1 /* 2 * Copyright 2008 Advanced Micro Devices, Inc. 3 * Copyright 2008 Red Hat Inc. 4 * Copyright 2009 Jerome Glisse. 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the "Software"), 8 * to deal in the Software without restriction, including without limitation 9 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 10 * and/or sell copies of the Software, and to permit persons to whom the 11 * Software is furnished to do so, subject to the following conditions: 12 * 13 * The above copyright notice and this permission notice shall be included in 14 * all copies or substantial portions of the Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 22 * OTHER DEALINGS IN THE SOFTWARE. 23 * 24 * Authors: Dave Airlie 25 * Alex Deucher 26 * Jerome Glisse 27 */ 28 #include <linux/power_supply.h> 29 #include <linux/kthread.h> 30 #include <linux/console.h> 31 #include <linux/slab.h> 32 #include <drm/drmP.h> 33 #include <drm/drm_crtc_helper.h> 34 #include <drm/drm_atomic_helper.h> 35 #include <drm/amdgpu_drm.h> 36 #include <linux/vgaarb.h> 37 #include <linux/vga_switcheroo.h> 38 #include <linux/efi.h> 39 #include "amdgpu.h" 40 #include "amdgpu_trace.h" 41 #include "amdgpu_i2c.h" 42 #include "atom.h" 43 #include "amdgpu_atombios.h" 44 #include "amdgpu_atomfirmware.h" 45 #include "amd_pcie.h" 46 #ifdef CONFIG_DRM_AMDGPU_SI 47 #include "si.h" 48 #endif 49 #ifdef CONFIG_DRM_AMDGPU_CIK 50 #include "cik.h" 51 #endif 52 #include "vi.h" 53 #include "soc15.h" 54 #include "bif/bif_4_1_d.h" 55 #include <linux/pci.h> 56 #include <linux/firmware.h> 57 #include "amdgpu_vf_error.h" 58 59 #include "amdgpu_amdkfd.h" 60 #include "amdgpu_pm.h" 61 62 MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin"); 63 MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin"); 64 MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin"); 65 66 #define AMDGPU_RESUME_MS 2000 67 68 static const char *amdgpu_asic_name[] = { 69 "TAHITI", 70 "PITCAIRN", 71 "VERDE", 72 "OLAND", 73 "HAINAN", 74 "BONAIRE", 75 "KAVERI", 76 "KABINI", 77 "HAWAII", 78 "MULLINS", 79 "TOPAZ", 80 "TONGA", 81 "FIJI", 82 "CARRIZO", 83 "STONEY", 84 "POLARIS10", 85 "POLARIS11", 86 "POLARIS12", 87 "VEGAM", 88 "VEGA10", 89 "VEGA12", 90 "VEGA20", 91 "RAVEN", 92 "LAST", 93 }; 94 95 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev); 96 97 /** 98 * amdgpu_device_is_px - Is the device is a dGPU with HG/PX power control 99 * 100 * @dev: drm_device pointer 101 * 102 * Returns true if the device is a dGPU with HG/PX power control, 103 * otherwise return false. 104 */ 105 bool amdgpu_device_is_px(struct drm_device *dev) 106 { 107 struct amdgpu_device *adev = dev->dev_private; 108 109 if (adev->flags & AMD_IS_PX) 110 return true; 111 return false; 112 } 113 114 /* 115 * MMIO register access helper functions. 116 */ 117 /** 118 * amdgpu_mm_rreg - read a memory mapped IO register 119 * 120 * @adev: amdgpu_device pointer 121 * @reg: dword aligned register offset 122 * @acc_flags: access flags which require special behavior 123 * 124 * Returns the 32 bit value from the offset specified. 125 */ 126 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, 127 uint32_t acc_flags) 128 { 129 uint32_t ret; 130 131 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev)) 132 return amdgpu_virt_kiq_rreg(adev, reg); 133 134 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX)) 135 ret = bus_space_read_4(adev->rmmio_bst, adev->rmmio_bsh, 136 reg * 4); 137 else { 138 unsigned long flags; 139 140 spin_lock_irqsave(&adev->mmio_idx_lock, flags); 141 bus_space_write_4(adev->rmmio_bst, adev->rmmio_bsh, 142 mmMM_INDEX * 4, reg * 4); 143 ret = bus_space_read_4(adev->rmmio_bst, adev->rmmio_bsh, 144 mmMM_DATA * 4); 145 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags); 146 } 147 trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret); 148 return ret; 149 } 150 151 /* 152 * MMIO register read with bytes helper functions 153 * @offset:bytes offset from MMIO start 154 * 155 */ 156 157 /** 158 * amdgpu_mm_rreg8 - read a memory mapped IO register 159 * 160 * @adev: amdgpu_device pointer 161 * @offset: byte aligned register offset 162 * 163 * Returns the 8 bit value from the offset specified. 164 */ 165 uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) { 166 if (offset < adev->rmmio_size) 167 return bus_space_read_1(adev->rmmio_bst, adev->rmmio_bsh, 168 offset); 169 BUG(); 170 } 171 172 /* 173 * MMIO register write with bytes helper functions 174 * @offset:bytes offset from MMIO start 175 * @value: the value want to be written to the register 176 * 177 */ 178 /** 179 * amdgpu_mm_wreg8 - read a memory mapped IO register 180 * 181 * @adev: amdgpu_device pointer 182 * @offset: byte aligned register offset 183 * @value: 8 bit value to write 184 * 185 * Writes the value specified to the offset specified. 186 */ 187 void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) { 188 if (offset < adev->rmmio_size) 189 bus_space_write_1(adev->rmmio_bst, adev->rmmio_bsh, 190 offset, value); 191 else 192 BUG(); 193 } 194 195 /** 196 * amdgpu_mm_wreg - write to a memory mapped IO register 197 * 198 * @adev: amdgpu_device pointer 199 * @reg: dword aligned register offset 200 * @v: 32 bit value to write to the register 201 * @acc_flags: access flags which require special behavior 202 * 203 * Writes the value specified to the offset specified. 204 */ 205 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v, 206 uint32_t acc_flags) 207 { 208 trace_amdgpu_mm_wreg(adev->pdev->device, reg, v); 209 210 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) { 211 adev->last_mm_index = v; 212 } 213 214 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev)) 215 return amdgpu_virt_kiq_wreg(adev, reg, v); 216 217 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX)) 218 bus_space_write_4(adev->rmmio_bst, adev->rmmio_bsh, 219 reg * 4, v); 220 else { 221 unsigned long flags; 222 223 spin_lock_irqsave(&adev->mmio_idx_lock, flags); 224 bus_space_write_4(adev->rmmio_bst, adev->rmmio_bsh, 225 mmMM_INDEX * 4, reg * 4); 226 bus_space_write_4(adev->rmmio_bst, adev->rmmio_bsh, 227 mmMM_DATA * 4, v); 228 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags); 229 } 230 231 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) { 232 udelay(500); 233 } 234 } 235 236 /** 237 * amdgpu_io_rreg - read an IO register 238 * 239 * @adev: amdgpu_device pointer 240 * @reg: dword aligned register offset 241 * 242 * Returns the 32 bit value from the offset specified. 243 */ 244 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg) 245 { 246 if ((reg * 4) < adev->rio_mem_size) 247 return bus_space_read_4(adev->rio_mem_bst, adev->rio_mem_bsh, reg); 248 else { 249 bus_space_write_4(adev->rio_mem_bst, adev->rio_mem_bsh, 250 mmMM_INDEX * 4, reg * 4); 251 return bus_space_read_4(adev->rio_mem_bst, adev->rio_mem_bsh, 252 mmMM_INDEX * 4); 253 } 254 } 255 256 /** 257 * amdgpu_io_wreg - write to an IO register 258 * 259 * @adev: amdgpu_device pointer 260 * @reg: dword aligned register offset 261 * @v: 32 bit value to write to the register 262 * 263 * Writes the value specified to the offset specified. 264 */ 265 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v) 266 { 267 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) { 268 adev->last_mm_index = v; 269 } 270 271 if ((reg * 4) < adev->rio_mem_size) 272 bus_space_write_4(adev->rio_mem_bst, adev->rio_mem_bsh, 273 reg * 4, v); 274 else { 275 bus_space_write_4(adev->rio_mem_bst, adev->rio_mem_bsh, 276 mmMM_INDEX * 4, reg * 4); 277 bus_space_write_4(adev->rio_mem_bst, adev->rio_mem_bsh, 278 mmMM_DATA * 4, v); 279 280 } 281 282 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) { 283 udelay(500); 284 } 285 } 286 287 /** 288 * amdgpu_mm_rdoorbell - read a doorbell dword 289 * 290 * @adev: amdgpu_device pointer 291 * @index: doorbell index 292 * 293 * Returns the value in the doorbell aperture at the 294 * requested doorbell index (CIK). 295 */ 296 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index) 297 { 298 if (index < adev->doorbell.num_doorbells) { 299 return bus_space_read_4(adev->doorbell.bst, adev->doorbell.bsh, 300 index * 4); 301 } else { 302 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index); 303 return 0; 304 } 305 } 306 307 /** 308 * amdgpu_mm_wdoorbell - write a doorbell dword 309 * 310 * @adev: amdgpu_device pointer 311 * @index: doorbell index 312 * @v: value to write 313 * 314 * Writes @v to the doorbell aperture at the 315 * requested doorbell index (CIK). 