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 = amdgpu_ucode_request(adev, &adev->gmc.fw, fw_name); 160 if (err) { 161 pr_err("cik_mc: Failed to load firmware \"%s\"\n", fw_name); 162 amdgpu_ucode_release(&adev->gmc.fw); 163 } 164 return err; 165 } 166 167 /** 168 * gmc_v7_0_mc_load_microcode - load MC ucode into the hw 169 * 170 * @adev: amdgpu_device pointer 171 * 172 * Load the GDDR MC ucode into the hw (CIK). 173 * Returns 0 on success, error on failure. 174 */ 175 static int gmc_v7_0_mc_load_microcode(struct amdgpu_device *adev) 176 { 177 const struct mc_firmware_header_v1_0 *hdr; 178 const __le32 *fw_data = NULL; 179 const __le32 *io_mc_regs = NULL; 180 u32 running; 181 int i, ucode_size, regs_size; 182 183 if (!adev->gmc.fw) 184 return -EINVAL; 185 186 hdr = (const struct mc_firmware_header_v1_0 *)adev->gmc.fw->data; 187 amdgpu_ucode_print_mc_hdr(&hdr->header); 188 189 adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version); 190 regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2); 191 io_mc_regs = (const __le32 *) 192 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes)); 193 ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4; 194 fw_data = (const __le32 *) 195 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes)); 196 197 running = REG_GET_FIELD(RREG32(mmMC_SEQ_SUP_CNTL), MC_SEQ_SUP_CNTL, RUN); 198 199 if (running == 0) { 200 /* reset the engine and set to writable */ 201 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008); 202 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000010); 203 204 /* load mc io regs */ 205 for (i = 0; i < regs_size; i++) { 206 WREG32(mmMC_SEQ_IO_DEBUG_INDEX, le32_to_cpup(io_mc_regs++)); 207 WREG32(mmMC_SEQ_IO_DEBUG_DATA, le32_to_cpup(io_mc_regs++)); 208 } 209 /* load the MC ucode */ 210 for (i = 0; i < ucode_size; i++) 211 WREG32(mmMC_SEQ_SUP_PGM, le32_to_cpup(fw_data++)); 212 213 /* put the engine back into the active state */ 214 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008); 215 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000004); 216 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000001); 217 218 /* wait for training to complete */ 219 for (i = 0; i < adev->usec_timeout; i++) { 220 if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL), 221 MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D0)) 222 break; 223 udelay(1); 224 } 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_D1)) 228 break; 229 udelay(1); 230 } 231 } 232 233 return 0; 234 } 235 236 static void gmc_v7_0_vram_gtt_location(struct amdgpu_device *adev, 237 struct amdgpu_gmc *mc) 238 { 239 u64 base = RREG32(mmMC_VM_FB_LOCATION) & 0xFFFF; 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 } 296 297 WREG32(mmBIF_FB_EN, BIF_FB_EN__FB_READ_EN_MASK | BIF_FB_EN__FB_WRITE_EN_MASK); 298 299 tmp = RREG32(mmHDP_MISC_CNTL); 300 tmp = REG_SET_FIELD(tmp, HDP_MISC_CNTL, FLUSH_INVALIDATE_CACHE, 0); 301 WREG32(mmHDP_MISC_CNTL, tmp); 302 303 tmp = RREG32(mmHDP_HOST_PATH_CNTL); 304 WREG32(mmHDP_HOST_PATH_CNTL, tmp); 305 } 306 307 /** 308 * gmc_v7_0_mc_init - initialize the memory controller driver params 309 * 310 * @adev: amdgpu_device pointer 311 * 312 * Look up the amount of vram, vram width, and decide how to place 313 * vram and gart within the GPU's physical address space (CIK). 