1 /* 2 * Copyright 2014 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 24 #include <linux/firmware.h> 25 #include <linux/module.h> 26 #include <linux/pci.h> 27 28 #include <drm/drm_cache.h> 29 #include "amdgpu.h" 30 #include "cikd.h" 31 #include "cik.h" 32 #include "gmc_v7_0.h" 33 #include "amdgpu_ucode.h" 34 #include "amdgpu_amdkfd.h" 35 #include "amdgpu_gem.h" 36 37 #include "bif/bif_4_1_d.h" 38 #include "bif/bif_4_1_sh_mask.h" 39 40 #include "gmc/gmc_7_1_d.h" 41 #include "gmc/gmc_7_1_sh_mask.h" 42 43 #include "oss/oss_2_0_d.h" 44 #include "oss/oss_2_0_sh_mask.h" 45 46 #include "dce/dce_8_0_d.h" 47 #include "dce/dce_8_0_sh_mask.h" 48 49 #include "amdgpu_atombios.h" 50 51 #include "ivsrcid/ivsrcid_vislands30.h" 52 53 static void gmc_v7_0_set_gmc_funcs(struct amdgpu_device *adev); 54 static void gmc_v7_0_set_irq_funcs(struct amdgpu_device *adev); 55 static int gmc_v7_0_wait_for_idle(void *handle); 56 57 MODULE_FIRMWARE("amdgpu/bonaire_mc.bin"); 58 MODULE_FIRMWARE("amdgpu/hawaii_mc.bin"); 59 MODULE_FIRMWARE("amdgpu/topaz_mc.bin"); 60 61 static const u32 golden_settings_iceland_a11[] = { 62 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff, 63 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff, 64 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff, 65 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff 66 }; 67 68 static const u32 iceland_mgcg_cgcg_init[] = { 69 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104 70 }; 71 72 static void gmc_v7_0_init_golden_registers(struct amdgpu_device *adev) 73 { 74 switch (adev->asic_type) { 75 case CHIP_TOPAZ: 76 amdgpu_device_program_register_sequence(adev, 77 iceland_mgcg_cgcg_init, 78 ARRAY_SIZE(iceland_mgcg_cgcg_init)); 79 amdgpu_device_program_register_sequence(adev, 80 golden_settings_iceland_a11, 81 ARRAY_SIZE(golden_settings_iceland_a11)); 82 break; 83 default: 84 break; 85 } 86 } 87 88 static void gmc_v7_0_mc_stop(struct amdgpu_device *adev) 89 { 90 u32 blackout; 91 92 gmc_v7_0_wait_for_idle((void *)adev); 93 94 blackout = RREG32(mmMC_SHARED_BLACKOUT_CNTL); 95 if (REG_GET_FIELD(blackout, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE) != 1) { 96 /* Block CPU access */ 97 WREG32(mmBIF_FB_EN, 0); 98 /* blackout the MC */ 99 blackout = REG_SET_FIELD(blackout, 100 MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 0); 101 WREG32(mmMC_SHARED_BLACKOUT_CNTL, blackout | 1); 102 } 103 /* wait for the MC to settle */ 104 udelay(100); 105 } 106 107 static void gmc_v7_0_mc_resume(struct amdgpu_device *adev) 108 { 109 u32 tmp; 110 111 /* unblackout the MC */ 112 tmp = RREG32(mmMC_SHARED_BLACKOUT_CNTL); 113 tmp = REG_SET_FIELD(tmp, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 0); 114 WREG32(mmMC_SHARED_BLACKOUT_CNTL, tmp); 115 /* allow CPU access */ 116 tmp = REG_SET_FIELD(0, BIF_FB_EN, FB_READ_EN, 1); 117 tmp = REG_SET_FIELD(tmp, BIF_FB_EN, FB_WRITE_EN, 1); 118 WREG32(mmBIF_FB_EN, tmp); 119 } 120 121 /** 122 * gmc_v7_0_init_microcode - load ucode images from disk 123 * 124 * @adev: amdgpu_device pointer 125 * 126 * Use the firmware interface to load the ucode images into 127 * the driver (not loaded into hw). 128 * Returns 0 on success, error on failure. 129 */ 130 static int gmc_v7_0_init_microcode(struct amdgpu_device *adev) 131 { 132 const char *chip_name; 133 char fw_name[30]; 134 int err; 135 136 DRM_DEBUG("\n"); 137 138 switch (adev->asic_type) { 139 case CHIP_BONAIRE: 140 chip_name = "bonaire"; 141 break; 142 case CHIP_HAWAII: 143 chip_name = "hawaii"; 144 break; 145 case CHIP_TOPAZ: 146 chip_name = "topaz"; 147 break; 148 case CHIP_KAVERI: 149 case CHIP_KABINI: 150 case CHIP_MULLINS: 151 return 0; 152 default: 153 return -EINVAL; 154 } 155 156 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mc.bin", chip_name); 157 158 err = amdgpu_ucode_request(adev, &adev->gmc.fw, fw_name); 159 if (err) { 160 pr_err("cik_mc: Failed to load firmware \"%s\"\n", fw_name); 161 amdgpu_ucode_release(&adev->gmc.fw); 162 } 163 return err; 164 } 165 166 /** 167 * gmc_v7_0_mc_load_microcode - load MC ucode into the hw 168 * 169 * @adev: amdgpu_device pointer 170 * 171 * Load the GDDR MC ucode into the hw (CIK). 172 * Returns 0 on success, error on failure. 173 */ 174 static int gmc_v7_0_mc_load_microcode(struct amdgpu_device *adev) 175 { 176 const struct mc_firmware_header_v1_0 *hdr; 177 const __le32 *fw_data = NULL; 178 const __le32 *io_mc_regs = NULL; 179 u32 running; 180 int i, ucode_size, regs_size; 181 182 if (!adev->gmc.fw) 183 return -EINVAL; 184 185 hdr = (const struct mc_firmware_header_v1_0 *)adev->gmc.fw->data; 186 amdgpu_ucode_print_mc_hdr(&hdr->header); 187 188 adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version); 189 regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2); 190 io_mc_regs = (const __le32 *) 191 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes)); 192 ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4; 193 fw_data = (const __le32 *) 194 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes)); 195 196 running = REG_GET_FIELD(RREG32(mmMC_SEQ_SUP_CNTL), MC_SEQ_SUP_CNTL, RUN); 197 198 if (running == 0) { 199 /* reset the engine and set to writable */ 200 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008); 201 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000010); 202 203 /* load mc io regs */ 204 for (i = 0; i < regs_size; i++) { 205 