058 1059 return 0; 1060 } 1061 1062 static int gmc_v10_0_resume(struct amdgpu_ip_block *ip_block) 1063 { 1064 int r; 1065 1066 r = gmc_v10_0_hw_init(ip_block); 1067 if (r) 1068 return r; 1069 1070 amdgpu_vmid_reset_all(ip_block->adev); 1071 1072 return 0; 1073 } 1074 1075 static bool gmc_v10_0_is_idle(struct amdgpu_ip_block *ip_block) 1076 { 1077 /* MC is always ready in GMC v10.*/ 1078 return true; 1079 } 1080 1081 static int gmc_v10_0_wait_for_idle(struct amdgpu_ip_block *ip_block) 1082 { 1083 /* There is no need to wait for MC idle in GMC v10.*/ 1084 return 0; 1085 } 1086 1087 static int gmc_v10_0_set_clockgating_state(struct amdgpu_ip_block *ip_block, 1088 enum amd_clockgating_state state) 1089 { 1090 int r; 1091 struct amdgpu_device *adev = ip_block->adev; 1092 1093 /* 1094 * The issue mmhub can't disconnect from DF with MMHUB clock gating being disabled 1095 * is a new problem observed at DF 3.0.3, however with the same suspend sequence not 1096 * seen any issue on the DF 3.0.2 series platform. 1097 */ 1098 if (adev->in_s0ix && 1099 amdgpu_ip_version(adev, DF_HWIP, 0) > IP_VERSION(3, 0, 2)) { 1100 dev_dbg(adev->dev, "keep mmhub clock gating being enabled for s0ix\n"); 1101 return 0; 1102 } 1103 1104 r = adev->mmhub.funcs->set_clockgating(adev, state); 1105 if (r) 1106 return r; 1107 1108 if (amdgpu_ip_version(adev, ATHUB_HWIP, 0) >= IP_VERSION(2, 1, 0)) 1109 return athub_v2_1_set_clockgating(adev, state); 1110 else 1111 return athub_v2_0_set_clockgating(adev, state); 1112 } 1113 1114 static void gmc_v10_0_get_clockgating_state(struct amdgpu_ip_block *ip_block, u64 *flags) 1115 { 1116 struct amdgpu_device *adev = ip_block->adev; 1117 1118 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(10, 1, 3) || 1119 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(10, 1, 4)) 1120 return; 1121 1122 adev->mmhub.funcs->get_clockgating(adev, flags); 1123 1124 if (amdgpu_ip_version(adev, ATHUB_HWIP, 0) >= IP_VERSION(2, 1, 0)) 1125 athub_v2_1_get_clockgating(adev, flags); 1126 else 1127 athub_v2_0_get_clockgating(adev, flags); 1128 } 1129 1130 static int gmc_v10_0_set_powergating_state(struct amdgpu_ip_block *ip_block, 1131 enum amd_powergating_state state) 1132 { 1133 return 0; 1134 } 1135 1136 const struct amd_ip_funcs gmc_v10_0_ip_funcs = { 1137 .name = "gmc_v10_0", 1138 .early_init = gmc_v10_0_early_init, 1139 .late_init = gmc_v10_0_late_init, 1140 .sw_init = gmc_v10_0_sw_init, 1141 .sw_fini = gmc_v10_0_sw_fini, 1142 .hw_init = gmc_v10_0_hw_init, 1143 .hw_fini = gmc_v10_0_hw_fini, 1144 .suspend = gmc_v10_0_suspend, 1145 .resume = gmc_v10_0_resume, 1146 .is_idle = gmc_v10_0_is_idle, 1147 .wait_for_idle = gmc_v10_0_wait_for_idle, 1148 .set_clockgating_state = gmc_v10_0_set_clockgating_state, 1149 .set_powergating_state = gmc_v10_0_set_powergating_state, 1150 .get_clockgating_state = gmc_v10_0_get_clockgating_state, 1151 }; 1152 1153 const struct amdgpu_ip_block_version gmc_v10_0_ip_block = { 1154 .type = AMD_IP_BLOCK_TYPE_GMC, 1155 .major = 10, 1156 .minor = 0, 1157 .rev = 0, 1158 .funcs = &gmc_v10_0_ip_funcs, 1159 }; 1160