1 /* 2 * Copyright 2021 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_ip.h" 30 #include "amdgpu_atomfirmware.h" 31 #include "gmc_v11_0.h" 32 #include "umc_v8_10.h" 33 #include "athub/athub_3_0_0_sh_mask.h" 34 #include "athub/athub_3_0_0_offset.h" 35 #include "dcn/dcn_3_2_0_offset.h" 36 #include "dcn/dcn_3_2_0_sh_mask.h" 37 #include "oss/osssys_6_0_0_offset.h" 38 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h" 39 #include "navi10_enum.h" 40 #include "soc15.h" 41 #include "soc15d.h" 42 #include "soc15_common.h" 43 #include "nbio_v4_3.h" 44 #include "gfxhub_v3_0.h" 45 #include "gfxhub_v3_0_3.h" 46 #include "gfxhub_v11_5_0.h" 47 #include "mmhub_v3_0.h" 48 #include "mmhub_v3_0_1.h" 49 #include "mmhub_v3_0_2.h" 50 #include "mmhub_v3_3.h" 51 #include "athub_v3_0.h" 52 53 54 static int gmc_v11_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_v11_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 && (adev->in_runpm || adev->in_suspend || 78 amdgpu_in_reset(adev))) 79 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB(0), false); 80 break; 81 case AMDGPU_IRQ_STATE_ENABLE: 82 /* MM HUB */ 83 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB0(0), true); 84 /* GFX HUB */ 85 /* This works because this interrupt is only 86 * enabled at init/resume and disabled in 87 * fini/suspend, so the overall state doesn't 88 * change over the course of suspend/resume. 89 */ 90 if (!adev->in_s0ix) 91 amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB(0), true); 92 break; 93 default: 94 break; 95 } 96 97 return 0; 98 } 99 100 static int gmc_v11_0_process_interrupt(struct amdgpu_device *adev, 101 struct amdgpu_irq_src *source, 102 struct amdgpu_iv_entry *entry) 103 { 104 uint32_t vmhub_index = entry->client_id == SOC21_IH_CLIENTID_VMC ? 105 AMDGPU_MMHUB0(0) : AMDGPU_GFXHUB(0); 106 struct amdgpu_vmhub *hub = &adev->vmhub[vmhub_index]; 107 bool retry_fault = !!(entry->src_data[1] & 108 AMDGPU_GMC9_FAULT_SOURCE_DATA_RETRY); 109 bool write_fault = !!(entry->src_data[1] & 110 AMDGPU_GMC9_FAULT_SOURCE_DATA_WRITE); 111 uint32_t status = 0; 112 uint32_t cam_index; 113 u64 addr; 114 115 addr = (u64)entry->src_data[0] << 12; 116 addr |= ((u64)entry->src_data[1] & 0xf) << 44; 117 118 if (retry_fault) { 119 cam_index = entry->src_data[2] & 0x3ff; 120 121 int ret = amdgpu_gmc_handle_retry_fault(adev, entry, addr, cam_index, 0, 122 write_fault); 123 /* Returning 1 here also prevents sending the IV to the KFD */ 124 if (ret == 1) 125 return 1; 126 } 127 128 if (!amdgpu_sriov_vf(adev)) { 129 /* 130 * Issue a dummy read to wait for the status register to 131 * be updated to avoid reading an incorrect value due to 132 * the new fast GRBM interface. 133 */ 134 if (entry->vmid_src == AMDGPU_GFXHUB(0)) 135 RREG32(hub->vm_l2_pro_fault_status); 136 137 status = RREG32(hub->vm_l2_pro_fault_status); 138 WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1); 139 140 amdgpu_vm_update_fault_cache(adev, entry->pasid, addr, status, 141 entry->vmid_src ? AMDGPU_MMHUB0(0) : AMDGPU_GFXHUB(0)); 142 } 143 144 if (printk_ratelimit()) { 145 struct amdgpu_task_info *task_info; 146 147 dev_err(adev->dev, 148 "[%s] page fault (src_id:%u ring:%u vmid:%u pasid:%u)\n", 149 entry->vmid_src ? "mmhub" : "gfxhub", 150 entry->src_id, entry->ring_id, entry->vmid, entry->pasid); 151 task_info = amdgpu_vm_get_task_info_pasid(adev, entry->pasid); 152 if (task_info) { 153 amdgpu_vm_print_task_info(adev, task_info); 154 amdgpu_vm_put_task_info(task_info); 155 } 156 157 dev_err(adev->dev, " in page starting at address 0x%016llx from client %d\n", 158 addr, entry->client_id); 159 160 /* Only print L2 fault status if the status register could be read and 161 * contains useful information 