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 #include <linux/firmware.h> 24 #include "drmP.h" 25 #include "amdgpu.h" 26 #include "gmc_v8_0.h" 27 #include "amdgpu_ucode.h" 28 29 #include "gmc/gmc_8_1_d.h" 30 #include "gmc/gmc_8_1_sh_mask.h" 31 32 #include "bif/bif_5_0_d.h" 33 #include "bif/bif_5_0_sh_mask.h" 34 35 #include "oss/oss_3_0_d.h" 36 #include "oss/oss_3_0_sh_mask.h" 37 38 #include "vid.h" 39 #include "vi.h" 40 41 42 static void gmc_v8_0_set_gart_funcs(struct amdgpu_device *adev); 43 static void gmc_v8_0_set_irq_funcs(struct amdgpu_device *adev); 44 static int gmc_v8_0_wait_for_idle(void *handle); 45 46 MODULE_FIRMWARE("amdgpu/tonga_mc.bin"); 47 MODULE_FIRMWARE("amdgpu/polaris11_mc.bin"); 48 MODULE_FIRMWARE("amdgpu/polaris10_mc.bin"); 49 50 static const u32 golden_settings_tonga_a11[] = 51 { 52 mmMC_ARB_WTM_GRPWT_RD, 0x00000003, 0x00000000, 53 mmMC_HUB_RDREQ_DMIF_LIMIT, 0x0000007f, 0x00000028, 54 mmMC_HUB_WDP_UMC, 0x00007fb6, 0x00000991, 55 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff, 56 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff, 57 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff, 58 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff, 59 }; 60 61 static const u32 tonga_mgcg_cgcg_init[] = 62 { 63 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104 64 }; 65 66 static const u32 golden_settings_fiji_a10[] = 67 { 68 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff, 69 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff, 70 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff, 71 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff, 72 }; 73 74 static const u32 fiji_mgcg_cgcg_init[] = 75 { 76 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104 77 }; 78 79 static const u32 golden_settings_polaris11_a11[] = 80 { 81 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff, 82 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff, 83 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff, 84 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff 85 }; 86 87 static const u32 golden_settings_polaris10_a11[] = 88 { 89 mmMC_ARB_WTM_GRPWT_RD, 0x00000003, 0x00000000, 90 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff, 91 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff, 92 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff, 93 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff 94 }; 95 96 static const u32 cz_mgcg_cgcg_init[] = 97 { 98 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104 99 }; 100 101 static const u32 stoney_mgcg_cgcg_init[] = 102 { 103 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104 104 }; 105 106 static const u32 golden_settings_stoney_common[] = 107 { 108 mmMC_HUB_RDREQ_UVD, MC_HUB_RDREQ_UVD__PRESCALE_MASK, 0x00000004, 109 mmMC_RD_GRP_OTH, MC_RD_GRP_OTH__UVD_MASK, 0x00600000 110 }; 111 112 static void gmc_v8_0_init_golden_registers(struct amdgpu_device *adev) 113 { 114 switch (adev->asic_type) { 115 case CHIP_FIJI: 116 amdgpu_program_register_sequence(adev, 117 fiji_mgcg_cgcg_init, 118 (const u32)ARRAY_SIZE(fiji_mgcg_cgcg_init)); 119 amdgpu_program_register_sequence(adev, 120 golden_settings_fiji_a10, 121 (const u32)ARRAY_SIZE(golden_settings_fiji_a10)); 122 break; 123 case CHIP_TONGA: 124 amdgpu_program_register_sequence(adev, 125 tonga_mgcg_cgcg_init, 126 (const u32)ARRAY_SIZE(tonga_mgcg_cgcg_init)); 127 amdgpu_program_register_sequence(adev, 128 golden_settings_tonga_a11, 129 (const u32)ARRAY_SIZE(golden_settings_tonga_a11)); 130 break; 131 case CHIP_POLARIS11: 132 amdgpu_program_register_sequence(adev, 133 golden_settings_polaris11_a11, 134 (const u32)ARRAY_SIZE(golden_settings_polaris11_a11)); 135 break; 136 case CHIP_POLARIS10: 137 amdgpu_program_register_sequence(adev, 138 golden_settings_polaris10_a11, 139 (const u32)ARRAY_SIZE(golden_settings_polaris10_a11)); 140 break; 141 case CHIP_CARRIZO: 142 amdgpu_program_register_sequence(adev, 143 cz_mgcg_cgcg_init, 144 (const u32)ARRAY_SIZE(cz_mgcg_cgcg_init)); 145 break; 146 case CHIP_STONEY: 147 amdgpu_program_register_sequence(adev, 148 stoney_mgcg_cgcg_init, 149 (const u32)ARRAY_SIZE(stoney_mgcg_cgcg_init)); 150 amdgpu_program_register_sequence(adev, 151 golden_settings_stoney_common, 152 (const u32)ARRAY_SIZE(golden_settings_stoney_common)); 153 break; 154 default: 155 break; 156 } 157 } 158 159 static void gmc_v8_0_mc_stop(struct amdgpu_device *adev, 160 struct amdgpu_mode_mc_save *save) 161 { 162 u32 blackout; 163 164 if (adev->mode_info.num_crtc) 165 amdgpu_display_stop_mc_access(adev, save); 166 167 gmc_v8_0_wait_for_idle(adev); 168 169 blackout = RREG32(mmMC_SHARED_BLACKOUT_CNTL); 170 if (REG_GET_FIELD(blackout, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE) != 1) { 171 /* Block CPU access */ 172 WREG32(mmBIF_FB_EN, 0); 173 /* blackout the MC */ 174 blackout = REG_SET_FIELD(blackout, 175 MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 1); 176 WREG32(mmMC_SHARED_BLACKOUT_CNTL, blackout); 177 } 178 /* wait for the MC to settle */ 179 udelay(100); 180 } 181 182 static void gmc_v8_0_mc_resume(struct amdgpu_device *adev, 183 struct amdgpu_mode_mc_save *save) 184 { 185 u32 tmp; 186 187 /* unblackout the MC */ 188 tmp = RREG32(mmMC_SHARED_BLACKOUT_CNTL); 189 tmp = REG_SET_FIELD(tmp, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 0); 190 WREG32(mmMC_SHARED_BLACKOUT_CNTL, tmp); 191 /* allow CPU access */ 192 tmp = REG_SET_FIELD(0, BIF_FB_EN, FB_READ_EN, 1); 193 tmp = REG_SET_FIELD(tmp, BIF_FB_EN, FB_WRITE_EN, 1); 194 WREG32(mmBIF_FB_EN, tmp); 195 196 if (adev->mode_info.num_crtc) 197 amdgpu_display_resume_mc_access(adev, save); 198 } 199 200 /** 201 * gmc_v8_0_init_microcode - load ucode images from disk 202 * 203 * @adev: amdgpu_device pointer 204 * 205 * Use the firmware interface to load the ucode images into 206 * the driver (not loaded into hw). 