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 { 63 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff, 64 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff, 65 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff, 66 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff 67 }; 68 69 static const u32 iceland_mgcg_cgcg_init[] = 70 { 71 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104 72 }; 73 74 static void gmc_v7_0_init_golden_registers(struct amdgpu_device *adev) 75 { 76 switch (adev->asic_type) { 77 case CHIP_TOPAZ: 78 amdgpu_device_program_register_sequence(adev, 79 iceland_mgcg_cgcg_init, 80 ARRAY_SIZE(iceland_mgcg_cgcg_init)); 81 amdgpu_device_program_register_sequence(adev, 82 golden_settings_iceland_a11, 83 ARRAY_SIZE(golden_settings_iceland_a11)); 84 break; 85 default: 86 break; 87 } 88 } 89 90 static void gmc_v7_0_mc_stop(struct amdgpu_device *adev) 91 { 92 u32 blackout; 93 94 gmc_v7_0_wait_for_idle((void *)adev); 95 96 blackout = RREG32(mmMC_SHARED_BLACKOUT_CNTL); 97 if (REG_GET_FIELD(blackout, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE) != 1) { 98 /* Block CPU access */ 99 WREG32(mmBIF_FB_EN, 0); 100 /* blackout the MC */ 101 blackout = REG_SET_FIELD(blackout, 102 MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 0); 103 WREG32(mmMC_SHARED_BLACKOUT_CNTL, blackout | 1); 104 } 105 /* wait for the MC to settle */ 106 udelay(100); 107 } 108 109 static void gmc_v7_0_mc_resume(struct amdgpu_device *adev) 110 { 111 u32 tmp; 112 113 /* unblackout the MC */ 114 tmp = RREG32(mmMC_SHARED_BLACKOUT_CNTL); 115 tmp = REG_SET_FIELD(tmp, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 0); 116 WREG32(mmMC_SHARED_BLACKOUT_CNTL, tmp); 117 /* allow CPU access */ 118 tmp = REG_SET_FIELD(0, BIF_FB_EN, FB_READ_EN, 1); 119 tmp = REG_SET_FIELD(tmp, BIF_FB_EN, FB_WRITE_EN, 1); 120 WREG32(mmBIF_FB_EN, tmp); 121 } 122 123 /** 124 * gmc_v7_0_init_microcode - load ucode images from disk 125 * 126 * @adev: amdgpu_device pointer 127 * 128 * Use the firmware interface to load the ucode images into 129 * the driver (not loaded into hw). 130 * Returns 0 on success, error on failure. 131 */ 132 static int gmc_v7_0_init_microcode(struct amdgpu_device *adev) 133 { 134 const char *chip_name; 135 char fw_name[30]; 136 int err; 137 138 DRM_DEBUG("\n"); 139 140 switch (adev->asic_type) { 141 case CHIP_BONAIRE: 142 chip_name = "bonaire"; 143 break; 144 case CHIP_HAWAII: 145 chip_name = "hawaii"; 146 break; 147 case CHIP_TOPAZ: 148 chip_name = "topaz"; 149 break; 150 case CHIP_KAVERI: 151 case CHIP_KABINI: 152 case CHIP_MULLINS: 153 return 0; 154 default: BUG(); 155 } 156 157 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mc.bin", chip_name); 158 159 err = request_firmware(&adev->gmc.fw, fw_name, adev->dev); 160 if (err) 161 goto out; 162 err = amdgpu_ucode_validate(adev->gmc.fw); 163 164 out: 165 if (err) { 166 pr_err("cik_mc: Failed to load firmware \"%s\"\n", fw_name); 167 release_firmware(adev->gmc.fw); 168 adev->gmc.fw = NULL; 169 } 170 return err; 171 } 172 173 /** 174 * gmc_v7_0_mc_load_microcode - load MC ucode into the hw 175 * 176 * @adev: amdgpu_device pointer 177 * 178 * Load the GDDR MC ucode into the hw (CIK). 179 * Returns 0 on success, error on failure. 180 */ 181 static int gmc_v7_0_mc_load_microcode(struct amdgpu_device *adev) 182 { 183 const struct mc_firmware_header_v1_0 *hdr; 184 const __le32 *fw_data = NULL; 185 const __le32 *io_mc_regs = NULL; 186 u32 running; 187 int i, ucode_size, regs_size; 188 189 if (!adev->gmc.fw) 190 return -EINVAL; 191 192 hdr = (const struct mc_firmware_header_v1_0 *)adev->gmc.fw->data; 193 amdgpu_ucode_print_mc_hdr(&hdr->header); 194 195 adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version); 196 regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2); 197 io_mc_regs = (const __le32 *) 198 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes)); 199 ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4; 200 fw_data = (const __le32 *) 201 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes)); 202 203 running = REG_GET_FIELD(RREG32(mmMC_SEQ_SUP_CNTL), MC_SEQ_SUP_CNTL, RUN); 204 205 if (running == 0) { 206 /* reset the engine and set to writable */ 207 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008); 208 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000010); 209 210 /* load mc io regs */ 211 for (i = 0; i < regs_size; i++) { 212 WREG32(mmMC_SEQ_IO_DEBUG_INDEX, le32_to_cpup(io_mc_regs++)); 213 WREG32(mmMC_SEQ_IO_DEBUG_DATA, le32_to_cpup(io_mc_regs++)); 214 } 215 /* load the MC ucode */ 216 for (i = 0; i < ucode_size; i++) 217 WREG32(mmMC_SEQ_SUP_PGM, le32_to_cpup(fw_data++)); 218 219 /* put the engine back into the active state */ 220 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008); 221 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000004); 222 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000001); 223 224 /* wait for training to complete */ 225 for (i = 0; i < adev->usec_timeout; i++) { 226 if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL), 227 MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D0)) 228 break; 229 udelay(1); 230 } 231 for (i = 0; i < adev->usec_timeout; i++) { 232 if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL), 233 MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D1)) 234 break; 235 udelay(1); 236 } 237 } 238 239 return 0; 240 } 241 242 static void gmc_v7_0_vram_gtt_location(struct amdgpu_device *adev, 243 struct amdgpu_gmc *mc) 244 { 245 u64 base = RREG32(mmMC_VM_FB_LOCATION) & 0xFFFF; 246 base <<= 24; 247 248 amdgpu_gmc_vram_location(adev, mc, base); 249 amdgpu_gmc_gart_location(adev, mc); 250 } 251 252 /** 253 * gmc_v7_0_mc_program - program the GPU memory controller 254 * 255 * @adev: amdgpu_device pointer 256 * 257 * Set the location of vram, gart, and AGP in the GPU's 258 * physical address space (CIK). 