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_vram_location(adev, mc, base); 243 amdgpu_gmc_gart_location(adev, mc); 244 } 245 246 /** 247 * gmc_v7_0_mc_program - program the GPU memory controller 248 * 249 * @adev: amdgpu_device pointer 250 * 251 * Set the location of vram, gart, and AGP in the GPU's 252 * physical address space (CIK). 253 */ 254 static void gmc_v7_0_mc_program(struct amdgpu_device *adev) 255 { 256 u32 tmp; 257 int i, j; 258 259 /* Initialize HDP */ 260 for (i = 0, j = 0; i < 32; i++, j += 0x6) { 261 WREG32((0xb05 + j), 0x00000000); 262 WREG32((0xb06 + j), 0x00000000); 263 WREG32((0xb07 + j), 0x00000000); 264 WREG32((0xb08 + j), 0x00000000); 265 WREG32((0xb09 + j), 0x00000000); 266 } 267 WREG32(mmHDP_REG_COHERENCY_FLUSH_CNTL, 0); 268 269 if (gmc_v7_0_wait_for_idle((void *)adev)) 270 dev_warn(adev->dev, "Wait for MC idle timedout !\n"); 271 272 if (adev->mode_info.num_crtc) { 273 /* Lockout access through VGA aperture*/ 274 tmp = RREG32(mmVGA_HDP_CONTROL); 275 tmp = REG_SET_FIELD(tmp, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1); 276 WREG32(mmVGA_HDP_CONTROL, tmp); 277 278 /* disable VGA render */ 279 tmp = RREG32(mmVGA_RENDER_CONTROL); 280 tmp = REG_SET_FIELD(tmp, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0); 281 WREG32(mmVGA_RENDER_CONTROL, tmp); 282 } 283 /* Update configuration */ 284 WREG32(mmMC_VM_SYSTEM_APERTURE_LOW_ADDR, 285 adev->gmc.vram_start >> 12); 286 WREG32(mmMC_VM_SYSTEM_APERTURE_HIGH_ADDR, 287 adev->gmc.vram_end >> 12); 288 WREG32(mmMC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 289 adev->mem_scratch.gpu_addr >> 12); 290 WREG32(mmMC_VM_AGP_BASE, 0); 291 WREG32(mmMC_VM_AGP_TOP, 0x0FFFFFFF); 292 WREG32(mmMC_VM_AGP_BOT, 0x0FFFFFFF); 293 if (gmc_v7_0_wait_for_idle((void *)adev)) 294 dev_warn(adev->dev, "Wait for MC idle timedout !\n"); 295 296 WREG32(mmBIF_FB_EN, BIF_FB_EN__FB_READ_EN_MASK | BIF_FB_EN__FB_WRITE_EN_MASK); 297 298 tmp = RREG32(mmHDP_MISC_CNTL); 299 tmp = REG_SET_FIELD(tmp, HDP_MISC_CNTL, FLUSH_INVALIDATE_CACHE, 0); 300 WREG32(mmHDP_MISC_CNTL, tmp); 301 302 tmp = RREG32(mmHDP_HOST_PATH_CNTL); 303 WREG32(mmHDP_HOST_PATH_CNTL, tmp); 304 } 305 306 /** 307 * gmc_v7_0_mc_init - initialize the memory controller driver params 308 * 309 * @adev: amdgpu_device pointer 310 * 311 * Look up the amount of vram, vram width, and decide how to place 312 * vram and gart within the GPU's physical address space (CIK). 313 * Returns 0 for success. 314 */ 315 static int gmc_v7_0_mc_init(struct amdgpu_device *adev) 316 { 317 int r; 318 319 adev->gmc.vram_width = amdgpu_atombios_get_vram_width(adev); 320 if (!adev->gmc.vram_width) { 321 u32 tmp; 322 int chansize, numchan; 323 324 /* Get VRAM informations */ 325 tmp = RREG32(mmMC_ARB_RAMCFG); 326 if (REG_GET_FIELD(tmp, MC_ARB_RAMCFG, CHANSIZE)) 327 chansize = 64; 328 else 329 chansize = 32; 330 331 tmp = RREG32(mmMC_SHARED_CHMAP); 332 switch (REG_GET_FIELD(tmp, MC_SHARED_CHMAP, NOOFCHAN)) { 333 case 0: 334 default: 335 numchan = 1; 336 break; 337 case 1: 338 numchan = 2; 339 break; 340 case 2: 341 numchan = 4; 342 break; 343 case 3: 344 numchan = 8; 345 break; 346 case 4: 347 numchan = 3; 348 break; 349 case 5: 350 numchan = 6; 351 break; 352 case 6: 353 numchan = 10; 354 break; 355 case 7: 356 numchan = 12; 357 break; 358 case 8: 359 numchan = 16; 360 break; 361 } 362 adev->gmc.vram_width = numchan * chansize; 363 } 364 /* size in MB on si */ 365 adev->gmc.mc_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL; 366 adev->gmc.real_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL; 367 368 if (!(adev->flags & AMD_IS_APU)) { 369 r = amdgpu_device_resize_fb_bar(adev); 370 if (r) 371 return r; 372 } 373 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0); 374 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0); 375 376 #ifdef CONFIG_X86_64 377 if ((adev->flags & AMD_IS_APU) && 378 adev->gmc.real_vram_size > adev->gmc.aper_size && 379 !amdgpu_passthrough(adev)) { 380 adev->gmc.aper_base = ((u64)RREG32(mmMC_VM_FB_OFFSET)) << 22; 381 adev->gmc.aper_size = adev->gmc.real_vram_size; 382 } 383 #endif 384 385 adev->gmc.visible_vram_size = adev->gmc.aper_size; 386 387 /* set the gart size */ 388 if (amdgpu_gart_size == -1) { 389 switch (adev->asic_type) { 390 case CHIP_TOPAZ: /* no MM engines */ 391 default: 392 adev->gmc.gart_size = 256ULL << 20; 393 break; 394 #ifdef CONFIG_DRM_AMDGPU_CIK 395 case CHIP_BONAIRE: /* UVD, VCE do not support GPUVM */ 396 case CHIP_HAWAII: /* UVD, VCE do not support GPUVM */ 397 case CHIP_KAVERI: /* UVD, VCE do not support GPUVM */ 398 case CHIP_KABINI: /* UVD, VCE do not support GPUVM */ 399 case CHIP_MULLINS: /* UVD, VCE do not support GPUVM */ 400 adev->gmc.gart_size = 1024ULL << 20; 401 break; 402 #endif 403 } 404 } else { 405 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20; 406 } 407 408 adev->gmc.gart_size += adev->pm.smu_prv_buffer_size; 409 gmc_v7_0_vram_gtt_location(adev, &adev->gmc); 410 411 return 0; 412 } 413 414 /** 415 * gmc_v7_0_flush_gpu_tlb_pasid - tlb flush via pasid 416 * 417 * @adev: amdgpu_device pointer 418 * @pasid: pasid to be flush 419 * @flush_type: type of flush 420 * @all_hub: flush all hubs 421 * @inst: is used to select which instance of KIQ to use for the invalidation 422 * 423 * Flush the TLB for the requested pasid. 424 */ 425 static int gmc_v7_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, 426 uint16_t pasid, uint32_t flush_type, 427 bool all_hub, uint32_t inst) 428 { 429 int vmid; 430 unsigned int tmp; 431 432 if (amdgpu_in_reset(adev)) 433 return -EIO; 434 435 for (vmid = 1; vmid < 16; vmid++) { 436 437 tmp = RREG32(mmATC_VMID0_PASID_MAPPING + vmid); 438 if ((tmp & ATC_VMID0_PASID_MAPPING__VALID_MASK) && 439 (tmp & ATC_VMID0_PASID_MAPPING__PASID_MASK) == pasid) { 440 WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid); 441 RREG32(mmVM_INVALIDATE_RESPONSE); 442 break; 443 } 444 } 445 446 return 0; 447 } 448 449 /* 450 * GART 451 * VMID 0 is the physical GPU addresses as used by the kernel. 452 * VMIDs 1-15 are used for userspace clients and are handled 453 * by the amdgpu vm/hsa code. 454 */ 455 456 /** 457 * gmc_v7_0_flush_gpu_tlb - gart tlb flush callback 458 * 459 * @adev: amdgpu_device pointer 460 * @vmid: vm instance to flush 461 * @vmhub: which hub to flush 462 * @flush_type: type of flush 463 * * 464 * Flush the TLB for the requested page table (CIK). 465 */ 466 static void gmc_v7_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, 467 uint32_t vmhub, uint32_t flush_type) 468 { 469 /* bits 0-15 are the VM contexts0-15 */ 470 WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid); 471 } 472 473 static uint64_t gmc_v7_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring, 474 unsigned int vmid, uint64_t pd_addr) 475 { 476 uint32_t reg; 477 478 if (vmid < 8) 479 reg = mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vmid; 480 else 481 reg = mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8; 482 amdgpu_ring_emit_wreg(ring, reg, pd_addr >> 12); 483 484 /* bits 0-15 are the VM contexts0-15 */ 485 amdgpu_ring_emit_wreg(ring, mmVM_INVALIDATE_REQUEST, 1 << vmid); 486 487 return pd_addr; 488 } 489 490 static void gmc_v7_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned int vmid, 491 unsigned int pasid) 492 { 493 amdgpu_ring_emit_wreg(ring, mmIH_VMID_0_LUT + vmid, pasid); 494 } 495 496 static void gmc_v7_0_get_vm_pde(struct amdgpu_device *adev, int level, 497 uint64_t *addr, uint64_t *flags) 498 { 499 BUG_ON(*addr & 0xFFFFFF0000000FFFULL); 500 } 501 502 static void gmc_v7_0_get_vm_pte(struct amdgpu_device *adev, 503 struct amdgpu_bo_va_mapping *mapping, 504 uint64_t *flags) 505 { 506 *flags &= ~AMDGPU_PTE_EXECUTABLE; 507 *flags &= ~AMDGPU_PTE_PRT; 508 } 509 510 /** 511 * gmc_v7_0_set_fault_enable_default - update VM fault handling 512 * 513 * @adev: amdgpu_device pointer 514 * @value: true redirects VM faults to the default page 515 */ 516 static void gmc_v7_0_set_fault_enable_default(struct amdgpu_device *adev, 517 bool value) 518 { 519 u32 tmp; 520 521 tmp = RREG32(mmVM_CONTEXT1_CNTL); 522 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 523 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 524 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 525 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 526 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 527 PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, value); 528 