WREG32(mmMC_SEQ_IO_DEBUG_INDEX, le32_to_cpup(io_mc_regs++)); 206 WREG32(mmMC_SEQ_IO_DEBUG_DATA, le32_to_cpup(io_mc_regs++)); 207 } 208 /* load the MC ucode */ 209 for (i = 0; i < ucode_size; i++) 210 WREG32(mmMC_SEQ_SUP_PGM, le32_to_cpup(fw_data++)); 211 212 /* put the engine back into the active state */ 213 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008); 214 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000004); 215 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000001); 216 217 /* wait for training to complete */ 218 for (i = 0; i < adev->usec_timeout; i++) { 219 if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL), 220 MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D0)) 221 break; 222 udelay(1); 223 } 224 for (i = 0; i < adev->usec_timeout; i++) { 225 if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL), 226 MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D1)) 227 break; 228 udelay(1); 229 } 230 } 231 232 return 0; 233 } 234 235 static void gmc_v7_0_vram_gtt_location(struct amdgpu_device *adev, 236 struct amdgpu_gmc *mc) 237 { 238 u64 base = RREG32(mmMC_VM_FB_LOCATION) & 0xFFFF; 239 240 base <<= 24; 241 242 amdgpu_gmc_set_agp_default(adev, mc); 243 amdgpu_gmc_vram_location(adev, mc, base); 244 amdgpu_gmc_gart_location(adev, mc, AMDGPU_GART_PLACEMENT_BEST_FIT); 245 } 246 247 /** 248 * gmc_v7_0_mc_program - program the GPU memory controller 249 * 250 * @adev: amdgpu_device pointer 251 * 252 * Set the location of vram, gart, and AGP in the GPU's 253 * physical address space (CIK). 254 */ 255 static void gmc_v7_0_mc_program(struct amdgpu_device *adev) 256 { 257 u32 tmp; 258 int i, j; 259 260 /* Initialize HDP */ 261 for (i = 0, j = 0; i < 32; i++, j += 0x6) { 262 WREG32((0xb05 + j), 0x00000000); 263 WREG32((0xb06 + j), 0x00000000); 264 WREG32((0xb07 + j), 0x00000000); 265 WREG32((0xb08 + j), 0x00000000); 266 WREG32((0xb09 + j), 0x00000000); 267 } 268 WREG32(mmHDP_REG_COHERENCY_FLUSH_CNTL, 0); 269 270 if (gmc_v7_0_wait_for_idle((void *)adev)) 271 dev_warn(adev->dev, "Wait for MC idle timedout !\n"); 272 273 if (adev->mode_info.num_crtc) { 274 /* Lockout access through VGA aperture*/ 275 tmp = RREG32(mmVGA_HDP_CONTROL); 276 tmp = REG_SET_FIELD(tmp, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1); 277 WREG32(mmVGA_HDP_CONTROL, tmp); 278 279 /* disable VGA render */ 280 tmp = RREG32(mmVGA_RENDER_CONTROL); 281 tmp = REG_SET_FIELD(tmp, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0); 282 WREG32(mmVGA_RENDER_CONTROL, tmp); 283 } 284 /* Update configuration */ 285 WREG32(mmMC_VM_SYSTEM_APERTURE_LOW_ADDR, 286 adev->gmc.vram_start >> 12); 287 WREG32(mmMC_VM_SYSTEM_APERTURE_HIGH_ADDR, 288 adev->gmc.vram_end >> 12); 289 WREG32(mmMC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 290 adev->mem_scratch.gpu_addr >> 12); 291 WREG32(mmMC_VM_AGP_BASE, 0); 292 WREG32(mmMC_VM_AGP_TOP, adev->gmc.agp_end >> 22); 293 WREG32(mmMC_VM_AGP_BOT, adev->gmc.agp_start >> 22); 294 if (gmc_v7_0_wait_for_idle((void *)adev)) 295 dev_warn(adev->dev, "Wait for MC idle timedout !\n"); 296 297 WREG32(mmBIF_FB_EN, BIF_FB_EN__FB_READ_EN_MASK | BIF_FB_EN__FB_WRITE_EN_MASK); 298 299 tmp = RREG32(mmHDP_MISC_CNTL); 300 tmp = REG_SET_FIELD(tmp, HDP_MISC_CNTL, FLUSH_INVALIDATE_CACHE, 0); 301 WREG32(mmHDP_MISC_CNTL, tmp); 302 303 tmp = RREG32(mmHDP_HOST_PATH_CNTL); 304 WREG32(mmHDP_HOST_PATH_CNTL, tmp); 305 } 306 307 /** 308 * gmc_v7_0_mc_init - initialize the memory controller driver params 309 * 310 * @adev: amdgpu_device pointer 311 * 312 * Look up the amount of vram, vram width, and decide how to place 313 * vram and gart within the GPU's physical address space (CIK). 314 * Returns 0 for success. 315 */ 316 static int gmc_v7_0_mc_init(struct amdgpu_device *adev) 317 { 318 int r; 319 320 adev->gmc.vram_width = amdgpu_atombios_get_vram_width(adev); 321 if (!adev->gmc.vram_width) { 322 u32 tmp; 323 int chansize, numchan; 324 325 /* Get VRAM informations */ 326 tmp = RREG32(mmMC_ARB_RAMCFG); 327 if (REG_GET_FIELD(tmp, MC_ARB_RAMCFG, CHANSIZE)) 328 chansize = 64; 329 else 330 chansize = 32; 331 332 tmp = RREG32(mmMC_SHARED_CHMAP); 333 switch (REG_GET_FIELD(tmp, MC_SHARED_CHMAP, NOOFCHAN)) { 334 case 0: 335 default: 336 numchan = 1; 337 break; 338 case 1: 339 numchan = 2; 340 break; 341 case 2: 342 numchan = 4; 343 break; 344 case 3: 345 numchan = 8; 346 break; 347 case 4: 348 numchan = 3; 349 break; 350 case 5: 351 numchan = 6; 352 break; 353 case 6: 354 numchan = 10; 355 break; 356 case 7: 357 numchan = 12; 358 break; 359 case 8: 360 numchan = 16; 361 break; 362 } 363 adev->gmc.vram_width = numchan * chansize; 364 } 365 /* size in MB on si */ 366 adev->gmc.mc_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL; 367 adev->gmc.real_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL; 368 369 if (!