314 * Returns 0 for success. 315 */ 316 static int gmc_v7_0_mc_init(struct amdgpu_device *adev) 317 { 318 int r; 319 320 adev->gmc.vram_width = amdgpu_atombios_get_vram_width(adev); 321 if (!adev->gmc.vram_width) { 322 u32 tmp; 323 int chansize, numchan; 324 325 /* Get VRAM informations */ 326 tmp = RREG32(mmMC_ARB_RAMCFG); 327 if (REG_GET_FIELD(tmp, MC_ARB_RAMCFG, CHANSIZE)) { 328 chansize = 64; 329 } else { 330 chansize = 32; 331 } 332 tmp = RREG32(mmMC_SHARED_CHMAP); 333 switch (REG_GET_FIELD(tmp, MC_SHARED_CHMAP, NOOFCHAN)) { 334 case 0: 335 default: 336 numchan = 1; 337 break; 338 case 1: 339 numchan = 2; 340 break; 341 case 2: 342 numchan = 4; 343 break; 344 case 3: 345 numchan = 8; 346 break; 347 case 4: 348 numchan = 3; 349 break; 350 case 5: 351 numchan = 6; 352 break; 353 case 6: 354 numchan = 10; 355 break; 356 case 7: 357 numchan = 12; 358 break; 359 case 8: 360 numchan = 16; 361 break; 362 } 363 adev->gmc.vram_width = numchan * chansize; 364 } 365 /* size in MB on si */ 366 adev->gmc.mc_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL; 367 adev->gmc.real_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL; 368 369 if (!(adev->flags & AMD_IS_APU)) { 370 r = amdgpu_device_resize_fb_bar(adev); 371 if (r) 372 return r; 373 } 374 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0); 375 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0); 376 377 #ifdef CONFIG_X86_64 378 if ((adev->flags & AMD_IS_APU) && 379 adev->gmc.real_vram_size > adev->gmc.aper_size && 380 !amdgpu_passthrough(adev)) { 381 adev->gmc.aper_base = ((u64)RREG32(mmMC_VM_FB_OFFSET)) << 22; 382 adev->gmc.aper_size = adev->gmc.real_vram_size; 383 } 384 #endif 385 386 adev->gmc.visible_vram_size = adev->gmc.aper_size; 387 388 /* set the gart size */ 389 if (amdgpu_gart_size == -1) { 390 switch (adev->asic_type) { 391 case CHIP_TOPAZ: /* no MM engines */ 392 default: 393 adev->gmc.gart_size = 256ULL << 20; 394 break; 395 #ifdef CONFIG_DRM_AMDGPU_CIK 396 case CHIP_BONAIRE: /* UVD, VCE do not support GPUVM */ 397 case CHIP_HAWAII: /* UVD, VCE do not support GPUVM */ 398 case CHIP_KAVERI: /* UVD, VCE do not support GPUVM */ 399 case CHIP_KABINI: /* UVD, VCE do not support GPUVM */ 400 case CHIP_MULLINS: /* UVD, VCE do not support GPUVM */ 401 adev->gmc.gart_size = 1024ULL << 20; 402 break; 403 #endif 404 } 405 } else { 406 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20; 407 } 408 409 adev->gmc.gart_size += adev->pm.smu_prv_buffer_size; 410 gmc_v7_0_vram_gtt_location(adev, &adev->gmc); 411 412 return 0; 413 } 414 415 /** 416 * gmc_v7_0_flush_gpu_tlb_pasid - tlb flush via pasid 417 * 418 * @adev: amdgpu_device pointer 419 * @pasid: pasid to be flush 420 * @flush_type: type of flush 421 * @all_hub: flush all hubs 422 * 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 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 vmid, 491 unsigned 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)(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 