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 529 VALID_PROTECTION_FAULT_ENABLE_DEFAULT, value); 530 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 531 READ_PROTECTION_FAULT_ENABLE_DEFAULT, value); 532 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 533 WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 534 WREG32(mmVM_CONTEXT1_CNTL, tmp); 535 } 536 537 /** 538 * gmc_v7_0_set_prt - set PRT VM fault 539 * 540 * @adev: amdgpu_device pointer 541 * @enable: enable/disable VM fault handling for PRT 542 */ 543 static void gmc_v7_0_set_prt(struct amdgpu_device *adev, bool enable) 544 { 545 uint32_t tmp; 546 547 if (enable && !adev->gmc.prt_warning) { 548 dev_warn(adev->dev, "Disabling VM faults because of PRT request!\n"); 549 adev->gmc.prt_warning = true; 550 } 551 552 tmp = RREG32(mmVM_PRT_CNTL); 553 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 554 CB_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable); 555 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 556 CB_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable); 557 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 558 TC_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable); 559 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 560 TC_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable); 561 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 562 L2_CACHE_STORE_INVALID_ENTRIES, enable); 563 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 564 L1_TLB_STORE_INVALID_ENTRIES, enable); 565 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 566 MASK_PDE0_FAULT, enable); 567 WREG32(mmVM_PRT_CNTL, tmp); 568 569 if (enable) { 570 uint32_t low = AMDGPU_VA_RESERVED_SIZE >> AMDGPU_GPU_PAGE_SHIFT; 571 uint32_t high = adev->vm_manager.max_pfn - 572 (AMDGPU_VA_RESERVED_SIZE >> AMDGPU_GPU_PAGE_SHIFT); 573 574 WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, low); 575 WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, low); 576 WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, low); 577 WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, low); 578 WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, high); 579 WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, high); 580 WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, high); 581 WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, high); 582 } else { 583 WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, 0xfffffff); 584 WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, 0xfffffff); 585 WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, 0xfffffff); 586 WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, 0xfffffff); 587 WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, 0x0); 588 WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, 0x0); 589 WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, 0x0); 590 WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, 0x0); 591 } 592 } 593 594 /** 595 * gmc_v7_0_gart_enable - gart enable 596 * 597 * @adev: amdgpu_device pointer 598 * 599 * This sets up the TLBs, programs the page tables for VMID0, 600 * sets up the hw for VMIDs 1-15 which are allocated on 601 * demand, and sets up the global locations for the LDS, GDS, 602 * and GPUVM for FSA64 clients (CIK). 603 * Returns 0 for success, errors for failure. 604 */ 605 static int gmc_v7_0_gart_enable(struct amdgpu_device *adev) 606 { 607 uint64_t table_addr; 608 u32 tmp, field; 609 int i; 610 611 if (adev->gart.bo == NULL) { 612 dev_err(adev->dev, "No VRAM object for PCIE GART.\n"); 613 return -EINVAL; 614 } 615 amdgpu_gtt_mgr_recover(&adev->mman.gtt_mgr); 616 table_addr = amdgpu_bo_gpu_offset(adev->gart.bo); 617 618 /* Setup TLB control */ 619 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL); 620 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 1); 621 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 1); 622 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE, 3); 623 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 1); 624 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_APERTURE_UNMAPPED_ACCESS, 0); 625 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp); 626 /* Setup L2 cache */ 627 tmp = RREG32(mmVM_L2_CNTL); 628 