1 /* 2 * Copyright 2023 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_v12_0.h" 31 #include "gmc_v12_1.h" 32 #include "athub/athub_4_1_0_sh_mask.h" 33 #include "athub/athub_4_1_0_offset.h" 34 #include "oss/osssys_7_0_0_offset.h" 35 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h" 36 #include "soc24_enum.h" 37 #include "soc24.h" 38 #include "soc15d.h" 39 #include "soc15_common.h" 40 #include "nbif_v6_3_1.h" 41 #include "gfxhub_v12_0.h" 42 #include "gfxhub_v12_1.h" 43 #include "mmhub_v4_1_0.h" 44 #include "mmhub_v4_2_0.h" 45 #include "athub_v4_1_0.h" 46 #include "umc_v8_14.h" 47 48 static int gmc_v12_0_ecc_interrupt_state(struct amdgpu_device *adev, 49 struct amdgpu_irq_src *src, 50 unsigned type, 51 enum amdgpu_interrupt_state state) 52 { 53 return 0; 54 } 55 56 static int gmc_v12_0_vm_fault_interrupt_state(struct amdgpu_device *adev, 57 struct amdgpu_irq_src *src, unsigned type, 58 enum amdgpu_interrupt_state state) 59 { 60 switch (state) { 61 case AMDGPU_IRQ_STATE_DISABLE: 62 /* MM HUB */ 63 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB0(0), false); 64 /* GFX HUB */ 65 /* This works because this interrupt is only 66 * enabled at init/resume and disabled in 67 * fini/suspend, so the overall state doesn't 68 * change over the course of suspend/resume. 69 */ 70 if (!adev->in_s0ix) 71 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB(0), false); 72 break; 73 case AMDGPU_IRQ_STATE_ENABLE: 74 /* MM HUB */ 75 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB0(0), true); 76 /* GFX HUB */ 77 /* This works because this interrupt is only 78 * enabled at init/resume and disabled in 79 * fini/suspend, so the overall state doesn't 80 * change over the course of suspend/resume. 81 */ 82 if (!adev->in_s0ix) 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_v12_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; 97 bool retry_fault = !!(entry->src_data[1] & 98 AMDGPU_GMC9_FAULT_SOURCE_DATA_RETRY); 99 bool write_fault = !!(entry->src_data[1] & 100 AMDGPU_GMC9_FAULT_SOURCE_DATA_WRITE); 101 uint32_t status = 0; 102 u64 addr; 103 104 addr = (u64)entry->src_data[0] << 12; 105 addr |= ((u64)entry->src_data[1] & 0xf) << 44; 106 107 if (entry->client_id == SOC21_IH_CLIENTID_VMC) 108 hub = &adev->vmhub[AMDGPU_MMHUB0(0)]; 109 else 110 hub = &adev->vmhub[AMDGPU_GFXHUB(0)]; 111 112 if (retry_fault) { 113 int ret = amdgpu_gmc_handle_retry_fault(adev, entry, addr, 0, 0, 114 write_fault); 115 /* Returning 1 here also prevents sending the IV to the KFD */ 116 if (ret == 1) 117 return 1; 118 } 119 120 if (!amdgpu_sriov_vf(adev)) { 121 /* 122 * Issue a dummy read to wait for the status register to 123 * be updated to avoid reading an incorrect value due to 124 * the new fast GRBM interface. 125 */ 126 if (entry->vmid_src == AMDGPU_GFXHUB(0)) 127 RREG32(hub->vm_l2_pro_fault_status); 128 129 status = RREG32(hub->vm_l2_pro_fault_status); 130 WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1); 131 132 amdgpu_vm_update_fault_cache(adev, entry->pasid, addr, status, 133 entry->vmid_src ? AMDGPU_MMHUB0(0) : AMDGPU_GFXHUB(0)); 134 } 135 136 if (printk_ratelimit()) { 137 struct amdgpu_task_info *task_info; 138 139 dev_err(adev->dev, 140 "[%s] page fault (src_id:%u ring:%u vmid:%u pasid:%u)\n", 141 entry->vmid_src ? "mmhub" : "gfxhub", 142 entry->src_id, entry->ring_id, entry->vmid, entry->pasid); 143 task_info = amdgpu_vm_get_task_info_pasid(adev, entry->pasid); 144 if (task_info) { 145 amdgpu_vm_print_task_info(adev, task_info); 146 amdgpu_vm_put_task_info(task_info); 147 } 148 149 dev_err(adev->dev, " in page starting at address 0x%016llx from client %d\n", 150 addr, entry->client_id); 151 152 /* Only print L2 fault status if the status register could be read and 153 * contains useful information 154 */ 155 if (status != 0) 156 hub->vmhub_funcs->print_l2_protection_fault_status(adev, status); 157 } 158 159 return 0; 160 } 161 162 static const struct amdgpu_irq_src_funcs gmc_v12_0_irq_funcs = { 163 .set = gmc_v12_0_vm_fault_interrupt_state, 164 .process = gmc_v12_0_process_interrupt, 165 }; 166 167 static const struct amdgpu_irq_src_funcs gmc_v12_0_ecc_funcs = { 168 .set = gmc_v12_0_ecc_interrupt_state, 169 .process = amdgpu_umc_process_ecc_irq, 170 }; 171 172 static void gmc_v12_0_set_irq_funcs(struct amdgpu_device *adev) 173 { 174 adev->gmc.vm_fault.num_types = 1; 175 adev->gmc.vm_fault.funcs = &gmc_v12_0_irq_funcs; 176 177 if (!amdgpu_sriov_vf(adev)) { 178 adev->gmc.ecc_irq.num_types = 1; 179 adev->gmc.ecc_irq.funcs = &gmc_v12_0_ecc_funcs; 180 } 181 } 182 183 /** 184 * gmc_v12_0_use_invalidate_semaphore - judge whether to use semaphore 185 * 186 * @adev: amdgpu_device pointer 187 * @vmhub: vmhub type 188 * 189 */ 190 static bool gmc_v12_0_use_invalidate_semaphore(struct amdgpu_device *adev, 191 uint32_t vmhub) 192 { 193 return ((vmhub == AMDGPU_MMHUB0(0)) && 194 (!amdgpu_sriov_vf(adev))); 195 } 196 197 static bool gmc_v12_0_get_vmid_pasid_mapping_info( 198 struct amdgpu_device *adev, 199 uint8_t vmid, uint16_t *p_pasid) 200 { 201 *p_pasid = RREG32(SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT) + vmid) & 0xffff; 202 203 return !!(*p_pasid); 204 } 205 206 /* 207 * GART 208 * VMID 0 is the physical GPU addresses as used by the kernel. 209 * VMIDs 1-15 are used for userspace clients and are handled 210 * by the amdgpu vm/hsa code. 211 */ 212 213 static void gmc_v12_0_flush_vm_hub(struct amdgpu_device *adev, uint32_t vmid, 214 unsigned int vmhub, uint32_t flush_type) 215 { 216 bool use_semaphore = gmc_v12_0_use_invalidate_semaphore(adev, vmhub); 217 struct amdgpu_vmhub *hub = &adev->vmhub[vmhub]; 218 u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type); 219 u32 tmp; 220 /* Use register 17 for GART */ 221 const unsigned eng = 17; 222 unsigned int i; 223 unsigned char hub_ip = 0; 224 225 hub_ip = (vmhub == AMDGPU_GFXHUB(0)) ? 226 GC_HWIP : MMHUB_HWIP; 227 228 spin_lock(&adev->gmc.invalidate_lock); 229 /* 230 * It may lose gpuvm invalidate acknowldege state across power-gating 231 * off cycle, add semaphore acquire before invalidation and semaphore 232 * release after invalidation to avoid entering power gated state 233 * to WA the Issue 234 */ 235 236 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 237 if (use_semaphore) { 238 for (i = 0; i < adev->usec_timeout; i++) { 239 /* a read return value of 1 means semaphore acuqire */ 240 tmp = RREG32_RLC_NO_KIQ(hub->vm_inv_eng0_sem + 241 hub->eng_distance * eng, hub_ip); 242 if (tmp & 0x1) 243 break; 244 udelay(1); 245 } 246 247 if (i >= adev->usec_timeout) 248 dev_err(adev->dev, 249 "Timeout waiting for sem acquire in VM flush!\n"); 250 } 251 252 WREG32_RLC_NO_KIQ(hub->vm_inv_eng0_req + hub->eng_distance * eng, inv_req, hub_ip); 253 254 /* Wait for ACK with a delay.*/ 255 for (i = 0; i < adev->usec_timeout; i++) { 256 tmp = RREG32_RLC_NO_KIQ(hub->vm_inv_eng0_ack + 257 hub->eng_distance * eng, hub_ip); 258 tmp &= 1 << vmid; 259 if (tmp) 260 break; 261 262 udelay(1); 263 } 264 265 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 266 if (use_semaphore) 267 /* 268 * add semaphore release after invalidation, 269 * write with 0 means semaphore release 270 */ 271 WREG32_RLC_NO_KIQ(hub->vm_inv_eng0_sem + 272 hub->eng_distance * eng, 0, hub_ip); 273 274 /* Issue additional private vm invalidation to MMHUB */ 275 if ((vmhub != AMDGPU_GFXHUB(0)) && 276 (hub->vm_l2_bank_select_reserved_cid2) && 277 !amdgpu_sriov_vf(adev)) { 278 inv_req = RREG32_NO_KIQ(hub->vm_l2_bank_select_reserved_cid2); 279 /* bit 25: RSERVED_CACHE_PRIVATE_INVALIDATION */ 280 inv_req |= (1 << 25); 281 /* Issue private invalidation */ 282 WREG32_NO_KIQ(hub->vm_l2_bank_select_reserved_cid2, inv_req); 283 /* Read back to ensure invalidation is done*/ 284 RREG32_NO_KIQ(hub->vm_l2_bank_select_reserved_cid2); 285 } 286 287 spin_unlock(&adev->gmc.invalidate_lock); 288 289 if (i < adev->usec_timeout) 290 return; 291 292 dev_err(adev->dev, "Timeout waiting for VM flush ACK!\n"); 293 } 294 295 /** 296 * gmc_v12_0_flush_gpu_tlb - gart tlb flush callback 297 * 298 * @adev: amdgpu_device pointer 299 * @vmid: vm instance to flush 300 * @vmhub: which hub to flush 301 * @flush_type: the flush type 302 * 303 * Flush the TLB for the requested page table. 