316 */ 317 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v) 318 { 319 if (index < adev->doorbell.num_doorbells) { 320 bus_space_write_4(adev->doorbell.bst, adev->doorbell.bsh, 321 index * 4, v); 322 } else { 323 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index); 324 } 325 } 326 327 /** 328 * amdgpu_mm_rdoorbell64 - read a doorbell Qword 329 * 330 * @adev: amdgpu_device pointer 331 * @index: doorbell index 332 * 333 * Returns the value in the doorbell aperture at the 334 * requested doorbell index (VEGA10+). 335 */ 336 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index) 337 { 338 if (index < adev->doorbell.num_doorbells) { 339 return bus_space_read_8(adev->doorbell.bst, adev->doorbell.bsh, 340 index * 4); 341 } else { 342 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index); 343 return 0; 344 } 345 } 346 347 /** 348 * amdgpu_mm_wdoorbell64 - write a doorbell Qword 349 * 350 * @adev: amdgpu_device pointer 351 * @index: doorbell index 352 * @v: value to write 353 * 354 * Writes @v to the doorbell aperture at the 355 * requested doorbell index (VEGA10+). 356 */ 357 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v) 358 { 359 if (index < adev->doorbell.num_doorbells) { 360 bus_space_write_8(adev->doorbell.bst, adev->doorbell.bsh, 361 index * 4, v); 362 } else { 363 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index); 364 } 365 } 366 367 /** 368 * amdgpu_invalid_rreg - dummy reg read function 369 * 370 * @adev: amdgpu device pointer 371 * @reg: offset of register 372 * 373 * Dummy register read function. Used for register blocks 374 * that certain asics don't have (all asics). 375 * Returns the value in the register. 376 */ 377 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg) 378 { 379 DRM_ERROR("Invalid callback to read register 0x%04X\n", reg); 380 BUG(); 381 return 0; 382 } 383 384 /** 385 * amdgpu_invalid_wreg - dummy reg write function 386 * 387 * @adev: amdgpu device pointer 388 * @reg: offset of register 389 * @v: value to write to the register 390 * 391 * Dummy register read function. Used for register blocks 392 * that certain asics don't have (all asics). 393 */ 394 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v) 395 { 396 DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n", 397 reg, v); 398 BUG(); 399 } 400 401 /** 402 * amdgpu_block_invalid_rreg - dummy reg read function 403 * 404 * @adev: amdgpu device pointer 405 * @block: offset of instance 406 * @reg: offset of register 407 * 408 * Dummy register read function. Used for register blocks 409 * that certain asics don't have (all asics). 410 * Returns the value in the register. 411 */ 412 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev, 413 uint32_t block, uint32_t reg) 414 { 415 DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n", 416 reg, block); 417 BUG(); 418 return 0; 419 } 420 421 /** 422 * amdgpu_block_invalid_wreg - dummy reg write function 423 * 424 * @adev: amdgpu device pointer 425 * @block: offset of instance 426 * @reg: offset of register 427 * @v: value to write to the register 428 * 429 * Dummy register read function. Used for register blocks 430 * that certain asics don't have (all asics). 431 */ 432 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev, 433 uint32_t block, 434 uint32_t reg, uint32_t v) 435 { 436 DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n", 437 reg, block, v); 438 BUG(); 439 } 440 441 /** 442 * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page 443 * 444 * @adev: amdgpu device pointer 445 * 446 * Allocates a scratch page of VRAM for use by various things in the 447 * driver. 448 */ 449 static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev) 450 { 451 return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE, 452 PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM, 453 &adev->vram_scratch.robj, 454 &adev->vram_scratch.gpu_addr, 455 (void **)&adev->vram_scratch.ptr); 456 } 457 458 /** 459 * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page 460 * 461 * @adev: amdgpu device pointer 462 * 463 * Frees the VRAM scratch page. 464 */ 465 static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev) 466 { 467 amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL); 468 } 469 470 /** 471 * amdgpu_device_program_register_sequence - program an array of registers. 472 * 473 * @adev: amdgpu_device pointer 474 * @registers: pointer to the register array 475 * @array_size: size of the register array 476 * 477 * Programs an array or registers with and and or masks. 478 * This is a helper for setting golden registers. 479 */ 480 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev, 481 const u32 *registers, 482 const u32 array_size) 483 { 484 u32 tmp, reg, and_mask, or_mask; 485 int i; 486 487 if (array_size % 3) 488 return; 489 490 for (i = 0; i < array_size; i +=3) { 491 reg = registers[i + 0]; 492 and_mask = registers[i + 1]; 493 or_mask = registers[i + 2]; 494 495 if (and_mask == 0xffffffff) { 496 tmp = or_mask; 497 } else { 498 tmp = RREG32(reg); 499 tmp &= ~and_mask; 500 tmp |= or_mask; 501 } 502 WREG32(reg, tmp); 503 } 504 } 505 506 /** 507 * amdgpu_device_pci_config_reset - reset the GPU 508 * 509 * @adev: amdgpu_device pointer 510 * 511 * Resets the GPU using the pci config reset sequence. 512 * Only applicable to asics prior to vega10. 513 */ 514 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev) 515 { 516 pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA); 517 } 518 519 /* 520 * GPU doorbell aperture helpers function. 521 */ 522 /** 523 * amdgpu_device_doorbell_init - Init doorbell driver information. 524 * 525 * @adev: amdgpu_device pointer 526 * 527 * Init doorbell driver information (CIK) 528 * Returns 0 on success, error on failure. 529 */ 530 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev) 531 { 532 /* No doorbell on SI hardware generation */ 533 if (adev->asic_type < CHIP_BONAIRE) { 534 adev->doorbell.base = 0; 535 adev->doorbell.size = 0; 536 adev->doorbell.num_doorbells = 0; 537 #ifdef __linux__ 538 adev->doorbell.ptr = NULL; 539 #endif 540 return 0; 541 } 542 543 #ifdef __linux__ 544 if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET) 545 return -EINVAL; 546 547 /* doorbell bar mapping */ 548 adev->doorbell.base = pci_resource_start(adev->pdev, 2); 549 adev->doorbell.size = pci_resource_len(adev->pdev, 2); 550 #endif 551 552 adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32), 553 AMDGPU_DOORBELL_MAX_ASSIGNMENT+1); 554 if (adev->doorbell.num_doorbells == 0) 555 return -EINVAL; 556 557 #ifdef __linux__ 558 adev->doorbell.ptr = ioremap(adev->doorbell.base, 559 adev->doorbell.num_doorbells * 560 sizeof(u32)); 561 if (adev->doorbell.ptr == NULL) 562 return -ENOMEM; 563 #endif 564 565 return 0; 566 } 567 568 /** 569 * amdgpu_device_doorbell_fini - Tear down doorbell driver information. 570 * 571 * @adev: amdgpu_device pointer 572 * 573 * Tear down doorbell driver information (CIK) 574 */ 575 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev) 576 { 577 #ifdef __linux__ 578 iounmap(adev->doorbell.ptr); 579 adev->doorbell.ptr = NULL; 580 #else 581 if (adev->doorbell.size > 0) 582 bus_space_unmap(adev->doorbell.bst, adev->doorbell.bsh, 583 adev->doorbell.size); 584 #endif 585 } 586 587 588 589 /* 590 * amdgpu_device_wb_*() 591 * Writeback is the method by which the GPU updates special pages in memory 592 * with the status of certain GPU events (fences, ring pointers,etc.). 593 */ 594 595 /** 596 * amdgpu_device_wb_fini - Disable Writeback and free memory 597 * 598 * @adev: amdgpu_device pointer 599 * 600 * Disables Writeback and frees the Writeback memory (all asics). 601 * Used at driver shutdown. 602 */ 603 static void amdgpu_device_wb_fini(struct amdgpu_device *adev) 604 { 605 if (adev->wb.wb_obj) { 606 amdgpu_bo_free_kernel(&adev->wb.wb_obj, 607 &adev->wb.gpu_addr, 608 (void **)&adev->wb.wb); 609 adev->wb.wb_obj = NULL; 610 } 611 } 612 613 /** 614 * amdgpu_device_wb_init- Init Writeback driver info and allocate memory 615 * 616 * @adev: amdgpu_device pointer 617 * 618 * Initializes writeback and allocates writeback memory (all asics). 619 * Used at driver startup. 620 * Returns 0 on success or an -error on failure. 