162 */ 163 if (status != 0) 164 hub->vmhub_funcs->print_l2_protection_fault_status(adev, status); 165 } 166 167 return 0; 168 } 169 170 static const struct amdgpu_irq_src_funcs gmc_v11_0_irq_funcs = { 171 .set = gmc_v11_0_vm_fault_interrupt_state, 172 .process = gmc_v11_0_process_interrupt, 173 }; 174 175 static const struct amdgpu_irq_src_funcs gmc_v11_0_ecc_funcs = { 176 .set = gmc_v11_0_ecc_interrupt_state, 177 .process = amdgpu_umc_process_ecc_irq, 178 }; 179 180 static void gmc_v11_0_set_irq_funcs(struct amdgpu_device *adev) 181 { 182 adev->gmc.vm_fault.num_types = 1; 183 adev->gmc.vm_fault.funcs = &gmc_v11_0_irq_funcs; 184 185 if (!amdgpu_sriov_vf(adev)) { 186 adev->gmc.ecc_irq.num_types = 1; 187 adev->gmc.ecc_irq.funcs = &gmc_v11_0_ecc_funcs; 188 } 189 } 190 191 /** 192 * gmc_v11_0_use_invalidate_semaphore - judge whether to use semaphore 193 * 194 * @adev: amdgpu_device pointer 195 * @vmhub: vmhub type 196 * 197 */ 198 static bool gmc_v11_0_use_invalidate_semaphore(struct amdgpu_device *adev, 199 uint32_t vmhub) 200 { 201 return ((vmhub == AMDGPU_MMHUB0(0)) && 202 (!amdgpu_sriov_vf(adev))); 203 } 204 205 static bool gmc_v11_0_get_vmid_pasid_mapping_info( 206 struct amdgpu_device *adev, 207 uint8_t vmid, uint16_t *p_pasid) 208 { 209 *p_pasid = RREG32(SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT) + vmid) & 0xffff; 210 211 return !!(*p_pasid); 212 } 213 214 /** 215 * gmc_v11_0_flush_gpu_tlb - gart tlb flush callback 216 * 217 * @adev: amdgpu_device pointer 218 * @vmid: vm instance to flush 219 * @vmhub: which hub to flush 220 * @flush_type: the flush type 221 * 222 * Flush the TLB for the requested page table. 223 */ 224 static void gmc_v11_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, 225 uint32_t vmhub, uint32_t flush_type) 226 { 227 bool use_semaphore = gmc_v11_0_use_invalidate_semaphore(adev, vmhub); 228 struct amdgpu_vmhub *hub = &adev->vmhub[vmhub]; 229 u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type); 230 /* Use register 17 for GART */ 231 const unsigned int eng = 17; 232 unsigned char hub_ip; 233 u32 sem, req, ack; 234 unsigned int i; 235 u32 tmp; 236 237 if ((vmhub == AMDGPU_GFXHUB(0)) && !adev->gfx.is_poweron) 238 return; 239 240 sem = hub->vm_inv_eng0_sem + hub->eng_distance * eng; 241 req = hub->vm_inv_eng0_req + hub->eng_distance * eng; 242 ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng; 243 244 /* flush hdp cache */ 245 amdgpu_device_flush_hdp(adev, NULL); 246 247 /* This is necessary for SRIOV as well as for GFXOFF to function 248 * properly under bare metal 249 */ 250 if ((adev->gfx.kiq[0].ring.sched.ready || adev->mes.ring[0].sched.ready) && 251 (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev))) { 252 amdgpu_gmc_fw_reg_write_reg_wait(adev, req, ack, inv_req, 253 1 << vmid, GET_INST(GC, 0)); 254 return; 255 } 256 257 /* This path is needed before KIQ/MES/GFXOFF are set up */ 258 hub_ip = (vmhub == AMDGPU_GFXHUB(0)) ? GC_HWIP : MMHUB_HWIP; 259 260 spin_lock(&adev->gmc.invalidate_lock); 261 /* 262 * It may lose gpuvm invalidate acknowldege state across power-gating 263 * off cycle, add semaphore acquire before invalidation and semaphore 264 * release after invalidation to avoid entering power gated state 265 * to WA the Issue 266 */ 267 268 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 269 if (use_semaphore) { 270 for (i = 0; i < adev->usec_timeout; i++) { 271 /* a read return value of 1 means semaphore acuqire */ 272 tmp = RREG32_RLC_NO_KIQ(sem, hub_ip); 273 if (tmp & 0x1) 274 break; 275 udelay(1); 276 } 277 278 if (i >= adev->usec_timeout) 279 DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n"); 280 } 281 282 WREG32_RLC_NO_KIQ(req, inv_req, hub_ip); 283 284 /* Wait for ACK with a delay.