304 */ 305 static void gmc_v12_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, 306 uint32_t vmhub, uint32_t flush_type) 307 { 308 if ((vmhub == AMDGPU_GFXHUB(0)) && !adev->gfx.is_poweron) 309 return; 310 311 /* flush hdp cache */ 312 amdgpu_device_flush_hdp(adev, NULL); 313 314 /* This is necessary for SRIOV as well as for GFXOFF to function 315 * properly under bare metal 316 */ 317 if ((adev->gfx.kiq[0].ring.sched.ready || adev->mes.ring[0].sched.ready) && 318 (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev))) { 319 struct amdgpu_vmhub *hub = &adev->vmhub[vmhub]; 320 const unsigned eng = 17; 321 u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type); 322 u32 req = hub->vm_inv_eng0_req + hub->eng_distance * eng; 323 u32 ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng; 324 325 amdgpu_gmc_fw_reg_write_reg_wait(adev, req, ack, inv_req, 326 1 << vmid, GET_INST(GC, 0)); 327 return; 328 } 329 330 gmc_v12_0_flush_vm_hub(adev, vmid, vmhub, 0); 331 return; 332 } 333 334 /** 335 * gmc_v12_0_flush_gpu_tlb_pasid - tlb flush via pasid 336 * 337 * @adev: amdgpu_device pointer 338 * @pasid: pasid to be flush 339 * @flush_type: the flush type 340 * @all_hub: flush all hubs 341 * @inst: is used to select which instance of KIQ to use for the invalidation 342 * 343 * Flush the TLB for the requested pasid. 344 */ 345 static void gmc_v12_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, 346 uint16_t pasid, uint32_t flush_type, 347 bool all_hub, uint32_t inst) 348 { 349 uint16_t queried; 350 int vmid, i; 351 352 if (adev->enable_uni_mes && adev->mes.ring[AMDGPU_MES_SCHED_PIPE].sched.ready && 353 (adev->mes.sched_version & AMDGPU_MES_VERSION_MASK) >= 0x84) { 354 struct mes_inv_tlbs_pasid_input input = {0}; 355 input.pasid = pasid; 356 input.flush_type = flush_type; 357 input.hub_id = AMDGPU_GFXHUB(0); 358 /* MES will invalidate all gc_hub for the device from master */ 359 adev->mes.funcs->invalidate_tlbs_pasid(&adev->mes, &input); 360 if (all_hub) { 361 /* Only need to invalidate mm_hub now, gfx12 only support one mmhub */ 362 input.hub_id = AMDGPU_MMHUB0(0); 363 adev->mes.funcs->invalidate_tlbs_pasid(&adev->mes, &input); 364 } 365 return; 366 } 367 368 for (vmid = 1; vmid < 16; vmid++) { 369 bool valid; 370 371 valid = gmc_v12_0_get_vmid_pasid_mapping_info(adev, vmid, 372 &queried); 373 if (!valid || queried != pasid) 374 continue; 375 376 if (all_hub) { 377 for_each_set_bit(i, adev->vmhubs_mask, 378 AMDGPU_MAX_VMHUBS) 379 gmc_v12_0_flush_gpu_tlb(adev, vmid, i, 380 flush_type); 381 } else { 382 gmc_v12_0_flush_gpu_tlb(adev, vmid, AMDGPU_GFXHUB(0), 383 flush_type); 384 } 385 } 386 } 387 388 static uint64_t gmc_v12_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring, 389 unsigned vmid, uint64_t pd_addr) 390 { 391 bool use_semaphore = gmc_v12_0_use_invalidate_semaphore(ring->adev, ring->vm_hub); 392 struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->vm_hub]; 393 uint32_t req = hub->vmhub_funcs->get_invalidate_req(vmid, 0); 394 unsigned eng = ring->vm_inv_eng; 395 396 /* 397 * It may lose gpuvm invalidate acknowldege state across power-gating 398 * off cycle, add semaphore acquire before invalidation and semaphore 399 * release after invalidation to avoid entering power gated state 400 * to WA the Issue 401 */ 402 403 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 404 if (use_semaphore) 405 /* a read return value of 1 means semaphore acuqire */ 406 amdgpu_ring_emit_reg_wait(ring, 407 hub->vm_inv_eng0_sem + 408 hub->eng_distance * eng, 0x1, 0x1); 409 410 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 + 411 (hub->ctx_addr_distance * vmid), 412 lower_32_bits(pd_addr)); 413 414 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + 415 (hub->ctx_addr_distance * vmid), 416 upper_32_bits(pd_addr)); 417 418 amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req + 419 hub->eng_distance * eng, 420 hub->vm_inv_eng0_ack + 421 hub->eng_distance * eng, 422 req, 1 << vmid); 423 424 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 425 if (use_semaphore) 426 /* 427 * add semaphore release after invalidation, 428 * write with 0 means semaphore release 429 */ 430 amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem + 431 