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.real_vram_size > adev->gmc.aper_size) { 386 adev->gmc.aper_base = ((u64)RREG32(mmMC_VM_FB_OFFSET)) << 22; 387 adev->gmc.aper_size = adev->gmc.real_vram_size; 388 } 389 #endif 390 391 /* In case the PCI BAR is larger than the actual amount of vram */ 392 adev->gmc.visible_vram_size = adev->gmc.aper_size; 393 if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size) 394 adev->gmc.visible_vram_size = adev->gmc.real_vram_size; 395 396 /* set the gart size */ 397 if (amdgpu_gart_size == -1) { 398 switch (adev->asic_type) { 399 case CHIP_TOPAZ: /* no MM engines */ 400 default: 401 adev->gmc.gart_size = 256ULL << 20; 402 break; 403 #ifdef CONFIG_DRM_AMDGPU_CIK 404 case CHIP_BONAIRE: /* UVD, VCE do not support GPUVM */ 405 case CHIP_HAWAII: /* UVD, VCE do not support GPUVM */ 406 case CHIP_KAVERI: /* UVD, VCE do not support GPUVM */ 407 case CHIP_KABINI: /* UVD, VCE do not support GPUVM */ 408 case CHIP_MULLINS: /* UVD, VCE do not support GPUVM */ 409 adev->gmc.gart_size = 1024ULL << 20; 410 break; 411 #endif 412 } 413 } else { 414 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20; 415 } 416 417 gmc_v7_0_vram_gtt_location(adev, &adev->gmc); 418 419 return 0; 420 } 421 422 /** 423 * gmc_v7_0_flush_gpu_tlb_pasid - tlb flush via pasid 424 * 425 * @adev: amdgpu_device pointer 426 * @pasid: pasid to be flush 427 * 428 * Flush the TLB for the requested pasid. 429 */ 430 static int gmc_v7_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, 431 uint16_t pasid, uint32_t flush_type, 432 bool all_hub) 433 { 434 int vmid; 435 unsigned int tmp; 436 437 if (adev->in_gpu_reset) 438 return -EIO; 439 440 for (vmid = 1; vmid < 16; vmid++) { 441 442 tmp = RREG32(mmATC_VMID0_PASID_MAPPING + vmid); 443 if ((tmp & ATC_VMID0_PASID_MAPPING__VALID_MASK) && 444 (tmp & ATC_VMID0_PASID_MAPPING__PASID_MASK) == pasid) { 445 WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid); 446 RREG32(mmVM_INVALIDATE_RESPONSE); 447 break; 448 } 449 } 450 451 return 0; 452 } 453 454 /* 455 * GART 456 * VMID 0 is the physical GPU addresses as used by the kernel. 457 * VMIDs 1-15 are used for userspace clients and are handled 458 * by the amdgpu vm/hsa code. 459 */ 460 461 /** 462 * gmc_v7_0_flush_gpu_tlb - gart tlb flush callback 463 * 464 * @adev: amdgpu_device pointer 465 * @vmid: vm instance to flush 466 * 467 * Flush the TLB for the requested page table (CIK). 468 */ 469 static void gmc_v7_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid, 470 uint32_t vmhub, uint32_t flush_type) 471 { 472 /* bits 0-15 are the VM contexts0-15 */ 473 WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid); 474 } 475 476 static uint64_t gmc_v7_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring, 477 unsigned vmid, uint64_t pd_addr) 478 { 479 uint32_t reg; 480 481 if (vmid < 8) 482 reg = mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vmid; 483 else 484 reg = mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8; 485 amdgpu_ring_emit_wreg(ring, reg, pd_addr >> 12); 486 487 /* bits 0-15 are the VM contexts0-15 */ 488 amdgpu_ring_emit_wreg(ring, mmVM_INVALIDATE_REQUEST, 1 << vmid); 489 490 return pd_addr; 491 } 492 493 static void gmc_v7_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid, 494 unsigned pasid) 495 { 496 amdgpu_ring_emit_wreg(ring, mmIH_VMID_0_LUT + vmid, pasid); 497 } 498 499 static void gmc_v7_0_get_vm_pde(struct amdgpu_device *adev, int level, 500 uint64_t *addr, uint64_t *flags) 501 { 502 BUG_ON(*addr & 0xFFFFFF0000000FFFULL); 503 } 504 505 static void gmc_v7_0_get_vm_pte(struct amdgpu_device *adev, 506 struct amdgpu_bo_va_mapping *mapping, 507 uint64_t *flags) 508 { 509 *flags &= ~AMDGPU_PTE_EXECUTABLE; 510 *flags &= ~AMDGPU_PTE_PRT; 511 } 512 513 /** 514 * gmc_v8_0_set_fault_enable_default - update VM fault handling 515 * 516 * @adev: amdgpu_device pointer 517 * @value: true redirects VM faults to the default page 518 */ 519 static void gmc_v7_0_set_fault_enable_default(struct amdgpu_device *adev, 520 bool value) 521 { 522 u32 tmp; 523 524 tmp = RREG32(mmVM_CONTEXT1_CNTL); 525 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 526 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 527 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 528 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 529 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 530 PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, value); 531 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 532 VALID_PROTECTION_FAULT_ENABLE_DEFAULT, value); 533 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 534 READ_PROTECTION_FAULT_ENABLE_DEFAULT, value); 535 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 536 WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 537 WREG32(mmVM_CONTEXT1_CNTL, tmp); 538 } 539 540 /** 541 * gmc_v7_0_set_prt - set PRT VM fault 542 * 543 * @adev: amdgpu_device pointer 544 * @enable: enable/disable VM fault handling for PRT 545 */ 546 static void gmc_v7_0_set_prt(struct amdgpu_device *adev, bool enable) 547 { 548 uint32_t tmp; 549 550 if (enable && !adev->gmc.prt_warning) { 551 dev_warn(adev->dev, "Disabling VM faults because of PRT request!\n"); 552 adev->gmc.prt_warning = true; 553 } 554 555 tmp = RREG32(mmVM_PRT_CNTL); 556 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 557 CB_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable); 558 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 559 CB_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable); 560 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 561 TC_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable); 562 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 563 TC_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable); 564 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 565 L2_CACHE_STORE_INVALID_ENTRIES, enable); 566 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 567 L1_TLB_STORE_INVALID_ENTRIES, enable); 568 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL, 569 MASK_PDE0_FAULT, enable); 570 WREG32(mmVM_PRT_CNTL, tmp); 571 572 if (enable) { 573 uint32_t low = AMDGPU_VA_RESERVED_SIZE >> AMDGPU_GPU_PAGE_SHIFT; 574 uint32_t high = adev->vm_manager.max_pfn - 575 (AMDGPU_VA_RESERVED_SIZE >> AMDGPU_GPU_PAGE_SHIFT); 576 577 WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, low); 578 WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, low); 579 WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, low); 580 WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, low); 581 WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, high); 582 WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, high); 583 WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, high); 584 WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, high); 585 } else { 586 WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, 0xfffffff); 587 WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, 0xfffffff); 588 WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, 0xfffffff); 589 WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, 0xfffffff); 590 WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, 0x0); 591 WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, 0x0); 592 WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, 0x0); 593 WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, 0x0); 594 } 595 } 596 597 /** 598 * gmc_v7_0_gart_enable - gart enable 599 * 600 * @adev: amdgpu_device pointer 601 * 602 * This sets up the TLBs, programs the page tables for VMID0, 603 * sets up the hw for VMIDs 1-15 which are allocated on 604 * demand, and sets up the global locations for the LDS, GDS, 605 * and GPUVM for FSA64 clients (CIK). 606 * Returns 0 for success, errors for failure. 607 */ 608 static int gmc_v7_0_gart_enable(struct amdgpu_device *adev) 609 { 610 uint64_t table_addr; 611 int r, i; 612 u32 tmp, field; 613 614 if (adev->gart.bo == NULL) { 615 dev_err(adev->dev, "No VRAM object for PCIE GART.