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 1); 629 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING, 1); 630 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE, 1); 631 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE, 1); 632 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, EFFECTIVE_L2_QUEUE_SIZE, 7); 633 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, CONTEXT1_IDENTITY_ACCESS_MODE, 1); 634 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_DEFAULT_PAGE_OUT_TO_SYSTEM_MEMORY, 1); 635 WREG32(mmVM_L2_CNTL, tmp); 636 tmp = REG_SET_FIELD(0, VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS, 1); 637 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_L2_CACHE, 1); 638 WREG32(mmVM_L2_CNTL2, tmp); 639 640 field = adev->vm_manager.fragment_size; 641 tmp = RREG32(mmVM_L2_CNTL3); 642 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY, 1); 643 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, BANK_SELECT, field); 644 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_FRAGMENT_SIZE, field); 645 WREG32(mmVM_L2_CNTL3, tmp); 646 /* setup context0 */ 647 WREG32(mmVM_CONTEXT0_PAGE_TABLE_START_ADDR, adev->gmc.gart_start >> 12); 648 WREG32(mmVM_CONTEXT0_PAGE_TABLE_END_ADDR, adev->gmc.gart_end >> 12); 649 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR, table_addr >> 12); 650 WREG32(mmVM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR, 651 (u32)(adev->dummy_page_addr >> 12)); 652 WREG32(mmVM_CONTEXT0_CNTL2, 0); 653 tmp = RREG32(mmVM_CONTEXT0_CNTL); 654 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, ENABLE_CONTEXT, 1); 655 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, PAGE_TABLE_DEPTH, 0); 656 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 657 WREG32(mmVM_CONTEXT0_CNTL, tmp); 658 659 WREG32(0x575, 0); 660 WREG32(0x576, 0); 661 WREG32(0x577, 0); 662 663 /* empty context1-15 */ 664 /* FIXME start with 4G, once using 2 level pt switch to full 665 * vm size space 666 */ 667 /* set vm size, must be a multiple of 4 */ 668 WREG32(mmVM_CONTEXT1_PAGE_TABLE_START_ADDR, 0); 669 WREG32(mmVM_CONTEXT1_PAGE_TABLE_END_ADDR, adev->vm_manager.max_pfn - 1); 670 for (i = 1; i < AMDGPU_NUM_VMID; i++) { 671 if (i < 8) 672 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + i, 673 table_addr >> 12); 674 else 675 WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + i - 8, 676 table_addr >> 12); 677 } 678 679 /* enable context1-15 */ 680 WREG32(mmVM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR, 681 (u32)(adev->dummy_page_addr >> 12)); 682 WREG32(mmVM_CONTEXT1_CNTL2, 4); 683 tmp = RREG32(mmVM_CONTEXT1_CNTL); 684 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, ENABLE_CONTEXT, 1); 685 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_DEPTH, 1); 686 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_BLOCK_SIZE, 687 adev->vm_manager.block_size - 9); 688 WREG32(mmVM_CONTEXT1_CNTL, tmp); 689 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) 690 gmc_v7_0_set_fault_enable_default(adev, false); 691 else 692 gmc_v7_0_set_fault_enable_default(adev, true); 693 694 if (adev->asic_type == CHIP_KAVERI) { 695 tmp = RREG32(mmCHUB_CONTROL); 696 tmp &= ~BYPASS_VM; 697 WREG32(mmCHUB_CONTROL, tmp); 698 } 699 700 gmc_v7_0_flush_gpu_tlb(adev, 0, 0, 0); 701 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n", 702 (unsigned int)(adev->gmc.gart_size >> 20), 703 (unsigned long long)table_addr); 704 return 0; 705 } 706 707 static int gmc_v7_0_gart_init(struct amdgpu_device *adev) 708 { 709 int r; 710 711 if (adev->gart.bo) { 712 WARN(1, "R600 PCIE GART already initialized\n"); 713 return 0; 714 } 715 /* Initialize common gart structure */ 716 r = amdgpu_gart_init(adev); 717 if (r) 718 return r; 719 adev->gart.table_size = adev->gart.num_gpu_pages * 8; 720 adev->gart.gart_pte_flags = 0; 721 return amdgpu_gart_table_vram_alloc(adev); 722 } 723 724 /** 725 * gmc_v7_0_gart_disable - gart disable 726 * 727 * @adev: amdgpu_device pointer 728 * 729 * This disables all VM page table (CIK). 