hub->eng_distance * eng, 0); 432 433 return pd_addr; 434 } 435 436 static void gmc_v12_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid, 437 unsigned pasid) 438 { 439 struct amdgpu_device *adev = ring->adev; 440 uint32_t reg; 441 442 if (ring->vm_hub == AMDGPU_GFXHUB(0)) 443 reg = SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT) + vmid; 444 else 445 reg = SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT_MM) + vmid; 446 447 amdgpu_ring_emit_wreg(ring, reg, pasid); 448 } 449 450 /* 451 * PTE format: 452 * 63 P 453 * 62:59 reserved 454 * 58 D 455 * 57 G 456 * 56 T 457 * 55:54 M 458 * 53:52 SW 459 * 51:48 reserved for future 460 * 47:12 4k physical page base address 461 * 11:7 fragment 462 * 6 write 463 * 5 read 464 * 4 exe 465 * 3 Z 466 * 2 snooped 467 * 1 system 468 * 0 valid 469 * 470 * PDE format: 471 * 63 P 472 * 62:58 block fragment size 473 * 57 reserved 474 * 56 A 475 * 55:54 M 476 * 53:52 reserved 477 * 51:48 reserved for future 478 * 47:6 physical base address of PD or PTE 479 * 5:3 reserved 480 * 2 C 481 * 1 system 482 * 0 valid 483 */ 484 485 static void gmc_v12_0_get_vm_pde(struct amdgpu_device *adev, int level, 486 uint64_t *addr, uint64_t *flags) 487 { 488 if (!(*flags & AMDGPU_PDE_PTE_GFX12) && !(*flags & AMDGPU_PTE_SYSTEM)) 489 *addr = adev->vm_manager.vram_base_offset + *addr - 490 adev->gmc.vram_start; 491 BUG_ON(*addr & 0xFFFF00000000003FULL); 492 493 if (!adev->gmc.translate_further) 494 return; 495 496 if (level == AMDGPU_VM_PDB1) { 497 /* Set the block fragment size */ 498 if (!(*flags & AMDGPU_PDE_PTE_GFX12)) 499 *flags |= AMDGPU_PDE_BFS_GFX12(0x9); 500 501 } else if (level == AMDGPU_VM_PDB0) { 502 if (*flags & AMDGPU_PDE_PTE_GFX12) 503 *flags &= ~AMDGPU_PDE_PTE_GFX12; 504 } 505 } 506 507 static void gmc_v12_0_get_vm_pte(struct amdgpu_device *adev, 508 struct amdgpu_vm *vm, 509 struct amdgpu_bo *bo, 510 uint32_t vm_flags, 511 uint64_t *flags) 512 { 513 if (vm_flags & AMDGPU_VM_PAGE_EXECUTABLE) 514 *flags |= AMDGPU_PTE_EXECUTABLE; 515 else 516 *flags &= ~AMDGPU_PTE_EXECUTABLE; 517 518 switch (vm_flags & AMDGPU_VM_MTYPE_MASK) { 519 case AMDGPU_VM_MTYPE_DEFAULT: 520 *flags = AMDGPU_PTE_MTYPE_GFX12(*flags, MTYPE_NC); 521 break; 522 case AMDGPU_VM_MTYPE_NC: 523 default: 524 *flags = AMDGPU_PTE_MTYPE_GFX12(*flags, MTYPE_NC); 525 break; 526 case AMDGPU_VM_MTYPE_UC: 527 *flags = AMDGPU_PTE_MTYPE_GFX12(*flags, MTYPE_UC); 528 break; 529 } 530 531 if (vm_flags & AMDGPU_VM_PAGE_NOALLOC) 532 *flags |= AMDGPU_PTE_NOALLOC; 533 else 534 *flags &= ~AMDGPU_PTE_NOALLOC; 535 536 if (vm_flags & AMDGPU_VM_PAGE_PRT) { 537 *flags |= AMDGPU_PTE_PRT_GFX12; 538 *flags |= AMDGPU_PTE_SNOOPED; 539 *flags |= AMDGPU_PTE_SYSTEM; 540 *flags |= AMDGPU_PTE_IS_PTE; 541 *flags &= ~AMDGPU_PTE_VALID; 542 } 543 544 if (bo && bo->flags & AMDGPU_GEM_CREATE_GFX12_DCC) 545 *flags |= AMDGPU_PTE_DCC; 546 547 if (bo && bo->flags & AMDGPU_GEM_CREATE_UNCACHED) 548 *flags = AMDGPU_PTE_MTYPE_GFX12(*flags, MTYPE_UC); 549 } 550 551 static unsigned gmc_v12_0_get_vbios_fb_size(struct amdgpu_device *adev) 552 { 553 return 0; 554 } 555 556 static unsigned int gmc_v12_0_get_dcc_alignment(struct amdgpu_device *adev) 557 { 558 unsigned int max_tex_channel_caches, alignment; 559 560 if (amdgpu_ip_version(adev, GC_HWIP, 0) != IP_VERSION(12, 0, 0) && 561 amdgpu_ip_version(adev, GC_HWIP, 0) != IP_VERSION(12, 0, 1)) 562 return 0; 563 564 max_tex_channel_caches = adev->gfx.config.max_texture_channel_caches; 565 if (is_power_of_2(max_tex_channel_caches)) 566 alignment = (unsigned int)(max_tex_channel_caches / SZ_4); 567 else 568 alignment = roundup_pow_of_two(max_tex_channel_caches); 569 570 return (unsigned int)(alignment * max_tex_channel_caches * SZ_1K); 571 } 572 573 static const struct amdgpu_gmc_funcs gmc_v12_0_gmc_funcs = { 574 .flush_gpu_tlb = gmc_v12_0_flush_gpu_tlb, 575 .flush_gpu_tlb_pasid = gmc_v12_0_flush_gpu_tlb_pasid, 576 .emit_flush_gpu_tlb = gmc_v12_0_emit_flush_gpu_tlb, 577 .emit_pasid_mapping = gmc_v12_0_emit_pasid_mapping, 578 .get_vm_pde = gmc_v12_0_get_vm_pde, 579 .get_vm_pte = gmc_v12_0_get_vm_pte, 580 .get_vbios_fb_size = gmc_v12_0_get_vbios_fb_size, 581 .get_dcc_alignment = gmc_v12_0_get_dcc_alignment, 582 }; 583 584 static void gmc_v12_0_set_gmc_funcs(struct amdgpu_device *adev) 585 { 586 adev->gmc.gmc_funcs = &gmc_v12_0_gmc_funcs; 587 } 588 589 static void gmc_v12_0_set_umc_funcs(struct amdgpu_device *adev) 590 { 591 switch (amdgpu_ip_version(adev, UMC_HWIP, 0)) { 592 case IP_VERSION(8, 14, 0): 593 adev->umc.channel_inst_num = UMC_V8_14_CHANNEL_INSTANCE_NUM; 594 adev->umc.umc_inst_num = UMC_V8_14_UMC_INSTANCE_NUM(adev); 595 adev->umc.node_inst_num = 0; 596 adev->umc.max_ras_err_cnt_per_query = UMC_V8_14_TOTAL_CHANNEL_NUM(adev); 597 adev->umc.channel_offs = UMC_V8_14_PER_CHANNEL_OFFSET; 598 adev->umc.ras = &umc_v8_14_ras; 599 break; 600 default: 601 break; 602 } 603 } 604 605 606 static void gmc_v12_0_set_mmhub_funcs(struct amdgpu_device *adev) 607 { 608 switch (amdgpu_ip_version(adev, MMHUB_HWIP, 0)) { 609 case IP_VERSION(4, 1, 0): 610 adev->mmhub.funcs = &mmhub_v4_1_0_funcs; 611 break; 612 case IP_VERSION(4, 2, 0): 613 adev->mmhub.funcs = &mmhub_v4_2_0_funcs; 614 break; 615 default: 616 break; 617 } 618 } 619 620 static void gmc_v12_0_set_gfxhub_funcs(struct amdgpu_device *adev) 621 { 622 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 623 case IP_VERSION(12, 0, 0): 624 case IP_VERSION(12, 0, 1): 625 adev->gfxhub.funcs = &gfxhub_v12_0_funcs; 626 break; 627 case IP_VERSION(12, 1, 0): 628 adev->gfxhub.funcs = &gfxhub_v12_1_funcs; 629 break; 630 default: 631 break; 632 } 633 } 634 635 static int gmc_v12_0_early_init(struct amdgpu_ip_block *ip_block) 636 { 637 struct amdgpu_device *adev = ip_block->adev; 638 639 if (adev->smuio.funcs && 640 adev->smuio.funcs->is_host_gpu_xgmi_supported) 641 adev->gmc.xgmi.connected_to_cpu = 642 adev->smuio.funcs->is_host_gpu_xgmi_supported(adev); 643 644 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 645 case IP_VERSION(12, 1, 0): 646 gmc_v12_1_set_gmc_funcs(adev); 647 gmc_v12_1_set_irq_funcs(adev); 648 adev->gmc.init_pte_flags = AMDGPU_PTE_IS_PTE; 649 break; 650 default: 651 gmc_v12_0_set_gmc_funcs(adev); 652 gmc_v12_0_set_irq_funcs(adev); 653 break; 654 } 655 gmc_v12_0_set_gfxhub_funcs(adev); 656 gmc_v12_0_set_mmhub_funcs(adev); 657 gmc_v12_0_set_umc_funcs(adev); 658 659 adev->gmc.shared_aperture_start = 0x2000000000000000ULL; 660 adev->gmc.shared_aperture_end = 661 adev->gmc.shared_aperture_start + (4ULL << 30) - 1; 662 663 adev->gmc.private_aperture_start = 0x1000000000000000ULL; 664 if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(12, 1, 0)) 665 adev->gmc.private_aperture_end = 666 adev->gmc.private_aperture_start + (1ULL << 57) - 1; 667 else 668 adev->gmc.private_aperture_end = 669 adev->gmc.private_aperture_start + (4ULL << 30) - 1; 670 671 adev->gmc.noretry_flags = AMDGPU_VM_NORETRY_FLAGS_TF; 672 673 return 0; 674 } 675 676 static int gmc_v12_0_late_init(struct amdgpu_ip_block *ip_block) 677 { 678 struct amdgpu_device *adev = ip_block->adev; 679 int r; 680 681 r = amdgpu_gmc_allocate_vm_inv_eng(adev); 682 if (r) 683 return r; 684 685 r = amdgpu_gmc_ras_late_init(adev); 686 if (r) 687 return r; 688 689 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0); 690 } 691 692 static void gmc_v12_0_vram_gtt_location(struct amdgpu_device *adev, 693 struct amdgpu_gmc *mc) 694 { 695 u64 base = 0; 696 697 base = adev->mmhub.funcs->get_fb_location(adev); 698 699 if (amdgpu_gmc_is_pdb0_enabled(adev)) { 700 amdgpu_gmc_sysvm_location(adev, mc); 701 } else { 702 amdgpu_gmc_set_agp_default(adev, mc); 703 amdgpu_gmc_vram_location(adev, &adev->gmc, base); 704 amdgpu_gmc_gart_location(adev, mc, AMDGPU_GART_PLACEMENT_LOW); 705 if (!amdgpu_sriov_vf(adev) && (amdgpu_agp == 1)) 706 amdgpu_gmc_agp_location(adev, mc); 707 } 708 /* base offset of vram pages */ 709 if (amdgpu_sriov_vf(adev)) 710 adev->vm_manager.vram_base_offset = 0; 711 else 712 adev->vm_manager.vram_base_offset = adev->mmhub.funcs->get_mc_fb_offset(adev); 713 714 adev->vm_manager.vram_base_offset += 715 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size; 716 } 717 718 /** 719 * gmc_v12_0_mc_init - initialize the memory controller driver params 720 * 721 * @adev: amdgpu_device pointer 722 * 723 * Look up the amount of vram, vram width, and decide how to place 724 * vram and gart within the GPU's physical address space. 