\n"); 616 return -EINVAL; 617 } 618 r = amdgpu_gart_table_vram_pin(adev); 619 if (r) 620 return r; 621 622 table_addr = amdgpu_bo_gpu_offset(adev->gart.bo); 623 624 /* Setup TLB control */ 625 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL); 626 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 1); 627 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 1); 628 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE, 3); 629 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 1); 630 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_APERTURE_UNMAPPED_ACCESS, 0); 631 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp); 632 /* Setup L2 cache */ 633 tmp = RREG32(mmVM_L2_CNTL); 634 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 1); 635 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING, 1); 636 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE, 1); 637 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE, 1); 638 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, EFFECTIVE_L2_QUEUE_SIZE, 7); 639 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, CONTEXT1_IDENTITY_ACCESS_MODE, 1); 640 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_DEFAULT_PAGE_OUT_TO_SYSTEM_MEMORY, 1); 641 WREG32(mmVM_L2_CNTL, tmp); 642 tmp = REG_SET_FIELD(0, VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS, 1); 643 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_L2_CACHE, 1); 644 WREG32(mmVM_L2_CNTL2, tmp); 645 646 field = adev->vm_manager.fragment_size; 647 tmp = RREG32(mmVM_L2_CNTL3); 648 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY, 1); 649 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, BANK_SELECT, field); 650 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_FRAGMENT_SIZE, field); 651 WREG32(mmVM_L2_CNTL3, tmp); 652 /* setup context0 */ 653 WREG32(mmVM_CONTEXT0_PAGE_TABLE_START_ADDR, adev->gmc.gart_start >> 12); 654 WREG32(mmVM_CONTEXT0_PAGE_TABLE_END_ADDR, adev->gmc.gart_end >> 12); 655 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR, table_addr >> 12); 656 WREG32(mmVM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR, 657 (u32)(adev->dummy_page_addr >> 12)); 658 WREG32(mmVM_CONTEXT0_CNTL2, 0); 659 tmp = RREG32(mmVM_CONTEXT0_CNTL); 660 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, ENABLE_CONTEXT, 1); 661 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, PAGE_TABLE_DEPTH, 0); 662 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 663 WREG32(mmVM_CONTEXT0_CNTL, tmp); 664 665 WREG32(0x575, 0); 666 WREG32(0x576, 0); 667 WREG32(0x577, 0); 668 669 /* empty context1-15 */ 670 /* FIXME start with 4G, once using 2 level pt switch to full 671 * vm size space 672 */ 673 /* set vm size, must be a multiple of 4 */ 674 WREG32(mmVM_CONTEXT1_PAGE_TABLE_START_ADDR, 0); 675 WREG32(mmVM_CONTEXT1_PAGE_TABLE_END_ADDR, adev->vm_manager.max_pfn - 1); 676 for (i = 1; i < 16; i++) { 677 if (i < 8) 678 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + i, 679 table_addr >> 12); 680 else 681 WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + i - 8, 682 table_addr >> 12); 683 } 684 685 /* enable context1-15 */ 686 WREG32(mmVM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR, 687 (u32)(adev->dummy_page_addr >> 12)); 688 WREG32(mmVM_CONTEXT1_CNTL2, 4); 689 tmp = RREG32(mmVM_CONTEXT1_CNTL); 690 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, ENABLE_CONTEXT, 1); 691 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_DEPTH, 1); 692 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_BLOCK_SIZE, 693 adev->vm_manager.block_size - 9); 694 WREG32(mmVM_CONTEXT1_CNTL, tmp); 695 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) 696 gmc_v7_0_set_fault_enable_default(adev, false); 697 else 698 gmc_v7_0_set_fault_enable_default(adev, true); 699 700 if (adev->asic_type == CHIP_KAVERI) { 701 tmp = RREG32(mmCHUB_CONTROL); 702 tmp &= ~BYPASS_VM; 703 WREG32(mmCHUB_CONTROL, tmp); 704 } 705 706 gmc_v7_0_flush_gpu_tlb(adev, 0, 0, 0); 707 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n", 708 (unsigned)(adev->gmc.gart_size >> 20), 709 (unsigned long long)table_addr); 710 adev->gart.ready = true; 711 return 0; 712 } 713 714 static int gmc_v7_0_gart_init(struct amdgpu_device *adev) 715 { 716 int r; 717 718 if (adev->gart.bo) { 719 WARN(1, "R600 PCIE GART already initialized\n"); 720 return 0; 721 } 722 /* Initialize common gart structure */ 723 r = amdgpu_gart_init(adev); 724 if (r) 725 return r; 726 adev->gart.table_size = adev->gart.num_gpu_pages * 8; 727 adev->gart.gart_pte_flags = 0; 728 return amdgpu_gart_table_vram_alloc(adev); 729 } 730 731 /** 732 * gmc_v7_0_gart_disable - gart disable 733 * 734 * @adev: amdgpu_device pointer 735 * 736 * This disables all VM page table (CIK). 