259 */ 260 static void gmc_v7_0_mc_program(struct amdgpu_device *adev) 261 { 262 u32 tmp; 263 int i, j; 264 265 /* Initialize HDP */ 266 for (i = 0, j = 0; i < 32; i++, j += 0x6) { 267 WREG32((0xb05 + j), 0x00000000); 268 WREG32((0xb06 + j), 0x00000000); 269 WREG32((0xb07 + j), 0x00000000); 270 WREG32((0xb08 + j), 0x00000000); 271 WREG32((0xb09 + j), 0x00000000); 272 } 273 WREG32(mmHDP_REG_COHERENCY_FLUSH_CNTL, 0); 274 275 if (gmc_v7_0_wait_for_idle((void *)adev)) { 276 dev_warn(adev->dev, "Wait for MC idle timedout !\n"); 277 } 278 if (adev->mode_info.num_crtc) { 279 /* Lockout access through VGA aperture*/ 280 tmp = RREG32(mmVGA_HDP_CONTROL); 281 tmp = REG_SET_FIELD(tmp, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1); 282 WREG32(mmVGA_HDP_CONTROL, tmp); 283 284 /* disable VGA render */ 285 tmp = RREG32(mmVGA_RENDER_CONTROL); 286 tmp = REG_SET_FIELD(tmp, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0); 287 WREG32(mmVGA_RENDER_CONTROL, tmp); 288 } 289 /* Update configuration */ 290 WREG32(mmMC_VM_SYSTEM_APERTURE_LOW_ADDR, 291 adev->gmc.vram_start >> 12); 292 WREG32(mmMC_VM_SYSTEM_APERTURE_HIGH_ADDR, 293 adev->gmc.vram_end >> 12); 294 WREG32(mmMC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 295 adev->vram_scratch.gpu_addr >> 12); 296 WREG32(mmMC_VM_AGP_BASE, 0); 297 WREG32(mmMC_VM_AGP_TOP, 0x0FFFFFFF); 298 WREG32(mmMC_VM_AGP_BOT, 0x0FFFFFFF); 299 if (gmc_v7_0_wait_for_idle((void *)adev)) { 300 dev_warn(adev->dev, "Wait for MC idle timedout !\n"); 301 } 302 303 WREG32(mmBIF_FB_EN, BIF_FB_EN__FB_READ_EN_MASK | BIF_FB_EN__FB_WRITE_EN_MASK); 304 305 tmp = RREG32(mmHDP_MISC_CNTL); 306 tmp = REG_SET_FIELD(tmp, HDP_MISC_CNTL, FLUSH_INVALIDATE_CACHE, 0); 307 WREG32(mmHDP_MISC_CNTL, tmp); 308 309 tmp = RREG32(mmHDP_HOST_PATH_CNTL); 310 WREG32(mmHDP_HOST_PATH_CNTL, tmp); 311 } 312 313 /** 314 * gmc_v7_0_mc_init - initialize the memory controller driver params 315 * 316 * @adev: amdgpu_device pointer 317 * 318 * Look up the amount of vram, vram width, and decide how to place 319 * vram and gart within the GPU's physical address space (CIK). 320 * Returns 0 for success. 321 */ 322 static int gmc_v7_0_mc_init(struct amdgpu_device *adev) 323 { 324 int r; 325 326 adev->gmc.vram_width = amdgpu_atombios_get_vram_width(adev); 327 if (!adev->gmc.vram_width) { 328 u32 tmp; 329 int chansize, numchan; 330 331 /* Get VRAM informations */ 332 tmp = RREG32(mmMC_ARB_RAMCFG); 333 if (REG_GET_FIELD(tmp, MC_ARB_RAMCFG, CHANSIZE)) { 334 chansize = 64; 335 } else { 336 chansize = 32; 337 } 338 tmp = RREG32(mmMC_SHARED_CHMAP); 339 switch (REG_GET_FIELD(tmp, MC_SHARED_CHMAP, NOOFCHAN)) { 340 case 0: 341 default: 342 numchan = 1; 343 break; 344 case 1: 345 numchan = 2; 346 break; 347 case 2: 348 numchan = 4; 349 break; 350 case 3: 351 numchan = 8; 352 break; 353 case 4: 354 numchan = 3; 355 break; 356 case 5: 357 numchan = 6; 358 break; 359 case 6: 360 numchan = 10; 361 break; 362 case 7: 363 numchan = 12; 364 break; 365 case 8: 366 numchan = 16; 367 break; 368 } 369 adev->gmc.vram_width = numchan * chansize; 370 } 371 /* size in MB on si */ 372 adev->gmc.mc_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL; 373 adev->gmc.real_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL; 374 375 if (!(adev->flags & AMD_IS_APU)) { 376 r = amdgpu_device_resize_fb_bar(adev); 377 if (r) 378 return r; 379 } 380 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0); 381 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0); 382 383 #ifdef CONFIG_X86_64 384 if (adev->flags & AMD_IS_APU) { 385 adev->gmc.aper_base = ((u64)RREG32(mmMC_VM_FB_OFFSET)) << 22; 386 adev->gmc.aper_size = adev->gmc.real_vram_size; 387 } 388 #endif 389 390 /* In case the PCI BAR is larger than the actual amount of vram */ 391 adev->gmc.visible_vram_size = adev->gmc.aper_size; 392 if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size) 393 adev->gmc.visible_vram_size = adev->gmc.real_vram_size; 394 395 /* set the gart size */ 396 if (amdgpu_gart_size == -1) { 397 switch (adev->asic_type) { 398 case CHIP_TOPAZ: /* no MM engines */ 399 default: 400 adev->gmc.gart_size = 256ULL << 20; 401 break; 402 #ifdef CONFIG_DRM_AMDGPU_CIK 403 case CHIP_BONAIRE: /* UVD, VCE do not support GPUVM */ 404 case CHIP_HAWAII: /* UVD, VCE do not support GPUVM */ 405 case CHIP_KAVERI: /* UVD, VCE do not support GPUVM */ 406 case CHIP_KABINI: /* UVD, VCE do not support GPUVM */ 407 case CHIP_MULLINS: /* UVD, VCE do not support GPUVM */ 408 adev->gmc.gart_size = 1024ULL << 20; 409 break; 410 #endif 411 } 412 } else { 413 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20; 414 } 415 416 gmc_v7_0_vram_gtt_location(adev, &adev->gmc); 417 418 return 0; 419 } 420 421 /** 422 * gmc_v7_0_flush_gpu_tlb_pasid - tlb flush via pasid 423 * 424 * @adev: amdgpu_device pointer 425 * @pasid: pasid to be flush 426 * 427 * Flush the TLB for the requested pasid. 