725 * Returns 0 for success. 726 */ 727 static int gmc_v12_0_mc_init(struct amdgpu_device *adev) 728 { 729 int r; 730 731 if (adev->gmc.xgmi.connected_to_cpu) 732 adev->gmc.mc_vram_size = 733 adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes; 734 else 735 adev->gmc.mc_vram_size = 736 adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL; 737 738 adev->gmc.real_vram_size = adev->gmc.mc_vram_size; 739 740 if (!(adev->flags & AMD_IS_APU) && 741 !adev->gmc.xgmi.connected_to_cpu) { 742 r = amdgpu_device_resize_fb_bar(adev); 743 if (r) 744 return r; 745 } 746 747 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0); 748 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0); 749 750 #ifdef CONFIG_X86_64 751 if (((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)) || 752 (adev->gmc.xgmi.connected_to_cpu)) { 753 adev->gmc.aper_base = 754 adev->mmhub.funcs->get_mc_fb_offset(adev) + 755 adev->gmc.xgmi.physical_node_id * 756 adev->gmc.xgmi.node_segment_size; 757 adev->gmc.aper_size = adev->gmc.real_vram_size; 758 } 759 #endif 760 /* In case the PCI BAR is larger than the actual amount of vram */ 761 adev->gmc.visible_vram_size = adev->gmc.aper_size; 762 if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size) 763 adev->gmc.visible_vram_size = adev->gmc.real_vram_size; 764 765 /* set the gart size */ 766 if (amdgpu_gart_size == -1) { 767 adev->gmc.gart_size = 512ULL << 20; 768 } else 769 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20; 770 771 gmc_v12_0_vram_gtt_location(adev, &adev->gmc); 772 773 return 0; 774 } 775 776 static int gmc_v12_0_gart_init(struct amdgpu_device *adev) 777 { 778 int r; 779 780 if (adev->gart.bo) { 781 WARN(1, "PCIE GART already initialized\n"); 782 return 0; 783 } 784 785 if (amdgpu_gmc_is_pdb0_enabled(adev)) { 786 adev->gmc.vmid0_page_table_depth = 1; 787 adev->gmc.vmid0_page_table_block_size = 12; 788 } else { 789 adev->gmc.vmid0_page_table_depth = 0; 790 adev->gmc.vmid0_page_table_block_size = 0; 791 } 792 793 /* Initialize common gart structure */ 794 r = amdgpu_gart_init(adev); 795 if (r) 796 return r; 797 798 adev->gart.table_size = adev->gart.num_gpu_pages * 8; 799 adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_GFX12(0ULL, MTYPE_UC) | 800 AMDGPU_PTE_EXECUTABLE | 801 AMDGPU_PTE_IS_PTE; 802 803 r = amdgpu_gart_table_vram_alloc(adev); 804 if (r) 805 return r; 806 807 if (amdgpu_gmc_is_pdb0_enabled(adev)) 808 r = amdgpu_gmc_pdb0_alloc(adev); 809 810 return r; 811 } 812 813 static int gmc_v12_0_sw_init(struct amdgpu_ip_block *ip_block) 814 { 815 int r, vram_width = 0, vram_type = 0, vram_vendor = 0; 816 struct amdgpu_device *adev = ip_block->adev; 817 int i; 818 819 adev->mmhub.funcs->init(adev); 820 821 adev->gfxhub.funcs->init(adev); 822 823 spin_lock_init(&adev->gmc.invalidate_lock); 824 825 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(12, 1, 0)) { 826 gmc_v12_1_init_vram_info(adev); 827 } else { 828 r = amdgpu_gmc_get_vram_info(adev, 829 &vram_width, &vram_type, &vram_vendor); 830 adev->gmc.vram_width = vram_width; 831 adev->gmc.vram_type = vram_type; 832 adev->gmc.vram_vendor = vram_vendor; 833 } 834 835 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 836 case IP_VERSION(12, 0, 0): 837 case IP_VERSION(12, 0, 1): 838 set_bit(AMDGPU_GFXHUB(0), adev->vmhubs_mask); 839 set_bit(AMDGPU_MMHUB0(0), adev->vmhubs_mask); 840 /* 841 * To fulfill 4-level page support, 842 * vm size is 256TB (48bit), maximum size, 843 * block size 512 (9bit) 844 */ 845 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 846 break; 847 case IP_VERSION(12, 1, 0): 848 bitmap_set(adev->vmhubs_mask, AMDGPU_GFXHUB(0), 849 NUM_XCC(adev->gfx.xcc_mask)); 850 for (i = 0; i < hweight32(adev->aid_mask); i++) 851 set_bit(AMDGPU_MMHUB0(i), adev->vmhubs_mask); 852 /* 853 * To fulfill 5-level page support, 854 * vm size is 128PetaByte (57bit), maximum size, 855 * block size 512 (9bit) 856 */ 857 amdgpu_vm_adjust_size(adev, 128 * 1024 * 1024, 9, 4, 57); 858 break; 859 default: 860 break; 861 } 862 863 /* This interrupt is VMC page fault.