207 * Returns 0 on success, error on failure. 208 */ 209 static int gmc_v8_0_init_microcode(struct amdgpu_device *adev) 210 { 211 const char *chip_name; 212 char fw_name[30]; 213 int err; 214 215 DRM_DEBUG("\n"); 216 217 switch (adev->asic_type) { 218 case CHIP_TONGA: 219 chip_name = "tonga"; 220 break; 221 case CHIP_POLARIS11: 222 chip_name = "polaris11"; 223 break; 224 case CHIP_POLARIS10: 225 chip_name = "polaris10"; 226 break; 227 case CHIP_FIJI: 228 case CHIP_CARRIZO: 229 case CHIP_STONEY: 230 return 0; 231 default: BUG(); 232 } 233 234 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mc.bin", chip_name); 235 err = request_firmware(&adev->mc.fw, fw_name, adev->dev); 236 if (err) 237 goto out; 238 err = amdgpu_ucode_validate(adev->mc.fw); 239 240 out: 241 if (err) { 242 printk(KERN_ERR 243 "mc: Failed to load firmware \"%s\"\n", 244 fw_name); 245 release_firmware(adev->mc.fw); 246 adev->mc.fw = NULL; 247 } 248 return err; 249 } 250 251 /** 252 * gmc_v8_0_mc_load_microcode - load MC ucode into the hw 253 * 254 * @adev: amdgpu_device pointer 255 * 256 * Load the GDDR MC ucode into the hw (CIK). 257 * Returns 0 on success, error on failure. 258 */ 259 static int gmc_v8_0_mc_load_microcode(struct amdgpu_device *adev) 260 { 261 const struct mc_firmware_header_v1_0 *hdr; 262 const __le32 *fw_data = NULL; 263 const __le32 *io_mc_regs = NULL; 264 u32 running; 265 int i, ucode_size, regs_size; 266 267 if (!adev->mc.fw) 268 return -EINVAL; 269 270 /* Skip MC ucode loading on SR-IOV capable boards. 271 * vbios does this for us in asic_init in that case. 272 * Skip MC ucode loading on VF, because hypervisor will do that 273 * for this adaptor. 274 */ 275 if (amdgpu_sriov_bios(adev)) 276 return 0; 277 278 hdr = (const struct mc_firmware_header_v1_0 *)adev->mc.fw->data; 279 amdgpu_ucode_print_mc_hdr(&hdr->header); 280 281 adev->mc.fw_version = le32_to_cpu(hdr->header.ucode_version); 282 regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2); 283 io_mc_regs = (const __le32 *) 284 (adev->mc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes)); 285 ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4; 286 fw_data = (const __le32 *) 287 (adev->mc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes)); 288 289 running = REG_GET_FIELD(RREG32(mmMC_SEQ_SUP_CNTL), MC_SEQ_SUP_CNTL, RUN); 290 291 if (running == 0) { 292 /* reset the engine and set to writable */ 293 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008); 294 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000010); 295 296 /* load mc io regs */ 297 for (i = 0; i < regs_size; i++) { 298 WREG32(mmMC_SEQ_IO_DEBUG_INDEX, le32_to_cpup(io_mc_regs++)); 299 WREG32(mmMC_SEQ_IO_DEBUG_DATA, le32_to_cpup(io_mc_regs++)); 300 } 301 /* load the MC ucode */ 302 for (i = 0; i < ucode_size; i++) 303 WREG32(mmMC_SEQ_SUP_PGM, le32_to_cpup(fw_data++)); 304 305 /* put the engine back into the active state */ 306 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008); 307 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000004); 308 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000001); 309 310 /* wait for training to complete */ 311 for (i = 0; i < adev->usec_timeout; i++) { 312 if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL), 313 MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D0)) 314 break; 315 udelay(1); 316 } 317 for (i = 0; i < adev->usec_timeout; i++) { 318 if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL), 319 MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D1)) 320 break; 321 udelay(1); 322 } 323 } 324 325 return 0; 326 } 327 328 static void gmc_v8_0_vram_gtt_location(struct amdgpu_device *adev, 329 struct amdgpu_mc *mc) 330 { 331 if (mc->mc_vram_size > 0xFFC0000000ULL) { 332 /* leave room for at least 1024M GTT */ 333 dev_warn(adev->dev, "limiting VRAM\n"); 334 mc->real_vram_size = 0xFFC0000000ULL; 335 mc->mc_vram_size = 0xFFC0000000ULL; 336 } 337 amdgpu_vram_location(adev, &adev->mc, 0); 338 adev->mc.gtt_base_align = 0; 339 amdgpu_gtt_location(adev, mc); 340 } 341 342 /** 343 * gmc_v8_0_mc_program - program the GPU memory controller 344 * 345 * @adev: amdgpu_device pointer 346 * 347 * Set the location of vram, gart, and AGP in the GPU's 348 * physical address space (CIK). 349 */ 350 static void gmc_v8_0_mc_program(struct amdgpu_device *adev) 351 { 352 struct amdgpu_mode_mc_save save; 353 u32 tmp; 354 int i, j; 355 356 /* Initialize HDP */ 357 for (i = 0, j = 0; i < 32; i++, j += 0x6) { 358 WREG32((0xb05 + j), 0x00000000); 359 WREG32((0xb06 + j), 0x00000000); 360 WREG32((0xb07 + j), 0x00000000); 361 WREG32((0xb08 + j), 0x00000000); 362 WREG32((0xb09 + j), 0x00000000); 363 } 364 WREG32(mmHDP_REG_COHERENCY_FLUSH_CNTL, 0); 365 366 if (adev->mode_info.num_crtc) 367 amdgpu_display_set_vga_render_state(adev, false); 368 369 gmc_v8_0_mc_stop(adev, &save); 370 if (gmc_v8_0_wait_for_idle((void *)adev)) { 371 dev_warn(adev->dev, "Wait for MC idle timedout !\n"); 372 } 373 /* Update configuration */ 374 WREG32(mmMC_VM_SYSTEM_APERTURE_LOW_ADDR, 375 adev->mc.vram_start >> 12); 376 WREG32(mmMC_VM_SYSTEM_APERTURE_HIGH_ADDR, 377 adev->mc.vram_end >> 12); 378 WREG32(mmMC_VM_SYSTEM_APERTURE_DEFAULT_ADDR, 379 adev->vram_scratch.gpu_addr >> 12); 380 tmp = ((adev->mc.vram_end >> 24) & 0xFFFF) << 16; 381 tmp |= ((adev->mc.vram_start >> 24) & 0xFFFF); 382 WREG32(mmMC_VM_FB_LOCATION, tmp); 383 /* XXX double check these! */ 384 WREG32(mmHDP_NONSURFACE_BASE, (adev->mc.vram_start >> 8)); 385 WREG32(mmHDP_NONSURFACE_INFO, (2 << 7) | (1 << 30)); 386 WREG32(mmHDP_NONSURFACE_SIZE, 0x3FFFFFFF); 387 WREG32(mmMC_VM_AGP_BASE, 0); 388 WREG32(mmMC_VM_AGP_TOP, 0x0FFFFFFF); 389 WREG32(mmMC_VM_AGP_BOT, 0x0FFFFFFF); 390 if (gmc_v8_0_wait_for_idle((void *)adev)) { 391 dev_warn(adev->dev, "Wait for MC idle timedout !