621 */ 622 static int amdgpu_device_wb_init(struct amdgpu_device *adev) 623 { 624 int r; 625 626 if (adev->wb.wb_obj == NULL) { 627 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */ 628 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8, 629 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT, 630 &adev->wb.wb_obj, &adev->wb.gpu_addr, 631 (void **)&adev->wb.wb); 632 if (r) { 633 dev_warn(adev->dev, "(%d) create WB bo failed\n", r); 634 return r; 635 } 636 637 adev->wb.num_wb = AMDGPU_MAX_WB; 638 memset(&adev->wb.used, 0, sizeof(adev->wb.used)); 639 640 /* clear wb memory */ 641 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8); 642 } 643 644 return 0; 645 } 646 647 /** 648 * amdgpu_device_wb_get - Allocate a wb entry 649 * 650 * @adev: amdgpu_device pointer 651 * @wb: wb index 652 * 653 * Allocate a wb slot for use by the driver (all asics). 654 * Returns 0 on success or -EINVAL on failure. 655 */ 656 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb) 657 { 658 unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb); 659 660 if (offset < adev->wb.num_wb) { 661 __set_bit(offset, adev->wb.used); 662 *wb = offset << 3; /* convert to dw offset */ 663 return 0; 664 } else { 665 return -EINVAL; 666 } 667 } 668 669 /** 670 * amdgpu_device_wb_free - Free a wb entry 671 * 672 * @adev: amdgpu_device pointer 673 * @wb: wb index 674 * 675 * Free a wb slot allocated for use by the driver (all asics) 676 */ 677 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb) 678 { 679 wb >>= 3; 680 if (wb < adev->wb.num_wb) 681 __clear_bit(wb, adev->wb.used); 682 } 683 684 /** 685 * amdgpu_device_vram_location - try to find VRAM location 686 * 687 * @adev: amdgpu device structure holding all necessary informations 688 * @mc: memory controller structure holding memory informations 689 * @base: base address at which to put VRAM 690 * 691 * Function will try to place VRAM at base address provided 692 * as parameter. 693 */ 694 void amdgpu_device_vram_location(struct amdgpu_device *adev, 695 struct amdgpu_gmc *mc, u64 base) 696 { 697 uint64_t limit = (uint64_t)amdgpu_vram_limit << 20; 698 699 mc->vram_start = base; 700 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1; 701 if (limit && limit < mc->real_vram_size) 702 mc->real_vram_size = limit; 703 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n", 704 mc->mc_vram_size >> 20, mc->vram_start, 705 mc->vram_end, mc->real_vram_size >> 20); 706 } 707 708 /** 709 * amdgpu_device_gart_location - try to find GART location 710 * 711 * @adev: amdgpu device structure holding all necessary informations 712 * @mc: memory controller structure holding memory informations 713 * 714 * Function will place try to place GART before or after VRAM. 715 * 716 * If GART size is bigger than space left then we ajust GART size. 717 * Thus function will never fails. 718 */ 719 void amdgpu_device_gart_location(struct amdgpu_device *adev, 720 struct amdgpu_gmc *mc) 721 { 722 u64 size_af, size_bf; 723 724 mc->gart_size += adev->pm.smu_prv_buffer_size; 725 726 size_af = adev->gmc.mc_mask - mc->vram_end; 727 size_bf = mc->vram_start; 728 if (size_bf > size_af) { 729 if (mc->gart_size > size_bf) { 730 dev_warn(adev->dev, "limiting GART\n"); 731 mc->gart_size = size_bf; 732 } 733 mc->gart_start = 0; 734 } else { 735 if (mc->gart_size > size_af) { 736 dev_warn(adev->dev, "limiting GART\n"); 737 mc->gart_size = size_af; 738 } 739 /* VCE doesn't like it when BOs cross a 4GB segment, so align 740 * the GART base on a 4GB boundary as well. 741 */ 742 mc->gart_start = roundup2(mc->vram_end + 1, 0x100000000ULL); 743 } 744 mc->gart_end = mc->gart_start + mc->gart_size - 1; 745 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n", 746 mc->gart_size >> 20, mc->gart_start, mc->gart_end); 747 } 748 749 /** 750 * amdgpu_device_resize_fb_bar - try to resize FB BAR 751 * 752 * @adev: amdgpu_device pointer 753 * 754 * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not 755 * to fail, but if any of the BARs is not accessible after the size we abort 756 * driver loading by returning -ENODEV. 757 */ 758 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev) 759 { 760 u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size); 761 u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1; 762 struct pci_bus *root; 763 struct resource *res; 764 unsigned i; 765 u16 cmd; 766 int r; 767 pcireg_t type; 768 769 /* XXX not right yet */ 770 STUB(); 771 return 0; 772 773 /* Bypass for VF */ 774 if (amdgpu_sriov_vf(adev)) 775 return 0; 776 #ifdef notyet 777 778 /* Check if the root BUS has 64bit memory resources */ 779 root = adev->pdev->bus; 780 while (root->parent) 781 root = root->parent; 782 783 pci_bus_for_each_resource(root, res, i) { 784 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) && 785 res->start > 0x100000000ull) 786 break; 787 } 788 789 /* Trying to resize is pointless without a root hub window above 4GB */ 790 if (!res) 791 return 0; 792 #endif 793 794 /* Disable memory decoding while we change the BAR addresses and size */ 795 pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd); 796 pci_write_config_word(adev->pdev, PCI_COMMAND, 797 cmd & ~PCI_COMMAND_MEMORY); 798 799 /* Free the VRAM and doorbell BAR, we most likely need to move both. */ 800 amdgpu_device_doorbell_fini(adev); 801 #ifdef __linux__ 802 if (adev->asic_type >= CHIP_BONAIRE) 803 pci_release_resource(adev->pdev, 2); 804 805 pci_release_resource(adev->pdev, 0); 806 #endif 807 808 r = pci_resize_resource(adev->pdev, 0, rbar_size); 809 if (r == -ENOSPC) 810 DRM_INFO("Not enough PCI address space for a large BAR."); 811 else if (r && r != -ENOTSUPP) 812 DRM_ERROR("Problem resizing BAR0 (%d).", r); 813 814 #ifdef __linux__ 815 pci_assign_unassigned_bus_resources(adev->pdev->bus); 816 #else 817 #define AMDGPU_PCI_MEM 0x10 818 819 type = pci_mapreg_type(adev->pc, adev->pa_tag, AMDGPU_PCI_MEM); 820 if (PCI_MAPREG_TYPE(type) != PCI_MAPREG_TYPE_MEM || 821 pci_mapreg_info(adev->pc, adev->pa_tag, AMDGPU_PCI_MEM, 822 type, NULL, &adev->fb_aper_size, NULL)) { 823 printf(": can't get frambuffer info\n"); 824 return -ENODEV; 825 } 826 #endif 827 /* When the doorbell or fb BAR isn't available we have no chance of 828 * using the device. 829 */ 830 r = amdgpu_device_doorbell_init(adev); 831 #ifdef notyet 832 if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET)) 833 #else 834 if (r) 835 #endif 836 return -ENODEV; 837 838 pci_write_config_word(adev->pdev, PCI_COMMAND, cmd); 839 840 return 0; 841 } 842 843 /* 844 * GPU helpers function. 845 */ 846 /** 847 * amdgpu_device_need_post - check if the hw need post or not 848 * 849 * @adev: amdgpu_device pointer 850 * 851 * Check if the asic has been initialized (all asics) at driver startup 852 * or post is needed if hw reset is performed. 853 * Returns true if need or false if not. 854 */ 855 bool amdgpu_device_need_post(struct amdgpu_device *adev) 856 { 857 uint32_t reg; 858 859 if (amdgpu_sriov_vf(adev)) 860 return false; 861 862 if (amdgpu_passthrough(adev)) { 863 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot 864 * some old smc fw still need driver do vPost otherwise gpu hang, while 865 * those smc fw version above 22.15 doesn't have this flaw, so we force 866 * vpost executed for smc version below 22.15 867 */ 868 if (adev->asic_type == CHIP_FIJI) { 869 int err; 870 uint32_t fw_ver; 871 err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev); 872 /* force vPost if error occured */ 873 if (err) 874 return true; 875 876 fw_ver = *((uint32_t *)adev->pm.fw->data + 69); 877 if (fw_ver < 0x00160e00) 878 return true; 879 } 880 } 881 882 if (adev->has_hw_reset) { 883 adev->has_hw_reset = false; 884 return true; 885 } 886 887 /* bios scratch used on CIK+ */ 888 if (adev->asic_type >= CHIP_BONAIRE) 889 return amdgpu_atombios_scratch_need_asic_init(adev); 890 891 /* check MEM_SIZE for older asics */ 892 reg = amdgpu_asic_get_config_memsize(adev); 893 894 if ((reg != 0) && (reg != 0xffffffff)) 895 return false; 896 897 return true; 898 } 899 900 /* if we get transitioned to only one device, take VGA back */ 901 /** 902 * amdgpu_device_vga_set_decode - enable/disable vga decode 903 * 904 * @cookie: amdgpu_device pointer 905 * @state: enable/disable vga decode 906 * 907 * Enable/disable vga decode (all asics). 908 * Returns VGA resource flags. 909 */ 910 #ifdef notyet 911 static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state) 912 { 913 struct amdgpu_device *adev = cookie; 914 amdgpu_asic_set_vga_state(adev, state); 915 if (state) 916 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM | 917 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM; 918 else 919 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM; 920 } 921 #endif 922 923 /** 924 * amdgpu_device_check_block_size - validate the vm block size 925 * 926 * @adev: amdgpu_device pointer 927 * 928 * Validates the vm block size specified via module parameter. 