730 */ 731 static void gmc_v7_0_gart_disable(struct amdgpu_device *adev) 732 { 733 u32 tmp; 734 735 /* Disable all tables */ 736 WREG32(mmVM_CONTEXT0_CNTL, 0); 737 WREG32(mmVM_CONTEXT1_CNTL, 0); 738 /* Setup TLB control */ 739 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL); 740 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 0); 741 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 0); 742 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 0); 743 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp); 744 /* Setup L2 cache */ 745 tmp = RREG32(mmVM_L2_CNTL); 746 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 0); 747 WREG32(mmVM_L2_CNTL, tmp); 748 WREG32(mmVM_L2_CNTL2, 0); 749 } 750 751 /** 752 * gmc_v7_0_vm_decode_fault - print human readable fault info 753 * 754 * @adev: amdgpu_device pointer 755 * @status: VM_CONTEXT1_PROTECTION_FAULT_STATUS register value 756 * @addr: VM_CONTEXT1_PROTECTION_FAULT_ADDR register value 757 * @mc_client: VM_CONTEXT1_PROTECTION_FAULT_MCCLIENT register value 758 * @pasid: debug logging only - no functional use 759 * 760 * Print human readable fault information (CIK). 761 */ 762 static void gmc_v7_0_vm_decode_fault(struct amdgpu_device *adev, u32 status, 763 u32 addr, u32 mc_client, unsigned int pasid) 764 { 765 u32 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, VMID); 766 u32 protections = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 767 PROTECTIONS); 768 char block[5] = { mc_client >> 24, (mc_client >> 16) & 0xff, 769 (mc_client >> 8) & 0xff, mc_client & 0xff, 0 }; 770 u32 mc_id; 771 772 mc_id = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 773 MEMORY_CLIENT_ID); 774 775 dev_err(adev->dev, "VM fault (0x%02x, vmid %d, pasid %d) at page %u, %s from '%s' (0x%08x) (%d)\n", 776 protections, vmid, pasid, addr, 777 REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 778 MEMORY_CLIENT_RW) ? 779 "write" : "read", block, mc_client, mc_id); 780 } 781 782 783 static const u32 mc_cg_registers[] = { 784 mmMC_HUB_MISC_HUB_CG, 785 mmMC_HUB_MISC_SIP_CG, 786 mmMC_HUB_MISC_VM_CG, 787 mmMC_XPB_CLK_GAT, 788 mmATC_MISC_CG, 789 mmMC_CITF_MISC_WR_CG, 790 mmMC_CITF_MISC_RD_CG, 791 mmMC_CITF_MISC_VM_CG, 792 mmVM_L2_CG, 793 }; 794 795 static const u32 mc_cg_ls_en[] = { 796 MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK, 797 MC_HUB_MISC_SIP_CG__MEM_LS_ENABLE_MASK, 798 MC_HUB_MISC_VM_CG__MEM_LS_ENABLE_MASK, 799 MC_XPB_CLK_GAT__MEM_LS_ENABLE_MASK, 800 ATC_MISC_CG__MEM_LS_ENABLE_MASK, 801 MC_CITF_MISC_WR_CG__MEM_LS_ENABLE_MASK, 802 MC_CITF_MISC_RD_CG__MEM_LS_ENABLE_MASK, 803 MC_CITF_MISC_VM_CG__MEM_LS_ENABLE_MASK, 804 VM_L2_CG__MEM_LS_ENABLE_MASK, 805 }; 806 807 static const u32 mc_cg_en[] = { 808 MC_HUB_MISC_HUB_CG__ENABLE_MASK, 809 MC_HUB_MISC_SIP_CG__ENABLE_MASK, 810 MC_HUB_MISC_VM_CG__ENABLE_MASK, 811 MC_XPB_CLK_GAT__ENABLE_MASK, 812 ATC_MISC_CG__ENABLE_MASK, 813 MC_CITF_MISC_WR_CG__ENABLE_MASK, 814 MC_CITF_MISC_RD_CG__ENABLE_MASK, 815 MC_CITF_MISC_VM_CG__ENABLE_MASK, 816 VM_L2_CG__ENABLE_MASK, 817 }; 818 819 static void gmc_v7_0_enable_mc_ls(struct amdgpu_device *adev, 820 bool enable) 821 { 822 int i; 823 u32 orig, data; 824 825 for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) { 826 orig = data = RREG32(mc_cg_registers[i]); 827 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_LS)) 828 data |= mc_cg_ls_en[i]; 829 else 830 data &= ~mc_cg_ls_en[i]; 831 if (data != orig) 832 WREG32(mc_cg_registers[i], data); 833 } 834 } 835 836 static void gmc_v7_0_enable_mc_mgcg(struct amdgpu_device *adev, 837 bool enable) 838 { 839 int i; 840 u32 orig, data; 841 842 for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) { 843 orig = data = RREG32(mc_cg_registers[i]); 844 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_MGCG)) 845 data |= mc_cg_en[i]; 846 else 847 data &= ~mc_cg_en[i]; 848 if (data != orig) 849 WREG32(mc_cg_registers[i], data); 850 } 851 } 852 853 static void gmc_v7_0_enable_bif_mgls(struct amdgpu_device *adev, 854 bool enable) 855 { 856 u32 orig, data; 857 858 orig = data = RREG32_PCIE(ixPCIE_CNTL2); 859 860 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_BIF_LS)) { 861 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_LS_EN, 1); 862 data = REG_SET_FIELD(data, PCIE_CNTL2, MST_MEM_LS_EN, 1); 863 data = REG_SET_FIELD(data, PCIE_CNTL2, REPLAY_MEM_LS_EN, 1); 864 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_AGGRESSIVE_LS_EN, 1); 865 } else { 866 