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; 944 945 return 0; 946 } 947 948 static int gmc_v7_0_late_init(void *handle) 949 { 950 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 951 952 if (amdgpu_vm_fault_stop != AMDGPU_VM_FAULT_STOP_ALWAYS) 953 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0); 954 else 955 return 0; 956 } 957 958 static unsigned gmc_v7_0_get_vbios_fb_size(struct amdgpu_device *adev) 959 { 960 u32 d1vga_control = RREG32(mmD1VGA_CONTROL); 961 unsigned size; 962 963 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) { 964 size = AMDGPU_VBIOS_VGA_ALLOCATION; 965 } else { 966 u32 viewport = RREG32(mmVIEWPORT_SIZE); 967 size = (REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_HEIGHT) * 968 REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_WIDTH) * 969 4); 970 } 971 972 return size; 973 } 974 975 static int gmc_v7_0_sw_init(void *handle) 976 { 977 int r; 978 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 979 980 set_bit(AMDGPU_GFXHUB(0), adev->vmhubs_mask); 981 982 if (adev->flags & AMD_IS_APU) { 983 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_UNKNOWN; 984 } else { 985 u32 tmp = RREG32(mmMC_SEQ_MISC0); 986 tmp &= MC_SEQ_MISC0__MT__MASK; 987 adev->gmc.vram_type = gmc_v7_0_convert_vram_type(tmp); 988 } 989 990 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_PAGE_INV_FAULT, &adev->gmc.vm_fault); 991 if (r) 992 return r; 993 994 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_MEM_PROT_FAULT, &adev->gmc.vm_fault); 995 if (r) 996 return r; 997 998 /* Adjust VM size here. 999 * Currently set to 4GB ((1 << 20) 4k pages). 1000 * Max GPUVM size for cayman and SI is 40 bits. 1001 */ 1002 amdgpu_vm_adjust_size(adev, 64, 9, 1, 40); 1003 1004 /* Set the internal MC address mask 1005 * This is the max address of the GPU's 1006 * internal address space. 1007 */ 1008 adev->gmc.mc_mask = 0xffffffffffULL; /* 40 bit MC */ 1009 1010 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(40)); 1011 if (r) { 1012 pr_warn("No suitable DMA available\n"); 1013 return r; 1014 } 1015 adev->need_swiotlb = drm_need_swiotlb(40); 1016 1017 r = gmc_v7_0_init_microcode(adev); 1018 if (r) { 1019 DRM_ERROR("Failed to load mc firmware!\n"); 1020 return r; 1021 } 1022 1023 r = gmc_v7_0_mc_init(adev); 1024 if (r) 1025 return r; 1026 1027 amdgpu_gmc_get_vbios_allocations(adev); 1028 1029 /* Memory manager */ 1030 r = amdgpu_bo_init(adev); 1031 if (r) 1032 return r; 1033 1034 r = gmc_v7_0_gart_init(adev); 1035 if (r) 1036 return r; 1037 1038 /* 1039 * number of VMs 1040 * VMID 0 is reserved for System 1041 * amdgpu graphics/compute will use VMIDs 1-7 1042 * amdkfd will use VMIDs 8-15 1043 */ 1044 adev->vm_manager.first_kfd_vmid = 8; 1045 amdgpu_vm_manager_init(adev); 1046 1047 /* base offset of vram pages */ 1048 if (adev->flags & AMD_IS_APU) { 1049 u64 tmp = RREG32(mmMC_VM_FB_OFFSET); 1050 1051 tmp <<= 22; 1052 adev->vm_manager.vram_base_offset = tmp; 1053 } else { 1054 adev->vm_manager.vram_base_offset = 0; 1055 } 1056 1057 adev->gmc.vm_fault_info = kmalloc(sizeof(struct kfd_vm_fault_info), 1058 GFP_KERNEL); 1059 