(adev->flags & AMD_IS_APU)) { 370 r = amdgpu_device_resize_fb_bar(adev); 371 if (r) 372 return r; 373 } 374 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0); 375 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0); 376 377 #ifdef CONFIG_X86_64 378 if ((adev->flags & AMD_IS_APU) && 379 adev->gmc.real_vram_size > adev->gmc.aper_size && 380 !amdgpu_passthrough(adev)) { 381 adev->gmc.aper_base = ((u64)RREG32(mmMC_VM_FB_OFFSET)) << 22; 382 adev->gmc.aper_size = adev->gmc.real_vram_size; 383 } 384 #endif 385 386 adev->gmc.visible_vram_size = adev->gmc.aper_size; 387 388 /* set the gart size */ 389 if (amdgpu_gart_size == -1) { 390 switch (adev->asic_type) { 391 case CHIP_TOPAZ: /* no MM engines */ 392 default: 393 adev->gmc.gart_size = 256ULL << 20; 394 break; 395 #ifdef CONFIG_DRM_AMDGPU_CIK 396 case CHIP_BONAIRE: /* UVD, VCE do not support GPUVM */ 397 case CHIP_HAWAII: /* UVD, VCE do not support GPUVM */ 398 case CHIP_KAVERI: /* UVD, VCE do not support GPUVM */ 399 case CHIP_KABINI: /* UVD, VCE do not support GPUVM */ 400 case CHIP_MULLINS: /* UVD, VCE do not support GPUVM */ 401 adev->gmc.gart_size = 1024ULL << 20; 402 break; 403 #endif 404 } 405 } else { 406 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20; 407 } 408 409 adev->gmc.gart_size += adev->pm.smu_prv_buffer_size; 410 gmc_v7_0_vram_gtt_location(adev, &adev->gmc); 411 412 return 0; 413 } 414 415 /** 416 * gmc_v7_0_flush_gpu_tlb_pasid - tlb flush via pasid 417 * 418 * @adev: amdgpu_device pointer 419 * @pasid: pasid to be flush 420 * @flush_type: type of flush 421 * @all_hub: flush all hubs 422 * @inst: is used to select which instance of KIQ to use for the invalidation 423 * 424 * Flush the TLB for the requested pasid. 425 */ 426 static void gmc_v7_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, 427 uint16_t pasid, uint32_t flush_type, 428 bool all_hub, uint32_t inst) 429 { 430 u32 mask = 0x0; 431 int vmid; 432 433 for (vmid = 1; vmid < 16; vmid++) { 434 u32 tmp = RREG32(mmATC_VMID0_PASID_MAPPING + vmid); 435 436 if ((tmp & ATC_VMID0_PASID_MAPPING__VALID_MASK) && 437 (tmp & ATC_VMID0_PASID_MAPPING__PASID_MASK) == pasid) 438 mask |= 1 << vmid; 439 } 440 441 WREG32(mmVM_INVALIDATE_REQUEST, mask); 442 RREG32(mmVM_INVALIDATE_RESPONSE); 443 } 444 445 /* 446 * GART 447 * VMID 0 is the physical GPU addresses as used by the kernel. 448 * VMIDs 1-15 are used for userspace clients and are handled 449 * by the amdgpu vm/hsa code. 450 */ 451 452 /** 453 * gmc_v7_0_flush_gpu_tlb - gart tlb flush callback 454 * 455 * @adev: amdgpu_device pointer 456 * @vmid: vm instance to flush 457 * @vmhub: which hub to flush 458 * @flush_type: type of flush 459 * * 460 * Flush the TLB for the requested page table (CIK). 461 */ 462 static void gmc_v7_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, 463 uint32_t vmhub, uint32_t flush_type) 464 { 465 /* bits 0-15 are the VM contexts0-15 */ 466 WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid); 467 } 468 469 static uint64_t gmc_v7_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring, 470 unsigned int vmid, uint64_t pd_addr) 471 { 472 uint32_t reg; 473 474 if (vmid < 8) 475 reg = mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vmid; 476 else 477 reg = mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8; 478 amdgpu_ring_emit_wreg(ring, reg, pd_addr >> 12); 479 480 /* bits 0-15 are the VM contexts0-15 */ 481 amdgpu_ring_emit_wreg(ring, mmVM_INVALIDATE_REQUEST, 1 << vmid); 482 483 return pd_addr; 484 } 485 486 static void gmc_v7_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned int vmid, 487 unsigned int pasid) 488 { 489 amdgpu_ring_emit_wreg(ring, mmIH_VMID_0_LUT + vmid, pasid); 490 } 491 492 static void gmc_v7_0_get_vm_pde(struct amdgpu_device *adev, int level, 493 uint64_t *addr, uint64_t *flags) 494 { 495 BUG_ON(*addr & 0xFFFFFF0000000FFFULL); 496 } 497 498 static void gmc_v7_0_get_vm_pte(struct amdgpu_device *adev, 499 struct amdgpu_bo_va_mapping *mapping, 500 uint64_t *flags) 501 { 502 *flags &= ~AMDGPU_PTE_EXECUTABLE; 503 *flags &= ~AMDGPU_PTE_PRT; 504 } 505 506 /** 507 * gmc_v7_0_set_fault_enable_default - update VM fault handling 508 * 509 * @adev: amdgpu_device pointer 510 * @value: true redirects VM faults to the default page 511 */ 512 static void gmc_v7_0_set_fault_enable_default(struct amdgpu_device *adev, 513 bool value) 514 { 515 u32 tmp; 516 517 tmp = RREG32(mmVM_CONTEXT1_CNTL); 518 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 519 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 520 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 521 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 522 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 523 PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, value); 524 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 525 VALID_PROTECTION_FAULT_ENABLE_DEFAULT, value); 526 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 527 READ_PROTECTION_FAULT_ENABLE_DEFAULT, value); 528 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 529 WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 530 WREG32(mmVM_CONTEXT1_CNTL, tmp); 531 } 532 533 /** 534 * gmc_v7_0_set_prt - set PRT VM fault 535 * 536 * @adev: amdgpu_device pointer 537 * @enable: enable/disable VM fault handling for PRT 538 */ 539 static void gmc_v7_0_set_prt(struct amdgpu_device *adev, bool enable) 540 { 541 uint32_t tmp; 542 543 if (enable && !adev->gmc.prt_warning) { 544 dev_warn(adev->dev, "Disabling VM faults because of PRT request!