737 */ 738 static void gmc_v7_0_gart_disable(struct amdgpu_device *adev) 739 { 740 u32 tmp; 741 742 /* Disable all tables */ 743 WREG32(mmVM_CONTEXT0_CNTL, 0); 744 WREG32(mmVM_CONTEXT1_CNTL, 0); 745 /* Setup TLB control */ 746 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL); 747 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 0); 748 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 0); 749 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 0); 750 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp); 751 /* Setup L2 cache */ 752 tmp = RREG32(mmVM_L2_CNTL); 753 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 0); 754 WREG32(mmVM_L2_CNTL, tmp); 755 WREG32(mmVM_L2_CNTL2, 0); 756 amdgpu_gart_table_vram_unpin(adev); 757 } 758 759 /** 760 * gmc_v7_0_vm_decode_fault - print human readable fault info 761 * 762 * @adev: amdgpu_device pointer 763 * @status: VM_CONTEXT1_PROTECTION_FAULT_STATUS register value 764 * @addr: VM_CONTEXT1_PROTECTION_FAULT_ADDR register value 765 * @mc_client: VM_CONTEXT1_PROTECTION_FAULT_MCCLIENT register value 766 * 767 * Print human readable fault information (CIK). 768 */ 769 static void gmc_v7_0_vm_decode_fault(struct amdgpu_device *adev, u32 status, 770 u32 addr, u32 mc_client, unsigned pasid) 771 { 772 u32 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, VMID); 773 u32 protections = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 774 PROTECTIONS); 775 char block[5] = { mc_client >> 24, (mc_client >> 16) & 0xff, 776 (mc_client >> 8) & 0xff, mc_client & 0xff, 0 }; 777 u32 mc_id; 778 779 mc_id = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 780 MEMORY_CLIENT_ID); 781 782 dev_err(adev->dev, "VM fault (0x%02x, vmid %d, pasid %d) at page %u, %s from '%s' (0x%08x) (%d)\n", 783 protections, vmid, pasid, addr, 784 REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 785 MEMORY_CLIENT_RW) ? 786 "write" : "read", block, mc_client, mc_id); 787 } 788 789 790 static const u32 mc_cg_registers[] = { 791 mmMC_HUB_MISC_HUB_CG, 792 mmMC_HUB_MISC_SIP_CG, 793 mmMC_HUB_MISC_VM_CG, 794 mmMC_XPB_CLK_GAT, 795 mmATC_MISC_CG, 796 mmMC_CITF_MISC_WR_CG, 797 mmMC_CITF_MISC_RD_CG, 798 mmMC_CITF_MISC_VM_CG, 799 mmVM_L2_CG, 800 }; 801 802 static const u32 mc_cg_ls_en[] = { 803 MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK, 804 MC_HUB_MISC_SIP_CG__MEM_LS_ENABLE_MASK, 805 MC_HUB_MISC_VM_CG__MEM_LS_ENABLE_MASK, 806 MC_XPB_CLK_GAT__MEM_LS_ENABLE_MASK, 807 ATC_MISC_CG__MEM_LS_ENABLE_MASK, 808 MC_CITF_MISC_WR_CG__MEM_LS_ENABLE_MASK, 809 MC_CITF_MISC_RD_CG__MEM_LS_ENABLE_MASK, 810 MC_CITF_MISC_VM_CG__MEM_LS_ENABLE_MASK, 811 VM_L2_CG__MEM_LS_ENABLE_MASK, 812 }; 813 814 static const u32 mc_cg_en[] = { 815 MC_HUB_MISC_HUB_CG__ENABLE_MASK, 816 MC_HUB_MISC_SIP_CG__ENABLE_MASK, 817 MC_HUB_MISC_VM_CG__ENABLE_MASK, 818 MC_XPB_CLK_GAT__ENABLE_MASK, 819 ATC_MISC_CG__ENABLE_MASK, 820 MC_CITF_MISC_WR_CG__ENABLE_MASK, 821 MC_CITF_MISC_RD_CG__ENABLE_MASK, 822 MC_CITF_MISC_VM_CG__ENABLE_MASK, 823 VM_L2_CG__ENABLE_MASK, 824 }; 825 826 static void gmc_v7_0_enable_mc_ls(struct amdgpu_device *adev, 827 bool enable) 828 { 829 int i; 830 u32 orig, data; 831 832 for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) { 833 orig = data = RREG32(mc_cg_registers[i]); 834 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_LS)) 835 data |= mc_cg_ls_en[i]; 836 else 837 data &= ~mc_cg_ls_en[i]; 838 if (data != orig) 839 WREG32(mc_cg_registers[i], data); 840 } 841 } 842 843 static void gmc_v7_0_enable_mc_mgcg(struct