if (!adev->gmc.vm_fault_info) 1060 return -ENOMEM; 1061 atomic_set(&adev->gmc.vm_fault_info_updated, 0); 1062 1063 return 0; 1064 } 1065 1066 static int gmc_v7_0_sw_fini(void *handle) 1067 { 1068 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1069 1070 amdgpu_gem_force_release(adev); 1071 amdgpu_vm_manager_fini(adev); 1072 kfree(adev->gmc.vm_fault_info); 1073 amdgpu_gart_table_vram_free(adev); 1074 amdgpu_bo_fini(adev); 1075 amdgpu_ucode_release(&adev->gmc.fw); 1076 1077 return 0; 1078 } 1079 1080 static int gmc_v7_0_hw_init(void *handle) 1081 { 1082 int r; 1083 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1084 1085 gmc_v7_0_init_golden_registers(adev); 1086 1087 gmc_v7_0_mc_program(adev); 1088 1089 if (!(adev->flags & AMD_IS_APU)) { 1090 r = gmc_v7_0_mc_load_microcode(adev); 1091 if (r) { 1092 DRM_ERROR("Failed to load MC firmware!\n"); 1093 return r; 1094 } 1095 } 1096 1097 r = gmc_v7_0_gart_enable(adev); 1098 if (r) 1099 return r; 1100 1101 if (amdgpu_emu_mode == 1) 1102 return amdgpu_gmc_vram_checking(adev); 1103 else 1104 return r; 1105 } 1106 1107 static int gmc_v7_0_hw_fini(void *handle) 1108 { 1109 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1110 1111 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0); 1112 gmc_v7_0_gart_disable(adev); 1113 1114 return 0; 1115 } 1116 1117 static int gmc_v7_0_suspend(void *handle) 1118 { 1119 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1120 1121 gmc_v7_0_hw_fini(adev); 1122 1123 return 0; 1124 } 1125 1126 static int gmc_v7_0_resume(void *handle) 1127 { 1128 int r; 1129 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1130 1131 r = gmc_v7_0_hw_init(adev); 1132 if (r) 1133 return r; 1134 1135 amdgpu_vmid_reset_all(adev); 1136 1137 return 0; 1138 } 1139 1140 static bool gmc_v7_0_is_idle(void *handle) 1141 { 1142 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1143 u32 tmp = RREG32(mmSRBM_STATUS); 1144 1145 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1146 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK | SRBM_STATUS__VMC_BUSY_MASK)) 1147 return false; 1148 1149 return true; 1150 } 1151 1152 static int gmc_v7_0_wait_for_idle(void *handle) 1153 { 1154 unsigned i; 1155 u32 tmp; 1156 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1157 1158 for (i = 0; i < adev->usec_timeout; i++) { 1159 /* read MC_STATUS */ 1160 tmp = RREG32(mmSRBM_STATUS) & (SRBM_STATUS__MCB_BUSY_MASK | 1161 SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1162 SRBM_STATUS__MCC_BUSY_MASK | 1163 SRBM_STATUS__MCD_BUSY_MASK | 1164 SRBM_STATUS__VMC_BUSY_MASK); 1165 if (!tmp) 1166 return 0; 1167 udelay(1); 1168 } 1169 return -ETIMEDOUT; 1170 1171 } 1172 1173 static int gmc_v7_0_soft_reset(void *handle) 1174 { 1175 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1176 u32 srbm_soft_reset = 0; 1177 u32 tmp = RREG32(mmSRBM_STATUS); 1178 1179 if (tmp & SRBM_STATUS__VMC_BUSY_MASK) 1180 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset, 1181 SRBM_SOFT_RESET, SOFT_RESET_VMC, 1); 1182 1183 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1184 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK)) { 1185 if (!