\n"); 392 } 393 gmc_v8_0_mc_resume(adev, &save); 394 395 WREG32(mmBIF_FB_EN, BIF_FB_EN__FB_READ_EN_MASK | BIF_FB_EN__FB_WRITE_EN_MASK); 396 397 tmp = RREG32(mmHDP_MISC_CNTL); 398 tmp = REG_SET_FIELD(tmp, HDP_MISC_CNTL, FLUSH_INVALIDATE_CACHE, 0); 399 WREG32(mmHDP_MISC_CNTL, tmp); 400 401 tmp = RREG32(mmHDP_HOST_PATH_CNTL); 402 WREG32(mmHDP_HOST_PATH_CNTL, tmp); 403 } 404 405 /** 406 * gmc_v8_0_mc_init - initialize the memory controller driver params 407 * 408 * @adev: amdgpu_device pointer 409 * 410 * Look up the amount of vram, vram width, and decide how to place 411 * vram and gart within the GPU's physical address space (CIK). 412 * Returns 0 for success. 413 */ 414 static int gmc_v8_0_mc_init(struct amdgpu_device *adev) 415 { 416 u32 tmp; 417 int chansize, numchan; 418 419 /* Get VRAM informations */ 420 tmp = RREG32(mmMC_ARB_RAMCFG); 421 if (REG_GET_FIELD(tmp, MC_ARB_RAMCFG, CHANSIZE)) { 422 chansize = 64; 423 } else { 424 chansize = 32; 425 } 426 tmp = RREG32(mmMC_SHARED_CHMAP); 427 switch (REG_GET_FIELD(tmp, MC_SHARED_CHMAP, NOOFCHAN)) { 428 case 0: 429 default: 430 numchan = 1; 431 break; 432 case 1: 433 numchan = 2; 434 break; 435 case 2: 436 numchan = 4; 437 break; 438 case 3: 439 numchan = 8; 440 break; 441 case 4: 442 numchan = 3; 443 break; 444 case 5: 445 numchan = 6; 446 break; 447 case 6: 448 numchan = 10; 449 break; 450 case 7: 451 numchan = 12; 452 break; 453 case 8: 454 numchan = 16; 455 break; 456 } 457 adev->mc.vram_width = numchan * chansize; 458 /* Could aper size report 0 ? */ 459 adev->mc.aper_base = pci_resource_start(adev->pdev, 0); 460 adev->mc.aper_size = pci_resource_len(adev->pdev, 0); 461 /* size in MB on si */ 462 adev->mc.mc_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL; 463 adev->mc.real_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL; 464 adev->mc.visible_vram_size = adev->mc.aper_size; 465 466 /* In case the PCI BAR is larger than the actual amount of vram */ 467 if (adev->mc.visible_vram_size > adev->mc.real_vram_size) 468 adev->mc.visible_vram_size = adev->mc.real_vram_size; 469 470 /* unless the user had overridden it, set the gart 471 * size equal to the 1024 or vram, whichever is larger. 472 */ 473 if (amdgpu_gart_size == -1) 474 adev->mc.gtt_size = amdgpu_ttm_get_gtt_mem_size(adev); 475 else 476 adev->mc.gtt_size = (uint64_t)amdgpu_gart_size << 20; 477 478 gmc_v8_0_vram_gtt_location(adev, &adev->mc); 479 480 return 0; 481 } 482 483 /* 484 * GART 485 * VMID 0 is the physical GPU addresses as used by the kernel. 486 * VMIDs 1-15 are used for userspace clients and are handled 487 * by the amdgpu vm/hsa code. 488 */ 489 490 /** 491 * gmc_v8_0_gart_flush_gpu_tlb - gart tlb flush callback 492 * 493 * @adev: amdgpu_device pointer 494 * @vmid: vm instance to flush 495 * 496 * Flush the TLB for the requested page table (CIK). 497 */ 498 static void gmc_v8_0_gart_flush_gpu_tlb(struct amdgpu_device *adev, 499 uint32_t vmid) 500 { 501 /* flush hdp cache */ 502 WREG32(mmHDP_MEM_COHERENCY_FLUSH_CNTL, 0); 503 504 /* bits 0-15 are the VM contexts0-15 */ 505 WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid); 506 } 507 508 /** 509 * gmc_v8_0_gart_set_pte_pde - update the page tables using MMIO 510 * 511 * @adev: amdgpu_device pointer 512 * @cpu_pt_addr: cpu address of the page table 513 * @gpu_page_idx: entry in the page table to update 514 * @addr: dst addr to write into pte/pde 515 * @flags: access flags 516 * 517 * Update the page tables using the CPU. 518 */ 519 static int gmc_v8_0_gart_set_pte_pde(struct amdgpu_device *adev, 520 void *cpu_pt_addr, 521 uint32_t gpu_page_idx, 522 uint64_t addr, 523 uint32_t flags) 524 { 525 void __iomem *ptr = (void *)cpu_pt_addr; 526 uint64_t value; 527 528 /* 529 * PTE format on VI: 530 * 63:40 reserved 531 * 39:12 4k physical page base address 532 * 11:7 fragment 533 * 6 write 534 * 5 read 535 * 4 exe 536 * 3 reserved 537 * 2 snooped 538 * 1 system 539 * 0 valid 540 * 541 * PDE format on VI: 542 * 63:59 block fragment size 543 * 58:40 reserved 544 * 39:1 physical base address of PTE 545 * bits 5:1 must be 0. 546 * 0 valid 547 */ 548 value = addr & 0x000000FFFFFFF000ULL; 549 value |= flags; 550 writeq(value, ptr + (gpu_page_idx * 8)); 551 552 return 0; 553 } 554 555 /** 556 * gmc_v8_0_set_fault_enable_default - update VM fault handling 557 * 558 * @adev: amdgpu_device pointer 559 * @value: true redirects VM faults to the default page 560 */ 561 static void gmc_v8_0_set_fault_enable_default(struct amdgpu_device *adev, 562 bool value) 563 { 564 u32 tmp; 565 566 tmp = RREG32(mmVM_CONTEXT1_CNTL); 567 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 568 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 569 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 570 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 571 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 572 PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, value); 573 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 574 VALID_PROTECTION_FAULT_ENABLE_DEFAULT, value); 575 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 576 READ_PROTECTION_FAULT_ENABLE_DEFAULT, value); 577 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 578 WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 579 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, 580 EXECUTE_PROTECTION_FAULT_ENABLE_DEFAULT, value); 581 WREG32(mmVM_CONTEXT1_CNTL, tmp); 582 } 583 584 /** 585 * gmc_v8_0_gart_enable - gart enable 586 * 587 * @adev: amdgpu_device pointer 588 * 589 * This sets up the TLBs, programs the page tables for VMID0, 590 * sets up the hw for VMIDs 1-15 which are allocated on 591 * demand, and sets up the global locations for the LDS, GDS, 592 * and GPUVM for FSA64 clients (CIK). 