\n"); 545 adev->gmc.prt_warning = true; 546 } 547 548 tmp = RREG32(mmVM_PRT_CNTL); 549 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 550 CB_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable); 551 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 552 CB_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable); 553 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 554 TC_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable); 555 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 556 TC_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable); 557 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 558 L2_CACHE_STORE_INVALID_ENTRIES, enable); 559 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 560 L1_TLB_STORE_INVALID_ENTRIES, enable); 561 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 562 MASK_PDE0_FAULT, enable); 563 WREG32(mmVM_PRT_CNTL, tmp); 564 565 if (enable) { 566 uint32_t low = AMDGPU_VA_RESERVED_BOTTOM >> 567 AMDGPU_GPU_PAGE_SHIFT; 568 uint32_t high = adev->vm_manager.max_pfn - 569 (AMDGPU_VA_RESERVED_TOP >> AMDGPU_GPU_PAGE_SHIFT); 570 571 WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, low); 572 WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, low); 573 WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, low); 574 WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, low); 575 WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, high); 576 WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, high); 577 WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, high); 578 WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, high); 579 } else { 580 WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, 0xfffffff); 581 WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, 0xfffffff); 582 WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, 0xfffffff); 583 WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, 0xfffffff); 584 WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, 0x0); 585 WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, 0x0); 586 WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, 0x0); 587 WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, 0x0); 588 } 589 } 590 591 /** 592 * gmc_v7_0_gart_enable - gart enable 593 * 594 * @adev: amdgpu_device pointer 595 * 596 * This sets up the TLBs, programs the page tables for VMID0, 597 * sets up the hw for VMIDs 1-15 which are allocated on 598 * demand, and sets up the global locations for the LDS, GDS, 599 * and GPUVM for FSA64 clients (CIK). 600 * Returns 0 for success, errors for failure. 601 */ 602 static int gmc_v7_0_gart_enable(struct amdgpu_device *adev) 603 { 604 uint64_t table_addr; 605 u32 tmp, field; 606 int i; 607 608 if (adev->gart.bo == NULL) { 609 dev_err(adev->dev, "No VRAM object for PCIE GART.\n"); 610 return -EINVAL; 611 } 612 amdgpu_gtt_mgr_recover(&adev->mman.gtt_mgr); 613 table_addr = amdgpu_bo_gpu_offset(adev->gart.bo); 614 615 /* Setup TLB control */ 616 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL); 617 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 1); 618 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 1); 619 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE, 3); 620 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 1); 621 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_APERTURE_UNMAPPED_ACCESS, 0); 622 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp); 623 /* Setup L2 cache */ 624 tmp = RREG32(mmVM_L2_CNTL); 625 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 1); 626 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING, 1); 627 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE, 1); 628 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE, 1); 629 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, EFFECTIVE_L2_QUEUE_SIZE, 7); 630 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, CONTEXT1_IDENTITY_ACCESS_MODE, 1); 631 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_DEFAULT_PAGE_OUT_TO_SYSTEM_MEMORY, 1); 632 WREG32(mmVM_L2_CNTL, tmp); 633 tmp = REG_SET_FIELD(0, VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS, 1); 634 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_L2_CACHE, 1); 635 WREG32(mmVM_L2_CNTL2, tmp); 636 637 field = adev->vm_manager.fragment_size; 638 tmp = RREG32(mmVM_L2_CNTL3); 639 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY, 1); 640 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, BANK_SELECT, field); 641 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_FRAGMENT_SIZE, field); 642 WREG32(mmVM_L2_CNTL3, tmp); 643 /* setup context0 */ 644 WREG32(mmVM_CONTEXT0_PAGE_TABLE_START_ADDR, adev->gmc.gart_start >> 12); 645 WREG32(mmVM_CONTEXT0_PAGE_TABLE_END_ADDR, adev->gmc.gart_end >> 12); 646 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR, table_addr >> 12); 647 WREG32(mmVM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR, 648 (u32)(adev->dummy_page_addr >> 12)); 649 WREG32(mmVM_CONTEXT0_CNTL2, 0); 650 tmp = RREG32(mmVM_CONTEXT0_CNTL); 651 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, ENABLE_CONTEXT, 1); 652 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, PAGE_TABLE_DEPTH, 0); 653 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 654 WREG32(mmVM_CONTEXT0_CNTL, tmp); 655 656 WREG32(0x575, 0); 657 WREG32(0x576, 0); 658 WREG32(0x577, 0); 659 660 /* empty context1-15 */ 661 /* FIXME start with 4G, once using 2 level pt switch to full 662 * vm size space 663 */ 664 /* set vm size, must be a multiple of 4 */ 665 WREG32(mmVM_CONTEXT1_PAGE_TABLE_START_ADDR, 0); 666 WREG32(mmVM_CONTEXT1_PAGE_TABLE_END_ADDR, adev->vm_manager.max_pfn - 1); 667 for (i = 1; i < AMDGPU_NUM_VMID; i++) { 668 if (i < 8) 669 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + i, 670 table_addr >> 12); 671 else 672 WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + i - 8, 673 table_addr >> 12); 674 } 675 676 /* enable context1-15 */ 677 WREG32(mmVM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR, 678 (u32)(adev->dummy_page_addr >> 12)); 679 WREG32(mmVM_CONTEXT1_CNTL2, 4); 680 tmp = RREG32(mmVM_CONTEXT1_CNTL); 681 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, ENABLE_CONTEXT, 1); 682 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_DEPTH, 1); 683 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_BLOCK_SIZE, 684 adev->vm_manager.block_size - 9); 685 WREG32(mmVM_CONTEXT1_CNTL, tmp); 686 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) 687 gmc_v7_0_set_fault_enable_default(adev, false); 688 else 689 gmc_v7_0_set_fault_enable_default(adev, true); 690 691 if (adev->asic_type == CHIP_KAVERI) { 692 tmp = RREG32(mmCHUB_CONTROL); 693 tmp &= ~BYPASS_VM; 694 WREG32(mmCHUB_CONTROL, tmp); 695 } 696 697 gmc_v7_0_flush_gpu_tlb(adev, 0, 0, 0); 698 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n", 699 (unsigned int)(adev->gmc.gart_size >> 20), 700 (unsigned long long)table_addr); 701 return 0; 702 } 703 704 static int gmc_v7_0_gart_init(struct amdgpu_device *adev) 705 { 706 int r; 707 708 if (adev->gart.bo) { 709 WARN(1, "R600 PCIE GART already initialized\n"); 710 return 0; 711 } 712 /* Initialize common gart structure */ 713 r = amdgpu_gart_init(adev); 714 if (r) 715 return r; 716 adev->gart.table_size = adev->gart.num_gpu_pages * 8; 717 adev->gart.gart_pte_flags = 0; 718 return amdgpu_gart_table_vram_alloc(adev); 719 } 720 721 /** 722 * gmc_v7_0_gart_disable - gart disable 723 * 724 * @adev: amdgpu_device pointer 725 * 726 * This disables all VM page table (CIK). 727 */ 728 static void gmc_v7_0_gart_disable(struct amdgpu_device *adev) 729 { 730 u32 tmp; 731 732 /* Disable all tables */ 733 WREG32(mmVM_CONTEXT0_CNTL, 0); 734 WREG32(mmVM_CONTEXT1_CNTL, 0); 735 /* Setup TLB control */ 736 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL); 737 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 0); 738 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 0); 739 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 0); 740 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp); 741 /* Setup L2 cache */ 742 tmp = RREG32(mmVM_L2_CNTL); 743 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 0); 744 WREG32(mmVM_L2_CNTL, tmp); 745 WREG32(mmVM_L2_CNTL2, 0); 746 } 747 748 /** 749 * gmc_v7_0_vm_decode_fault - print human readable fault info 750 * 751 * @adev: amdgpu_device pointer 752 * @status: VM_CONTEXT1_PROTECTION_FAULT_STATUS register value 753 * @addr: VM_CONTEXT1_PROTECTION_FAULT_ADDR register value 754 * @mc_client: VM_CONTEXT1_PROTECTION_FAULT_MCCLIENT register value 755 * @pasid: debug logging only - no functional use 756 * 757 * Print human readable fault information (CIK). 758 */ 759 static void gmc_v7_0_vm_decode_fault(struct amdgpu_device *adev, u32 status, 760 u32 addr, u32 mc_client, unsigned int pasid) 761 { 762 u32 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, VMID); 763 u32 protections = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 764 PROTECTIONS); 765 char block[5] = { mc_client >> 24, (mc_client >> 16) & 0xff, 766 (mc_client >> 8) & 0xff, mc_client & 0xff, 0 }; 767 u32 mc_id; 768 769 mc_id = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 770 MEMORY_CLIENT_ID); 771 772 dev_err(adev->dev, "VM fault (0x%02x, vmid %d, pasid %d) at page %u, %s from '%s' (0x%08x) (%d)\n", 773 protections, vmid, pasid, addr, 774 REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 775 MEMORY_CLIENT_RW) ? 776 "write" : "read", block, mc_client, mc_id); 777 } 778 779 780 static const u32 mc_cg_registers[] = { 781 mmMC_HUB_MISC_HUB_CG, 782 mmMC_HUB_MISC_SIP_CG, 783 mmMC_HUB_MISC_VM_CG, 784 mmMC_XPB_CLK_GAT, 785 mmATC_MISC_CG, 786 mmMC_CITF_MISC_WR_CG, 787 mmMC_CITF_MISC_RD_CG, 788 mmMC_CITF_MISC_VM_CG, 789 mmVM_L2_CG, 790 }; 791 792 static const u32 mc_cg_ls_en[] = { 793 MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK, 794 MC_HUB_MISC_SIP_CG__MEM_LS_ENABLE_MASK, 795 MC_HUB_MISC_VM_CG__MEM_LS_ENABLE_MASK, 796 MC_XPB_CLK_GAT__MEM_LS_ENABLE_MASK, 797 ATC_MISC_CG__MEM_LS_ENABLE_MASK, 798 MC_CITF_MISC_WR_CG__MEM_LS_ENABLE_MASK, 799 MC_CITF_MISC_RD_CG__MEM_LS_ENABLE_MASK, 800 MC_CITF_MISC_VM_CG__MEM_LS_ENABLE_MASK, 801 VM_L2_CG__MEM_LS_ENABLE_MASK, 802 }; 803 804 static const u32 mc_cg_en[] = { 805 MC_HUB_MISC_HUB_CG__ENABLE_MASK, 806 MC_HUB_MISC_SIP_CG__ENABLE_MASK, 807 MC_HUB_MISC_VM_CG__ENABLE_MASK, 808 MC_XPB_CLK_GAT__ENABLE_MASK, 809 ATC_MISC_CG__ENABLE_MASK, 810 MC_CITF_MISC_WR_CG__ENABLE_MASK, 811 MC_CITF_MISC_RD_CG__ENABLE_MASK, 812 MC_CITF_MISC_VM_CG__ENABLE_MASK, 813 VM_L2_CG__ENABLE_MASK, 814 }; 815 816 static void gmc_v7_0_enable_mc_ls(struct amdgpu_device *adev, 817 bool enable) 818 { 819 int i; 820 u32 orig, data; 821 822 for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) { 823 orig = data = RREG32(mc_cg_registers[i]); 824 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_LS)) 825 data |= mc_cg_ls_en[i]; 826 else 827 data &= ~mc_cg_ls_en[i]; 828 if (data != orig) 829 WREG32(mc_cg_registers[i], data); 830 } 831 } 832 833 static void gmc_v7_0_enable_mc_mgcg(struct amdgpu_device *adev, 834 bool enable) 835 { 836 int i; 837 u32 orig, data; 838 839 for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) { 840 orig = data = RREG32(mc_cg_registers[i]); 841 