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_LS_EN, 0); 867 data = REG_SET_FIELD(data, PCIE_CNTL2, MST_MEM_LS_EN, 0); 868 data = REG_SET_FIELD(data, PCIE_CNTL2, REPLAY_MEM_LS_EN, 0); 869 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_AGGRESSIVE_LS_EN, 0); 870 } 871 872 if (orig != data) 873 WREG32_PCIE(ixPCIE_CNTL2, data); 874 } 875 876 static void gmc_v7_0_enable_hdp_mgcg(struct amdgpu_device *adev, 877 bool enable) 878 { 879 u32 orig, data; 880 881 orig = data = RREG32(mmHDP_HOST_PATH_CNTL); 882 883 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_HDP_MGCG)) 884 data = REG_SET_FIELD(data, HDP_HOST_PATH_CNTL, CLOCK_GATING_DIS, 0); 885 else 886 data = REG_SET_FIELD(data, HDP_HOST_PATH_CNTL, CLOCK_GATING_DIS, 1); 887 888 if (orig != data) 889 WREG32(mmHDP_HOST_PATH_CNTL, data); 890 } 891 892 static void gmc_v7_0_enable_hdp_ls(struct amdgpu_device *adev, 893 bool enable) 894 { 895 u32 orig, data; 896 897 orig = data = RREG32(mmHDP_MEM_POWER_LS); 898 899 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_HDP_LS)) 900 data = REG_SET_FIELD(data, HDP_MEM_POWER_LS, LS_ENABLE, 1); 901 else 902 data = REG_SET_FIELD(data, HDP_MEM_POWER_LS, LS_ENABLE, 0); 903 904 if (orig != data) 905 WREG32(mmHDP_MEM_POWER_LS, data); 906 } 907 908 static int gmc_v7_0_convert_vram_type(int mc_seq_vram_type) 909 { 910 switch (mc_seq_vram_type) { 911 case MC_SEQ_MISC0__MT__GDDR1: 912 return AMDGPU_VRAM_TYPE_GDDR1; 913 case MC_SEQ_MISC0__MT__DDR2: 914 return AMDGPU_VRAM_TYPE_DDR2; 915 case MC_SEQ_MISC0__MT__GDDR3: 916 return AMDGPU_VRAM_TYPE_GDDR3; 917 case MC_SEQ_MISC0__MT__GDDR4: 918 return AMDGPU_VRAM_TYPE_GDDR4; 919 case MC_SEQ_MISC0__MT__GDDR5: 920 return AMDGPU_VRAM_TYPE_GDDR5; 921 case MC_SEQ_MISC0__MT__HBM: 922 return AMDGPU_VRAM_TYPE_HBM; 923 case MC_SEQ_MISC0__MT__DDR3: 924 return AMDGPU_VRAM_TYPE_DDR3; 925 default: 926 return AMDGPU_VRAM_TYPE_UNKNOWN; 927 } 928 } 929 930 static int gmc_v7_0_early_init(void *handle) 931 { 932 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 933 934 gmc_v7_0_set_gmc_funcs(adev); 935 gmc_v7_0_set_irq_funcs(adev); 936 937 adev->gmc.shared_aperture_start = 0x2000000000000000ULL; 938 adev->gmc.shared_aperture_end = 939 adev->gmc.shared_aperture_start + (4ULL << 30) - 1; 940 adev->gmc.private_aperture_start = 941 adev->gmc.shared_aperture_end + 1; 942 adev->gmc.private_aperture_end = 943 adev->gmc.private_aperture_start + (4ULL << 30) - 1; 944 adev->gmc.noretry_flags = AMDGPU_VM_NORETRY_FLAGS_TF; 945 946 return 0; 947 } 948 949 static int gmc_v7_0_late_init(void *handle) 950 { 951 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 952 953 if (amdgpu_vm_fault_stop != AMDGPU_VM_FAULT_STOP_ALWAYS) 954 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0); 955 else 956 return 0; 957 } 958 959 static unsigned int gmc_v7_0_get_vbios_fb_size(struct amdgpu_device *adev) 960 { 961 u32 d1vga_control = RREG32(mmD1VGA_CONTROL); 962 unsigned int size; 963 964 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) { 965 size = AMDGPU_VBIOS_VGA_ALLOCATION; 966 } else { 967 u32 viewport = RREG32(mmVIEWPORT_SIZE); 968 969 size = (REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_HEIGHT) * 970 REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_WIDTH) * 971 4); 972 } 973 974 return size; 975 } 976 977 static int gmc_v7_0_sw_init(void *handle) 978 { 979 int r; 980 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 981 982 set_bit(AMDGPU_GFXHUB(0), adev->vmhubs_mask); 983 984 if (adev->flags & AMD_IS_APU) { 985 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_UNKNOWN; 986 } else { 987 u32 tmp = RREG32(mmMC_SEQ_MISC0); 988 989 tmp &= MC_SEQ_MISC0__MT__MASK; 990 adev->gmc.vram_type = gmc_v7_0_convert_vram_type(tmp); 991 } 992 993 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_PAGE_INV_FAULT, &adev->gmc.vm_fault); 994 if (r) 995 return r; 996 997 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_MEM_PROT_FAULT, &adev->gmc.vm_fault); 998 if (r) 999 return r; 1000 1001 /* Adjust VM size here. 1002 * Currently set to 4GB ((1 << 20) 4k pages). 1003 * Max GPUVM size for cayman and SI is 40 bits. 1004 */ 1005 amdgpu_vm_adjust_size(adev, 64, 9, 1, 40); 1006 1007 /* Set the internal MC address mask 1008 * This is the max address of the GPU's 1009 * internal address space. 