929 * The vm block size defines number of bits in page table versus page directory, 930 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the 931 * page table and the remaining bits are in the page directory. 932 */ 933 static void amdgpu_device_check_block_size(struct amdgpu_device *adev) 934 { 935 /* defines number of bits in page table versus page directory, 936 * a page is 4KB so we have 12 bits offset, minimum 9 bits in the 937 * page table and the remaining bits are in the page directory */ 938 if (amdgpu_vm_block_size == -1) 939 return; 940 941 if (amdgpu_vm_block_size < 9) { 942 dev_warn(adev->dev, "VM page table size (%d) too small\n", 943 amdgpu_vm_block_size); 944 amdgpu_vm_block_size = -1; 945 } 946 } 947 948 /** 949 * amdgpu_device_check_vm_size - validate the vm size 950 * 951 * @adev: amdgpu_device pointer 952 * 953 * Validates the vm size in GB specified via module parameter. 954 * The VM size is the size of the GPU virtual memory space in GB. 955 */ 956 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev) 957 { 958 /* no need to check the default value */ 959 if (amdgpu_vm_size == -1) 960 return; 961 962 if (amdgpu_vm_size < 1) { 963 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n", 964 amdgpu_vm_size); 965 amdgpu_vm_size = -1; 966 } 967 } 968 969 static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev) 970 { 971 #ifdef __linux__ 972 struct sysinfo si; 973 #endif 974 bool is_os_64 = (sizeof(void *) == 8) ? true : false; 975 uint64_t total_memory; 976 uint64_t dram_size_seven_GB = 0x1B8000000; 977 uint64_t dram_size_three_GB = 0xB8000000; 978 979 if (amdgpu_smu_memory_pool_size == 0) 980 return; 981 982 if (!is_os_64) { 983 DRM_WARN("Not 64-bit OS, feature not supported\n"); 984 goto def_value; 985 } 986 #ifdef __linux__ 987 si_meminfo(&si); 988 total_memory = (uint64_t)si.totalram * si.mem_unit; 989 #else 990 total_memory = ptoa(physmem); 991 #endif 992 993 if ((amdgpu_smu_memory_pool_size == 1) || 994 (amdgpu_smu_memory_pool_size == 2)) { 995 if (total_memory < dram_size_three_GB) 996 goto def_value1; 997 } else if ((amdgpu_smu_memory_pool_size == 4) || 998 (amdgpu_smu_memory_pool_size == 8)) { 999 if (total_memory < dram_size_seven_GB) 1000 goto def_value1; 1001 } else { 1002 DRM_WARN("Smu memory pool size not supported\n"); 1003 goto def_value; 1004 } 1005 adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28; 1006 1007 return; 1008 1009 def_value1: 1010 DRM_WARN("No enough system memory\n"); 1011 def_value: 1012 adev->pm.smu_prv_buffer_size = 0; 1013 } 1014 1015 /** 1016 * amdgpu_device_check_arguments - validate module params 1017 * 1018 * @adev: amdgpu_device pointer 1019 * 1020 * Validates certain module parameters and updates 1021 * the associated values used by the driver (all asics). 1022 */ 1023 static void amdgpu_device_check_arguments(struct amdgpu_device *adev) 1024 { 1025 if (amdgpu_sched_jobs < 4) { 1026 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n", 1027 amdgpu_sched_jobs); 1028 amdgpu_sched_jobs = 4; 1029 } else if (!is_power_of_2(amdgpu_sched_jobs)){ 1030 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n", 1031 amdgpu_sched_jobs); 1032 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs); 1033 } 1034 1035 if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) { 1036 /* gart size must be greater or equal to 32M */ 1037 dev_warn(adev->dev, "gart size (%d) too small\n", 1038 amdgpu_gart_size); 1039 amdgpu_gart_size = -1; 1040 } 1041 1042 if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) { 1043 /* gtt size must be greater or equal to 32M */ 1044 dev_warn(adev->dev, "gtt size (%d) too small\n", 1045 amdgpu_gtt_size); 1046 amdgpu_gtt_size = -1; 1047 } 1048 1049 /* valid range is between 4 and 9 inclusive */ 1050 if (amdgpu_vm_fragment_size != -1 && 1051 (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) { 1052 dev_warn(adev->dev, "valid range is between 4 and 9\n"); 1053 amdgpu_vm_fragment_size = -1; 1054 } 1055 1056 amdgpu_device_check_smu_prv_buffer_size(adev); 1057 1058 amdgpu_device_check_vm_size(adev); 1059 1060 amdgpu_device_check_block_size(adev); 1061 1062 if (amdgpu_vram_page_split != -1 && (amdgpu_vram_page_split < 16 || 1063 !is_power_of_2(amdgpu_vram_page_split))) { 1064 dev_warn(adev->dev, "invalid VRAM page split (%d)\n", 1065 amdgpu_vram_page_split); 1066 amdgpu_vram_page_split = 1024; 1067 } 1068 1069 if (amdgpu_lockup_timeout == 0) { 1070 dev_warn(adev->dev, "lockup_timeout msut be > 0, adjusting to 10000\n"); 1071 amdgpu_lockup_timeout = 10000; 1072 } 1073 1074 adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type); 1075 } 1076 1077 #ifdef __linux__ 1078 /** 1079 * amdgpu_switcheroo_set_state - set switcheroo state 1080 * 1081 * @pdev: pci dev pointer 1082 * @state: vga_switcheroo state 1083 * 1084 * Callback for the switcheroo driver. Suspends or resumes the 1085 * the asics before or after it is powered up using ACPI methods. 1086 */ 1087 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state) 1088 { 1089 struct drm_device *dev = pci_get_drvdata(pdev); 1090 1091 if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF) 1092 return; 1093 1094 if (state == VGA_SWITCHEROO_ON) { 1095 pr_info("amdgpu: switched on\n"); 1096 /* don't suspend or resume card normally */ 1097 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING; 1098 1099 amdgpu_device_resume(dev, true, true); 1100 1101 dev->switch_power_state = DRM_SWITCH_POWER_ON; 1102 drm_kms_helper_poll_enable(dev); 1103 } else { 1104 pr_info("amdgpu: switched off\n"); 1105 drm_kms_helper_poll_disable(dev); 1106 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING; 1107 amdgpu_device_suspend(dev, true, true); 1108 dev->switch_power_state = DRM_SWITCH_POWER_OFF; 1109 } 1110 } 1111 1112 /** 1113 * amdgpu_switcheroo_can_switch - see if switcheroo state can change 1114 * 1115 * @pdev: pci dev pointer 1116 * 1117 * Callback for the switcheroo driver. Check of the switcheroo 1118 * state can be changed. 1119 * Returns true if the state can be changed, false if not. 1120 */ 1121 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev) 1122 { 1123 struct drm_device *dev = pci_get_drvdata(pdev); 1124 1125 /* 1126 * FIXME: open_count is protected by drm_global_mutex but that would lead to 1127 * locking inversion with the driver load path. And the access here is 1128 * completely racy anyway. So don't bother with locking for now. 1129 */ 1130 return dev->open_count == 0; 1131 } 1132 1133 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = { 1134 .set_gpu_state = amdgpu_switcheroo_set_state, 1135 .reprobe = NULL, 1136 .can_switch = amdgpu_switcheroo_can_switch, 1137 }; 1138 #endif /* __linux__ */ 1139 1140 /** 1141 * amdgpu_device_ip_set_clockgating_state - set the CG state 1142 * 1143 * @dev: amdgpu_device pointer 1144 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) 1145 * @state: clockgating state (gate or ungate) 1146 * 1147 * Sets the requested clockgating state for all instances of 1148 * the hardware IP specified. 1149 * Returns the error code from the last instance. 1150 */ 1151 int amdgpu_device_ip_set_clockgating_state(void *dev, 1152 enum amd_ip_block_type block_type, 1153 enum amd_clockgating_state state) 1154 { 1155 struct amdgpu_device *adev = dev; 1156 int i, r = 0; 1157 1158 for (i = 0; i < adev->num_ip_blocks; i++) { 1159 if (!adev->ip_blocks[i].status.valid) 1160 continue; 1161 if (adev->ip_blocks[i].version->type != block_type) 1162 continue; 1163 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state) 1164 continue; 1165 r = adev->ip_blocks[i].version->funcs->set_clockgating_state( 1166 (void *)adev, state); 1167 if (r) 1168 DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n", 1169 adev->ip_blocks[i].version->funcs->name, r); 1170 } 1171 return r; 1172 } 1173 1174 /** 1175 * amdgpu_device_ip_set_powergating_state - set the PG state 1176 * 1177 * @dev: amdgpu_device pointer 1178 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) 1179 * @state: powergating state (gate or ungate) 1180 * 1181 * Sets the requested powergating state for all instances of 1182 * the hardware IP specified. 1183 * Returns the error code from the last instance. 