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_MGCG)) 842 data |= mc_cg_en[i]; 843 else 844 data &= ~mc_cg_en[i]; 845 if (data != orig) 846 WREG32(mc_cg_registers[i], data); 847 } 848 } 849 850 static void gmc_v7_0_enable_bif_mgls(struct amdgpu_device *adev, 851 bool enable) 852 { 853 u32 orig, data; 854 855 orig = data = RREG32_PCIE(ixPCIE_CNTL2); 856 857 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_BIF_LS)) { 858 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_LS_EN, 1); 859 data = REG_SET_FIELD(data, PCIE_CNTL2, MST_MEM_LS_EN, 1); 860 data = REG_SET_FIELD(data, PCIE_CNTL2, REPLAY_MEM_LS_EN, 1); 861 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_AGGRESSIVE_LS_EN, 1); 862 } else { 863 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_LS_EN, 0); 864 data = REG_SET_FIELD(data, PCIE_CNTL2, MST_MEM_LS_EN, 0); 865 data = REG_SET_FIELD(data, PCIE_CNTL2, REPLAY_MEM_LS_EN, 0); 866 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_AGGRESSIVE_LS_EN, 0); 867 } 868 869 if (orig != data) 870 WREG32_PCIE(ixPCIE_CNTL2, data); 871 } 872 873 static void gmc_v7_0_enable_hdp_mgcg(struct amdgpu_device *adev, 874 bool enable) 875 { 876 u32 orig, data; 877 878 orig = data = RREG32(mmHDP_HOST_PATH_CNTL); 879 880 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_HDP_MGCG)) 881 data = REG_SET_FIELD(data, HDP_HOST_PATH_CNTL, CLOCK_GATING_DIS, 0); 882 else 883 data = REG_SET_FIELD(data, HDP_HOST_PATH_CNTL, CLOCK_GATING_DIS, 1); 884 885 if (orig != data) 886 WREG32(mmHDP_HOST_PATH_CNTL, data); 887 } 888 889 static void gmc_v7_0_enable_hdp_ls(struct amdgpu_device *adev, 890 bool enable) 891 { 892 u32 orig, data; 893 894 orig = data = RREG32(mmHDP_MEM_POWER_LS); 895 896 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_HDP_LS)) 897 data = REG_SET_FIELD(data, HDP_MEM_POWER_LS, LS_ENABLE, 1); 898 else 899 data = REG_SET_FIELD(data, HDP_MEM_POWER_LS, LS_ENABLE, 0); 900 901 if (orig != data) 902 WREG32(mmHDP_MEM_POWER_LS, data); 903 } 904 905 static int gmc_v7_0_convert_vram_type(int mc_seq_vram_type) 906 { 907 switch (mc_seq_vram_type) { 908 case MC_SEQ_MISC0__MT__GDDR1: 909 return AMDGPU_VRAM_TYPE_GDDR1; 910 case MC_SEQ_MISC0__MT__DDR2: 911 return AMDGPU_VRAM_TYPE_DDR2; 912 case MC_SEQ_MISC0__MT__GDDR3: 913 return AMDGPU_VRAM_TYPE_GDDR3; 914 case MC_SEQ_MISC0__MT__GDDR4: 915 return AMDGPU_VRAM_TYPE_GDDR4; 916 case MC_SEQ_MISC0__MT__GDDR5: 917 return AMDGPU_VRAM_TYPE_GDDR5; 918 case MC_SEQ_MISC0__MT__HBM: 919 return AMDGPU_VRAM_TYPE_HBM; 920 case MC_SEQ_MISC0__MT__DDR3: 921 return AMDGPU_VRAM_TYPE_DDR3; 922 default: 923 return AMDGPU_VRAM_TYPE_UNKNOWN; 924 } 925 } 926 927 static int gmc_v7_0_early_init(void *handle) 928 { 929 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 930 931 gmc_v7_0_set_gmc_funcs(adev); 932 gmc_v7_0_set_irq_funcs(adev); 933 934 adev->gmc.shared_aperture_start = 0x2000000000000000ULL; 935 adev->gmc.shared_aperture_end = 936 adev->gmc.shared_aperture_start + (4ULL << 30) - 1; 937 adev->gmc.private_aperture_start = 938 adev->gmc.shared_aperture_end + 1; 939 adev->gmc.private_aperture_end = 940 adev->gmc.private_aperture_start + (4ULL << 30) - 1; 941 adev->gmc.noretry_flags = AMDGPU_VM_NORETRY_FLAGS_TF; 942 943 return 0; 944 } 945 946 static int gmc_v7_0_late_init(void *handle) 947 { 948 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 949 950 if (amdgpu_vm_fault_stop != AMDGPU_VM_FAULT_STOP_ALWAYS) 951 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0); 952 else 953 return 0; 954 } 955 956 static unsigned int gmc_v7_0_get_vbios_fb_size(struct amdgpu_device *adev) 957 { 958 u32 d1vga_control = RREG32(mmD1VGA_CONTROL); 959 unsigned int size; 960 961 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) { 962 size = AMDGPU_VBIOS_VGA_ALLOCATION; 963 } else { 964 u32 viewport = RREG32(mmVIEWPORT_SIZE); 965 966 size = (REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_HEIGHT) * 967 REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_WIDTH) * 968 4); 969 } 970 971 return size; 972 } 973 974 static int gmc_v7_0_sw_init(void *handle) 975 { 976 int r; 977 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 978 979 set_bit(AMDGPU_GFXHUB(0), adev->vmhubs_mask); 980 981 if (adev->flags & AMD_IS_APU) { 982 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_UNKNOWN; 983 } else { 984 u32 tmp = RREG32(mmMC_SEQ_MISC0); 985 986 tmp &= MC_SEQ_MISC0__MT__MASK; 987 adev->gmc.vram_type = gmc_v7_0_convert_vram_type(tmp); 988 } 989 990 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_PAGE_INV_FAULT, &adev->gmc.vm_fault); 991 if (r) 992 return r; 993 994 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_MEM_PROT_FAULT, &adev->gmc.vm_fault); 995 if (r) 996 return r; 997 998 /* Adjust VM size here. 999 * Currently set to 4GB ((1 << 20) 4k pages). 1000 * Max GPUVM size for cayman and SI is 40 bits. 1001 */ 1002 amdgpu_vm_adjust_size(adev, 64, 9, 1, 40); 1003 1004 /* Set the internal MC address mask 1005 * This is the max address of the GPU's 1006 * internal address space. 1007 */ 1008 adev->gmc.mc_mask = 0xffffffffffULL; /* 40 bit MC */ 1009 1010 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(40)); 1011 if (r) { 1012 pr_warn("No suitable DMA available\n"); 1013 return r; 1014 } 1015 adev->need_swiotlb = drm_need_swiotlb(40); 1016 1017 r = gmc_v7_0_init_microcode(adev); 1018 if (r) { 1019 DRM_ERROR("Failed to load mc firmware!