593 * Returns 0 for success, errors for failure. 594 */ 595 static int gmc_v8_0_gart_enable(struct amdgpu_device *adev) 596 { 597 int r, i; 598 u32 tmp; 599 600 if (adev->gart.robj == NULL) { 601 dev_err(adev->dev, "No VRAM object for PCIE GART.\n"); 602 return -EINVAL; 603 } 604 r = amdgpu_gart_table_vram_pin(adev); 605 if (r) 606 return r; 607 /* Setup TLB control */ 608 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL); 609 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 1); 610 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 1); 611 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE, 3); 612 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 1); 613 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_APERTURE_UNMAPPED_ACCESS, 0); 614 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp); 615 /* Setup L2 cache */ 616 tmp = RREG32(mmVM_L2_CNTL); 617 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 1); 618 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING, 1); 619 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE, 1); 620 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE, 1); 621 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, EFFECTIVE_L2_QUEUE_SIZE, 7); 622 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, CONTEXT1_IDENTITY_ACCESS_MODE, 1); 623 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_DEFAULT_PAGE_OUT_TO_SYSTEM_MEMORY, 1); 624 WREG32(mmVM_L2_CNTL, tmp); 625 tmp = RREG32(mmVM_L2_CNTL2); 626 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS, 1); 627 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_L2_CACHE, 1); 628 WREG32(mmVM_L2_CNTL2, tmp); 629 tmp = RREG32(mmVM_L2_CNTL3); 630 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY, 1); 631 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, BANK_SELECT, 4); 632 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_FRAGMENT_SIZE, 4); 633 WREG32(mmVM_L2_CNTL3, tmp); 634 /* XXX: set to enable PTE/PDE in system memory */ 635 tmp = RREG32(mmVM_L2_CNTL4); 636 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PDE_REQUEST_PHYSICAL, 0); 637 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PDE_REQUEST_SHARED, 0); 638 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PDE_REQUEST_SNOOP, 0); 639 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PTE_REQUEST_PHYSICAL, 0); 640 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PTE_REQUEST_SHARED, 0); 641 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PTE_REQUEST_SNOOP, 0); 642 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PDE_REQUEST_PHYSICAL, 0); 643 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PDE_REQUEST_SHARED, 0); 644 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PDE_REQUEST_SNOOP, 0); 645 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PTE_REQUEST_PHYSICAL, 0); 646 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PTE_REQUEST_SHARED, 0); 647 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PTE_REQUEST_SNOOP, 0); 648 WREG32(mmVM_L2_CNTL4, tmp); 649 /* setup context0 */ 650 WREG32(mmVM_CONTEXT0_PAGE_TABLE_START_ADDR, adev->mc.gtt_start >> 12); 651 WREG32(mmVM_CONTEXT0_PAGE_TABLE_END_ADDR, adev->mc.gtt_end >> 12); 652 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR, adev->gart.table_addr >> 12); 653 WREG32(mmVM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR, 654 (u32)(adev->dummy_page.addr >> 12)); 655 WREG32(mmVM_CONTEXT0_CNTL2, 0); 656 tmp = RREG32(mmVM_CONTEXT0_CNTL); 657 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, ENABLE_CONTEXT, 1); 658 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, PAGE_TABLE_DEPTH, 0); 659 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 660 WREG32(mmVM_CONTEXT0_CNTL, tmp); 661 662 WREG32(mmVM_L2_CONTEXT1_IDENTITY_APERTURE_LOW_ADDR, 0); 663 WREG32(mmVM_L2_CONTEXT1_IDENTITY_APERTURE_HIGH_ADDR, 0); 664 WREG32(mmVM_L2_CONTEXT_IDENTITY_PHYSICAL_OFFSET, 0); 665 666 /* empty context1-15 */ 667 /* FIXME start with 4G, once using 2 level pt switch to full 668 * vm size space 669 */ 670 /* set vm size, must be a multiple of 4 */ 671 WREG32(mmVM_CONTEXT1_PAGE_TABLE_START_ADDR, 0); 672 WREG32(mmVM_CONTEXT1_PAGE_TABLE_END_ADDR, adev->vm_manager.max_pfn - 1); 673 for (i = 1; i < 16; i++) { 674 if (i < 8) 675 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + i, 676 adev->gart.table_addr >> 12); 677 else 678 WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + i - 8, 679 adev->gart.table_addr >> 12); 680 } 681 682 /* enable context1-15 */ 683 WREG32(mmVM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR, 684 (u32)(adev->dummy_page.addr >> 12)); 685 WREG32(mmVM_CONTEXT1_CNTL2, 4); 686 tmp = RREG32(mmVM_CONTEXT1_CNTL); 687 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, ENABLE_CONTEXT, 1); 688 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_DEPTH, 1); 689 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 690 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 691 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 692 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, VALID_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 693 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, READ_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 