1184 */ 1185 int amdgpu_device_ip_set_powergating_state(void *dev, 1186 enum amd_ip_block_type block_type, 1187 enum amd_powergating_state state) 1188 { 1189 struct amdgpu_device *adev = dev; 1190 int i, r = 0; 1191 1192 for (i = 0; i < adev->num_ip_blocks; i++) { 1193 if (!adev->ip_blocks[i].status.valid) 1194 continue; 1195 if (adev->ip_blocks[i].version->type != block_type) 1196 continue; 1197 if (!adev->ip_blocks[i].version->funcs->set_powergating_state) 1198 continue; 1199 r = adev->ip_blocks[i].version->funcs->set_powergating_state( 1200 (void *)adev, state); 1201 if (r) 1202 DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n", 1203 adev->ip_blocks[i].version->funcs->name, r); 1204 } 1205 return r; 1206 } 1207 1208 /** 1209 * amdgpu_device_ip_get_clockgating_state - get the CG state 1210 * 1211 * @adev: amdgpu_device pointer 1212 * @flags: clockgating feature flags 1213 * 1214 * Walks the list of IPs on the device and updates the clockgating 1215 * flags for each IP. 1216 * Updates @flags with the feature flags for each hardware IP where 1217 * clockgating is enabled. 1218 */ 1219 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev, 1220 u32 *flags) 1221 { 1222 int i; 1223 1224 for (i = 0; i < adev->num_ip_blocks; i++) { 1225 if (!adev->ip_blocks[i].status.valid) 1226 continue; 1227 if (adev->ip_blocks[i].version->funcs->get_clockgating_state) 1228 adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags); 1229 } 1230 } 1231 1232 /** 1233 * amdgpu_device_ip_wait_for_idle - wait for idle 1234 * 1235 * @adev: amdgpu_device pointer 1236 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) 1237 * 1238 * Waits for the request hardware IP to be idle. 1239 * Returns 0 for success or a negative error code on failure. 1240 */ 1241 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev, 1242 enum amd_ip_block_type block_type) 1243 { 1244 int i, r; 1245 1246 for (i = 0; i < adev->num_ip_blocks; i++) { 1247 if (!adev->ip_blocks[i].status.valid) 1248 continue; 1249 if (adev->ip_blocks[i].version->type == block_type) { 1250 r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev); 1251 if (r) 1252 return r; 1253 break; 1254 } 1255 } 1256 return 0; 1257 1258 } 1259 1260 /** 1261 * amdgpu_device_ip_is_idle - is the hardware IP idle 1262 * 1263 * @adev: amdgpu_device pointer 1264 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) 1265 * 1266 * Check if the hardware IP is idle or not. 1267 * Returns true if it the IP is idle, false if not. 1268 */ 1269 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev, 1270 enum amd_ip_block_type block_type) 1271 { 1272 int i; 1273 1274 for (i = 0; i < adev->num_ip_blocks; i++) { 1275 if (!adev->ip_blocks[i].status.valid) 1276 continue; 1277 if (adev->ip_blocks[i].version->type == block_type) 1278 return adev->ip_blocks[i].version->funcs->is_idle((void *)adev); 1279 } 1280 return true; 1281 1282 } 1283 1284 /** 1285 * amdgpu_device_ip_get_ip_block - get a hw IP pointer 1286 * 1287 * @adev: amdgpu_device pointer 1288 * @type: Type of hardware IP (SMU, GFX, UVD, etc.) 1289 * 1290 * Returns a pointer to the hardware IP block structure 1291 * if it exists for the asic, otherwise NULL. 1292 */ 1293 struct amdgpu_ip_block * 1294 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev, 1295 enum amd_ip_block_type type) 1296 { 1297 int i; 1298 1299 for (i = 0; i < adev->num_ip_blocks; i++) 1300 if (adev->ip_blocks[i].version->type == type) 1301 return &adev->ip_blocks[i]; 1302 1303 return NULL; 1304 } 1305 1306 /** 1307 * amdgpu_device_ip_block_version_cmp 1308 * 1309 * @adev: amdgpu_device pointer 1310 * @type: enum amd_ip_block_type 1311 * @major: major version 1312 * @minor: minor version 1313 * 1314 * return 0 if equal or greater 1315 * return 1 if smaller or the ip_block doesn't exist 1316 */ 1317 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev, 1318 enum amd_ip_block_type type, 1319 u32 major, u32 minor) 1320 { 1321 struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type); 1322 1323 if (ip_block && ((ip_block->version->major > major) || 1324 ((ip_block->version->major == major) && 1325 (ip_block->version->minor >= minor)))) 1326 return 0; 1327 1328 return 1; 1329 } 1330 1331 /** 1332 * amdgpu_device_ip_block_add 1333 * 1334 * @adev: amdgpu_device pointer 1335 * @ip_block_version: pointer to the IP to add 1336 * 1337 * Adds the IP block driver information to the collection of IPs 1338 * on the asic. 1339 */ 1340 int amdgpu_device_ip_block_add(struct amdgpu_device *adev, 1341 const struct amdgpu_ip_block_version *ip_block_version) 1342 { 1343 if (!ip_block_version) 1344 return -EINVAL; 1345 1346 DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks, 1347 ip_block_version->funcs->name); 1348 1349 adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version; 1350 1351 return 0; 1352 } 1353 1354 /** 1355 * amdgpu_device_enable_virtual_display - enable virtual display feature 1356 * 1357 * @adev: amdgpu_device pointer 1358 * 1359 * Enabled the virtual display feature if the user has enabled it via 1360 * the module parameter virtual_display. This feature provides a virtual 1361 * display hardware on headless boards or in virtualized environments. 1362 * This function parses and validates the configuration string specified by 1363 * the user and configues the virtual display configuration (number of 1364 * virtual connectors, crtcs, etc.) specified. 1365 */ 1366 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev) 1367 { 1368 adev->enable_virtual_display = false; 1369 1370 #ifdef notyet 1371 if (amdgpu_virtual_display) { 1372 struct drm_device *ddev = adev->ddev; 1373 const char *pci_address_name = pci_name(ddev->pdev); 1374 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname; 1375 1376 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL); 1377 pciaddstr_tmp = pciaddstr; 1378 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) { 1379 pciaddname = strsep(&pciaddname_tmp, ","); 1380 if (!strcmp("all", pciaddname) 1381 || !strcmp(pci_address_name, pciaddname)) { 1382 long num_crtc; 1383 int res = -1; 1384 1385 adev->enable_virtual_display = true; 1386 1387 if (pciaddname_tmp) 1388 res = kstrtol(pciaddname_tmp, 10, 1389 &num_crtc); 1390 1391 if (!res) { 1392 if (num_crtc < 1) 1393 num_crtc = 1; 1394 if (num_crtc > 6) 1395 num_crtc = 6; 1396 adev->mode_info.num_crtc = num_crtc; 1397 } else { 1398 adev->mode_info.num_crtc = 1; 1399 } 1400 break; 1401 } 1402 } 1403 1404 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n", 1405 amdgpu_virtual_display, pci_address_name, 1406 adev->enable_virtual_display, adev->mode_info.num_crtc); 1407 1408 kfree(pciaddstr); 1409 } 1410 #endif 1411 } 1412 1413 /** 1414 * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware 1415 * 1416 * @adev: amdgpu_device pointer 1417 * 1418 * Parses the asic configuration parameters specified in the gpu info 1419 * firmware and makes them availale to the driver for use in configuring 1420 * the asic. 1421 * Returns 0 on success, -EINVAL on failure. 1422 */ 1423 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev) 1424 { 1425 const char *chip_name; 1426 char fw_name[30]; 1427 int err; 1428 const struct gpu_info_firmware_header_v1_0 *hdr; 1429 1430 adev->firmware.gpu_info_fw = NULL; 1431 1432 switch (adev->asic_type) { 1433 case CHIP_TOPAZ: 1434 case CHIP_TONGA: 1435 case CHIP_FIJI: 1436 case CHIP_POLARIS10: 1437 case CHIP_POLARIS11: 1438 case CHIP_POLARIS12: 1439 case CHIP_VEGAM: 1440 case CHIP_CARRIZO: 1441 case CHIP_STONEY: 1442 #ifdef CONFIG_DRM_AMDGPU_SI 1443 case CHIP_VERDE: 1444 case CHIP_TAHITI: 1445 case CHIP_PITCAIRN: 1446 case CHIP_OLAND: 1447 case CHIP_HAINAN: 1448 #endif 1449 #ifdef CONFIG_DRM_AMDGPU_CIK 1450 case CHIP_BONAIRE: 1451 case CHIP_HAWAII: 1452 case CHIP_KAVERI: 1453 case CHIP_KABINI: 1454 case CHIP_MULLINS: 1455 #endif 1456 case CHIP_VEGA20: 1457 default: 1458 return 0; 1459 case CHIP_VEGA10: 1460 chip_name = "vega10"; 1461 break; 1462 case CHIP_VEGA12: 1463 chip_name = "vega12"; 1464 break; 1465 case CHIP_RAVEN: 1466 chip_name = "raven"; 1467 break; 1468 } 1469 1470 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name); 1471 err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev); 1472 if (err) { 1473 dev_err(adev->dev, 1474 "Failed to load gpu_info firmware \"%s\"\n", 1475 fw_name); 1476 goto out; 1477 } 1478 err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw); 1479 if (err) { 1480 dev_err(adev->dev, 1481 "Failed to validate gpu_info firmware \"%s\"\n", 1482 fw_name); 1483 goto out; 1484 } 1485 1486 hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data; 1487 amdgpu_ucode_print_gpu_info_hdr(&hdr->header); 