\n"); 1020 return r; 1021 } 1022 1023 r = gmc_v7_0_mc_init(adev); 1024 if (r) 1025 return r; 1026 1027 amdgpu_gmc_get_vbios_allocations(adev); 1028 1029 /* Memory manager */ 1030 r = amdgpu_bo_init(adev); 1031 if (r) 1032 return r; 1033 1034 r = gmc_v7_0_gart_init(adev); 1035 if (r) 1036 return r; 1037 1038 /* 1039 * number of VMs 1040 * VMID 0 is reserved for System 1041 * amdgpu graphics/compute will use VMIDs 1-7 1042 * amdkfd will use VMIDs 8-15 1043 */ 1044 adev->vm_manager.first_kfd_vmid = 8; 1045 amdgpu_vm_manager_init(adev); 1046 1047 /* base offset of vram pages */ 1048 if (adev->flags & AMD_IS_APU) { 1049 u64 tmp = RREG32(mmMC_VM_FB_OFFSET); 1050 1051 tmp <<= 22; 1052 adev->vm_manager.vram_base_offset = tmp; 1053 } else { 1054 adev->vm_manager.vram_base_offset = 0; 1055 } 1056 1057 adev->gmc.vm_fault_info = kmalloc(sizeof(struct kfd_vm_fault_info), 1058 GFP_KERNEL); 1059 if (!adev->gmc.vm_fault_info) 1060 return -ENOMEM; 1061 atomic_set(&adev->gmc.vm_fault_info_updated, 0); 1062 1063 return 0; 1064 } 1065 1066 static int gmc_v7_0_sw_fini(void *handle) 1067 { 1068 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1069 1070 amdgpu_gem_force_release(adev); 1071 amdgpu_vm_manager_fini(adev); 1072 kfree(adev->gmc.vm_fault_info); 1073 amdgpu_gart_table_vram_free(adev); 1074 amdgpu_bo_fini(adev); 1075 amdgpu_ucode_release(&adev->gmc.fw); 1076 1077 return 0; 1078 } 1079 1080 static int gmc_v7_0_hw_init(void *handle) 1081 { 1082 int r; 1083 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1084 1085 gmc_v7_0_init_golden_registers(adev); 1086 1087 gmc_v7_0_mc_program(adev); 1088 1089 if (!(adev->flags & AMD_IS_APU)) { 1090 r = gmc_v7_0_mc_load_microcode(adev); 1091 if (r) { 1092 DRM_ERROR("Failed to load MC firmware!\n"); 1093 return r; 1094 } 1095 } 1096 1097 r = gmc_v7_0_gart_enable(adev); 1098 if (r) 1099 return r; 1100 1101 if (amdgpu_emu_mode == 1) 1102 return amdgpu_gmc_vram_checking(adev); 1103 1104 return 0; 1105 } 1106 1107 static int gmc_v7_0_hw_fini(void *handle) 1108 { 1109 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1110 1111 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0); 1112 gmc_v7_0_gart_disable(adev); 1113 1114 return 0; 1115 } 1116 1117 static int gmc_v7_0_suspend(void *handle) 1118 { 1119 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1120 1121 gmc_v7_0_hw_fini(adev); 1122 1123 return 0; 1124 } 1125 1126 static int gmc_v7_0_resume(void *handle) 1127 { 1128 int r; 1129 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1130 1131 r = gmc_v7_0_hw_init(adev); 1132 if (r) 1133 return r; 1134 1135 amdgpu_vmid_reset_all(adev); 1136 1137 return 0; 1138 } 1139 1140 static bool gmc_v7_0_is_idle(void *handle) 1141 { 1142 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1143 u32 tmp = RREG32(mmSRBM_STATUS); 1144 1145 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1146 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK | SRBM_STATUS__VMC_BUSY_MASK)) 1147 return false; 1148 1149 return true; 1150 } 1151 1152 static int gmc_v7_0_wait_for_idle(void *handle) 1153 { 1154 unsigned int i; 1155 u32 tmp; 1156 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1157 1158 for (i = 0; i < adev->usec_timeout; i++) { 1159 /* read MC_STATUS */ 1160 tmp = RREG32(mmSRBM_STATUS) & (SRBM_STATUS__MCB_BUSY_MASK | 1161 SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1162 SRBM_STATUS__MCC_BUSY_MASK | 1163 SRBM_STATUS__MCD_BUSY_MASK | 1164 SRBM_STATUS__VMC_BUSY_MASK); 1165 if (!tmp) 1166 return 0; 1167 udelay(1); 1168 } 1169 return -ETIMEDOUT; 1170 1171 } 1172 1173 static int gmc_v7_0_soft_reset(void *handle) 1174 { 1175 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1176 u32 srbm_soft_reset = 0; 1177 u32 tmp = RREG32(mmSRBM_STATUS); 1178 1179 if (tmp & SRBM_STATUS__VMC_BUSY_MASK) 1180 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset, 1181 SRBM_SOFT_RESET, SOFT_RESET_VMC, 1); 1182 1183 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1184 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK)) { 1185 if (!(adev->flags & AMD_IS_APU)) 1186 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset, 1187 SRBM_SOFT_RESET, SOFT_RESET_MC, 1); 1188 } 1189 1190 if (srbm_soft_reset) { 1191 gmc_v7_0_mc_stop(adev); 1192 if (gmc_v7_0_wait_for_idle((void *)adev)) 1193 dev_warn(adev->dev, "Wait for GMC idle timed out !\n"); 1194 1195 tmp = RREG32(mmSRBM_SOFT_RESET); 1196 tmp |= srbm_soft_reset; 1197 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp); 1198 WREG32(mmSRBM_SOFT_RESET, tmp); 1199 tmp = RREG32(mmSRBM_SOFT_RESET); 1200 1201 udelay(50); 1202 1203 tmp &= ~srbm_soft_reset; 1204 WREG32(mmSRBM_SOFT_RESET, tmp); 1205 tmp = RREG32(mmSRBM_SOFT_RESET); 1206 1207 /* Wait a little for things to settle down */ 1208 udelay(50); 1209 1210 gmc_v7_0_mc_resume(adev); 1211 udelay(50); 1212 } 1213 1214 return 0; 1215 } 1216 1217 static int gmc_v7_0_vm_fault_interrupt_state(struct amdgpu_device *adev, 1218 struct amdgpu_irq_src *src, 1219 unsigned int type, 1220 enum amdgpu_interrupt_state state) 1221 { 1222 u32 tmp; 1223 u32 bits = (VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1224 VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1225 VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1226 VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1227 VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1228 VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK); 1229 1230 