694 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 695 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, EXECUTE_PROTECTION_FAULT_ENABLE_DEFAULT, 1); 696 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_BLOCK_SIZE, 697 amdgpu_vm_block_size - 9); 698 WREG32(mmVM_CONTEXT1_CNTL, tmp); 699 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) 700 gmc_v8_0_set_fault_enable_default(adev, false); 701 else 702 gmc_v8_0_set_fault_enable_default(adev, true); 703 704 gmc_v8_0_gart_flush_gpu_tlb(adev, 0); 705 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n", 706 (unsigned)(adev->mc.gtt_size >> 20), 707 (unsigned long long)adev->gart.table_addr); 708 adev->gart.ready = true; 709 return 0; 710 } 711 712 static int gmc_v8_0_gart_init(struct amdgpu_device *adev) 713 { 714 int r; 715 716 if (adev->gart.robj) { 717 WARN(1, "R600 PCIE GART already initialized\n"); 718 return 0; 719 } 720 /* Initialize common gart structure */ 721 r = amdgpu_gart_init(adev); 722 if (r) 723 return r; 724 adev->gart.table_size = adev->gart.num_gpu_pages * 8; 725 return amdgpu_gart_table_vram_alloc(adev); 726 } 727 728 /** 729 * gmc_v8_0_gart_disable - gart disable 730 * 731 * @adev: amdgpu_device pointer 732 * 733 * This disables all VM page table (CIK). 734 */ 735 static void gmc_v8_0_gart_disable(struct amdgpu_device *adev) 736 { 737 u32 tmp; 738 739 /* Disable all tables */ 740 WREG32(mmVM_CONTEXT0_CNTL, 0); 741 WREG32(mmVM_CONTEXT1_CNTL, 0); 742 /* Setup TLB control */ 743 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL); 744 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 0); 745 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 0); 746 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 0); 747 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp); 748 /* Setup L2 cache */ 749 tmp = RREG32(mmVM_L2_CNTL); 750 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 0); 751 WREG32(mmVM_L2_CNTL, tmp); 752 WREG32(mmVM_L2_CNTL2, 0); 753 amdgpu_gart_table_vram_unpin(adev); 754 } 755 756 /** 757 * gmc_v8_0_gart_fini - vm fini callback 758 * 759 * @adev: amdgpu_device pointer 760 * 761 * Tears down the driver GART/VM setup (CIK). 762 */ 763 static void gmc_v8_0_gart_fini(struct amdgpu_device *adev) 764 { 765 amdgpu_gart_table_vram_free(adev); 766 amdgpu_gart_fini(adev); 767 } 768 769 /* 770 * vm 771 * VMID 0 is the physical GPU addresses as used by the kernel. 772 * VMIDs 1-15 are used for userspace clients and are handled 773 * by the amdgpu vm/hsa code. 774 */ 775 /** 776 * gmc_v8_0_vm_init - cik vm init callback 777 * 778 * @adev: amdgpu_device pointer 779 * 780 * Inits cik specific vm parameters (number of VMs, base of vram for 781 * VMIDs 1-15) (CIK). 782 * Returns 0 for success. 783 */ 784 static int gmc_v8_0_vm_init(struct amdgpu_device *adev) 785 { 786 /* 787 * number of VMs 788 * VMID 0 is reserved for System 789 * amdgpu graphics/compute will use VMIDs 1-7 790 * amdkfd will use VMIDs 8-15 791 */ 792 adev->vm_manager.num_ids = AMDGPU_NUM_OF_VMIDS; 793 amdgpu_vm_manager_init(adev); 794 795 /* base offset of vram pages */ 796 if (adev->flags & AMD_IS_APU) { 797 u64 tmp = RREG32(mmMC_VM_FB_OFFSET); 798 tmp <<= 22; 799 adev->vm_manager.vram_base_offset = tmp; 800 } else 801 adev->vm_manager.vram_base_offset = 0; 802 803 return 0; 804 } 805 806 /** 807 * gmc_v8_0_vm_fini - cik vm fini callback 808 * 809 * @adev: amdgpu_device pointer 810 * 811 * Tear down any asic specific VM setup (CIK). 812 */ 813 static void gmc_v8_0_vm_fini(struct amdgpu_device *adev) 814 { 815 } 816 817 /** 818 * gmc_v8_0_vm_decode_fault - print human readable fault info 819 * 820 * @adev: amdgpu_device pointer 821 * @status: VM_CONTEXT1_PROTECTION_FAULT_STATUS register value 822 * @addr: VM_CONTEXT1_PROTECTION_FAULT_ADDR register value 823 * 824 * Print human readable fault information (CIK). 825 */ 826 static void gmc_v8_0_vm_decode_fault(struct amdgpu_device *adev, 827 u32 status, u32 addr, u32 mc_client) 828 { 829 u32 mc_id; 830 u32 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, VMID); 831 u32 protections = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 832 PROTECTIONS); 833 char block[5] = { mc_client >> 24, (mc_client >> 16) & 0xff, 834 (mc_client >> 8) & 0xff, mc_client & 0xff, 0 }; 835 836 mc_id = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 837 MEMORY_CLIENT_ID); 838 839 printk("VM fault (0x%02x, vmid %d) at page %u, %s from '%s' (0x%08x) (%d)\n", 840 protections, vmid, addr, 841 REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, 842 MEMORY_CLIENT_RW) ? 843 "write" : "read", block, mc_client, mc_id); 844 } 845 846 static int gmc_v8_0_convert_vram_type(int mc_seq_vram_type) 847 { 848 switch (mc_seq_vram_type) { 849 case MC_SEQ_MISC0__MT__GDDR1: 850 return AMDGPU_VRAM_TYPE_GDDR1; 851 case MC_SEQ_MISC0__MT__DDR2: 852 return AMDGPU_VRAM_TYPE_DDR2; 853 case MC_SEQ_MISC0__MT__GDDR3: 854 return AMDGPU_VRAM_TYPE_GDDR3; 855 case MC_SEQ_MISC0__MT__GDDR4: 856 return AMDGPU_VRAM_TYPE_GDDR4; 857 case MC_SEQ_MISC0__MT__GDDR5: 858 return AMDGPU_VRAM_TYPE_GDDR5; 859 case MC_SEQ_MISC0__MT__HBM: 860 return AMDGPU_VRAM_TYPE_HBM; 861 case MC_SEQ_MISC0__MT__DDR3: 862 return AMDGPU_VRAM_TYPE_DDR3; 863 default: 864 return AMDGPU_VRAM_TYPE_UNKNOWN; 865 } 866 } 867 868 static int gmc_v8_0_early_init(void *handle) 869 { 870 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 871 872 gmc_v8_0_set_gart_funcs(adev); 873 gmc_v8_0_set_irq_funcs(adev); 874 875 return 0; 876 } 877 878 static int gmc_v8_0_late_init(void *handle) 879 { 880 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 881 882 if (amdgpu_vm_fault_stop != AMDGPU_VM_FAULT_STOP_ALWAYS) 883 return