1488 1489 switch (hdr->version_major) { 1490 case 1: 1491 { 1492 const struct gpu_info_firmware_v1_0 *gpu_info_fw = 1493 (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data + 1494 le32_to_cpu(hdr->header.ucode_array_offset_bytes)); 1495 1496 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se); 1497 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh); 1498 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se); 1499 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se); 1500 adev->gfx.config.max_texture_channel_caches = 1501 le32_to_cpu(gpu_info_fw->gc_num_tccs); 1502 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs); 1503 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds); 1504 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth); 1505 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth); 1506 adev->gfx.config.double_offchip_lds_buf = 1507 le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer); 1508 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size); 1509 adev->gfx.cu_info.max_waves_per_simd = 1510 le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd); 1511 adev->gfx.cu_info.max_scratch_slots_per_cu = 1512 le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu); 1513 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size); 1514 break; 1515 } 1516 default: 1517 dev_err(adev->dev, 1518 "Unsupported gpu_info table %d\n", hdr->header.ucode_version); 1519 err = -EINVAL; 1520 goto out; 1521 } 1522 out: 1523 return err; 1524 } 1525 1526 /** 1527 * amdgpu_device_ip_early_init - run early init for hardware IPs 1528 * 1529 * @adev: amdgpu_device pointer 1530 * 1531 * Early initialization pass for hardware IPs. The hardware IPs that make 1532 * up each asic are discovered each IP's early_init callback is run. This 1533 * is the first stage in initializing the asic. 1534 * Returns 0 on success, negative error code on failure. 1535 */ 1536 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev) 1537 { 1538 int i, r; 1539 1540 amdgpu_device_enable_virtual_display(adev); 1541 1542 switch (adev->asic_type) { 1543 case CHIP_TOPAZ: 1544 case CHIP_TONGA: 1545 case CHIP_FIJI: 1546 case CHIP_POLARIS10: 1547 case CHIP_POLARIS11: 1548 case CHIP_POLARIS12: 1549 case CHIP_VEGAM: 1550 case CHIP_CARRIZO: 1551 case CHIP_STONEY: 1552 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY) 1553 adev->family = AMDGPU_FAMILY_CZ; 1554 else 1555 adev->family = AMDGPU_FAMILY_VI; 1556 1557 r = vi_set_ip_blocks(adev); 1558 if (r) 1559 return r; 1560 break; 1561 #ifdef CONFIG_DRM_AMDGPU_SI 1562 case CHIP_VERDE: 1563 case CHIP_TAHITI: 1564 case CHIP_PITCAIRN: 1565 case CHIP_OLAND: 1566 case CHIP_HAINAN: 1567 adev->family = AMDGPU_FAMILY_SI; 1568 r = si_set_ip_blocks(adev); 1569 if (r) 1570 return r; 1571 break; 1572 #endif 1573 #ifdef CONFIG_DRM_AMDGPU_CIK 1574 case CHIP_BONAIRE: 1575 case CHIP_HAWAII: 1576 case CHIP_KAVERI: 1577 case CHIP_KABINI: 1578 case CHIP_MULLINS: 1579 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII)) 1580 adev->family = AMDGPU_FAMILY_CI; 1581 else 1582 adev->family = AMDGPU_FAMILY_KV; 1583 1584 r = cik_set_ip_blocks(adev); 1585 if (r) 1586 return r; 1587 break; 1588 #endif 1589 case CHIP_VEGA10: 1590 case CHIP_VEGA12: 1591 case CHIP_VEGA20: 1592 case CHIP_RAVEN: 1593 if (adev->asic_type == CHIP_RAVEN) 1594 adev->family = AMDGPU_FAMILY_RV; 1595 else 1596 adev->family = AMDGPU_FAMILY_AI; 1597 1598 r = soc15_set_ip_blocks(adev); 1599 if (r) 1600 return r; 1601 break; 1602 default: 1603 /* FIXME: not supported yet */ 1604 return -EINVAL; 1605 } 1606 1607 r = amdgpu_device_parse_gpu_info_fw(adev); 1608 if (r) 1609 return r; 1610 1611 amdgpu_amdkfd_device_probe(adev); 1612 1613 if (amdgpu_sriov_vf(adev)) { 1614 r = amdgpu_virt_request_full_gpu(adev, true); 1615 if (r) 1616 return -EAGAIN; 1617 } 1618 1619 adev->powerplay.pp_feature = amdgpu_pp_feature_mask; 1620 1621 for (i = 0; i < adev->num_ip_blocks; i++) { 1622 if ((amdgpu_ip_block_mask & (1 << i)) == 0) { 1623 DRM_ERROR("disabled ip block: %d <%s>\n", 1624 i, adev->ip_blocks[i].version->funcs->name); 1625 adev->ip_blocks[i].status.valid = false; 1626 } else { 1627 if (adev->ip_blocks[i].version->funcs->early_init) { 1628 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev); 1629 if (r == -ENOENT) { 1630 adev->ip_blocks[i].status.valid = false; 1631 } else if (r) { 1632 DRM_ERROR("early_init of IP block <%s> failed %d\n", 1633 adev->ip_blocks[i].version->funcs->name, r); 1634 return r; 1635 } else { 1636 adev->ip_blocks[i].status.valid = true; 1637 } 1638 } else { 1639 adev->ip_blocks[i].status.valid = true; 1640 } 1641 } 1642 } 1643 1644 adev->cg_flags &= amdgpu_cg_mask; 1645 adev->pg_flags &= amdgpu_pg_mask; 1646 1647 return 0; 1648 } 1649 1650 /** 1651 * amdgpu_device_ip_init - run init for hardware IPs 1652 * 1653 * @adev: amdgpu_device pointer 1654 * 1655 * Main initialization pass for hardware IPs. The list of all the hardware 1656 * IPs that make up the asic is walked and the sw_init and hw_init callbacks 1657 * are run. sw_init initializes the software state associated with each IP 1658 * and hw_init initializes the hardware associated with each IP. 1659 * Returns 0 on success, negative error code on failure. 1660 */ 1661 static int amdgpu_device_ip_init(struct amdgpu_device *adev) 1662 { 1663 int i, r; 1664 1665 for (i = 0; i < adev->num_ip_blocks; i++) { 1666 if (!adev->ip_blocks[i].status.valid) 1667 continue; 1668 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev); 1669 if (r) { 1670 DRM_ERROR("sw_init of IP block <%s> failed %d\n", 1671 adev->ip_blocks[i].version->funcs->name, r); 1672 return r; 1673 } 1674 adev->ip_blocks[i].status.sw = true; 1675 1676 /* need to do gmc hw init early so we can allocate gpu mem */ 1677 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) { 1678 r = amdgpu_device_vram_scratch_init(adev); 1679 if (r) { 1680 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r); 1681 return r; 1682 } 1683 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev); 1684 if (r) { 1685 DRM_ERROR("hw_init %d failed %d\n", i, r); 1686 return r; 1687 } 1688 r = amdgpu_device_wb_init(adev); 1689 if (r) { 1690 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r); 1691 return r; 1692 } 1693 adev->ip_blocks[i].status.hw = true; 1694 1695 /* right after GMC hw init, we create CSA */ 1696 if (amdgpu_sriov_vf(adev)) { 1697 r = amdgpu_allocate_static_csa(adev); 1698 if (r) { 1699 DRM_ERROR("allocate CSA failed %d\n", r); 1700 return r; 1701 } 1702 } 1703 } 1704 } 1705 1706 for (i = 0; i < adev->num_ip_blocks; i++) { 1707 if (!adev->ip_blocks[i].status.sw) 1708 continue; 1709 if (adev->ip_blocks[i].status.hw) 1710 continue; 1711 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev); 1712 if (r) { 1713 DRM_ERROR("hw_init of IP block <%s> failed %d\n", 1714 adev->ip_blocks[i].version->funcs->name, r); 1715 return r; 1716 } 1717 adev->ip_blocks[i].status.hw = true; 1718 } 1719 1720 amdgpu_amdkfd_device_init(adev); 1721 1722 if (amdgpu_sriov_vf(adev)) { 1723 amdgpu_virt_init_data_exchange(adev); 1724 amdgpu_virt_release_full_gpu(adev, true); 1725 } 1726 1727 return 0; 1728 } 1729 1730 /** 1731 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer 1732 * 1733 * @adev: amdgpu_device pointer 1734 * 1735 * Writes a reset magic value to the gart pointer in VRAM. The driver calls 1736 * this function before a GPU reset. If the value is retained after a 1737 * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents. 1738 */ 1739 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev) 1740 { 1741 memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM); 1742 } 1743 1744 /** 1745 * amdgpu_device_check_vram_lost - check if vram is valid 1746 * 1747 * @adev: amdgpu_device pointer 1748 * 1749 * Checks the reset magic value written to the gart pointer in VRAM. 1750 * The driver calls this after a GPU reset to see if the contents of 1751 * VRAM is lost or now. 1752 * returns true if vram is lost, false if not. 1753 */ 1754 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev) 1755 { 1756 return !!memcmp(adev->gart.ptr, adev->reset_magic, 1757 AMDGPU_RESET_MAGIC_NUM); 1758 } 1759 1760 /** 1761 * amdgpu_device_ip_late_set_cg_state - late init for clockgating 1762 * 1763 * @adev: amdgpu_device pointer 1764 * 1765 * Late initialization pass enabling clockgating for hardware IPs. 1766 * The list of all the hardware IPs that make up the asic is walked and the 1767 * set_clockgating_state callbacks are run. This stage is run late 1768 * in the init process. 