*/ 285 for (i = 0; i < adev->usec_timeout; i++) { 286 tmp = RREG32_RLC_NO_KIQ(ack, hub_ip); 287 tmp &= 1 << vmid; 288 if (tmp) 289 break; 290 291 udelay(1); 292 } 293 294 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 295 if (use_semaphore) 296 WREG32_RLC_NO_KIQ(sem, 0, hub_ip); 297 298 /* Issue additional private vm invalidation to MMHUB */ 299 if ((vmhub != AMDGPU_GFXHUB(0)) && 300 (hub->vm_l2_bank_select_reserved_cid2) && 301 !amdgpu_sriov_vf(adev)) { 302 inv_req = RREG32_NO_KIQ(hub->vm_l2_bank_select_reserved_cid2); 303 /* bit 25: RSERVED_CACHE_PRIVATE_INVALIDATION */ 304 inv_req |= (1 << 25); 305 /* Issue private invalidation */ 306 WREG32_NO_KIQ(hub->vm_l2_bank_select_reserved_cid2, inv_req); 307 /* Read back to ensure invalidation is done*/ 308 RREG32_NO_KIQ(hub->vm_l2_bank_select_reserved_cid2); 309 } 310 311 spin_unlock(&adev->gmc.invalidate_lock); 312 313 if (i >= adev->usec_timeout) 314 dev_err(adev->dev, "Timeout waiting for VM flush ACK!\n"); 315 } 316 317 /** 318 * gmc_v11_0_flush_gpu_tlb_pasid - tlb flush via pasid 319 * 320 * @adev: amdgpu_device pointer 321 * @pasid: pasid to be flush 322 * @flush_type: the flush type 323 * @all_hub: flush all hubs 324 * @inst: is used to select which instance of KIQ to use for the invalidation 325 * 326 * Flush the TLB for the requested pasid. 327 */ 328 static void gmc_v11_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, 329 uint16_t pasid, uint32_t flush_type, 330 bool all_hub, uint32_t inst) 331 { 332 uint16_t queried; 333 int vmid, i; 334 335 for (vmid = 1; vmid < 16; vmid++) { 336 bool valid; 337 338 valid = gmc_v11_0_get_vmid_pasid_mapping_info(adev, vmid, 339 &queried); 340 if (!valid || queried != pasid) 341 continue; 342 343 if (all_hub) { 344 for_each_set_bit(i, adev->vmhubs_mask, 345 AMDGPU_MAX_VMHUBS) 346 gmc_v11_0_flush_gpu_tlb(adev, vmid, i, 347 flush_type); 348 } else { 349 gmc_v11_0_flush_gpu_tlb(adev, vmid, AMDGPU_GFXHUB(0), 350 flush_type); 351 } 352 } 353 } 354 355 static uint64_t gmc_v11_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring, 356 unsigned int vmid, uint64_t pd_addr) 357 { 358 bool use_semaphore = gmc_v11_0_use_invalidate_semaphore(ring->adev, ring->vm_hub); 359 struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->vm_hub]; 360 uint32_t req = hub->vmhub_funcs->get_invalidate_req(vmid, 0); 361 unsigned int eng = ring->vm_inv_eng; 362 363 /* 364 * It may lose gpuvm invalidate acknowldege state across power-gating 365 * off cycle, add semaphore acquire before invalidation and semaphore 366 * release after invalidation to avoid entering power gated state 367 * to WA the Issue 368 */ 369 370 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 371 if (use_semaphore) 372 /* a read return value of 1 means semaphore acuqire */ 373 amdgpu_ring_emit_reg_wait(ring, 374 hub->vm_inv_eng0_sem + 375 hub->eng_distance * eng, 0x1, 0x1); 376 377 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 + 378 (hub->ctx_addr_distance * vmid), 379 lower_32_bits(pd_addr)); 380 381 amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + 382 (hub->ctx_addr_distance * vmid), 383 upper_32_bits(pd_addr)); 384 385 amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req + 386 hub->eng_distance * eng, 387 hub->vm_inv_eng0_ack + 388 hub->eng_distance * eng, 389 req, 1 << vmid); 390 391 /* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */ 392 if (use_semaphore) 393 /* 394 * add semaphore release after invalidation, 395 * write with 0 means semaphore release 396 */ 397 amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem + 398 hub->eng_distance * eng, 0); 399 400 return pd_addr; 401 } 402 403 static void