428 */ 429 static int gmc_v7_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, 430 uint16_t pasid, uint32_t flush_type, 431 bool all_hub) 432 { 433 int vmid; 434 unsigned int tmp; 435 436 if (adev->in_gpu_reset) 437 return -EIO; 438 439 for (vmid = 1; vmid < 16; vmid++) { 440 441 tmp = RREG32(mmATC_VMID0_PASID_MAPPING + vmid); 442 if ((tmp & ATC_VMID0_PASID_MAPPING__VALID_MASK) && 443 (tmp & ATC_VMID0_PASID_MAPPING__PASID_MASK) == pasid) { 444 WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid); 445 RREG32(mmVM_INVALIDATE_RESPONSE); 446 break; 447 } 448 } 449 450 return 0; 451 } 452 453 /* 454 * GART 455 * VMID 0 is the physical GPU addresses as used by the kernel. 456 * VMIDs 1-15 are used for userspace clients and are handled 457 * by the amdgpu vm/hsa code. 458 */ 459 460 /** 461 * gmc_v7_0_flush_gpu_tlb - gart tlb flush callback 462 * 463 * @adev: amdgpu_device pointer 464 * @vmid: vm instance to flush 465 * 466 * Flush the TLB for the requested page table (CIK). 467 */ 468 static void gmc_v7_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, 469 uint32_t vmhub, uint32_t flush_type) 470 { 471 /* bits 0-15 are the VM contexts0-15 */ 472 WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid); 473 } 474 475 static uint64_t gmc_v7_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring, 476 unsigned vmid, uint64_t pd_addr) 477 { 478 uint32_t reg; 479 480 if (vmid < 8) 481 reg = mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vmid; 482 else 483 reg = mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8; 484 amdgpu_ring_emit_wreg(ring, reg, pd_addr >> 12); 485 486 /* bits 0-15 are the VM contexts0-15 */ 487 amdgpu_ring_emit_wreg(ring, mmVM_INVALIDATE_REQUEST, 1 << vmid); 488 489 return pd_addr; 490 } 491 492 static void gmc_v7_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid, 493 unsigned pasid) 494 { 495 amdgpu_ring_emit_wreg(ring, mmIH_VMID_0_LUT + vmid, pasid); 496 } 497 498 static void gmc_v7_0_get_vm_pde(struct amdgpu_device *adev, int level, 499 uint64_t *addr, uint64_t *flags) 500 { 501 BUG_ON(*addr & 0xFFFFFF0000000FFFULL); 502 } 503 504 static void gmc_v7_0_get_vm_pte(struct amdgpu_device *adev, 505 struct amdgpu_bo_va_mapping *mapping, 506 uint64_t *flags) 507 { 508 *flags &= ~AMDGPU_PTE_EXECUTABLE; 509 *flags &= ~AMDGPU_PTE_PRT; 510 } 511 512 /** 513 * gmc_v8_0_set_fault_enable_default - update VM fault handling 514 * 515 * @adev: amdgpu_device pointer 516 * @value: true redirects VM faults to the default page 517 */ 518 static void gmc_v7_0_set_fault_enable_default(struct amdgpu_device *adev, 519 bool value) 520 { 521 u32 tmp; 522 523 tmp = RREG32(mmVM_CONTEXT1_CNTL); 524 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 525 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 526 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 527 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 528 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 529 PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, value); 530 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 531 VALID_PROTECTION_FAULT_ENABLE_DEFAULT, value); 532 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 533 READ_PROTECTION_FAULT_ENABLE_DEFAULT, value); 534 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 535 WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 536 WREG32(mmVM_CONTEXT1_CNTL, tmp); 537 } 538 539 /** 540 * gmc_v7_0_set_prt - set PRT VM fault 541 * 542 * @adev: amdgpu_device pointer 543 * @enable: enable/disable VM fault handling for PRT 544 */ 545 static void gmc_v7_0_set_prt(struct amdgpu_device *adev, bool enable) 546 { 547 uint32_t tmp; 548 549 if (enable && !adev->gmc.prt_warning) { 550 dev_warn(adev->dev, "Disabling VM faults because of PRT request!\n"); 551 adev->gmc.prt_warning = true; 552 } 553 554 tmp = RREG32(mmVM_PRT_CNTL); 555 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 556 CB_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable); 557 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 558 CB_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable); 559 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 560 TC_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable); 561 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 562 TC_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable); 563 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 564 L2_CACHE_STORE_INVALID_ENTRIES, enable); 565 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 566 L1_TLB_STORE_INVALID_ENTRIES, enable); 567 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 568 MASK_PDE0_FAULT, enable); 569 WREG32(mmVM_PRT_CNTL, tmp); 570 571 if (enable) { 572 uint32_t low = AMDGPU_VA_RESERVED_SIZE >> AMDGPU_GPU_PAGE_SHIFT; 573 uint32_t high = adev->vm_manager.max_pfn - 574 (AMDGPU_VA_RESERVED_SIZE >> AMDGPU_GPU_PAGE_SHIFT); 575 576 WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, low); 577 WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, low); 578 WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, low); 579 WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, low); 580 WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, high); 581 WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, high); 582 WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, high); 583 WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, high); 584 } else { 585 WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, 0xfffffff); 586 WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, 0xfffffff); 587 WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, 0xfffffff); 588 WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, 0xfffffff); 589 WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, 0x0); 590 WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, 0x0); 591 WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, 0x0); 592 WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, 0x0); 593 } 594 } 595 596 /** 597 * gmc_v7_0_gart_enable - gart enable 598 * 599 * @adev: amdgpu_device pointer 600 * 601 * This sets up the TLBs, programs the page tables for VMID0, 602 * sets up the hw for VMIDs 1-15 which are allocated on 603 * demand, and sets up the global locations for the LDS, GDS, 604 * and GPUVM for FSA64 clients (CIK). 605 * Returns 0 for success, errors for failure. 606 */ 607 static int gmc_v7_0_gart_enable(struct amdgpu_device *adev) 608 { 609 uint64_t table_addr; 610 int r, i; 611 u32 tmp, field; 612 613 if (adev->gart.bo == NULL) { 614 dev_err(adev->dev, "No VRAM object for PCIE GART.\n"); 615 return -EINVAL; 616 } 617 r = amdgpu_gart_table_vram_pin(adev); 618 if (r) 619 return r; 620 621 table_addr = amdgpu_bo_gpu_offset(adev->gart.bo); 622 623 /* Setup TLB control */ 624 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL); 625 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 1); 626 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 1); 627 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE, 3); 628 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 1); 629 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_APERTURE_UNMAPPED_ACCESS, 0); 630 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp); 631 /* Setup L2 cache */ 632 tmp = RREG32(mmVM_L2_CNTL); 633 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 1); 634 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING, 1); 635 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE, 1); 636 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE, 1); 637 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, EFFECTIVE_L2_QUEUE_SIZE, 7); 638 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, CONTEXT1_IDENTITY_ACCESS_MODE, 1); 639 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_DEFAULT_PAGE_OUT_TO_SYSTEM_MEMORY, 1); 640 WREG32(mmVM_L2_CNTL, tmp); 641 tmp = REG_SET_FIELD(0, VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS, 1); 642 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_L2_CACHE, 1); 643 WREG32(mmVM_L2_CNTL2, tmp); 644 645 field = adev->vm_manager.fragment_size; 646 tmp = RREG32(mmVM_L2_CNTL3); 647 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY, 1); 648 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, BANK_SELECT, field); 649 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_FRAGMENT_SIZE, field); 650 WREG32(mmVM_L2_CNTL3, tmp); 651 /* setup context0 */ 652 WREG32(mmVM_CONTEXT0_PAGE_TABLE_START_ADDR, adev->gmc.gart_start >> 12); 653 WREG32(mmVM_CONTEXT0_PAGE_TABLE_END_ADDR, adev->gmc.gart_end >> 12); 654 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR, table_addr >> 12); 655 WREG32(mmVM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR, 656 (u32)(adev->dummy_page_addr >> 12)); 657 WREG32(mmVM_CONTEXT0_CNTL2, 0); 658 tmp = RREG32(mmVM_CONTEXT0_CNTL); 659 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, ENABLE_CONTEXT, 1); 660 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, PAGE_TABLE_DEPTH, 0); 661 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 662 WREG32(mmVM_CONTEXT0_CNTL, tmp); 663 664 WREG32(0x575, 0); 665 WREG32(0x576, 0); 666 WREG32(0x577, 0); 667 668 /* empty context1-15 */ 669 /* FIXME start with 4G, once using 2 level pt switch to full 670 * vm size space 671 */ 672 /* set vm size, must be a multiple of 4 */ 673 WREG32(mmVM_CONTEXT1_PAGE_TABLE_START_ADDR, 0); 674 WREG32(mmVM_CONTEXT1_PAGE_TABLE_END_ADDR, adev->vm_manager.max_pfn - 1); 675 for (i = 1; i < 16; i++) { 676 if (i < 8) 677 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + i, 678 table_addr >> 12); 679 else 680 WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + i - 8, 681 table_addr >> 12); 682 } 683 684 /* enable context1-15 */ 685 WREG32(mmVM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR, 686 (u32)(adev->dummy_page_addr >> 12)); 687 WREG32(mmVM_CONTEXT1_CNTL2, 4); 688 tmp = RREG32(mmVM_CONTEXT1_CNTL); 689 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, ENABLE_CONTEXT, 1); 690 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_DEPTH, 