amdgpu_device *adev, 844 bool enable) 845 { 846 int i; 847 u32 orig, data; 848 849 for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) { 850 orig = data = RREG32(mc_cg_registers[i]); 851 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_MGCG)) 852 data |= mc_cg_en[i]; 853 else 854 data &= ~mc_cg_en[i]; 855 if (data != orig) 856 WREG32(mc_cg_registers[i], data); 857 } 858 } 859 860 static void gmc_v7_0_enable_bif_mgls(struct amdgpu_device *adev, 861 bool enable) 862 { 863 u32 orig, data; 864 865 orig = data = RREG32_PCIE(ixPCIE_CNTL2); 866 867 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_BIF_LS)) { 868 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_LS_EN, 1); 869 data = REG_SET_FIELD(data, PCIE_CNTL2, MST_MEM_LS_EN, 1); 870 data = REG_SET_FIELD(data, PCIE_CNTL2, REPLAY_MEM_LS_EN, 1); 871 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_AGGRESSIVE_LS_EN, 1); 872 } else { 873 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_LS_EN, 0); 874 data = REG_SET_FIELD(data, PCIE_CNTL2, MST_MEM_LS_EN, 0); 875 data = REG_SET_FIELD(data, PCIE_CNTL2, REPLAY_MEM_LS_EN, 0); 876 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_AGGRESSIVE_LS_EN, 0); 877 } 878 879 if (orig != data) 880 WREG32_PCIE(ixPCIE_CNTL2, data); 881 } 882 883 static void gmc_v7_0_enable_hdp_mgcg(struct amdgpu_device *adev, 884 bool enable) 885 { 886 u32 orig, data; 887 888 orig = data = RREG32(mmHDP_HOST_PATH_CNTL); 889 890 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_HDP_MGCG)) 891 data = REG_SET_FIELD(data, HDP_HOST_PATH_CNTL, CLOCK_GATING_DIS, 0); 892 else 893 data = REG_SET_FIELD(data, HDP_HOST_PATH_CNTL, CLOCK_GATING_DIS, 1); 894 895 if (orig != data) 896 WREG32(mmHDP_HOST_PATH_CNTL, data); 897 } 898 899 static void gmc_v7_0_enable_hdp_ls(struct amdgpu_device *adev, 900 bool enable) 901 { 902 u32 orig, data; 903 904 orig = data = RREG32(mmHDP_MEM_POWER_LS); 905 906 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_HDP_LS)) 907 data = REG_SET_FIELD(data, HDP_MEM_POWER_LS, LS_ENABLE, 1); 908 else 909 data = REG_SET_FIELD(data, HDP_MEM_POWER_LS, LS_ENABLE, 0); 910 911 if (orig != data) 912 WREG32(mmHDP_MEM_POWER_LS, data); 913 } 914 915 static int gmc_v7_0_convert_vram_type(int mc_seq_vram_type) 916 { 917 switch (mc_seq_vram_type) { 918 case MC_SEQ_MISC0__MT__GDDR1: 919 return AMDGPU_VRAM_TYPE_GDDR1; 920 case MC_SEQ_MISC0__MT__DDR2: 921 return AMDGPU_VRAM_TYPE_DDR2; 922 case MC_SEQ_MISC0__MT__GDDR3: 923 return AMDGPU_VRAM_TYPE_GDDR3; 924 case MC_SEQ_MISC0__MT__GDDR4: 925 return AMDGPU_VRAM_TYPE_GDDR4; 926 case MC_SEQ_MISC0__MT__GDDR5: 927 return AMDGPU_VRAM_TYPE_GDDR5; 928 case MC_SEQ_MISC0__MT__HBM: 929 return AMDGPU_VRAM_TYPE_HBM; 930 case MC_SEQ_MISC0__MT__DDR3: 931 return AMDGPU_VRAM_TYPE_DDR3; 932 default: 933 return AMDGPU_VRAM_TYPE_UNKNOWN; 934 } 935 } 936 937 static int gmc_v7_0_early_init(void *handle) 938 { 939 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 940 941 gmc_v7_0_set_gmc_funcs(adev); 942 gmc_v7_0_set_irq_funcs(adev); 943 944 adev->gmc.shared_aperture_start = 0x2000000000000000ULL; 945 adev->gmc.shared_aperture_end = 946 adev->gmc.shared_aperture_start + (4ULL << 30) - 1; 947 adev->gmc.private_aperture_start = 948 adev->gmc.shared_aperture_end + 1; 949 adev->gmc.private_aperture_end = 950 adev->gmc.private_aperture_start + (4ULL << 30) - 1; 951 952 return 0; 953 } 954 955 static int gmc_v7_0_late_init(void *handle) 956 { 957 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 958 959 amdgpu_bo_late_init(adev); 960 961 if (amdgpu_vm_fault_stop != AMDGPU_VM_FAULT_STOP_ALWAYS) 962 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0); 963 else 964 return 0; 965 } 966 967 static unsigned gmc_v7_0_get_vbios_fb_size(struct amdgpu_device *adev) 968 { 969 u32 d1vga_control = RREG32(mmD1VGA_CONTROL); 970 unsigned size; 971 972 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) { 973 size = 9 * 1024 * 1024; /* reserve 8MB for vga emulator and 1 MB for FB */ 974 } else { 975 u32 viewport = RREG32(mmVIEWPORT_SIZE); 976 size = (REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_HEIGHT) * 