*/ 864 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_VMC, 865 VMC_1_0__SRCID__VM_FAULT, 866 &adev->gmc.vm_fault); 867 868 if (r) 869 return r; 870 871 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(12, 1, 0)) { 872 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_UTCL2, 873 UTCL2_1_0__SRCID__FAULT, 874 &adev->gmc.vm_fault); 875 if (r) 876 return r; 877 /* Add GCVM UTCL2 Retry fault */ 878 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_UTCL2, 879 UTCL2_1_0__SRCID__RETRY, 880 &adev->gmc.vm_fault); 881 if (r) 882 return r; 883 884 /* Add MMVM UTCL2 Retry fault */ 885 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_VMC, 886 VMC_1_0__SRCID__VM_RETRY, 887 &adev->gmc.vm_fault); 888 } else { 889 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GFX, 890 UTCL2_1_0__SRCID__FAULT, 891 &adev->gmc.vm_fault); 892 } 893 if (r) 894 return r; 895 896 if (!amdgpu_sriov_vf(adev)) { 897 /* interrupt sent to DF. */ 898 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(12, 0, 0)) 899 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_DF, 0, 900 &adev->gmc.ecc_irq); 901 else 902 r = amdgpu_irq_add_id(adev, SOC_V1_0_IH_CLIENTID_DF, 0, 903 &adev->gmc.ecc_irq); 904 905 if (r) 906 return r; 907 } 908 909 /* 910 * Set the internal MC address mask This is the max address of the GPU's 911 * internal address space. 912 */ 913 adev->gmc.mc_mask = AMDGPU_GMC_HOLE_MASK; 914 915 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44)); 916 if (r) { 917 drm_warn(adev_to_drm(adev), "No suitable DMA available.\n"); 918 return r; 919 } 920 921 adev->need_swiotlb = drm_need_swiotlb(44); 922 923 r = gmc_v12_0_mc_init(adev); 924 if (r) 925 return r; 926 927 amdgpu_gmc_get_vbios_allocations(adev); 928 929 #ifdef HAVE_ACPI_DEV_GET_FIRST_MATCH_DEV 930 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(12, 1, 0)) { 931 r = amdgpu_gmc_init_mem_ranges(adev); 932 if (r) 933 return r; 934 } 935 #endif 936 937 /* Memory manager */ 938 r = amdgpu_bo_init(adev); 939 if (r) 940 return r; 941 942 r = gmc_v12_0_gart_init(adev); 943 if (r) 944 return r; 945 946 /* 947 * number of VMs 948 * VMID 0 is reserved for System 949 * amdgpu graphics/compute will use VMIDs 1-7 950 * amdkfd will use VMIDs 8-15 951 */ 952 adev->vm_manager.first_kfd_vmid = 953 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(12, 1, 0) ? 954 3 : 8; 955 adev->vm_manager.first_kfd_vmid = 956 adev->gfx.disable_kq ? 1 : (adev->vm_manager.first_kfd_vmid); 957 958 amdgpu_vm_manager_init(adev); 959 960 r = amdgpu_gmc_ras_sw_init(adev); 961 if (r) 962 return r; 963 964 return 0; 965 } 966 967 /** 968 * gmc_v12_0_gart_fini - vm fini callback 969 * 970 * @adev: amdgpu_device pointer 971 * 972 * Tears down the driver GART/VM setup (CIK). 973 */ 974 static void gmc_v12_0_gart_fini(struct amdgpu_device *adev) 975 { 976 amdgpu_gart_table_vram_free(adev); 977 } 978 979 static int gmc_v12_0_sw_fini(struct amdgpu_ip_block *ip_block) 980 { 981 struct amdgpu_device *adev = ip_block->adev; 982 983 amdgpu_vm_manager_fini(adev); 984 gmc_v12_0_gart_fini(adev); 985 amdgpu_gem_force_release(adev); 986 amdgpu_bo_free_kernel(&adev->gmc.pdb0_bo, NULL, &adev->gmc.ptr_pdb0); 987 amdgpu_bo_fini(adev); 988 989 return 0; 990 } 991 992 static void gmc_v12_0_init_golden_registers(struct amdgpu_device *adev) 993 { 994 } 995 996 /** 997 * gmc_v12_0_gart_enable - gart enable 998 * 999 * @adev: amdgpu_device pointer 1000 */ 1001 static int gmc_v12_0_gart_enable(struct amdgpu_device *adev) 1002 { 1003 int r; 1004 bool value; 1005 1006 if (adev->gmc.xgmi.connected_to_cpu) 1007 amdgpu_gmc_init_pdb0(adev); 1008 1009 if (adev->gart.bo == NULL) { 1010 dev_err(adev->dev, "No VRAM object for PCIE GART.\n"); 1011 return -EINVAL; 1012 } 1013 1014 amdgpu_gtt_mgr_recover(&adev->mman.gtt_mgr); 1015 1016 r = adev->mmhub.funcs->gart_enable(adev); 1017 if (r) 1018 return r; 1019 1020 /* Flush HDP after it is initialized */ 1021 amdgpu_device_flush_hdp(adev, NULL); 1022 1023 value = amdgpu_vm_fault_stop != AMDGPU_VM_FAULT_STOP_ALWAYS; 1024 1025 adev->mmhub.funcs->set_fault_enable_default(adev, value); 1026 adev->gmc.gmc_funcs->flush_gpu_tlb(adev, 0, AMDGPU_MMHUB0(0), 0); 1027 1028 drm_info(adev_to_drm(adev), "PCIE GART of %uM enabled (table at 0x%016llX).