1); 691 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_BLOCK_SIZE, 692 adev->vm_manager.block_size - 9); 693 WREG32(mmVM_CONTEXT1_CNTL, tmp); 694 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) 695 gmc_v7_0_set_fault_enable_default(adev, false); 696 else 697 gmc_v7_0_set_fault_enable_default(adev, true); 698 699 if (adev->asic_type == CHIP_KAVERI) { 700 tmp = RREG32(mmCHUB_CONTROL); 701 tmp &= ~BYPASS_VM; 702 WREG32(mmCHUB_CONTROL, tmp); 703 } 704 705 gmc_v7_0_flush_gpu_tlb(adev, 0, 0, 0); 706 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n", 707 (unsigned)(adev->gmc.gart_size >> 20), 708 (unsigned long long)table_addr); 709 adev->gart.ready = true; 710 return 0; 711 } 712 713 static int gmc_v7_0_gart_init(struct amdgpu_device *adev) 714 { 715 int r; 716 717 if (adev->gart.bo) { 718 WARN(1, "R600 PCIE GART already initialized\n"); 719 return 0; 720 } 721 /* Initialize common gart structure */ 722 r = amdgpu_gart_init(adev); 723 if (r) 724 return r; 725 adev->gart.table_size = adev->gart.num_gpu_pages * 8; 726 adev->gart.gart_pte_flags = 0; 727 return amdgpu_gart_table_vram_alloc(adev); 728 } 729 730 /** 731 * gmc_v7_0_gart_disable - gart disable 732 * 733 * @adev: amdgpu_device pointer 734 * 735 * This disables all VM page table (CIK). 736 */ 737 static void gmc_v7_0_gart_disable(struct amdgpu_device *adev) 738 { 739 u32 tmp; 740 741 /* Disable all tables */ 742 WREG32(mmVM_CONTEXT0_CNTL, 0); 743 WREG32(mmVM_CONTEXT1_CNTL, 0); 744 /* Setup TLB control */ 745 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL); 746 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 0); 747 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 0); 748 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 0); 749 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp); 750 /* Setup L2 cache */ 751 tmp = RREG32(mmVM_L2_CNTL); 752 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 0); 753 WREG32(mmVM_L2_CNTL, tmp); 754 WREG32(mmVM_L2_CNTL2, 0); 755 amdgpu_gart_table_vram_unpin(adev); 756 } 757 758 /** 759 * gmc_v7_0_vm_decode_fault - print human readable fault info 760 * 761 * @adev: amdgpu_device pointer 762 * @status: VM_CONTEXT1_PROTECTION_FAULT_STATUS register value 763 * @addr: VM_CONTEXT1_PROTECTION_FAULT_ADDR register value 764 * 765 * Print human readable fault information (CIK). 766 */ 767 static void gmc_v7_0_vm_decode_fault(struct amdgpu_device *adev, u32 status, 768 u32 addr, u32 mc_client, unsigned pasid) 769 { 770 u32 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, VMID); 771 u32 protections = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 772 PROTECTIONS); 773 char block[5] = { mc_client >> 24, (mc_client >> 16) & 0xff, 774 (mc_client >> 8) & 0xff, mc_client & 0xff, 0 }; 775 u32 mc_id; 776 777 mc_id = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 778 MEMORY_CLIENT_ID); 779 780 dev_err(adev->dev, "VM fault (0x%02x, vmid %d, pasid %d) at page %u, %s from '%s' (0x%08x) (%d)\n", 781 protections, vmid, pasid, addr, 782 REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 783 MEMORY_CLIENT_RW) ? 784 "write" : "read", block, mc_client, mc_id); 785 } 786 787 788 static const u32 mc_cg_registers[] = { 789 mmMC_HUB_MISC_HUB_CG, 790 mmMC_HUB_MISC_SIP_CG, 791 mmMC_HUB_MISC_VM_CG, 792 mmMC_XPB_CLK_GAT, 793 mmATC_MISC_CG, 794 mmMC_CITF_MISC_WR_CG, 795 mmMC_CITF_MISC_RD_CG, 796 mmMC_CITF_MISC_VM_CG, 797 mmVM_L2_CG, 798 }; 799 800 static const u32 mc_cg_ls_en[] = { 801 MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK, 802 MC_HUB_MISC_SIP_CG__MEM_LS_ENABLE_MASK, 803 MC_HUB_MISC_VM_CG__MEM_LS_ENABLE_MASK, 804 MC_XPB_CLK_GAT__MEM_LS_ENABLE_MASK, 805 ATC_MISC_CG__MEM_LS_ENABLE_MASK, 806 MC_CITF_MISC_WR_CG__MEM_LS_ENABLE_MASK, 807 MC_CITF_MISC_RD_CG__MEM_LS_ENABLE_MASK, 808 MC_CITF_MISC_VM_CG__MEM_LS_ENABLE_MASK, 809 VM_L2_CG__MEM_LS_ENABLE_MASK, 810 }; 811 812 static const u32 mc_cg_en[] = { 813 MC_HUB_MISC_HUB_CG__ENABLE_MASK, 814 MC_HUB_MISC_SIP_CG__ENABLE_MASK, 815 MC_HUB_MISC_VM_CG__ENABLE_MASK, 816 MC_XPB_CLK_GAT__ENABLE_MASK, 817 ATC_MISC_CG__ENABLE_MASK, 818 MC_CITF_MISC_WR_CG__ENABLE_MASK, 819 MC_CITF_MISC_RD_CG__ENABLE_MASK, 820 MC_CITF_MISC_VM_CG__ENABLE_MASK, 821 VM_L2_CG__ENABLE_MASK, 822 }; 823 824 static void gmc_v7_0_enable_mc_ls(struct amdgpu_device *adev, 825 bool enable) 826 { 827 int i; 828 u32 orig, data; 829 830 for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) { 831 orig = data = RREG32(mc_cg_registers[i]); 832 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_LS)) 833 data |= mc_cg_ls_en[i]; 834 else 835 data &= ~mc_cg_ls_en[i]; 836 if (data != orig) 837 WREG32(mc_cg_registers[i], data); 838 } 839 } 840 841 static void gmc_v7_0_enable_mc_mgcg(struct amdgpu_device *adev, 842 bool enable) 843 { 844 int i; 845 u32 orig, data; 846 847 for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) { 848 orig = data = RREG32(mc_cg_registers[i]); 849 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_MGCG)) 850 data |= mc_cg_en[i]; 851 else 852 data &= ~mc_cg_en[i]; 853 if (data != orig) 854 WREG32(mc_cg_registers[i], data); 855 } 856 } 857 858 static void gmc_v7_0_enable_bif_mgls(struct amdgpu_device *adev, 859 bool enable) 860 { 861 u32 