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