1010 */ 1011 adev->gmc.mc_mask = 0xffffffffffULL; /* 40 bit MC */ 1012 1013 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(40)); 1014 if (r) { 1015 pr_warn("No suitable DMA available\n"); 1016 return r; 1017 } 1018 adev->need_swiotlb = drm_need_swiotlb(40); 1019 1020 r = gmc_v7_0_init_microcode(adev); 1021 if (r) { 1022 DRM_ERROR("Failed to load mc firmware!\n"); 1023 return r; 1024 } 1025 1026 r = gmc_v7_0_mc_init(adev); 1027 if (r) 1028 return r; 1029 1030 amdgpu_gmc_get_vbios_allocations(adev); 1031 1032 /* Memory manager */ 1033 r = amdgpu_bo_init(adev); 1034 if (r) 1035 return r; 1036 1037 r = gmc_v7_0_gart_init(adev); 1038 if (r) 1039 return r; 1040 1041 /* 1042 * number of VMs 1043 * VMID 0 is reserved for System 1044 * amdgpu graphics/compute will use VMIDs 1-7 1045 * amdkfd will use VMIDs 8-15 1046 */ 1047 adev->vm_manager.first_kfd_vmid = 8; 1048 amdgpu_vm_manager_init(adev); 1049 1050 /* base offset of vram pages */ 1051 if (adev->flags & AMD_IS_APU) { 1052 u64 tmp = RREG32(mmMC_VM_FB_OFFSET); 1053 1054 tmp <<= 22; 1055 adev->vm_manager.vram_base_offset = tmp; 1056 } else { 1057 adev->vm_manager.vram_base_offset = 0; 1058 } 1059 1060 adev->gmc.vm_fault_info = kmalloc(sizeof(struct kfd_vm_fault_info), 1061 GFP_KERNEL); 1062 if (!adev->gmc.vm_fault_info) 1063 return -ENOMEM; 1064 atomic_set(&adev->gmc.vm_fault_info_updated, 0); 1065 1066 return 0; 1067 } 1068 1069 static int gmc_v7_0_sw_fini(void *handle) 1070 { 1071 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1072 1073 amdgpu_gem_force_release(adev); 1074 amdgpu_vm_manager_fini(adev); 1075 kfree(adev->gmc.vm_fault_info); 1076 amdgpu_gart_table_vram_free(adev); 1077 amdgpu_bo_fini(adev); 1078 amdgpu_ucode_release(&adev->gmc.fw); 1079 1080 return 0; 1081 } 1082 1083 static int gmc_v7_0_hw_init(void *handle) 1084 { 1085 int r; 1086 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1087 1088 gmc_v7_0_init_golden_registers(adev); 1089 1090 gmc_v7_0_mc_program(adev); 1091 1092 if (!(adev->flags & AMD_IS_APU)) { 1093 r = gmc_v7_0_mc_load_microcode(adev); 1094 if (r) { 1095 DRM_ERROR("Failed to load MC firmware!\n"); 1096 return r; 1097 } 1098 } 1099 1100 r = gmc_v7_0_gart_enable(adev); 1101 if (r) 1102 return r; 1103 1104 if (amdgpu_emu_mode == 1) 1105 return amdgpu_gmc_vram_checking(adev); 1106 else 1107 return r; 1108 } 1109 1110 static int gmc_v7_0_hw_fini(void *handle) 1111 { 1112 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1113 1114 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0); 1115 gmc_v7_0_gart_disable(adev); 1116 1117 return 0; 1118 } 1119 1120 static int gmc_v7_0_suspend(void *handle) 1121 { 1122 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1123 1124 gmc_v7_0_hw_fini(adev); 1125 1126 return 0; 1127 } 1128 1129 static int gmc_v7_0_resume(void *handle) 1130 { 1131 int r; 1132 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1133 1134 r = gmc_v7_0_hw_init(adev); 1135 if (r) 1136 return r; 1137 1138 amdgpu_vmid_reset_all(adev); 1139 1140 return 0; 1141 } 1142 1143 static bool gmc_v7_0_is_idle(void *handle) 1144 { 1145 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1146 u32 tmp = RREG32(mmSRBM_STATUS); 1147 1148 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1149 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK | SRBM_STATUS__VMC_BUSY_MASK)) 1150 return false; 1151 1152 return true; 1153 } 1154 1155 static int gmc_v7_0_wait_for_idle(void *handle) 1156 { 1157 unsigned int i; 1158 u32 tmp; 1159 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1160 1161 for (i = 0; i < adev->usec_timeout; i++) { 1162 /* read MC_STATUS */ 1163 tmp = RREG32(mmSRBM_STATUS) & (SRBM_STATUS__MCB_BUSY_MASK | 1164 SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1165 SRBM_STATUS__MCC_BUSY_MASK | 1166 SRBM_STATUS__MCD_BUSY_MASK | 1167 SRBM_STATUS__VMC_BUSY_MASK); 1168 if (!tmp) 1169 return 0; 1170 udelay(1); 1171 } 1172 return -ETIMEDOUT; 1173 1174 } 1175 1176 static int gmc_v7_0_soft_reset(void *handle) 1177 { 1178 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1179 u32 srbm_soft_reset = 0; 1180 u32 tmp = RREG32(mmSRBM_STATUS); 1181 1182 if (tmp & SRBM_STATUS__VMC_BUSY_MASK) 1183 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset, 1184 SRBM_SOFT_RESET, SOFT_RESET_VMC, 1); 1185 1186 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1187 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK)) { 1188 if (!