orig, data; 862 863 orig = data = RREG32_PCIE(ixPCIE_CNTL2); 864 865 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_BIF_LS)) { 866 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_LS_EN, 1); 867 data = REG_SET_FIELD(data, PCIE_CNTL2, MST_MEM_LS_EN, 1); 868 data = REG_SET_FIELD(data, PCIE_CNTL2, REPLAY_MEM_LS_EN, 1); 869 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_AGGRESSIVE_LS_EN, 1); 870 } else { 871 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_LS_EN, 0); 872 data = REG_SET_FIELD(data, PCIE_CNTL2, MST_MEM_LS_EN, 0); 873 data = REG_SET_FIELD(data, PCIE_CNTL2, REPLAY_MEM_LS_EN, 0); 874 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_AGGRESSIVE_LS_EN, 0); 875 } 876 877 if (orig != data) 878 WREG32_PCIE(ixPCIE_CNTL2, data); 879 } 880 881 static void gmc_v7_0_enable_hdp_mgcg(struct amdgpu_device *adev, 882 bool enable) 883 { 884 u32 orig, data; 885 886 orig = data = RREG32(mmHDP_HOST_PATH_CNTL); 887 888 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_HDP_MGCG)) 889 data = REG_SET_FIELD(data, HDP_HOST_PATH_CNTL, CLOCK_GATING_DIS, 0); 890 else 891 data = REG_SET_FIELD(data, HDP_HOST_PATH_CNTL, CLOCK_GATING_DIS, 1); 892 893 if (orig != data) 894 WREG32(mmHDP_HOST_PATH_CNTL, data); 895 } 896 897 static void gmc_v7_0_enable_hdp_ls(struct amdgpu_device *adev, 898 bool enable) 899 { 900 u32 orig, data; 901 902 orig = data = RREG32(mmHDP_MEM_POWER_LS); 903 904 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_HDP_LS)) 905 data = REG_SET_FIELD(data, HDP_MEM_POWER_LS, LS_ENABLE, 1); 906 else 907 data = REG_SET_FIELD(data, HDP_MEM_POWER_LS, LS_ENABLE, 0); 908 909 if (orig != data) 910 WREG32(mmHDP_MEM_POWER_LS, data); 911 } 912 913 static int gmc_v7_0_convert_vram_type(int mc_seq_vram_type) 914 { 915 switch (mc_seq_vram_type) { 916 case MC_SEQ_MISC0__MT__GDDR1: 917 return AMDGPU_VRAM_TYPE_GDDR1; 918 case MC_SEQ_MISC0__MT__DDR2: 919 return AMDGPU_VRAM_TYPE_DDR2; 920 case MC_SEQ_MISC0__MT__GDDR3: 921 return AMDGPU_VRAM_TYPE_GDDR3; 922 case MC_SEQ_MISC0__MT__GDDR4: 923 return AMDGPU_VRAM_TYPE_GDDR4; 924 case MC_SEQ_MISC0__MT__GDDR5: 925 return AMDGPU_VRAM_TYPE_GDDR5; 926 case MC_SEQ_MISC0__MT__HBM: 927 return AMDGPU_VRAM_TYPE_HBM; 928 case MC_SEQ_MISC0__MT__DDR3: 929 return AMDGPU_VRAM_TYPE_DDR3; 930 default: 931 return AMDGPU_VRAM_TYPE_UNKNOWN; 932 } 933 } 934 935 static int gmc_v7_0_early_init(void *handle) 936 { 937 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 938 939 gmc_v7_0_set_gmc_funcs(adev); 940 gmc_v7_0_set_irq_funcs(adev); 941 942 adev->gmc.shared_aperture_start = 0x2000000000000000ULL; 943 adev->gmc.shared_aperture_end = 944 adev->gmc.shared_aperture_start + (4ULL << 30) - 1; 945 adev->gmc.private_aperture_start = 946 adev->gmc.shared_aperture_end + 1; 947 adev->gmc.private_aperture_end = 948 adev->gmc.private_aperture_start + (4ULL << 30) - 1; 949 950 return 0; 951 } 952 953 static int gmc_v7_0_late_init(void *handle) 954 { 955 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 956 957 amdgpu_bo_late_init(adev); 958 959 if (amdgpu_vm_fault_stop != AMDGPU_VM_FAULT_STOP_ALWAYS) 960 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0); 961 else 962 return 0; 963 } 964 965 static unsigned gmc_v7_0_get_vbios_fb_size(struct amdgpu_device *adev) 966 { 967 u32 d1vga_control = RREG32(mmD1VGA_CONTROL); 968 unsigned size; 969 970 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) { 971 size = 9 * 1024 * 1024; /* reserve 8MB for vga emulator and 1 MB for FB */ 972 } else { 973 u32 viewport = RREG32(mmVIEWPORT_SIZE); 974 size = (REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_HEIGHT) * 975 REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_WIDTH) * 976 4); 977 } 978 /* return 0 if the pre-OS buffer uses up most of vram */ 979 if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024)) 980 return 0; 981 return size; 982 } 983 984 static int gmc_v7_0_sw_init(void *handle) 985 { 986 int r; 987 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 988 989 adev->num_vmhubs = 1; 990 991 if (adev->flags & AMD_IS_APU) { 992 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_UNKNOWN; 993 } else { 994 u32 tmp = RREG32(mmMC_SEQ_MISC0); 995 tmp &= MC_SEQ_MISC0__MT__MASK; 996 adev->gmc.vram_type = gmc_v7_0_convert_vram_type(tmp); 997 } 998 999 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_PAGE_INV_FAULT, &adev->gmc.vm_fault); 1000 if (r) 1001 return r; 1002 1003 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_MEM_PROT_FAULT, &adev->gmc.vm_fault); 1004 if (r) 1005 return r; 1006 1007 /* Adjust VM size here. 1008 * Currently set to 4GB ((1 << 20) 4k pages). 1009 * Max GPUVM size for cayman and SI is 40 bits. 1010 */ 1011 amdgpu_vm_adjust_size(adev, 64, 9, 1, 40); 1012 1013 /* Set the internal MC address mask 1014 * This is the max address of the GPU's 1015 * internal address space. 1016 */ 1017 adev->gmc.mc_mask = 0xffffffffffULL; /* 40 bit MC */ 1018 1019 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(40)); 1020 if (r) { 1021 pr_warn("amdgpu: No suitable DMA available\n"); 1022 return r; 1023 } 1024 adev->need_swiotlb = drm_need_swiotlb(40); 1025 1026 r = gmc_v7_0_init_microcode(adev); 1027 if (r) { 1028 DRM_ERROR("Failed to load mc firmware!