amdgpu_irq_get(adev, &adev->mc.vm_fault, 0); 884 else 885 return 0; 886 } 887 888 #define mmMC_SEQ_MISC0_FIJI 0xA71 889 890 static int gmc_v8_0_sw_init(void *handle) 891 { 892 int r; 893 int dma_bits; 894 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 895 896 if (adev->flags & AMD_IS_APU) { 897 adev->mc.vram_type = AMDGPU_VRAM_TYPE_UNKNOWN; 898 } else { 899 u32 tmp; 900 901 if (adev->asic_type == CHIP_FIJI) 902 tmp = RREG32(mmMC_SEQ_MISC0_FIJI); 903 else 904 tmp = RREG32(mmMC_SEQ_MISC0); 905 tmp &= MC_SEQ_MISC0__MT__MASK; 906 adev->mc.vram_type = gmc_v8_0_convert_vram_type(tmp); 907 } 908 909 r = amdgpu_irq_add_id(adev, 146, &adev->mc.vm_fault); 910 if (r) 911 return r; 912 913 r = amdgpu_irq_add_id(adev, 147, &adev->mc.vm_fault); 914 if (r) 915 return r; 916 917 /* Adjust VM size here. 918 * Currently set to 4GB ((1 << 20) 4k pages). 919 * Max GPUVM size for cayman and SI is 40 bits. 920 */ 921 adev->vm_manager.max_pfn = amdgpu_vm_size << 18; 922 923 /* Set the internal MC address mask 924 * This is the max address of the GPU's 925 * internal address space. 926 */ 927 adev->mc.mc_mask = 0xffffffffffULL; /* 40 bit MC */ 928 929 /* set DMA mask + need_dma32 flags. 930 * PCIE - can handle 40-bits. 931 * IGP - can handle 40-bits 932 * PCI - dma32 for legacy pci gart, 40 bits on newer asics 933 */ 934 adev->need_dma32 = false; 935 dma_bits = adev->need_dma32 ? 32 : 40; 936 r = pci_set_dma_mask(adev->pdev, DMA_BIT_MASK(dma_bits)); 937 if (r) { 938 adev->need_dma32 = true; 939 dma_bits = 32; 940 printk(KERN_WARNING "amdgpu: No suitable DMA available.\n"); 941 } 942 r = pci_set_consistent_dma_mask(adev->pdev, DMA_BIT_MASK(dma_bits)); 943 if (r) { 944 pci_set_consistent_dma_mask(adev->pdev, DMA_BIT_MASK(32)); 945 printk(KERN_WARNING "amdgpu: No coherent DMA available.\n"); 946 } 947 948 r = gmc_v8_0_init_microcode(adev); 949 if (r) { 950 DRM_ERROR("Failed to load mc firmware!\n"); 951 return r; 952 } 953 954 r = amdgpu_ttm_global_init(adev); 955 if (r) { 956 return r; 957 } 958 959 r = gmc_v8_0_mc_init(adev); 960 if (r) 961 return r; 962 963 /* Memory manager */ 964 r = amdgpu_bo_init(adev); 965 if (r) 966 return r; 967 968 r = gmc_v8_0_gart_init(adev); 969 if (r) 970 return r; 971 972 if (!adev->vm_manager.enabled) { 973 r = gmc_v8_0_vm_init(adev); 974 if (r) { 975 dev_err(adev->dev, "vm manager initialization failed (%d).\n", r); 976 return r; 977 } 978 adev->vm_manager.enabled = true; 979 } 980 981 return r; 982 } 983 984 static int gmc_v8_0_sw_fini(void *handle) 985 { 986 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 987 988 if (adev->vm_manager.enabled) { 989 amdgpu_vm_manager_fini(adev); 990 gmc_v8_0_vm_fini(adev); 991 adev->vm_manager.enabled = false; 992 } 993 gmc_v8_0_gart_fini(adev); 994 amdgpu_gem_force_release(adev); 995 amdgpu_bo_fini(adev); 996 997 return 0; 998 } 999 1000 static int gmc_v8_0_hw_init(void *handle) 1001 { 1002 int r; 1003 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1004 1005 gmc_v8_0_init_golden_registers(adev); 1006 1007 gmc_v8_0_mc_program(adev); 1008 1009 if (adev->asic_type == CHIP_TONGA) { 1010 r = gmc_v8_0_mc_load_microcode(adev); 1011 if (r) { 1012 DRM_ERROR("Failed to load MC firmware!\n"); 1013 return r; 1014 } 1015 } 1016 1017 r = gmc_v8_0_gart_enable(adev); 1018 if (r) 1019 return r; 1020 1021 return r; 1022 } 1023 1024 static int gmc_v8_0_hw_fini(void *handle) 1025 { 1026 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1027 1028 amdgpu_irq_put(adev, &adev->mc.vm_fault, 0); 1029 gmc_v8_0_gart_disable(adev); 1030 1031 return 0; 1032 } 1033 1034 static int gmc_v8_0_suspend(void *handle) 1035 { 1036 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1037 1038 if (adev->vm_manager.enabled) { 1039 gmc_v8_0_vm_fini(adev); 1040 adev->vm_manager.enabled = false; 1041 } 1042 gmc_v8_0_hw_fini(adev); 1043 1044 return 0; 1045 } 1046 1047 static int gmc_v8_0_resume(void *handle) 1048 { 1049 int r; 1050 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1051 1052 r = gmc_v8_0_hw_init(adev); 1053 if (r) 1054 return r; 1055 1056 if (!adev->vm_manager.enabled) { 1057 r = gmc_v8_0_vm_init(adev); 1058 if (r) { 1059 dev_err(adev->dev, "vm manager initialization failed (%d).\n", r); 1060 return r; 1061 } 1062 adev->vm_manager.enabled = true; 1063 } 1064 1065 return r; 1066 } 1067 1068 static bool gmc_v8_0_is_idle(void *handle) 1069 { 1070 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1071 u32 tmp = RREG32(mmSRBM_STATUS); 1072 1073 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1074 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK | SRBM_STATUS__VMC_BUSY_MASK)) 1075 return false; 1076 1077 return true; 1078 } 1079 1080 static int gmc_v8_0_wait_for_idle(void *handle) 1081 { 1082 unsigned i; 1083 u32 tmp; 1084 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1085 1086 for (i = 0; i < adev->usec_timeout; i++) { 1087 /* read MC_STATUS */ 1088 tmp = RREG32(mmSRBM_STATUS) & (SRBM_STATUS__MCB_BUSY_MASK | 1089 SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1090 SRBM_STATUS__MCC_BUSY_MASK | 1091 SRBM_STATUS__MCD_BUSY_MASK | 1092 SRBM_STATUS__VMC_BUSY_MASK | 1093 SRBM_STATUS__VMC1_BUSY_MASK); 1094 if (!tmp) 1095 return 0; 1096 udelay(1); 1097 } 1098 return -ETIMEDOUT; 1099 1100 } 1101 1102 static bool gmc_v8_0_check_soft_reset(void *handle) 1103 { 1104 u32 srbm_soft_reset = 0; 1105 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1106 u32 tmp = RREG32(mmSRBM_STATUS); 1107 1108 if (tmp & SRBM_STATUS__VMC_BUSY_MASK) 1109 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset, 1110 SRBM_SOFT_RESET, SOFT_RESET_VMC, 1); 1111 1112 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK | 1113 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK)) { 1114 if (!