(adev->flags & AMD_IS_APU)) 1186 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset, 1187 SRBM_SOFT_RESET, SOFT_RESET_MC, 1); 1188 } 1189 1190 if (srbm_soft_reset) { 1191 gmc_v7_0_mc_stop(adev); 1192 if (gmc_v7_0_wait_for_idle((void *)adev)) { 1193 dev_warn(adev->dev, "Wait for GMC idle timed out !\n"); 1194 } 1195 1196 1197 tmp = RREG32(mmSRBM_SOFT_RESET); 1198 tmp |= srbm_soft_reset; 1199 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp); 1200 WREG32(mmSRBM_SOFT_RESET, tmp); 1201 tmp = RREG32(mmSRBM_SOFT_RESET); 1202 1203 udelay(50); 1204 1205 tmp &= ~srbm_soft_reset; 1206 WREG32(mmSRBM_SOFT_RESET, tmp); 1207 tmp = RREG32(mmSRBM_SOFT_RESET); 1208 1209 /* Wait a little for things to settle down */ 1210 udelay(50); 1211 1212 gmc_v7_0_mc_resume(adev); 1213 udelay(50); 1214 } 1215 1216 return 0; 1217 } 1218 1219 static int gmc_v7_0_vm_fault_interrupt_state(struct amdgpu_device *adev, 1220 struct amdgpu_irq_src *src, 1221 unsigned type, 1222 enum amdgpu_interrupt_state state) 1223 { 1224 u32 tmp; 1225 u32 bits = (VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1226 VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1227 VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1228 VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1229 VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1230 VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK); 1231 1232 switch (state) { 1233 case AMDGPU_IRQ_STATE_DISABLE: 1234 /* system context */ 1235 tmp = RREG32(mmVM_CONTEXT0_CNTL); 1236 tmp &= ~bits; 1237 WREG32(mmVM_CONTEXT0_CNTL, tmp); 1238 /* VMs */ 1239 tmp = RREG32(mmVM_CONTEXT1_CNTL); 1240 tmp &= ~bits; 1241 WREG32(mmVM_CONTEXT1_CNTL, tmp); 1242 break; 1243 case AMDGPU_IRQ_STATE_ENABLE: 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 default: 1254 break; 1255 } 1256 1257 return 0; 1258 } 1259 1260 static int gmc_v7_0_process_interrupt(struct amdgpu_device *adev, 1261 struct amdgpu_irq_src *source, 1262 struct amdgpu_iv_entry *entry) 1263 { 1264 u32 addr, status, mc_client, vmid; 1265 1266 addr = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_ADDR); 1267 status = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_STATUS); 1268 mc_client = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_MCCLIENT); 1269 /* reset addr and status */ 1270 WREG32_P(mmVM_CONTEXT1_CNTL2, 1, ~1); 1271 1272 if (!addr && !status) 1273 return 0; 1274 1275 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_FIRST) 1276 gmc_v7_0_set_fault_enable_default(adev, false); 1277 1278 if (printk_ratelimit()) { 1279 dev_err(adev->dev, "GPU fault detected: %d 0x%08x\n", 1280 entry->src_id, entry->src_data[0]); 1281 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n", 1282 addr); 1283 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n", 1284 status); 1285 gmc_v7_0_vm_decode_fault(adev, status, addr, mc_client, 1286 entry->pasid); 1287 } 1288 1289 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 1290 VMID); 1291 if (amdgpu_amdkfd_is_kfd_vmid(adev, vmid) 1292 && !atomic_read(&adev->gmc.vm_fault_info_updated)) { 1293 struct kfd_vm_fault_info *info = adev->gmc.vm_fault_info; 