gmc_v11_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned int vmid, 404 unsigned int pasid) 405 { 406 struct amdgpu_device *adev = ring->adev; 407 uint32_t reg; 408 409 if (ring->vm_hub == AMDGPU_GFXHUB(0)) 410 reg = SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT) + vmid; 411 else 412 reg = SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT_MM) + vmid; 413 414 amdgpu_ring_emit_wreg(ring, reg, pasid); 415 } 416 417 /* 418 * PTE format: 419 * 63:59 reserved 420 * 58:57 reserved 421 * 56 F 422 * 55 L 423 * 54 reserved 424 * 53:52 SW 425 * 51 T 426 * 50:48 mtype 427 * 47:12 4k physical page base address 428 * 11:7 fragment 429 * 6 write 430 * 5 read 431 * 4 exe 432 * 3 Z 433 * 2 snooped 434 * 1 system 435 * 0 valid 436 * 437 * PDE format: 438 * 63:59 block fragment size 439 * 58:55 reserved 440 * 54 P 441 * 53:48 reserved 442 * 47:6 physical base address of PD or PTE 443 * 5:3 reserved 444 * 2 C 445 * 1 system 446 * 0 valid 447 */ 448 449 static void gmc_v11_0_get_vm_pde(struct amdgpu_device *adev, int level, 450 uint64_t *addr, uint64_t *flags) 451 { 452 if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM)) 453 *addr = amdgpu_gmc_vram_mc2pa(adev, *addr); 454 BUG_ON(*addr & 0xFFFF00000000003FULL); 455 456 if (!adev->gmc.translate_further) 457 return; 458 459 if (level == AMDGPU_VM_PDB1) { 460 /* Set the block fragment size */ 461 if (!(*flags & AMDGPU_PDE_PTE)) 462 *flags |= AMDGPU_PDE_BFS(0x9); 463 464 } else if (level == AMDGPU_VM_PDB0) { 465 if (*flags & AMDGPU_PDE_PTE) 466 *flags &= ~AMDGPU_PDE_PTE; 467 else 468 *flags |= AMDGPU_PTE_TF; 469 } 470 } 471 472 static void gmc_v11_0_get_vm_pte(struct amdgpu_device *adev, 473 struct amdgpu_vm *vm, 474 struct amdgpu_bo *bo, 475 uint32_t vm_flags, 476 uint64_t *flags) 477 { 478 if (vm_flags & AMDGPU_VM_PAGE_EXECUTABLE) 479 *flags |= AMDGPU_PTE_EXECUTABLE; 480 else 481 *flags &= ~AMDGPU_PTE_EXECUTABLE; 482 483 switch (vm_flags & AMDGPU_VM_MTYPE_MASK) { 484 case AMDGPU_VM_MTYPE_DEFAULT: 485 case AMDGPU_VM_MTYPE_NC: 486 default: 487 *flags = AMDGPU_PTE_MTYPE_NV10(*flags, MTYPE_NC); 488 break; 489 case AMDGPU_VM_MTYPE_WC: 490 *flags = AMDGPU_PTE_MTYPE_NV10(*flags, MTYPE_WC); 491 break; 492 case AMDGPU_VM_MTYPE_CC: 493 *flags = AMDGPU_PTE_MTYPE_NV10(*flags, MTYPE_CC); 494 break; 495 case AMDGPU_VM_MTYPE_UC: 496 *flags = AMDGPU_PTE_MTYPE_NV10(*flags, MTYPE_UC); 497 break; 498 } 499 500 if (vm_flags & AMDGPU_VM_PAGE_NOALLOC) 501 *flags |= AMDGPU_PTE_NOALLOC; 502 else 503 *flags &= ~AMDGPU_PTE_NOALLOC; 504 505 if (vm_flags & AMDGPU_VM_PAGE_PRT) { 506 *flags |= AMDGPU_PTE_PRT; 507 *flags |= AMDGPU_PTE_SNOOPED; 508 *flags |= AMDGPU_PTE_LOG; 509 *flags |= AMDGPU_PTE_SYSTEM; 510 *flags &= ~AMDGPU_PTE_VALID; 511 } 512 513 if (bo && bo->flags & (AMDGPU_GEM_CREATE_COHERENT | 514 AMDGPU_GEM_CREATE_EXT_COHERENT | 515 AMDGPU_GEM_CREATE_UNCACHED)) 516 *flags = AMDGPU_PTE_MTYPE_NV10(*flags, MTYPE_UC); 517 } 518 519 static unsigned int gmc_v11_0_get_vbios_fb_size(struct amdgpu_device *adev) 520 { 521 u32 d1vga_control = RREG32_SOC15(DCE, 0, regD1VGA_CONTROL); 522 unsigned int size; 523 524 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) { 525 size = AMDGPU_VBIOS_VGA_ALLOCATION; 526 } else { 527 u32 viewport; 528 u32 pitch; 529 530 viewport = RREG32_SOC15(DCE, 0, regHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION); 531 pitch = RREG32_SOC15(DCE, 0, regHUBPREQ0_DCSURF_SURFACE_PITCH); 532 size = (REG_GET_FIELD(viewport, 533 HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) * 534 REG_GET_FIELD(pitch, HUBPREQ0_DCSURF_SURFACE_PITCH, PITCH) * 535 4); 536 } 537 538 return size; 539 } 540 541 static const struct amdgpu_gmc_funcs gmc_v11_0_gmc_funcs = { 542 .flush_gpu_tlb = gmc_v11_0_flush_gpu_tlb, 543 .flush_gpu_tlb_pasid = gmc_v11_0_flush_gpu_tlb_pasid, 544 .emit_flush_gpu_tlb = gmc_v11_0_emit_flush_gpu_tlb, 545 .emit_pasid_mapping = gmc_v11_0_emit_pasid_mapping, 546 .get_vm_pde = gmc_v11_0_get_vm_pde, 547 .get_vm_pte = gmc_v11_0_get_vm_pte, 548 .get_vbios_fb_size = gmc_v11_0_get_vbios_fb_size, 549 }; 550 551 static void gmc_v11_0_set_gmc_funcs(struct amdgpu_device *adev) 552 { 553 adev->gmc.gmc_funcs = &gmc_v11_0_gmc_funcs; 554 } 555 556 static void gmc_v11_0_set_umc_funcs(struct amdgpu_device *adev) 557 { 558 switch (amdgpu_ip_version(adev, UMC_HWIP, 0)) { 559 case IP_VERSION(8, 10, 0): 560 adev->umc.channel_inst_num = UMC_V8_10_CHANNEL_INSTANCE_NUM; 561 adev->umc.umc_inst_num = UMC_V8_10_UMC_INSTANCE_NUM; 562 adev->umc.max_ras_err_cnt_per_query = UMC_V8_10_TOTAL_CHANNEL_NUM(adev); 563 adev->umc.channel_offs = UMC_V8_10_PER_CHANNEL_OFFSET; 564 adev->umc.retire_unit = UMC_V8_10_NA_COL_2BITS_POWER_OF_2_NUM; 565 if (adev->umc.node_inst_num == 4) 566 adev->umc.channel_idx_tbl = &umc_v8_10_channel_idx_tbl_ext0[0][0][0]; 567 else 568 adev->umc.channel_idx_tbl = &umc_v8_10_channel_idx_tbl[0][0][0]; 569 adev->umc.ras = &umc_v8_10_ras; 570 break; 571 case IP_VERSION(8, 11, 0): 572 break; 573 default: 574 break; 575 } 576 } 577 578 579 static void gmc_v11_0_set_mmhub_funcs(struct amdgpu_device *adev) 580 { 581 switch (amdgpu_ip_version(adev, MMHUB_HWIP, 0)) { 582 case IP_VERSION(3, 0, 1): 583 adev->mmhub.funcs = &mmhub_v3_0_1_funcs; 584 break; 585 case IP_VERSION(3, 0, 2): 586 adev->mmhub.funcs = &mmhub_v3_0_2_funcs; 587 break; 588 case IP_VERSION(3, 3, 0): 589 case IP_VERSION(3, 3, 1): 590 case IP_VERSION(3, 3, 2): 591 case IP_VERSION(3, 4, 0): 592 case IP_VERSION(3, 4, 2): 593 adev->mmhub.funcs = &mmhub_v3_3_funcs; 594 break; 595 default: 596 adev->mmhub.funcs = &mmhub_v3_0_funcs; 597 break; 598 } 599 } 600 601 static void gmc_v11_0_set_gfxhub_funcs(struct amdgpu_device *adev) 602 { 603 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 604 case IP_VERSION(11, 0, 3): 605 adev->gfxhub.funcs = &gfxhub_v3_0_3_funcs; 606 break; 607 case IP_VERSION(11, 5, 0): 608 case IP_VERSION(11, 5, 1): 609 case IP_VERSION(11, 5, 2): 610 case IP_VERSION(11, 5, 3): 611 case IP_VERSION(11, 5, 4): 612 case IP_VERSION(11, 5, 6): 613 case IP_VERSION(11, 7, 0): 614 case IP_VERSION(11, 7, 1): 615 adev->gfxhub.funcs = &gfxhub_v11_5_0_funcs; 616 break; 617 default: 618 adev->gfxhub.funcs = &gfxhub_v3_0_funcs; 619 break; 620 } 621 } 622 623 static int gmc_v11_0_early_init(struct amdgpu_ip_block *ip_block) 624 { 625 struct amdgpu_device *adev = ip_block->adev; 626 627 gmc_v11_0_set_gfxhub_funcs(adev); 628 gmc_v11_0_set_mmhub_funcs(adev); 629 gmc_v11_0_set_gmc_funcs(adev); 630 gmc_v11_0_set_irq_funcs(adev); 631 gmc_v11_0_set_umc_funcs(adev); 632 633 adev->gmc.shared_aperture_start = 0x2000000000000000ULL; 634 adev->gmc.shared_aperture_end = 635 adev->gmc.shared_aperture_start + (4ULL << 30) - 1; 636 adev->gmc.private_aperture_start = 0x1000000000000000ULL; 637 adev->gmc.private_aperture_end = 638 adev->gmc.private_aperture_start + (4ULL << 30) - 1; 639 adev->gmc.noretry_flags = AMDGPU_VM_NORETRY_FLAGS_TF; 640 641 return 0; 642 } 643 644 static int gmc_v11_0_late_init(struct amdgpu_ip_block *ip_block) 645 { 646 struct amdgpu_device *adev = ip_block->adev; 647 int r; 648 649 r = amdgpu_gmc_allocate_vm_inv_eng(adev); 650 if (r) 651 return r; 652 653 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0); 654 } 655 656 static void gmc_v11_0_vram_gtt_location(struct amdgpu_device *adev, 657 struct amdgpu_gmc *mc) 658 { 659 u64 base = 0; 660 661 base = adev->mmhub.funcs->get_fb_location(adev); 662 663 amdgpu_gmc_set_agp_default(adev, mc); 664 amdgpu_gmc_vram_location(adev, &adev->gmc, base); 