977 REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_WIDTH) * 978 4); 979 } 980 /* return 0 if the pre-OS buffer uses up most of vram */ 981 if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024)) 982 return 0; 983 return size; 984 } 985 986 static int gmc_v7_0_sw_init(void *handle) 987 { 988 int r; 989 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 990 991 adev->num_vmhubs = 1; 992 993 if (adev->flags & AMD_IS_APU) { 994 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_UNKNOWN; 995 } else { 996 u32 tmp = RREG32(mmMC_SEQ_MISC0); 997 tmp &= MC_SEQ_MISC0__MT__MASK; 998 adev->gmc.vram_type = gmc_v7_0_convert_vram_type(tmp); 999 } 1000 1001 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_PAGE_INV_FAULT, &adev->gmc.vm_fault); 1002 if (r) 1003 return r; 1004 1005 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_MEM_PROT_FAULT, &adev->gmc.vm_fault); 1006 if (r) 1007 return r; 1008 1009 /* Adjust VM size here. 1010 * Currently set to 4GB ((1 << 20) 4k pages). 1011 * Max GPUVM size for cayman and SI is 40 bits. 1012 */ 1013 amdgpu_vm_adjust_size(adev, 64, 9, 1, 40); 1014 1015 /* Set the internal MC address mask 1016 * This is the max address of the GPU's 1017 * internal address space. 1018 */ 1019 adev->gmc.mc_mask = 0xffffffffffULL; /* 40 bit MC */ 1020 1021 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(40)); 1022 if (r) { 1023 pr_warn("No suitable DMA available\n"); 1024 return r; 1025 } 1026 adev->need_swiotlb = drm_need_swiotlb(40); 1027 1028 r = gmc_v7_0_init_microcode(adev); 1029 if (r) { 1030 DRM_ERROR("Failed to load mc firmware!\n"); 1031 return r; 1032 } 1033 1034 r = gmc_v7_0_mc_init(adev); 1035 if (r) 1036 return r; 1037 1038 adev->gmc.stolen_size = gmc_v7_0_get_vbios_fb_size(adev); 1039 1040 /* Memory manager */ 1041 r = amdgpu_bo_init(adev); 1042 if (r) 1043 return r; 1044 1045 r = gmc_v7_0_gart_init(adev); 1046 if (r) 1047 return r; 1048 1049 /* 1050 * number of VMs 1051 * VMID 0 is reserved for System 1052 * amdgpu graphics/compute will use VMIDs 1-7 1053 * amdkfd will use VMIDs 8-15 1054 */ 1055 adev->vm_manager.first_kfd_vmid = 8; 1056 amdgpu_vm_manager_init(adev); 1057 1058 /* base offset of vram pages */ 1059 if (adev->flags & AMD_IS_APU) { 1060 u64 tmp = RREG32(mmMC_VM_FB_OFFSET); 1061 1062 tmp <<= 22; 1063 adev->vm_manager.vram_base_offset = tmp; 1064 } else { 1065 adev->vm_manager.vram_base_offset = 0; 1066 } 1067 1068 adev->gmc.vm_fault_info = kmalloc(sizeof(struct kfd_vm_fault_info), 1069 GFP_KERNEL); 1070 if (!adev->gmc.vm_fault_info) 1071 return -ENOMEM; 1072 atomic_set(&adev->gmc.vm_fault_info_updated, 0); 1073 1074 return 0; 1075 } 1076 1077 static int gmc_v7_0_sw_fini(void *handle) 1078 { 1079 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1080 1081 amdgpu_gem_force_release(adev); 1082 amdgpu_vm_manager_fini(adev); 1083 kfree(adev->gmc.vm_fault_info); 1084 amdgpu_gart_table_vram_free(adev); 1085 amdgpu_bo_fini(adev); 1086 amdgpu_gart_fini(adev); 1087 release_firmware(adev->gmc.fw); 1088 adev->gmc.fw = NULL; 1089 1090 return 0; 1091 } 1092 1093 static int gmc_v7_0_hw_init(void *handle) 1094 { 1095 int r; 1096 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1097 1098 gmc_v7_0_init_golden_registers(adev); 1099 1100 gmc_v7_0_mc_program(adev); 1101 1102 if (!(adev->flags & AMD_IS_APU)) { 1103 r = gmc_v7_0_mc_load_microcode(adev); 1104 if (r) { 1105 DRM_ERROR("Failed to load MC firmware!\n"); 1106 return r; 1107 } 1108 } 1109 1110 r = gmc_v7_0_gart_enable(adev); 1111 if (r) 1112 return r; 1113 1114 return r; 1115 } 1116 1117 static int gmc_v7_0_hw_fini(void *handle) 1118 { 1119 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1120 1121 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0); 1122 gmc_v7_0_gart_disable(adev); 1123 1124 return 0; 1125 } 1126 1127 static int gmc_v7_0_suspend(void *handle) 1128 { 1129 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1130 1131 gmc_v7_0_hw_fini(adev); 1132 1133 return 0; 1134 } 1135 1136 static int gmc_v7_0_resume(void *handle) 1137 { 1138 int r; 