1769 * Returns 0 on success, negative error code on failure. 1770 */ 1771 static int amdgpu_device_ip_late_set_cg_state(struct amdgpu_device *adev) 1772 { 1773 int i = 0, r; 1774 1775 if (amdgpu_emu_mode == 1) 1776 return 0; 1777 1778 for (i = 0; i < adev->num_ip_blocks; i++) { 1779 if (!adev->ip_blocks[i].status.valid) 1780 continue; 1781 /* skip CG for VCE/UVD, it's handled specially */ 1782 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD && 1783 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE && 1784 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN && 1785 adev->ip_blocks[i].version->funcs->set_clockgating_state) { 1786 /* enable clockgating to save power */ 1787 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev, 1788 AMD_CG_STATE_GATE); 1789 if (r) { 1790 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n", 1791 adev->ip_blocks[i].version->funcs->name, r); 1792 return r; 1793 } 1794 } 1795 } 1796 1797 return 0; 1798 } 1799 1800 static int amdgpu_device_ip_late_set_pg_state(struct amdgpu_device *adev) 1801 { 1802 int i = 0, r; 1803 1804 if (amdgpu_emu_mode == 1) 1805 return 0; 1806 1807 for (i = 0; i < adev->num_ip_blocks; i++) { 1808 if (!adev->ip_blocks[i].status.valid) 1809 continue; 1810 /* skip CG for VCE/UVD, it's handled specially */ 1811 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD && 1812 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE && 1813 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN && 1814 adev->ip_blocks[i].version->funcs->set_powergating_state) { 1815 /* enable powergating to save power */ 1816 r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev, 1817 AMD_PG_STATE_GATE); 1818 if (r) { 1819 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n", 1820 adev->ip_blocks[i].version->funcs->name, r); 1821 return r; 1822 } 1823 } 1824 } 1825 return 0; 1826 } 1827 1828 /** 1829 * amdgpu_device_ip_late_init - run late init for hardware IPs 1830 * 1831 * @adev: amdgpu_device pointer 1832 * 1833 * Late initialization pass for hardware IPs. The list of all the hardware 1834 * IPs that make up the asic is walked and the late_init callbacks are run. 1835 * late_init covers any special initialization that an IP requires 1836 * after all of the have been initialized or something that needs to happen 1837 * late in the init process. 1838 * Returns 0 on success, negative error code on failure. 1839 */ 1840 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev) 1841 { 1842 int i = 0, r; 1843 1844 for (i = 0; i < adev->num_ip_blocks; i++) { 1845 if (!adev->ip_blocks[i].status.valid) 1846 continue; 1847 if (adev->ip_blocks[i].version->funcs->late_init) { 1848 r = adev->ip_blocks[i].version->funcs->late_init((void *)adev); 1849 if (r) { 1850 DRM_ERROR("late_init of IP block <%s> failed %d\n", 1851 adev->ip_blocks[i].version->funcs->name, r); 1852 return r; 1853 } 1854 adev->ip_blocks[i].status.late_initialized = true; 1855 } 1856 } 1857 1858 amdgpu_device_ip_late_set_cg_state(adev); 1859 amdgpu_device_ip_late_set_pg_state(adev); 1860 1861 queue_delayed_work(system_wq, &adev->late_init_work, 1862 msecs_to_jiffies(AMDGPU_RESUME_MS)); 1863 1864 amdgpu_device_fill_reset_magic(adev); 1865 1866 return 0; 1867 } 1868 1869 /** 1870 * amdgpu_device_ip_fini - run fini for hardware IPs 1871 * 1872 * @adev: amdgpu_device pointer 1873 * 1874 * Main teardown pass for hardware IPs. The list of all the hardware 1875 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks 1876 * are run. hw_fini tears down the hardware associated with each IP 1877 * and sw_fini tears down any software state associated with each IP. 1878 * Returns 0 on success, negative error code on failure. 1879 */ 1880 static int amdgpu_device_ip_fini(struct amdgpu_device *adev) 1881 { 1882 int i, r; 1883 1884 amdgpu_amdkfd_device_fini(adev); 1885 /* need to disable SMC first */ 1886 for (i = 0; i < adev->num_ip_blocks; i++) { 1887 if (!adev->ip_blocks[i].status.hw) 1888 continue; 1889 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC && 1890 adev->ip_blocks[i].version->funcs->set_clockgating_state) { 1891 /* ungate blocks before hw fini so that we can shutdown the blocks safely */ 1892 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev, 1893 AMD_CG_STATE_UNGATE); 1894 if (r) { 1895 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n", 1896 adev->ip_blocks[i].version->funcs->name, r); 1897 return r; 1898 } 1899 if (adev->powerplay.pp_funcs->set_powergating_by_smu) 1900 amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false); 1901 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev); 1902 /* XXX handle errors */ 1903 if (r) { 1904 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n", 1905 adev->ip_blocks[i].version->funcs->name, r); 1906 } 1907 adev->ip_blocks[i].status.hw = false; 1908 break; 1909 } 1910 } 1911 1912 for (i = adev->num_ip_blocks - 1; i >= 0; i--) { 1913 if (!adev->ip_blocks[i].status.hw) 1914 continue; 1915 1916 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD && 1917 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE && 1918 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN && 1919 adev->ip_blocks[i].version->funcs->set_clockgating_state) { 1920 /* ungate blocks before hw fini so that we can shutdown the blocks safely */ 1921 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev, 1922 AMD_CG_STATE_UNGATE); 1923 if (r) { 1924 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n", 1925 adev->ip_blocks[i].version->funcs->name, r); 1926 return r; 1927 } 1928 } 1929 1930 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev); 1931 /* XXX handle errors */ 1932 if (r) { 1933 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n", 1934 adev->ip_blocks[i].version->funcs->name, r); 1935 } 1936 1937 adev->ip_blocks[i].status.hw = false; 1938 } 1939 1940 1941 for (i = adev->num_ip_blocks - 1; i >= 0; i--) { 1942 if (!adev->ip_blocks[i].status.sw) 1943 continue; 1944 1945 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) { 1946 amdgpu_free_static_csa(adev); 1947 amdgpu_device_wb_fini(adev); 1948 amdgpu_device_vram_scratch_fini(adev); 1949 } 1950 1951 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev); 1952 /* XXX handle errors */ 1953 if (r) { 1954 DRM_DEBUG("sw_fini of IP block <%s> failed %d\n", 1955 adev->ip_blocks[i].version->funcs->name, r); 1956 } 1957 adev->ip_blocks[i].status.sw = false; 1958 adev->ip_blocks[i].status.valid = false; 1959 } 1960 1961 for (i = adev->num_ip_blocks - 1; i >= 0; i--) { 1962 if (!adev->ip_blocks[i].status.late_initialized) 1963 continue; 1964 if (adev->ip_blocks[i].version->funcs->late_fini) 1965 adev->ip_blocks[i].version->funcs->late_fini((void *)adev); 1966 adev->ip_blocks[i].status.late_initialized = false; 1967 } 1968 1969 if (amdgpu_sriov_vf(adev)) 1970 if (amdgpu_virt_release_full_gpu(adev, false)) 1971 DRM_ERROR("failed to release exclusive mode on fini\n"); 1972 1973 return 0; 1974 } 1975 1976 /** 1977 * amdgpu_device_ip_late_init_func_handler - work handler for clockgating 1978 * 1979 * @work: work_struct 1980 * 1981 * Work handler for amdgpu_device_ip_late_set_cg_state. We put the 1982 * clockgating setup into a worker thread to speed up driver init and 1983 * resume from suspend. 1984 */ 1985 static void amdgpu_device_ip_late_init_func_handler(struct work_struct *work) 1986 { 1987 struct amdgpu_device *adev = 1988 container_of(work, struct amdgpu_device, late_init_work.work); 1989 int r; 1990 1991 r = amdgpu_ib_ring_tests(adev); 1992 if (r) 1993 DRM_ERROR("ib ring test failed (%d).\n", r); 1994 } 1995 1996 /** 1997 * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1) 1998 * 1999 * @adev: amdgpu_device pointer 2000 * 2001 * Main suspend function for hardware IPs. The list of all the hardware 2002 * IPs that make up the asic is walked, clockgating is disabled and the 2003 * suspend callbacks are run. suspend puts the hardware and software state 2004 * in each IP into a state suitable for suspend. 2005 * Returns 0 on success, negative error code on failure. 