1294 u32 protections = REG_GET_FIELD(status, 1295 VM_CONTEXT1_PROTECTION_FAULT_STATUS, 1296 PROTECTIONS); 1297 1298 info->vmid = vmid; 1299 info->mc_id = REG_GET_FIELD(status, 1300 VM_CONTEXT1_PROTECTION_FAULT_STATUS, 1301 MEMORY_CLIENT_ID); 1302 info->status = status; 1303 info->page_addr = addr; 1304 info->prot_valid = protections & 0x7 ? true : false; 1305 info->prot_read = protections & 0x8 ? true : false; 1306 info->prot_write = protections & 0x10 ? true : false; 1307 info->prot_exec = protections & 0x20 ? true : false; 1308 mb(); 1309 atomic_set(&adev->gmc.vm_fault_info_updated, 1); 1310 } 1311 1312 return 0; 1313 } 1314 1315 static int gmc_v7_0_set_clockgating_state(void *handle, 1316 enum amd_clockgating_state state) 1317 { 1318 bool gate = false; 1319 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1320 1321 if (state == AMD_CG_STATE_GATE) 1322 gate = true; 1323 1324 if (!(adev->flags & AMD_IS_APU)) { 1325 gmc_v7_0_enable_mc_mgcg(adev, gate); 1326 gmc_v7_0_enable_mc_ls(adev, gate); 1327 } 1328 gmc_v7_0_enable_bif_mgls(adev, gate); 1329 gmc_v7_0_enable_hdp_mgcg(adev, gate); 1330 gmc_v7_0_enable_hdp_ls(adev, gate); 1331 1332 return 0; 1333 } 1334 1335 static int gmc_v7_0_set_powergating_state(void *handle, 1336 enum amd_powergating_state state) 1337 { 1338 return 0; 1339 } 1340 1341 static const struct amd_ip_funcs gmc_v7_0_ip_funcs = { 1342 .name = "gmc_v7_0", 1343 .early_init = gmc_v7_0_early_init, 1344 .late_init = gmc_v7_0_late_init, 1345 .sw_init = gmc_v7_0_sw_init, 1346 .sw_fini = gmc_v7_0_sw_fini, 1347 .hw_init = gmc_v7_0_hw_init, 1348 .hw_fini = gmc_v7_0_hw_fini, 1349 .suspend = gmc_v7_0_suspend, 1350 .resume = gmc_v7_0_resume, 1351 .is_idle = gmc_v7_0_is_idle, 1352 .wait_for_idle = gmc_v7_0_wait_for_idle, 1353 .soft_reset = gmc_v7_0_soft_reset, 1354 .set_clockgating_state = gmc_v7_0_set_clockgating_state, 1355 .set_powergating_state = gmc_v7_0_set_powergating_state, 1356 }; 1357 1358 static const struct amdgpu_gmc_funcs gmc_v7_0_gmc_funcs = { 1359 .flush_gpu_tlb = gmc_v7_0_flush_gpu_tlb, 1360 .flush_gpu_tlb_pasid = gmc_v7_0_flush_gpu_tlb_pasid, 1361 .emit_flush_gpu_tlb = gmc_v7_0_emit_flush_gpu_tlb, 1362 .emit_pasid_mapping = gmc_v7_0_emit_pasid_mapping, 1363 .set_prt = gmc_v7_0_set_prt, 1364 .get_vm_pde = gmc_v7_0_get_vm_pde, 1365 .get_vm_pte = gmc_v7_0_get_vm_pte, 1366 .get_vbios_fb_size = gmc_v7_0_get_vbios_fb_size, 1367 }; 1368 1369 static const struct amdgpu_irq_src_funcs gmc_v7_0_irq_funcs = { 1370 .set = gmc_v7_0_vm_fault_interrupt_state, 1371 .process = gmc_v7_0_process_interrupt, 1372 }; 1373 1374 static void gmc_v7_0_set_gmc_funcs(struct amdgpu_device *adev) 1375 { 1376 adev->gmc.gmc_funcs = &gmc_v7_0_gmc_funcs; 1377 } 1378 1379 static void gmc_v7_0_set_irq_funcs(struct amdgpu_device *adev) 1380 { 1381 adev->gmc.vm_fault.num_types = 1; 1382 adev->gmc.vm_fault.funcs = &gmc_v7_0_irq_funcs; 1383 } 1384 1385 const struct amdgpu_ip_block_version gmc_v7_0_ip_block = 1386 { 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 { 1396 .type = AMD_IP_BLOCK_TYPE_GMC, 1397 .major = 7, 1398 .minor = 4, 1399 .rev = 0, 1400 .funcs = &gmc_v7_0_ip_funcs, 1401 }; 1402