\n", 1029 (unsigned)(adev->gmc.gart_size >> 20), 1030 (adev->gmc.pdb0_bo) ? (unsigned long long)amdgpu_bo_gpu_offset(adev->gmc.pdb0_bo) : 1031 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo)); 1032 1033 return 0; 1034 } 1035 1036 static int gmc_v12_0_hw_init(struct amdgpu_ip_block *ip_block) 1037 { 1038 int r; 1039 struct amdgpu_device *adev = ip_block->adev; 1040 1041 /* The sequence of these two function calls matters.*/ 1042 gmc_v12_0_init_golden_registers(adev); 1043 1044 r = gmc_v12_0_gart_enable(adev); 1045 if (r) 1046 return r; 1047 1048 if (adev->umc.funcs && adev->umc.funcs->init_registers) 1049 adev->umc.funcs->init_registers(adev); 1050 1051 return 0; 1052 } 1053 1054 /** 1055 * gmc_v12_0_gart_disable - gart disable 1056 * 1057 * @adev: amdgpu_device pointer 1058 * 1059 * This disables all VM page table. 1060 */ 1061 static void gmc_v12_0_gart_disable(struct amdgpu_device *adev) 1062 { 1063 adev->mmhub.funcs->gart_disable(adev); 1064 } 1065 1066 static int gmc_v12_0_hw_fini(struct amdgpu_ip_block *ip_block) 1067 { 1068 struct amdgpu_device *adev = ip_block->adev; 1069 1070 if (amdgpu_sriov_vf(adev)) { 1071 /* full access mode, so don't touch any GMC register */ 1072 DRM_DEBUG("For SRIOV client, shouldn't do anything.\n"); 1073 return 0; 1074 } 1075 1076 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0); 1077 1078 if (adev->gmc.ecc_irq.funcs && 1079 amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__UMC)) 1080 amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0); 1081 1082 gmc_v12_0_gart_disable(adev); 1083 1084 return 0; 1085 } 1086 1087 static int gmc_v12_0_suspend(struct amdgpu_ip_block *ip_block) 1088 { 1089 gmc_v12_0_hw_fini(ip_block); 1090 1091 return 0; 1092 } 1093 1094 static int gmc_v12_0_resume(struct amdgpu_ip_block *ip_block) 1095 { 1096 int r; 1097 1098 r = gmc_v12_0_hw_init(ip_block); 1099 if (r) 1100 return r; 1101 1102 amdgpu_vmid_reset_all(ip_block->adev); 1103 1104 return 0; 1105 } 1106 1107 static bool gmc_v12_0_is_idle(struct amdgpu_ip_block *ip_block) 1108 { 1109 /* MC is always ready in GMC v11.*/ 1110 return true; 1111 } 1112 1113 static int gmc_v12_0_wait_for_idle(struct amdgpu_ip_block *ip_block) 1114 { 1115 /* There is no need to wait for MC idle in GMC v11.*/ 1116 return 0; 1117 } 1118 1119 static int gmc_v12_0_set_clockgating_state(struct amdgpu_ip_block *ip_block, 1120 enum amd_clockgating_state state) 1121 { 1122 int r; 1123 struct amdgpu_device *adev = ip_block->adev; 1124 1125 r = adev->mmhub.funcs->set_clockgating(adev, state); 1126 if (r) 1127 return r; 1128 1129 if (amdgpu_ip_version(adev, GC_HWIP, 0) != IP_VERSION(12, 1, 0)) 1130 return athub_v4_1_0_set_clockgating(adev, state); 1131 else 1132 return 0; 1133 } 1134 1135 static void gmc_v12_0_get_clockgating_state(struct amdgpu_ip_block *ip_block, u64 *flags) 1136 { 1137 struct amdgpu_device *adev = ip_block->adev; 1138 1139 adev->mmhub.funcs->get_clockgating(adev, flags); 1140 1141 if (amdgpu_ip_version(adev, GC_HWIP, 0) != IP_VERSION(12, 1, 0)) 1142 athub_v4_1_0_get_clockgating(adev, flags); 1143 } 1144 1145 static int gmc_v12_0_set_powergating_state(struct amdgpu_ip_block *ip_block, 1146 enum amd_powergating_state state) 1147 { 1148 return 0; 1149 } 1150 1151 const struct amd_ip_funcs gmc_v12_0_ip_funcs = { 1152 .name = "gmc_v12_0", 1153 .early_init = gmc_v12_0_early_init, 1154 .sw_init = gmc_v12_0_sw_init, 1155 .hw_init = gmc_v12_0_hw_init, 1156 .late_init = gmc_v12_0_late_init, 1157 .sw_fini = gmc_v12_0_sw_fini, 1158 .hw_fini = gmc_v12_0_hw_fini, 1159 .suspend = gmc_v12_0_suspend, 1160 .resume = gmc_v12_0_resume, 1161 .is_idle = gmc_v12_0_is_idle, 1162 .wait_for_idle = gmc_v12_0_wait_for_idle, 1163 .set_clockgating_state = gmc_v12_0_set_clockgating_state, 1164 .set_powergating_state = gmc_v12_0_set_powergating_state, 1165 .get_clockgating_state = gmc_v12_0_get_clockgating_state, 1166 }; 1167 1168 const struct amdgpu_ip_block_version gmc_v12_0_ip_block = { 1169 .type = AMD_IP_BLOCK_TYPE_GMC, 1170 .major = 12, 1171 .minor = 0, 1172 .rev = 0, 1173 .funcs = &gmc_v12_0_ip_funcs, 1174 }; 1175