665 amdgpu_gmc_gart_location(adev, mc, AMDGPU_GART_PLACEMENT_HIGH); 666 if (!amdgpu_sriov_vf(adev) && 667 (amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(11, 5, 0)) && 668 (amdgpu_agp == 1)) 669 amdgpu_gmc_agp_location(adev, mc); 670 671 /* base offset of vram pages */ 672 if (amdgpu_sriov_vf(adev)) 673 adev->vm_manager.vram_base_offset = 0; 674 else 675 adev->vm_manager.vram_base_offset = adev->mmhub.funcs->get_mc_fb_offset(adev); 676 } 677 678 /** 679 * gmc_v11_0_mc_init - initialize the memory controller driver params 680 * 681 * @adev: amdgpu_device pointer 682 * 683 * Look up the amount of vram, vram width, and decide how to place 684 * vram and gart within the GPU's physical address space. 685 * Returns 0 for success. 686 */ 687 static int gmc_v11_0_mc_init(struct amdgpu_device *adev) 688 { 689 int r; 690 691 /* size in MB on si */ 692 adev->gmc.mc_vram_size = 693 adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL; 694 adev->gmc.real_vram_size = adev->gmc.mc_vram_size; 695 696 if (!(adev->flags & AMD_IS_APU)) { 697 r = amdgpu_device_resize_fb_bar(adev); 698 if (r) 699 return r; 700 } 701 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0); 702 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0); 703 704 #ifdef CONFIG_X86_64 705 if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)) { 706 adev->gmc.aper_base = adev->mmhub.funcs->get_mc_fb_offset(adev); 707 adev->gmc.aper_size = adev->gmc.real_vram_size; 708 } 709 #endif 710 /* In case the PCI BAR is larger than the actual amount of vram */ 711 adev->gmc.visible_vram_size = adev->gmc.aper_size; 712 if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size) 713 adev->gmc.visible_vram_size = adev->gmc.real_vram_size; 714 715 /* set the gart size */ 716 amdgpu_gmc_set_gart_size(adev, SZ_512M); 717 718 gmc_v11_0_vram_gtt_location(adev, &adev->gmc); 719 720 return 0; 721 } 722 723 static int gmc_v11_0_gart_init(struct amdgpu_device *adev) 724 { 725 int r; 726 727 if (adev->gart.bo) { 728 WARN(1, "PCIE GART already initialized\n"); 729 return 0; 730 } 731 732 /* Initialize common gart structure */ 733 r = amdgpu_gart_init(adev); 734 if (r) 735 return r; 736 737 adev->gart.table_size = adev->gart.num_gpu_pages * 8; 738 adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_NV10(0ULL, MTYPE_UC) | 739 AMDGPU_PTE_EXECUTABLE; 740 741 return amdgpu_gart_table_vram_alloc(adev); 742 } 743 744 static int gmc_v11_0_sw_init(struct amdgpu_ip_block *ip_block) 745 { 746 int r, vram_width = 0, vram_type = 0, vram_vendor = 0; 747 struct amdgpu_device *adev = ip_block->adev; 748 749 adev->mmhub.funcs->init(adev); 750 751 adev->gfxhub.funcs->init(adev); 752 753 spin_lock_init(&adev->gmc.invalidate_lock); 754 755 r = amdgpu_gmc_get_vram_info(adev, 756 &vram_width, &vram_type, &vram_vendor); 757 adev->gmc.vram_width = vram_width; 758 759 adev->gmc.vram_type = vram_type; 760 adev->gmc.vram_vendor = vram_vendor; 761 762 /* The mall_size is already calculated as mall_size_per_umc * num_umc. 763 * However, for gfx1151, which features a 2-to-1 UMC mapping, 764 * the result must be multiplied by 2 to determine the actual mall size. 765 */ 766 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 767 case IP_VERSION(11, 5, 1): 768 adev->gmc.mall_size *= 2; 769 break; 770 default: 771 break; 772 } 773 774 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 775 case IP_VERSION(11, 0, 0): 776 case IP_VERSION(11, 0, 1): 777 case IP_VERSION(11, 0, 2): 778 case IP_VERSION(11, 0, 3): 779 case IP_VERSION(11, 0, 4): 780 case IP_VERSION(11, 5, 0): 781 case IP_VERSION(11, 5, 1): 782 case IP_VERSION(11, 5, 2): 783 case IP_VERSION(11, 5, 3): 784 case IP_VERSION(11, 5, 4): 785 case IP_VERSION(11, 5, 6): 786 case IP_VERSION(11, 7, 0): 787 case IP_VERSION(11, 7, 1): 788 set_bit(AMDGPU_GFXHUB(0), adev->vmhubs_mask); 789 set_bit(AMDGPU_MMHUB0(0), adev->vmhubs_mask); 790 /* 791 * To fulfill 4-level page support, 792 * vm size is 256TB (48bit), maximum size, 793 * block size 512 (9bit) 794 */ 795 amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48); 796 break; 797 default: 798 break; 799 } 800 801 /* This interrupt is VMC page fault.