\n"); 1029 return r; 1030 } 1031 1032 r = gmc_v7_0_mc_init(adev); 1033 if (r) 1034 return r; 1035 1036 adev->gmc.stolen_size = gmc_v7_0_get_vbios_fb_size(adev); 1037 1038 /* Memory manager */ 1039 r = amdgpu_bo_init(adev); 1040 if (r) 1041 return r; 1042 1043 r = gmc_v7_0_gart_init(adev); 1044 if (r) 1045 return r; 1046 1047 /* 1048 * number of VMs 1049 * VMID 0 is reserved for System 1050 * amdgpu graphics/compute will use VMIDs 1-7 1051 * amdkfd will use VMIDs 8-15 1052 */ 1053 adev->vm_manager.id_mgr[0].num_ids = AMDGPU_NUM_OF_VMIDS; 1054 amdgpu_vm_manager_init(adev); 1055 1056 /* base offset of vram pages */ 1057 if (adev->flags & AMD_IS_APU) { 1058 u64 tmp = RREG32(mmMC_VM_FB_OFFSET); 1059 1060 tmp <<= 22; 1061 adev->vm_manager.vram_base_offset = tmp; 1062 } else { 1063 adev->vm_manager.vram_base_offset = 0; 1064 } 1065 1066 adev->gmc.vm_fault_info = kmalloc(sizeof(struct kfd_vm_fault_info), 1067 GFP_KERNEL); 1068 if (!adev->gmc.vm_fault_info) 1069 return -ENOMEM; 1070 atomic_set(&adev->gmc.vm_fault_info_updated, 0); 1071 1072 return 0; 1073 } 1074 1075 static int gmc_v7_0_sw_fini(void *handle) 1076 { 1077 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1078 1079 amdgpu_gem_force_release(adev); 1080 amdgpu_vm_manager_fini(adev); 1081 kfree(adev->gmc.vm_fault_info); 1082 amdgpu_gart_table_vram_free(adev); 1083 amdgpu_bo_fini(adev); 1084 amdgpu_gart_fini(adev); 1085 release_firmware(adev->gmc.fw); 1086 adev->gmc.fw = NULL; 1087 1088 return 0; 1089 } 1090 1091 static int gmc_v7_0_hw_init(void *handle) 1092 { 1093 int r; 1094 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1095 1096 gmc_v7_0_init_golden_registers(adev); 1097 1098 gmc_v7_0_mc_program(adev); 1099 1100 if (!(adev->flags & AMD_IS_APU)) { 1101 r = gmc_v7_0_mc_load_microcode(adev); 1102 if (r) { 1103 DRM_ERROR("Failed to load MC firmware!\n"); 1104 return r; 1105 } 1106 } 1107 1108 r = gmc_v7_0_gart_enable(adev); 1109 if (r) 1110 return r; 1111 1112 return r; 1113 } 1114 1115 static int gmc_v7_0_hw_fini(void *handle) 1116 { 1117 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1118 1119 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0); 1120 gmc_v7_0_gart_disable(adev); 1121 1122 return 0; 1123 } 1124 1125 static int gmc_v7_0_suspend(void *handle) 1126 { 1127 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1128 1129 gmc_v7_0_hw_fini(adev); 1130 1131 return 0; 1132 } 1133 1134 static int gmc_v7_0_resume(void *handle) 1135 { 1136 int r; 1137 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1138 1139 r = gmc_v7_0_hw_init(adev); 1140 if (r) 1141 return r; 1142 1143 amdgpu_vmid_reset_all(adev); 1144 1145 return 0; 1146 } 1147 1148 static bool gmc_v7_0_is_idle(void *handle) 1149 { 1150 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1151 u32 tmp = RREG32(mmSRBM_STATUS); 1152 1153 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1154 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK | SRBM_STATUS__VMC_BUSY_MASK)) 1155 return false; 1156 1157 return true; 1158 } 1159 1160 static int gmc_v7_0_wait_for_idle(void *handle) 1161 { 1162 unsigned i; 1163 u32 tmp; 1164 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1165 1166 for (i = 0; i < adev->usec_timeout; i++) { 1167 /* read MC_STATUS */ 1168 tmp = RREG32(mmSRBM_STATUS) & (SRBM_STATUS__MCB_BUSY_MASK | 1169 SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1170 SRBM_STATUS__MCC_BUSY_MASK | 1171 SRBM_STATUS__MCD_BUSY_MASK | 1172 SRBM_STATUS__VMC_BUSY_MASK); 1173 if (!tmp) 1174 return 0; 1175 udelay(1); 1176 } 1177 return -ETIMEDOUT; 1178 1179 } 1180 1181 static int gmc_v7_0_soft_reset(void *handle) 1182 { 1183 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1184 u32 srbm_soft_reset = 0; 1185 u32 tmp = RREG32(mmSRBM_STATUS); 1186 1187 if (tmp & SRBM_STATUS__VMC_BUSY_MASK) 1188 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset, 1189 SRBM_SOFT_RESET, SOFT_RESET_VMC, 1); 1190 1191 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1192 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK)) { 1193 if (!(adev->flags & AMD_IS_APU)) 1194 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset, 1195 SRBM_SOFT_RESET, SOFT_RESET_MC, 1); 1196 } 1197 1198 if (srbm_soft_reset) { 1199 gmc_v7_0_mc_stop(adev); 1200 if (gmc_v7_0_wait_for_idle((void *)adev)) { 1201 dev_warn(adev->dev, "Wait for GMC idle timed out !\n"); 1202 } 1203 1204 1205 tmp = RREG32(mmSRBM_SOFT_RESET); 1206 tmp |= srbm_soft_reset; 1207 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp); 1208 WREG32(mmSRBM_SOFT_RESET, tmp); 1209 tmp = RREG32(mmSRBM_SOFT_RESET); 1210 1211 udelay(50); 1212 1213 tmp &= ~srbm_soft_reset; 1214 WREG32(mmSRBM_SOFT_RESET, tmp); 1215 tmp = RREG32(mmSRBM_SOFT_RESET); 1216 1217 /* Wait a little for things to settle down */ 1218 udelay(50); 1219 1220 gmc_v7_0_mc_resume(adev); 1221 udelay(50); 1222 } 1223 1224 return 0; 1225 } 1226 1227 static int gmc_v7_0_vm_fault_interrupt_state(struct amdgpu_device *adev, 1228 struct amdgpu_irq_src *src, 1229 unsigned type, 1230 enum amdgpu_interrupt_state state) 1231 { 1232 u32 tmp; 1233 u32 bits = (VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1234 VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1235 VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1236 VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1237 VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1238 VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK); 1239 1240 switch (state) { 1241 case AMDGPU_IRQ_STATE_DISABLE: 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 case AMDGPU_IRQ_STATE_ENABLE: 1252 /* system context */ 1253 tmp = RREG32(mmVM_CONTEXT0_CNTL); 1254 tmp |= bits; 1255 WREG32(mmVM_CONTEXT0_CNTL, tmp); 1256 /* VMs */ 1257 tmp = RREG32(mmVM_CONTEXT1_CNTL); 1258 tmp |= bits; 1259 WREG32(mmVM_CONTEXT1_CNTL, tmp); 1260 break; 1261 default: 1262 break; 1263 } 1264 1265 return 0; 1266 } 1267 1268 static int gmc_v7_0_process_interrupt(struct amdgpu_device *adev, 1269 struct amdgpu_irq_src *source, 1270 struct amdgpu_iv_entry *entry) 1271 { 1272 u32 addr, status, mc_client, vmid; 1273 1274 addr = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_ADDR); 1275 status = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_STATUS); 1276 mc_client = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_MCCLIENT); 1277 /* reset addr and status */ 1278 WREG32_P(mmVM_CONTEXT1_CNTL2, 1, ~1); 1279 1280 if (!addr && !status) 1281 return 0; 1282 1283 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_FIRST) 1284 gmc_v7_0_set_fault_enable_default(adev, false); 1285 1286 if (printk_ratelimit()) { 1287 dev_err(adev->dev, "GPU fault detected: %d 0x%08x\n", 1288 entry->src_id, entry->src_data[0]); 1289 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n", 1290 addr); 1291 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n", 1292 status); 1293 gmc_v7_0_vm_decode_fault(adev, status, addr, mc_client, 1294 entry->pasid); 1295 } 1296 1297 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 1298 VMID); 1299 if (amdgpu_amdkfd_is_kfd_vmid(adev, vmid) 1300 && !atomic_read(&adev->gmc.vm_fault_info_updated)) { 1301 struct kfd_vm_fault_info *info = adev->gmc.vm_fault_info; 1302 u32 protections = REG_GET_FIELD(status, 1303 VM_CONTEXT1_PROTECTION_FAULT_STATUS, 1304 PROTECTIONS); 1305 1306 info->vmid = vmid; 1307 info->mc_id = REG_GET_FIELD(status, 1308 VM_CONTEXT1_PROTECTION_FAULT_STATUS, 1309 MEMORY_CLIENT_ID); 1310 info->status = status; 1311 info->page_addr = addr; 1312 info->prot_valid = protections & 0x7 ? true : false; 1313 info->prot_read = protections & 0x8 ? true : false; 1314 info->prot_write = protections & 0x10 ? true : false; 1315 info->prot_exec = protections & 0x20 ? true : false; 1316 mb(); 1317 atomic_set(&adev->gmc.vm_fault_info_updated, 1); 1318 } 1319 1320 return 0; 1321 } 1322 1323 static int gmc_v7_0_set_clockgating_state(void *handle, 1324 enum amd_clockgating_state state) 1325 { 1326 bool gate = false; 1327 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1328 1329 if (state == AMD_CG_STATE_GATE) 1330 gate = true; 1331 1332 if (!(adev->flags & AMD_IS_APU)) { 1333 gmc_v7_0_enable_mc_mgcg(adev, gate); 1334 gmc_v7_0_enable_mc_ls(adev, gate); 1335 } 1336 gmc_v7_0_enable_bif_mgls(adev, gate); 1337 gmc_v7_0_enable_hdp_mgcg(adev, gate); 1338 gmc_v7_0_enable_hdp_ls(adev, gate); 1339 1340 return 0; 1341 } 1342 1343 static int gmc_v7_0_set_powergating_state(void *handle, 1344 enum amd_powergating_state state) 1345 { 1346 return 0; 1347 } 1348 1349 static const struct amd_ip_funcs gmc_v7_0_ip_funcs = { 1350 .name = "gmc_v7_0", 1351 .early_init = gmc_v7_0_early_init, 1352 .late_init = gmc_v7_0_late_init, 1353 .sw_init = gmc_v7_0_sw_init, 1354 .sw_fini = gmc_v7_0_sw_fini, 1355 .hw_init = gmc_v7_0_hw_init, 1356 .hw_fini = gmc_v7_0_hw_fini, 1357 .suspend = gmc_v7_0_suspend, 1358 .resume = gmc_v7_0_resume, 1359 .is_idle = gmc_v7_0_is_idle, 1360 .wait_for_idle = gmc_v7_0_wait_for_idle, 1361 .soft_reset = gmc_v7_0_soft_reset, 1362 .set_clockgating_state = gmc_v7_0_set_clockgating_state, 1363 .set_powergating_state = gmc_v7_0_set_powergating_state, 1364 }; 1365 1366 static const struct amdgpu_gmc_funcs gmc_v7_0_gmc_funcs = { 1367 .flush_gpu_tlb = gmc_v7_0_flush_gpu_tlb, 1368 .flush_gpu_tlb_pasid = gmc_v7_0_flush_gpu_tlb_pasid, 1369 .emit_flush_gpu_tlb = gmc_v7_0_emit_flush_gpu_tlb, 1370 .emit_pasid_mapping = gmc_v7_0_emit_pasid_mapping, 1371 .set_prt = gmc_v7_0_set_prt, 1372 .get_vm_pde = gmc_v7_0_get_vm_pde, 1373 .get_vm_pte = gmc_v7_0_get_vm_pte 1374 }; 1375 1376 static const struct amdgpu_irq_src_funcs gmc_v7_0_irq_funcs = { 1377 .set = gmc_v7_0_vm_fault_interrupt_state, 1378 .process = gmc_v7_0_process_interrupt, 1379 }; 1380 1381 static void gmc_v7_0_set_gmc_funcs(struct amdgpu_device *adev) 1382 { 1383 adev->gmc.gmc_funcs = &gmc_v7_0_gmc_funcs; 1384 } 1385 1386 static void gmc_v7_0_set_irq_funcs(struct amdgpu_device *adev) 1387 { 1388 adev->gmc.vm_fault.num_types = 1; 1389 adev->gmc.vm_fault.funcs = &gmc_v7_0_irq_funcs; 1390 } 1391 1392 const struct amdgpu_ip_block_version gmc_v7_0_ip_block = 1393 { 1394 .type = AMD_IP_BLOCK_TYPE_GMC, 1395 .major = 7, 1396 .minor = 0, 1397 .rev = 0, 1398 .funcs = &gmc_v7_0_ip_funcs, 1399 }; 1400 1401 const struct amdgpu_ip_block_version gmc_v7_4_ip_block = 1402 { 1403 .type = AMD_IP_BLOCK_TYPE_GMC, 1404 .major = 7, 1405 .minor = 4, 1406 .rev = 0, 1407 .funcs = &gmc_v7_0_ip_funcs, 1408 }; 1409