switch (state) { 1231 case AMDGPU_IRQ_STATE_DISABLE: 1232 /* system context */ 1233 tmp = RREG32(mmVM_CONTEXT0_CNTL); 1234 tmp &= ~bits; 1235 WREG32(mmVM_CONTEXT0_CNTL, tmp); 1236 /* VMs */ 1237 tmp = RREG32(mmVM_CONTEXT1_CNTL); 1238 tmp &= ~bits; 1239 WREG32(mmVM_CONTEXT1_CNTL, tmp); 1240 break; 1241 case AMDGPU_IRQ_STATE_ENABLE: 1242 /* system context */ 1243 tmp = RREG32(mmVM_CONTEXT0_CNTL); 1244 tmp |= bits; 1245 WREG32(mmVM_CONTEXT0_CNTL, tmp); 1246 /* VMs */ 1247 tmp = RREG32(mmVM_CONTEXT1_CNTL); 1248 tmp |= bits; 1249 WREG32(mmVM_CONTEXT1_CNTL, tmp); 1250 break; 1251 default: 1252 break; 1253 } 1254 1255 return 0; 1256 } 1257 1258 static int gmc_v7_0_process_interrupt(struct amdgpu_device *adev, 1259 struct amdgpu_irq_src *source, 1260 struct amdgpu_iv_entry *entry) 1261 { 1262 u32 addr, status, mc_client, vmid; 1263 1264 addr = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_ADDR); 1265 status = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_STATUS); 1266 mc_client = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_MCCLIENT); 1267 /* reset addr and status */ 1268 WREG32_P(mmVM_CONTEXT1_CNTL2, 1, ~1); 1269 1270 if (!addr && !status) 1271 return 0; 1272 1273 amdgpu_vm_update_fault_cache(adev, entry->pasid, 1274 ((u64)addr) << AMDGPU_GPU_PAGE_SHIFT, status, AMDGPU_GFXHUB(0)); 1275 1276 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_FIRST) 1277 gmc_v7_0_set_fault_enable_default(adev, false); 1278 1279 if (printk_ratelimit()) { 1280 dev_err(adev->dev, "GPU fault detected: %d 0x%08x\n", 1281 entry->src_id, entry->src_data[0]); 1282 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n", 1283 addr); 1284 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n", 1285 status); 1286 gmc_v7_0_vm_decode_fault(adev, status, addr, mc_client, 1287 entry->pasid); 1288 } 1289 1290 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 1291 VMID); 1292 if (amdgpu_amdkfd_is_kfd_vmid(adev, vmid) 1293 && !atomic_read(&adev->gmc.vm_fault_info_updated)) { 1294 struct kfd_vm_fault_info *info = adev->gmc.vm_fault_info; 1295 u32 protections = REG_GET_FIELD(status, 1296 VM_CONTEXT1_PROTECTION_FAULT_STATUS, 1297 PROTECTIONS); 1298 1299 info->vmid = vmid; 1300 info->mc_id = REG_GET_FIELD(status, 1301 VM_CONTEXT1_PROTECTION_FAULT_STATUS, 1302 MEMORY_CLIENT_ID); 1303 info->status = status; 1304 info->page_addr = addr; 1305 info->prot_valid = protections & 0x7 ? true : false; 1306 info->prot_read = protections & 0x8 ? true : false; 1307 info->prot_write = protections & 0x10 ? true : false; 1308 info->prot_exec = protections & 0x20 ? true : false; 1309 mb(); 1310 atomic_set(&adev->gmc.vm_fault_info_updated, 1); 1311 } 1312 1313 return 0; 1314 } 1315 1316 static int gmc_v7_0_set_clockgating_state(void *handle, 1317 enum amd_clockgating_state state) 1318 { 1319 bool gate = false; 1320 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1321 1322 if (state == AMD_CG_STATE_GATE) 1323 gate = true; 1324 1325 if (!(adev->flags & AMD_IS_APU)) { 1326 gmc_v7_0_enable_mc_mgcg(adev, gate); 1327 gmc_v7_0_enable_mc_ls(adev, gate); 1328 } 1329 gmc_v7_0_enable_bif_mgls(adev, gate); 1330 gmc_v7_0_enable_hdp_mgcg(adev, gate); 1331 gmc_v7_0_enable_hdp_ls(adev, gate); 1332 1333 return 0; 1334 } 1335 1336 static int gmc_v7_0_set_powergating_state(void *handle, 1337 enum amd_powergating_state state) 1338 { 1339 return 0; 1340 } 1341 1342 static const struct amd_ip_funcs gmc_v7_0_ip_funcs = { 1343 .name = "gmc_v7_0", 1344 .early_init = gmc_v7_0_early_init, 1345 .late_init = gmc_v7_0_late_init, 1346 .sw_init = gmc_v7_0_sw_init, 1347 .sw_fini = gmc_v7_0_sw_fini, 1348 .hw_init = gmc_v7_0_hw_init, 1349 .hw_fini = gmc_v7_0_hw_fini, 1350 .suspend = gmc_v7_0_suspend, 1351 .resume = gmc_v7_0_resume, 1352 .is_idle = gmc_v7_0_is_idle, 1353 .wait_for_idle = gmc_v7_0_wait_for_idle, 1354 .soft_reset = gmc_v7_0_soft_reset, 1355 .set_clockgating_state = gmc_v7_0_set_clockgating_state, 1356 .set_powergating_state = gmc_v7_0_set_powergating_state, 1357 .dump_ip_state = NULL, 1358 .print_ip_state = NULL, 1359 }; 1360 1361 static const struct amdgpu_gmc_funcs gmc_v7_0_gmc_funcs = { 1362 .flush_gpu_tlb = gmc_v7_0_flush_gpu_tlb, 1363 .flush_gpu_tlb_pasid = gmc_v7_0_flush_gpu_tlb_pasid, 1364 .emit_flush_gpu_tlb = gmc_v7_0_emit_flush_gpu_tlb, 1365 .emit_pasid_mapping = gmc_v7_0_emit_pasid_mapping, 1366 .set_prt = gmc_v7_0_set_prt, 1367 .get_vm_pde = gmc_v7_0_get_vm_pde, 1368 .get_vm_pte = gmc_v7_0_get_vm_pte, 1369 .get_vbios_fb_size = gmc_v7_0_get_vbios_fb_size, 1370 }; 1371 1372 static const struct amdgpu_irq_src_funcs gmc_v7_0_irq_funcs = { 1373 .set = gmc_v7_0_vm_fault_interrupt_state, 1374 .process = gmc_v7_0_process_interrupt, 1375 }; 1376 1377 static void gmc_v7_0_set_gmc_funcs(struct amdgpu_device *adev) 1378 { 1379 adev->gmc.gmc_funcs = &gmc_v7_0_gmc_funcs; 1380 } 1381 1382 static void gmc_v7_0_set_irq_funcs(struct amdgpu_device *adev) 1383 { 1384 adev->gmc.vm_fault.num_types = 1; 1385 adev->gmc.vm_fault.funcs = &gmc_v7_0_irq_funcs; 1386 } 1387 1388 const struct amdgpu_ip_block_version gmc_v7_0_ip_block = { 1389 .type = AMD_IP_BLOCK_TYPE_GMC, 1390 .major = 7, 1391 .minor = 0, 1392 .rev = 0, 1393 .funcs = &gmc_v7_0_ip_funcs, 1394 }; 1395 1396 const struct amdgpu_ip_block_version gmc_v7_4_ip_block = { 1397 .type = AMD_IP_BLOCK_TYPE_GMC, 1398 .major = 7, 1399 .minor = 4, 1400 .rev = 0, 1401 .funcs = &gmc_v7_0_ip_funcs, 1402 }; 1403