*/ 802 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_VMC, 803 VMC_1_0__SRCID__VM_FAULT, 804 &adev->gmc.vm_fault); 805 806 if (r) 807 return r; 808 809 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GFX, 810 UTCL2_1_0__SRCID__FAULT, 811 &adev->gmc.vm_fault); 812 if (r) 813 return r; 814 815 if (!amdgpu_sriov_vf(adev)) { 816 /* interrupt sent to DF. */ 817 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_DF, 0, 818 &adev->gmc.ecc_irq); 819 if (r) 820 return r; 821 } 822 823 /* 824 * Set the internal MC address mask This is the max address of the GPU's 825 * internal address space. 826 */ 827 adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */ 828 adev->gmc.pte_addr_mask = 0x0000FFFFFFFFF000ULL; /* 48 bit PA */ 829 830 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44)); 831 if (r) { 832 drm_warn(adev_to_drm(adev), "No suitable DMA available.\n"); 833 return r; 834 } 835 836 adev->need_swiotlb = drm_need_swiotlb(44); 837 838 r = gmc_v11_0_mc_init(adev); 839 if (r) 840 return r; 841 842 /* Memory manager */ 843 r = amdgpu_bo_init(adev); 844 if (r) 845 return r; 846 847 r = gmc_v11_0_gart_init(adev); 848 if (r) 849 return r; 850 851 /* 852 * number of VMs 853 * VMID 0 is reserved for System 854 * amdgpu graphics/compute will use VMIDs 1-7 855 * amdkfd will use VMIDs 8-15 856 */ 857 adev->vm_manager.first_kfd_vmid = adev->gfx.disable_kq ? 1 : 8; 858 859 amdgpu_vmid_mgr_set_vmid_mask(adev, 860 GENMASK(adev->vm_manager.first_kfd_vmid - 1, 1), 861 false); 862 amdgpu_vmid_mgr_set_vmid_mask(adev, GENMASK(AMDGPU_NUM_VMID - 1, 1), true); 863 864 amdgpu_vm_manager_init(adev); 865 866 r = amdgpu_gmc_ras_sw_init(adev); 867 if (r) 868 return r; 869 870 return 0; 871 } 872 873 /** 874 * gmc_v11_0_gart_fini - vm fini callback 875 * 876 * @adev: amdgpu_device pointer 877 * 878 * Tears down the driver GART/VM setup (CIK). 879 */ 880 static void gmc_v11_0_gart_fini(struct amdgpu_device *adev) 881 { 882 amdgpu_gart_table_vram_free(adev); 883 } 884 885 static int gmc_v11_0_sw_fini(struct amdgpu_ip_block *ip_block) 886 { 887 struct amdgpu_device *adev = ip_block->adev; 888 889 amdgpu_vm_manager_fini(adev); 890 gmc_v11_0_gart_fini(adev); 891 amdgpu_gem_force_release(adev); 892 amdgpu_bo_fini(adev); 893 894 return 0; 895 } 896 897 static void gmc_v11_0_init_golden_registers(struct amdgpu_device *adev) 898 { 899 if (amdgpu_sriov_vf(adev)) { 900 struct amdgpu_vmhub *hub = &adev->vmhub[AMDGPU_MMHUB0(0)]; 901 902 WREG32(hub->vm_contexts_disable, 0); 903 return; 904 } 905 } 906 907 /** 908 * gmc_v11_0_gart_enable - gart enable 909 * 910 * @adev: amdgpu_device pointer 911 */ 912 static int gmc_v11_0_gart_enable(struct amdgpu_device *adev) 913 { 914 int r; 915 bool value; 916 917 if (adev->gart.bo == NULL) { 918 dev_err(adev->dev, "No VRAM object for PCIE GART.\n"); 919 return -EINVAL; 920 } 921 922 amdgpu_gtt_mgr_recover(&adev->mman.gtt_mgr); 923 924 r = adev->mmhub.funcs->gart_enable(adev); 925 if (r) 926 return r; 927 928 /* Flush HDP after it is initialized */ 929 amdgpu_device_flush_hdp(adev, NULL); 930 931 value = amdgpu_vm_fault_stop != AMDGPU_VM_FAULT_STOP_ALWAYS; 932 933 adev->mmhub.funcs->set_fault_enable_default(adev, value); 934 gmc_v11_0_flush_gpu_tlb(adev, 0, AMDGPU_MMHUB0(0), 0); 935 936 drm_info(adev_to_drm(adev), "PCIE GART of %uM enabled (table at 0x%016llX).