(adev->flags & AMD_IS_APU)) 1115 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset, 1116 SRBM_SOFT_RESET, SOFT_RESET_MC, 1); 1117 } 1118 if (srbm_soft_reset) { 1119 adev->mc.srbm_soft_reset = srbm_soft_reset; 1120 return true; 1121 } else { 1122 adev->mc.srbm_soft_reset = 0; 1123 return false; 1124 } 1125 } 1126 1127 static int gmc_v8_0_pre_soft_reset(void *handle) 1128 { 1129 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1130 1131 if (!adev->mc.srbm_soft_reset) 1132 return 0; 1133 1134 gmc_v8_0_mc_stop(adev, &adev->mc.save); 1135 if (gmc_v8_0_wait_for_idle(adev)) { 1136 dev_warn(adev->dev, "Wait for GMC idle timed out !\n"); 1137 } 1138 1139 return 0; 1140 } 1141 1142 static int gmc_v8_0_soft_reset(void *handle) 1143 { 1144 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1145 u32 srbm_soft_reset; 1146 1147 if (!adev->mc.srbm_soft_reset) 1148 return 0; 1149 srbm_soft_reset = adev->mc.srbm_soft_reset; 1150 1151 if (srbm_soft_reset) { 1152 u32 tmp; 1153 1154 tmp = RREG32(mmSRBM_SOFT_RESET); 1155 tmp |= srbm_soft_reset; 1156 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp); 1157 WREG32(mmSRBM_SOFT_RESET, tmp); 1158 tmp = RREG32(mmSRBM_SOFT_RESET); 1159 1160 udelay(50); 1161 1162 tmp &= ~srbm_soft_reset; 1163 WREG32(mmSRBM_SOFT_RESET, tmp); 1164 tmp = RREG32(mmSRBM_SOFT_RESET); 1165 1166 /* Wait a little for things to settle down */ 1167 udelay(50); 1168 } 1169 1170 return 0; 1171 } 1172 1173 static int gmc_v8_0_post_soft_reset(void *handle) 1174 { 1175 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1176 1177 if (!adev->mc.srbm_soft_reset) 1178 return 0; 1179 1180 gmc_v8_0_mc_resume(adev, &adev->mc.save); 1181 return 0; 1182 } 1183 1184 static int gmc_v8_0_vm_fault_interrupt_state(struct amdgpu_device *adev, 1185 struct amdgpu_irq_src *src, 1186 unsigned type, 1187 enum amdgpu_interrupt_state state) 1188 { 1189 u32 tmp; 1190 u32 bits = (VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1191 VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1192 VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1193 VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1194 VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1195 VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK | 1196 VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK); 1197 1198 switch (state) { 1199 case AMDGPU_IRQ_STATE_DISABLE: 1200 /* system context */ 1201 tmp = RREG32(mmVM_CONTEXT0_CNTL); 1202 tmp &= ~bits; 1203 WREG32(mmVM_CONTEXT0_CNTL, tmp); 1204 /* VMs */ 1205 tmp = RREG32(mmVM_CONTEXT1_CNTL); 1206 tmp &= ~bits; 1207 WREG32(mmVM_CONTEXT1_CNTL, tmp); 1208 break; 1209 case AMDGPU_IRQ_STATE_ENABLE: 1210 /* system context */ 1211 tmp = RREG32(mmVM_CONTEXT0_CNTL); 1212 tmp |= bits; 1213 WREG32(mmVM_CONTEXT0_CNTL, tmp); 1214 /* VMs */ 1215 tmp = RREG32(mmVM_CONTEXT1_CNTL); 1216 tmp |= bits; 1217 WREG32(mmVM_CONTEXT1_CNTL, tmp); 1218 break; 1219 default: 1220 break; 1221 } 1222 1223 return 0; 1224 } 1225 1226 static int gmc_v8_0_process_interrupt(struct amdgpu_device *adev, 1227 struct amdgpu_irq_src *source, 1228 struct amdgpu_iv_entry *entry) 1229 { 1230 u32 addr, status, mc_client; 1231 1232 addr = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_ADDR); 1233 status = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_STATUS); 1234 mc_client = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_MCCLIENT); 1235 /* reset addr and status */ 1236 WREG32_P(mmVM_CONTEXT1_CNTL2, 1, ~1); 1237 1238 if (!addr && !status) 1239 return 0; 1240 1241 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_FIRST) 1242 gmc_v8_0_set_fault_enable_default(adev, false); 1243 1244 dev_err(adev->dev, "GPU fault detected: %d 0x%08x\n", 1245 entry->src_id, entry->src_data); 1246 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n", 1247 addr); 1248 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n", 1249 status); 1250 gmc_v8_0_vm_decode_fault(adev, status, addr, mc_client); 1251 1252 return 0; 1253 } 1254 1255 static void fiji_update_mc_medium_grain_clock_gating(struct amdgpu_device *adev, 1256 bool enable) 1257 { 1258 uint32_t data; 1259 1260 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_MGCG)) { 1261 data = RREG32(mmMC_HUB_MISC_HUB_CG); 1262 data |= MC_HUB_MISC_HUB_CG__ENABLE_MASK; 1263 WREG32(mmMC_HUB_MISC_HUB_CG, data); 1264 1265 data = RREG32(mmMC_HUB_MISC_SIP_CG); 1266 data |= MC_HUB_MISC_SIP_CG__ENABLE_MASK; 1267 WREG32(mmMC_HUB_MISC_SIP_CG, data); 1268 1269 data = RREG32(mmMC_HUB_MISC_VM_CG); 1270 data |= MC_HUB_MISC_VM_CG__ENABLE_MASK; 1271 WREG32(mmMC_HUB_MISC_VM_CG, data); 1272 1273 data = RREG32(mmMC_XPB_CLK_GAT); 1274 data |= MC_XPB_CLK_GAT__ENABLE_MASK; 1275 WREG32(mmMC_XPB_CLK_GAT, data); 1276 1277 data = RREG32(mmATC_MISC_CG); 1278 data |= ATC_MISC_CG__ENABLE_MASK; 1279 WREG32(mmATC_MISC_CG, data); 1280 1281 data = RREG32(mmMC_CITF_MISC_WR_CG); 1282 data |= MC_CITF_MISC_WR_CG__ENABLE_MASK; 1283 WREG32(mmMC_CITF_MISC_WR_CG, data); 1284 1285 data = RREG32(mmMC_CITF_MISC_RD_CG); 1286 data |= MC_CITF_MISC_RD_CG__ENABLE_MASK; 1287 WREG32(mmMC_CITF_MISC_RD_CG, data); 1288 1289 data = RREG32(mmMC_CITF_MISC_VM_CG); 1290 data |= MC_CITF_MISC_VM_CG__ENABLE_MASK; 1291 WREG32(mmMC_CITF_MISC_VM_CG, data); 1292 1293 data = RREG32(mmVM_L2_CG); 1294 data |= VM_L2_CG__ENABLE_MASK; 1295 WREG32(mmVM_L2_CG, data); 1296 } else { 1297 data = RREG32(mmMC_HUB_MISC_HUB_CG); 1298 data &= ~MC_HUB_MISC_HUB_CG__ENABLE_MASK; 1299 WREG32(mmMC_HUB_MISC_HUB_CG, data); 1300 1301 data = RREG32(mmMC_HUB_MISC_SIP_CG); 1302 data &= ~MC_HUB_MISC_SIP_CG__ENABLE_MASK; 1303 WREG32(mmMC_HUB_MISC_SIP_CG, data); 