1139 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1140 1141 r = gmc_v7_0_hw_init(adev); 1142 if (r) 1143 return r; 1144 1145 amdgpu_vmid_reset_all(adev); 1146 1147 return 0; 1148 } 1149 1150 static bool gmc_v7_0_is_idle(void *handle) 1151 { 1152 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1153 u32 tmp = RREG32(mmSRBM_STATUS); 1154 1155 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1156 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK | SRBM_STATUS__VMC_BUSY_MASK)) 1157 return false; 1158 1159 return true; 1160 } 1161 1162 static int gmc_v7_0_wait_for_idle(void *handle) 1163 { 1164 unsigned i; 1165 u32 tmp; 1166 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1167 1168 for (i = 0; i < adev->usec_timeout; i++) { 1169 /* read MC_STATUS */ 1170 tmp = RREG32(mmSRBM_STATUS) & (SRBM_STATUS__MCB_BUSY_MASK | 1171 SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1172 SRBM_STATUS__MCC_BUSY_MASK | 1173 SRBM_STATUS__MCD_BUSY_MASK | 1174 SRBM_STATUS__VMC_BUSY_MASK); 1175 if (!tmp) 1176 return 0; 1177 udelay(1); 1178 } 1179 return -ETIMEDOUT; 1180 1181 } 1182 1183 static int gmc_v7_0_soft_reset(void *handle) 1184 { 1185 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1186 u32 srbm_soft_reset = 0; 1187 u32 tmp = RREG32(mmSRBM_STATUS); 1188 1189 if (tmp & SRBM_STATUS__VMC_BUSY_MASK) 1190 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset, 1191 SRBM_SOFT_RESET, SOFT_RESET_VMC, 1); 1192 1193 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1194 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK)) { 1195 if (!(adev->flags & AMD_IS_APU)) 1196 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset, 1197 SRBM_SOFT_RESET, SOFT_RESET_MC, 1); 1198 } 1199 1200 if (srbm_soft_reset) { 1201 gmc_v7_0_mc_stop(adev); 1202 if (gmc_v7_0_wait_for_idle((void *)adev)) { 1203 dev_warn(adev->dev, "Wait for GMC idle timed out !\n"); 1204 } 1205 1206 1207 tmp = RREG32(mmSRBM_SOFT_RESET); 1208 tmp |= srbm_soft_reset; 1209 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp); 1210 WREG32(mmSRBM_SOFT_RESET, tmp); 1211 tmp = RREG32(mmSRBM_SOFT_RESET); 1212 1213 udelay(50); 1214 1215 tmp &= ~srbm_soft_reset; 1216 WREG32(mmSRBM_SOFT_RESET, tmp); 1217 tmp = RREG32(mmSRBM_SOFT_RESET); 1218 1219 /* Wait a little for things to settle down */ 1220 udelay(50); 1221 1222 gmc_v7_0_mc_resume(adev); 1223 udelay(50); 1224 } 1225 1226 return 0; 1227 } 1228 1229 static int gmc_v7_0_vm_fault_interrupt_state(struct amdgpu_device *adev, 1230 struct amdgpu_irq_src *src, 1231 unsigned type, 1232 enum amdgpu_interrupt_state state) 1233 { 1234 u32 tmp; 1235 u32 bits = (VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1236 VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1237 VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1238 VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1239 VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1240 VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK); 1241 1242 switch (state) { 1243 case AMDGPU_IRQ_STATE_DISABLE: 1244 /* system context */ 1245 tmp = RREG32(mmVM_CONTEXT0_CNTL); 1246 tmp &= ~bits; 1247 WREG32(mmVM_CONTEXT0_CNTL, tmp); 1248 /* VMs */ 1249 tmp = RREG32(mmVM_CONTEXT1_CNTL); 1250 tmp &= ~bits; 1251 WREG32(mmVM_CONTEXT1_CNTL, tmp); 1252 break; 1253 case AMDGPU_IRQ_STATE_ENABLE: 1254 /* system context */ 1255 tmp = RREG32(mmVM_CONTEXT0_CNTL); 1256 tmp |= bits; 1257 WREG32(mmVM_CONTEXT0_CNTL, tmp); 1258 /* VMs */ 1259 tmp = RREG32(mmVM_CONTEXT1_CNTL); 1260 tmp |= bits; 1261 WREG32(mmVM_CONTEXT1_CNTL, tmp); 1262 break; 1263 default: 1264 break; 1265 } 1266 1267 return 0; 1268 } 1269 1270 static int gmc_v7_0_process_interrupt(struct amdgpu_device *adev, 1271 struct amdgpu_irq_src *source, 1272 struct amdgpu_iv_entry *entry) 1273 { 1274 u32 addr, status, mc_client, vmid; 1275 1276 addr = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_ADDR); 1277 status = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_STATUS); 1278 mc_client = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_MCCLIENT); 1279 /* reset addr and