\n", 937 (unsigned int)(adev->gmc.gart_size >> 20), 938 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo)); 939 940 return 0; 941 } 942 943 static int gmc_v11_0_hw_init(struct amdgpu_ip_block *ip_block) 944 { 945 struct amdgpu_device *adev = ip_block->adev; 946 int r; 947 948 adev->gmc.flush_pasid_uses_kiq = !amdgpu_emu_mode; 949 950 /* The sequence of these two function calls matters.*/ 951 gmc_v11_0_init_golden_registers(adev); 952 953 r = gmc_v11_0_gart_enable(adev); 954 if (r) 955 return r; 956 957 if (adev->umc.funcs && adev->umc.funcs->init_registers) 958 adev->umc.funcs->init_registers(adev); 959 960 return 0; 961 } 962 963 /** 964 * gmc_v11_0_gart_disable - gart disable 965 * 966 * @adev: amdgpu_device pointer 967 * 968 * This disables all VM page table. 969 */ 970 static void gmc_v11_0_gart_disable(struct amdgpu_device *adev) 971 { 972 adev->mmhub.funcs->gart_disable(adev); 973 } 974 975 static int gmc_v11_0_hw_fini(struct amdgpu_ip_block *ip_block) 976 { 977 struct amdgpu_device *adev = ip_block->adev; 978 979 if (amdgpu_sriov_vf(adev)) { 980 /* full access mode, so don't touch any GMC register */ 981 DRM_DEBUG("For SRIOV client, shouldn't do anything.\n"); 982 return 0; 983 } 984 985 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0); 986 987 if (adev->gmc.ecc_irq.funcs && 988 amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__UMC)) 989 amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0); 990 991 gmc_v11_0_gart_disable(adev); 992 993 return 0; 994 } 995 996 static int gmc_v11_0_suspend(struct amdgpu_ip_block *ip_block) 997 { 998 gmc_v11_0_hw_fini(ip_block); 999 1000 return 0; 1001 } 1002 1003 static int gmc_v11_0_resume(struct amdgpu_ip_block *ip_block) 1004 { 1005 int r; 1006 1007 r = gmc_v11_0_hw_init(ip_block); 1008 if (r) 1009 return r; 1010 1011 amdgpu_vmid_reset_all(ip_block->adev); 1012 1013 return 0; 1014 } 1015 1016 static int gmc_v11_0_wait_for_idle(struct amdgpu_ip_block *ip_block) 1017 { 1018 /* There is no need to wait for MC idle in GMC v11.*/ 1019 return 0; 1020 } 1021 1022 static int gmc_v11_0_set_clockgating_state(struct amdgpu_ip_block *ip_block, 1023 enum amd_clockgating_state state) 1024 { 1025 int r; 1026 struct amdgpu_device *adev = ip_block->adev; 1027 1028 r = adev->mmhub.funcs->set_clockgating(adev, state); 1029 if (r) 1030 return r; 1031 1032 return athub_v3_0_set_clockgating(adev, state); 1033 } 1034 1035 static void gmc_v11_0_get_clockgating_state(struct amdgpu_ip_block *ip_block, u64 *flags) 1036 { 1037 struct amdgpu_device *adev = ip_block->adev; 1038 1039 adev->mmhub.funcs->get_clockgating(adev, flags); 1040 1041 athub_v3_0_get_clockgating(adev, flags); 1042 } 1043 1044 static int gmc_v11_0_set_powergating_state(struct amdgpu_ip_block *ip_block, 1045 enum amd_powergating_state state) 1046 { 1047 return 0; 1048 } 1049 1050 const struct amd_ip_funcs gmc_v11_0_ip_funcs = { 1051 .name = "gmc_v11_0", 1052 .early_init = gmc_v11_0_early_init, 1053 .sw_init = gmc_v11_0_sw_init, 1054 .hw_init = gmc_v11_0_hw_init, 1055 .late_init = gmc_v11_0_late_init, 1056 .sw_fini = gmc_v11_0_sw_fini, 1057 .hw_fini = gmc_v11_0_hw_fini, 1058 .suspend = gmc_v11_0_suspend, 1059 .resume = gmc_v11_0_resume, 1060 .wait_for_idle = gmc_v11_0_wait_for_idle, 1061 .set_clockgating_state = gmc_v11_0_set_clockgating_state, 1062 .set_powergating_state = gmc_v11_0_set_powergating_state, 1063 .get_clockgating_state = gmc_v11_0_get_clockgating_state, 1064 }; 1065 1066 const struct amdgpu_ip_block_version gmc_v11_0_ip_block = { 1067 .type = AMD_IP_BLOCK_TYPE_GMC, 1068 .major = 11, 1069 .minor = 0, 1070 .rev = 0, 1071 .funcs = &gmc_v11_0_ip_funcs, 1072 }; 1073