1304 1305 data = RREG32(mmMC_HUB_MISC_VM_CG); 1306 data &= ~MC_HUB_MISC_VM_CG__ENABLE_MASK; 1307 WREG32(mmMC_HUB_MISC_VM_CG, data); 1308 1309 data = RREG32(mmMC_XPB_CLK_GAT); 1310 data &= ~MC_XPB_CLK_GAT__ENABLE_MASK; 1311 WREG32(mmMC_XPB_CLK_GAT, data); 1312 1313 data = RREG32(mmATC_MISC_CG); 1314 data &= ~ATC_MISC_CG__ENABLE_MASK; 1315 WREG32(mmATC_MISC_CG, data); 1316 1317 data = RREG32(mmMC_CITF_MISC_WR_CG); 1318 data &= ~MC_CITF_MISC_WR_CG__ENABLE_MASK; 1319 WREG32(mmMC_CITF_MISC_WR_CG, data); 1320 1321 data = RREG32(mmMC_CITF_MISC_RD_CG); 1322 data &= ~MC_CITF_MISC_RD_CG__ENABLE_MASK; 1323 WREG32(mmMC_CITF_MISC_RD_CG, data); 1324 1325 data = RREG32(mmMC_CITF_MISC_VM_CG); 1326 data &= ~MC_CITF_MISC_VM_CG__ENABLE_MASK; 1327 WREG32(mmMC_CITF_MISC_VM_CG, data); 1328 1329 data = RREG32(mmVM_L2_CG); 1330 data &= ~VM_L2_CG__ENABLE_MASK; 1331 WREG32(mmVM_L2_CG, data); 1332 } 1333 } 1334 1335 static void fiji_update_mc_light_sleep(struct amdgpu_device *adev, 1336 bool enable) 1337 { 1338 uint32_t data; 1339 1340 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_LS)) { 1341 data = RREG32(mmMC_HUB_MISC_HUB_CG); 1342 data |= MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK; 1343 WREG32(mmMC_HUB_MISC_HUB_CG, data); 1344 1345 data = RREG32(mmMC_HUB_MISC_SIP_CG); 1346 data |= MC_HUB_MISC_SIP_CG__MEM_LS_ENABLE_MASK; 1347 WREG32(mmMC_HUB_MISC_SIP_CG, data); 1348 1349 data = RREG32(mmMC_HUB_MISC_VM_CG); 1350 data |= MC_HUB_MISC_VM_CG__MEM_LS_ENABLE_MASK; 1351 WREG32(mmMC_HUB_MISC_VM_CG, data); 1352 1353 data = RREG32(mmMC_XPB_CLK_GAT); 1354 data |= MC_XPB_CLK_GAT__MEM_LS_ENABLE_MASK; 1355 WREG32(mmMC_XPB_CLK_GAT, data); 1356 1357 data = RREG32(mmATC_MISC_CG); 1358 data |= ATC_MISC_CG__MEM_LS_ENABLE_MASK; 1359 WREG32(mmATC_MISC_CG, data); 1360 1361 data = RREG32(mmMC_CITF_MISC_WR_CG); 1362 data |= MC_CITF_MISC_WR_CG__MEM_LS_ENABLE_MASK; 1363 WREG32(mmMC_CITF_MISC_WR_CG, data); 1364 1365 data = RREG32(mmMC_CITF_MISC_RD_CG); 1366 data |= MC_CITF_MISC_RD_CG__MEM_LS_ENABLE_MASK; 1367 WREG32(mmMC_CITF_MISC_RD_CG, data); 1368 1369 data = RREG32(mmMC_CITF_MISC_VM_CG); 1370 data |= MC_CITF_MISC_VM_CG__MEM_LS_ENABLE_MASK; 1371 WREG32(mmMC_CITF_MISC_VM_CG, data); 1372 1373 data = RREG32(mmVM_L2_CG); 1374 data |= VM_L2_CG__MEM_LS_ENABLE_MASK; 1375 WREG32(mmVM_L2_CG, data); 1376 } else { 1377 data = RREG32(mmMC_HUB_MISC_HUB_CG); 1378 data &= ~MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK; 1379 WREG32(mmMC_HUB_MISC_HUB_CG, data); 1380 1381 data = RREG32(mmMC_HUB_MISC_SIP_CG); 1382 data &= ~MC_HUB_MISC_SIP_CG__MEM_LS_ENABLE_MASK; 1383 WREG32(mmMC_HUB_MISC_SIP_CG, data); 1384 1385 data = RREG32(mmMC_HUB_MISC_VM_CG); 1386 data &= ~MC_HUB_MISC_VM_CG__MEM_LS_ENABLE_MASK; 1387 WREG32(mmMC_HUB_MISC_VM_CG, data); 1388 1389 data = RREG32(mmMC_XPB_CLK_GAT); 1390 data &= ~MC_XPB_CLK_GAT__MEM_LS_ENABLE_MASK; 1391 WREG32(mmMC_XPB_CLK_GAT, data); 1392 1393 data = RREG32(mmATC_MISC_CG); 1394 data &= ~ATC_MISC_CG__MEM_LS_ENABLE_MASK; 1395 WREG32(mmATC_MISC_CG, data); 1396 1397 data = RREG32(mmMC_CITF_MISC_WR_CG); 1398 data &= ~MC_CITF_MISC_WR_CG__MEM_LS_ENABLE_MASK; 1399 WREG32(mmMC_CITF_MISC_WR_CG, data); 1400 1401 data = RREG32(mmMC_CITF_MISC_RD_CG); 1402 data &= ~MC_CITF_MISC_RD_CG__MEM_LS_ENABLE_MASK; 1403 WREG32(mmMC_CITF_MISC_RD_CG, data); 1404 1405 data = RREG32(mmMC_CITF_MISC_VM_CG); 1406 data &= ~MC_CITF_MISC_VM_CG__MEM_LS_ENABLE_MASK; 1407 WREG32(mmMC_CITF_MISC_VM_CG, data); 1408 1409 data = RREG32(mmVM_L2_CG); 1410 data &= ~VM_L2_CG__MEM_LS_ENABLE_MASK; 1411 WREG32(mmVM_L2_CG, data); 1412 } 1413 } 1414 1415 static int gmc_v8_0_set_clockgating_state(void *handle, 1416 enum amd_clockgating_state state) 1417 { 1418 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1419 1420 switch (adev->asic_type) { 1421 case CHIP_FIJI: 1422 fiji_update_mc_medium_grain_clock_gating(adev, 1423 state == AMD_CG_STATE_GATE ? true : false); 1424 fiji_update_mc_light_sleep(adev, 1425 state == AMD_CG_STATE_GATE ? true : false); 1426 break; 1427 default: 1428 break; 1429 } 1430 return 0; 1431 } 1432 1433 static int gmc_v8_0_set_powergating_state(void *handle, 1434 enum amd_powergating_state state) 1435 { 1436 return 0; 1437 } 1438 1439 const struct amd_ip_funcs gmc_v8_0_ip_funcs = { 1440 .name = "gmc_v8_0", 1441 .early_init = gmc_v8_0_early_init, 1442 .late_init = gmc_v8_0_late_init, 1443 .sw_init = gmc_v8_0_sw_init, 1444 .sw_fini = gmc_v8_0_sw_fini, 1445 .hw_init = gmc_v8_0_hw_init, 1446 .hw_fini = gmc_v8_0_hw_fini, 1447 .suspend = gmc_v8_0_suspend, 1448 .resume = gmc_v8_0_resume, 1449 .is_idle = gmc_v8_0_is_idle, 1450 .wait_for_idle = gmc_v8_0_wait_for_idle, 1451 .check_soft_reset = gmc_v8_0_check_soft_reset, 1452 .pre_soft_reset = gmc_v8_0_pre_soft_reset, 1453 .soft_reset = gmc_v8_0_soft_reset, 1454 .post_soft_reset = gmc_v8_0_post_soft_reset, 1455 .set_clockgating_state = gmc_v8_0_set_clockgating_state, 1456 .set_powergating_state = gmc_v8_0_set_powergating_state, 1457 }; 1458 1459 static const struct amdgpu_gart_funcs gmc_v8_0_gart_funcs = { 1460 .flush_gpu_tlb = gmc_v8_0_gart_flush_gpu_tlb, 1461 .set_pte_pde = gmc_v8_0_gart_set_pte_pde, 1462 }; 1463 1464 static const struct amdgpu_irq_src_funcs gmc_v8_0_irq_funcs = { 1465 .set = gmc_v8_0_vm_fault_interrupt_state, 1466 .process = gmc_v8_0_process_interrupt, 1467 }; 1468 1469 static void gmc_v8_0_set_gart_funcs(struct amdgpu_device *adev) 1470 { 1471 if (adev->gart.gart_funcs == NULL) 1472 adev->gart.gart_funcs = &gmc_v8_0_gart_funcs; 1473 } 1474 1475 static void gmc_v8_0_set_irq_funcs(struct amdgpu_device *adev) 1476 { 1477 adev->mc.vm_fault.num_types = 1; 1478 adev->mc.vm_fault.funcs = &gmc_v8_0_irq_funcs; 1479 } 1480