status */ 1280 WREG32_P(mmVM_CONTEXT1_CNTL2, 1, ~1); 1281 1282 if (!addr && !status) 1283 return 0; 1284 1285 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_FIRST) 1286 gmc_v7_0_set_fault_enable_default(adev, false); 1287 1288 if (printk_ratelimit()) { 1289 dev_err(adev->dev, "GPU fault detected: %d 0x%08x\n", 1290 entry->src_id, entry->src_data[0]); 1291 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n", 1292 addr); 1293 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n", 1294 status); 1295 gmc_v7_0_vm_decode_fault(adev, status, addr, mc_client, 1296 entry->pasid); 1297 } 1298 1299 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 1300 VMID); 1301 if (amdgpu_amdkfd_is_kfd_vmid(adev, vmid) 1302 && !atomic_read(&adev->gmc.vm_fault_info_updated)) { 1303 struct kfd_vm_fault_info *info = adev->gmc.vm_fault_info; 1304 u32 protections = REG_GET_FIELD(status, 1305 VM_CONTEXT1_PROTECTION_FAULT_STATUS, 1306 PROTECTIONS); 1307 1308 info->vmid = vmid; 1309 info->mc_id = REG_GET_FIELD(status, 1310 VM_CONTEXT1_PROTECTION_FAULT_STATUS, 1311 MEMORY_CLIENT_ID); 1312 info->status = status; 1313 info->page_addr = addr; 1314 info->prot_valid = protections & 0x7 ? true : false; 1315 info->prot_read = protections & 0x8 ? true : false; 1316 info->prot_write = protections & 0x10 ? true : false; 1317 info->prot_exec = protections & 0x20 ? true : false; 1318 mb(); 1319 atomic_set(&adev->gmc.vm_fault_info_updated, 1); 1320 } 1321 1322 return 0; 1323 } 1324 1325 static int gmc_v7_0_set_clockgating_state(void *handle, 1326 enum amd_clockgating_state state) 1327 { 1328 bool gate = false; 1329 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1330 1331 if (state == AMD_CG_STATE_GATE) 1332 gate = true; 1333 1334 if (!(adev->flags & AMD_IS_APU)) { 1335 gmc_v7_0_enable_mc_mgcg(adev, gate); 1336 gmc_v7_0_enable_mc_ls(adev, gate); 1337 } 1338 gmc_v7_0_enable_bif_mgls(adev, gate); 1339 gmc_v7_0_enable_hdp_mgcg(adev, gate); 1340 gmc_v7_0_enable_hdp_ls(adev, gate); 1341 1342 return 0; 1343 } 1344 1345 static int gmc_v7_0_set_powergating_state(void *handle, 1346 enum amd_powergating_state state) 1347 { 1348 return 0; 1349 } 1350 1351 static const struct amd_ip_funcs gmc_v7_0_ip_funcs = { 1352 .name = "gmc_v7_0", 1353 .early_init = gmc_v7_0_early_init, 1354 .late_init = gmc_v7_0_late_init, 1355 .sw_init = gmc_v7_0_sw_init, 1356 .sw_fini = gmc_v7_0_sw_fini, 1357 .hw_init = gmc_v7_0_hw_init, 1358 .hw_fini = gmc_v7_0_hw_fini, 1359 .suspend = gmc_v7_0_suspend, 1360 .resume = gmc_v7_0_resume, 1361 .is_idle = gmc_v7_0_is_idle, 1362 .wait_for_idle = gmc_v7_0_wait_for_idle, 1363 .soft_reset = gmc_v7_0_soft_reset, 1364 .set_clockgating_state = gmc_v7_0_set_clockgating_state, 1365 .set_powergating_state = gmc_v7_0_set_powergating_state, 1366 }; 1367 1368 static const struct amdgpu_gmc_funcs gmc_v7_0_gmc_funcs = { 1369 .flush_gpu_tlb = gmc_v7_0_flush_gpu_tlb, 1370 .flush_gpu_tlb_pasid = gmc_v7_0_flush_gpu_tlb_pasid, 1371 .emit_flush_gpu_tlb = gmc_v7_0_emit_flush_gpu_tlb, 1372 .emit_pasid_mapping = gmc_v7_0_emit_pasid_mapping, 1373 .set_prt = gmc_v7_0_set_prt, 1374 .get_vm_pde = gmc_v7_0_get_vm_pde, 1375 .get_vm_pte = gmc_v7_0_get_vm_pte 1376 }; 1377 1378 static const struct amdgpu_irq_src_funcs gmc_v7_0_irq_funcs = { 1379 .set = gmc_v7_0_vm_fault_interrupt_state, 1380 .process = gmc_v7_0_process_interrupt, 1381 }; 1382 1383 static void gmc_v7_0_set_gmc_funcs(struct amdgpu_device *adev) 1384 { 1385 adev->gmc.gmc_funcs = &gmc_v7_0_gmc_funcs; 1386 } 1387 1388 static void gmc_v7_0_set_irq_funcs(struct amdgpu_device *adev) 1389 { 1390 adev->gmc.vm_fault.num_types = 1; 1391 adev->gmc.vm_fault.funcs = &gmc_v7_0_irq_funcs; 1392 } 1393 1394 const struct amdgpu_ip_block_version gmc_v7_0_ip_block = 1395 { 1396 .type = AMD_IP_BLOCK_TYPE_GMC, 1397 .major = 7, 1398 .minor = 0, 1399 .rev = 0, 1400 .funcs = &gmc_v7_0_ip_funcs, 1401 }; 1402 1403 const struct amdgpu_ip_block_version gmc_v7_4_ip_block = 1404 { 1405 .type = AMD_IP_BLOCK_TYPE_GMC, 1406 .major = 7, 1407 .minor = 4, 1408 .rev = 0, 1409 .funcs = &gmc_v7_0_ip_funcs, 1410 }; 1411