(adev->flags & AMD_IS_APU)) 1189 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset, 1190 SRBM_SOFT_RESET, SOFT_RESET_MC, 1); 1191 } 1192 1193 if (srbm_soft_reset) { 1194 gmc_v7_0_mc_stop(adev); 1195 if (gmc_v7_0_wait_for_idle((void *)adev)) 1196 dev_warn(adev->dev, "Wait for GMC idle timed out !\n"); 1197 1198 tmp = RREG32(mmSRBM_SOFT_RESET); 1199 tmp |= srbm_soft_reset; 1200 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp); 1201 WREG32(mmSRBM_SOFT_RESET, tmp); 1202 tmp = RREG32(mmSRBM_SOFT_RESET); 1203 1204 udelay(50); 1205 1206 tmp &= ~srbm_soft_reset; 1207 WREG32(mmSRBM_SOFT_RESET, tmp); 1208 tmp = RREG32(mmSRBM_SOFT_RESET); 1209 1210 /* Wait a little for things to settle down */ 1211 udelay(50); 1212 1213 gmc_v7_0_mc_resume(adev); 1214 udelay(50); 1215 } 1216 1217 return 0; 1218 } 1219 1220 static int gmc_v7_0_vm_fault_interrupt_state(struct amdgpu_device *adev, 1221 struct amdgpu_irq_src *src, 1222 unsigned int type, 1223 enum amdgpu_interrupt_state state) 1224 { 1225 u32 tmp; 1226 u32 bits = (VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1227 VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1228 VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1229 VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1230 VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1231 VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK); 1232 1233 switch (state) { 1234 case AMDGPU_IRQ_STATE_DISABLE: 1235 /* system context */ 1236 tmp = RREG32(mmVM_CONTEXT0_CNTL); 1237 tmp &= ~bits; 1238 WREG32(mmVM_CONTEXT0_CNTL, tmp); 1239 /* VMs */ 1240 tmp = RREG32(mmVM_CONTEXT1_CNTL); 1241 tmp &= ~bits; 1242 WREG32(mmVM_CONTEXT1_CNTL, tmp); 1243 break; 1244 case AMDGPU_IRQ_STATE_ENABLE: 1245 /* system context */ 1246 tmp = RREG32(mmVM_CONTEXT0_CNTL); 1247 tmp |= bits; 1248 WREG32(mmVM_CONTEXT0_CNTL, tmp); 1249 /* VMs */ 1250 tmp = RREG32(mmVM_CONTEXT1_CNTL); 1251 tmp |= bits; 1252 WREG32(mmVM_CONTEXT1_CNTL, tmp); 1253 break; 1254 default: 1255 break; 1256 } 1257 1258 return 0; 1259 } 1260 1261 static int gmc_v7_0_process_interrupt(struct amdgpu_device *adev, 1262 struct amdgpu_irq_src *source, 1263 struct amdgpu_iv_entry *entry) 1264 { 1265 u32 addr, status, mc_client, vmid; 1266 1267 addr = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_ADDR); 1268 status = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_STATUS); 1269 mc_client = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_MCCLIENT); 1270 /* reset addr and status */ 1271 WREG32_P(mmVM_CONTEXT1_CNTL2, 1, ~1); 1272 1273 if (!addr && !status) 1274 return 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 }; 1358 1359 static const struct amdgpu_gmc_funcs gmc_v7_0_gmc_funcs = { 1360 .flush_gpu_tlb = gmc_v7_0_flush_gpu_tlb, 1361 .flush_gpu_tlb_pasid = gmc_v7_0_flush_gpu_tlb_pasid, 1362 .emit_flush_gpu_tlb = gmc_v7_0_emit_flush_gpu_tlb, 1363 .emit_pasid_mapping = gmc_v7_0_emit_pasid_mapping, 1364 .set_prt = gmc_v7_0_set_prt, 1365 .get_vm_pde = gmc_v7_0_get_vm_pde, 1366 .get_vm_pte = gmc_v7_0_get_vm_pte, 1367 .get_vbios_fb_size = gmc_v7_0_get_vbios_fb_size, 1368 }; 1369 1370 static const struct amdgpu_irq_src_funcs gmc_v7_0_irq_funcs = { 1371 .set = gmc_v7_0_vm_fault_interrupt_state, 1372 .process = gmc_v7_0_process_interrupt, 1373 }; 1374 1375 static void gmc_v7_0_set_gmc_funcs(struct amdgpu_device *adev) 1376 { 1377 adev->gmc.gmc_funcs = &gmc_v7_0_gmc_funcs; 1378 } 1379 1380 static void gmc_v7_0_set_irq_funcs(struct amdgpu_device *adev) 1381 { 1382 adev->gmc.vm_fault.num_types = 1; 1383 adev->gmc.vm_fault.funcs = &gmc_v7_0_irq_funcs; 1384 } 1385 1386 const struct amdgpu_ip_block_version gmc_v7_0_ip_block = { 1387 .type = AMD_IP_BLOCK_TYPE_GMC, 1388 .major = 7, 1389 .minor = 0, 1390 .rev = 0, 1391 .funcs = &gmc_v7_0_ip_funcs, 1392 }; 1393 1394 const struct amdgpu_ip_block_version gmc_v7_4_ip_block = { 1395 .type = AMD_IP_BLOCK_TYPE_GMC, 1396 .major = 7, 1397 .minor = 4, 1398 .rev = 0, 1399 .funcs = &gmc_v7_0_ip_funcs, 1400 }; 1401