2006 */ 2007 static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev) 2008 { 2009 int i, r; 2010 2011 if (amdgpu_sriov_vf(adev)) 2012 amdgpu_virt_request_full_gpu(adev, false); 2013 2014 for (i = adev->num_ip_blocks - 1; i >= 0; i--) { 2015 if (!adev->ip_blocks[i].status.valid) 2016 continue; 2017 /* displays are handled separately */ 2018 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) { 2019 /* ungate blocks so that suspend can properly shut them down */ 2020 if (adev->ip_blocks[i].version->funcs->set_clockgating_state) { 2021 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev, 2022 AMD_CG_STATE_UNGATE); 2023 if (r) { 2024 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n", 2025 adev->ip_blocks[i].version->funcs->name, r); 2026 } 2027 } 2028 /* XXX handle errors */ 2029 r = adev->ip_blocks[i].version->funcs->suspend(adev); 2030 /* XXX handle errors */ 2031 if (r) { 2032 DRM_ERROR("suspend of IP block <%s> failed %d\n", 2033 adev->ip_blocks[i].version->funcs->name, r); 2034 } 2035 } 2036 } 2037 2038 if (amdgpu_sriov_vf(adev)) 2039 amdgpu_virt_release_full_gpu(adev, false); 2040 2041 return 0; 2042 } 2043 2044 /** 2045 * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2) 2046 * 2047 * @adev: amdgpu_device pointer 2048 * 2049 * Main suspend function for hardware IPs. The list of all the hardware 2050 * IPs that make up the asic is walked, clockgating is disabled and the 2051 * suspend callbacks are run. suspend puts the hardware and software state 2052 * in each IP into a state suitable for suspend. 2053 * Returns 0 on success, negative error code on failure. 2054 */ 2055 static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev) 2056 { 2057 int i, r; 2058 2059 if (amdgpu_sriov_vf(adev)) 2060 amdgpu_virt_request_full_gpu(adev, false); 2061 2062 /* ungate SMC block first */ 2063 r = amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_SMC, 2064 AMD_CG_STATE_UNGATE); 2065 if (r) { 2066 DRM_ERROR("set_clockgating_state(ungate) SMC failed %d\n", r); 2067 } 2068 2069 /* call smu to disable gfx off feature first when suspend */ 2070 if (adev->powerplay.pp_funcs->set_powergating_by_smu) 2071 amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false); 2072 2073 for (i = adev->num_ip_blocks - 1; i >= 0; i--) { 2074 if (!adev->ip_blocks[i].status.valid) 2075 continue; 2076 /* displays are handled in phase1 */ 2077 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) 2078 continue; 2079 /* ungate blocks so that suspend can properly shut them down */ 2080 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_SMC && 2081 adev->ip_blocks[i].version->funcs->set_clockgating_state) { 2082 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev, 2083 AMD_CG_STATE_UNGATE); 2084 if (r) { 2085 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n", 2086 adev->ip_blocks[i].version->funcs->name, r); 2087 } 2088 } 2089 /* XXX handle errors */ 2090 r = adev->ip_blocks[i].version->funcs->suspend(adev); 2091 /* XXX handle errors */ 2092 if (r) { 2093 DRM_ERROR("suspend of IP block <%s> failed %d\n", 2094 adev->ip_blocks[i].version->funcs->name, r); 2095 } 2096 } 2097 2098 if (amdgpu_sriov_vf(adev)) 2099 amdgpu_virt_release_full_gpu(adev, false); 2100 2101 return 0; 2102 } 2103 2104 /** 2105 * amdgpu_device_ip_suspend - run suspend for hardware IPs 2106 * 2107 * @adev: amdgpu_device pointer 2108 * 2109 * Main suspend function for hardware IPs. The list of all the hardware 2110 * IPs that make up the asic is walked, clockgating is disabled and the 2111 * suspend callbacks are run. suspend puts the hardware and software state 2112 * in each IP into a state suitable for suspend. 2113 * Returns 0 on success, negative error code on failure. 2114 */ 2115 int amdgpu_device_ip_suspend(struct amdgpu_device *adev) 2116 { 2117 int r; 2118 2119 r = amdgpu_device_ip_suspend_phase1(adev); 2120 if (r) 2121 return r; 2122 r = amdgpu_device_ip_suspend_phase2(adev); 2123 2124 return r; 2125 } 2126 2127 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev) 2128 { 2129 int i, r; 2130 2131 static enum amd_ip_block_type ip_order[] = { 2132 AMD_IP_BLOCK_TYPE_GMC, 2133 AMD_IP_BLOCK_TYPE_COMMON, 2134 AMD_IP_BLOCK_TYPE_PSP, 2135 AMD_IP_BLOCK_TYPE_IH, 2136 }; 2137 2138 for (i = 0; i < ARRAY_SIZE(ip_order); i++) { 2139 int j; 2140 struct amdgpu_ip_block *block; 2141 2142 for (j = 0; j < adev->num_ip_blocks; j++) { 2143 block = &adev->ip_blocks[j]; 2144 2145 if (block->version->type != ip_order[i] || 2146 !block->status.valid) 2147 continue; 2148 2149 r = block->version->funcs->hw_init(adev); 2150 DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"succeeded"); 2151 if (r) 2152 return r; 2153 } 2154 } 2155 2156 return 0; 2157 } 2158 2159 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev) 2160 { 2161 int i, r; 2162 2163 static enum amd_ip_block_type ip_order[] = { 2164 AMD_IP_BLOCK_TYPE_SMC, 2165 AMD_IP_BLOCK_TYPE_DCE, 2166 AMD_IP_BLOCK_TYPE_GFX, 2167 AMD_IP_BLOCK_TYPE_SDMA, 2168 AMD_IP_BLOCK_TYPE_UVD, 2169 AMD_IP_BLOCK_TYPE_VCE 2170 }; 2171 2172 for (i = 0; i < ARRAY_SIZE(ip_order); i++) { 2173 int j; 2174 struct amdgpu_ip_block *block; 2175 2176 for (j = 0; j < adev->num_ip_blocks; j++) { 2177 block = &adev->ip_blocks[j]; 2178 2179 if (block->version->type != ip_order[i] || 2180 !block->status.valid) 2181 continue; 2182 2183 r = block->version->funcs->hw_init(adev); 2184 DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"succeeded"); 2185 if (r) 2186 return r; 2187 } 2188 } 2189 2190 return 0; 2191 } 2192 2193 /** 2194 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs 2195 * 2196 * @adev: amdgpu_device pointer 2197 * 2198 * First resume function for hardware IPs. The list of all the hardware 2199 * IPs that make up the asic is walked and the resume callbacks are run for 2200 * COMMON, GMC, and IH. resume puts the hardware into a functional state 2201 * after a suspend and updates the software state as necessary. This 2202 * function is also used for restoring the GPU after a GPU reset. 2203 * Returns 0 on success, negative error code on failure. 2204 */ 2205 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev) 2206 { 2207 int i, r; 2208 2209 for (i = 0; i < adev->num_ip_blocks; i++) { 2210 if (!adev->ip_blocks[i].status.valid) 2211 continue; 2212 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || 2213 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || 2214 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) { 2215 r = adev->ip_blocks[i].version->funcs->resume(adev); 2216 if (r) { 2217 DRM_ERROR("resume of IP block <%s> failed %d\n", 2218 adev->ip_blocks[i].version->funcs->name, r); 2219 return r; 2220 } 2221 } 2222 } 2223 2224 return 0; 2225 } 2226 2227 /** 2228 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs 2229 * 2230 * @adev: amdgpu_device pointer 2231 * 2232 * First resume function for hardware IPs. The list of all the hardware 2233 * IPs that make up the asic is walked and the resume callbacks are run for 2234 * all blocks except COMMON, GMC, and IH. resume puts the hardware into a 2235 * functional state after a suspend and updates the software state as 2236 * necessary. This function is also used for restoring the GPU after a GPU 2237 * reset. 2238 * Returns 0 on success, negative error code on failure. 2239 */ 2240 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev) 2241 { 2242 int i, r; 2243 2244 for (i = 0; i < adev->num_ip_blocks; i++) { 2245 if (!adev->ip_blocks[i].status.valid) 2246 continue; 2247 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || 2248 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || 2249 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) 2250 continue; 2251 r = adev->ip_blocks[i].version->funcs->resume(adev); 2252 if (r) { 2253 DRM_ERROR("resume of IP block <%s> failed %d\n", 2254 adev->ip_blocks[i].version->funcs->name, r); 2255 return r; 2256 } 2257 } 2258 2259 return 0; 2260 } 2261 2262 /** 2263 * amdgpu_device_ip_resume - run resume for hardware IPs 2264 * 2265 * @adev: amdgpu_device pointer 2266 * 2267 * Main resume function for hardware IPs. The hardware IPs 2268 * are split into two resume functions because they are 2269 * are also used in in recovering from a GPU reset and some additional 2270 * steps need to be take between them. In this case (S3/S4) they are 2271 * run sequentially. 2272 * Returns 0 on success, negative error code on failure. 2273 */ 2274 static int amdgpu_device_ip_resume(struct amdgpu_device *adev) 2275 { 2276 int r; 2277 2278 r = amdgpu_device_ip_resume_phase1(adev); 2279 if (r) 2280 return r; 2281 r = amdgpu_device_ip_resume_phase2(adev); 2282 2283 return r; 2284 } 2285 2286 /** 2287 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV 2288 * 2289 * @adev: amdgpu_device pointer 2290 * 2291 * Query the VBIOS data tables to determine if the board supports SR-IOV. 2292 */ 2293 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev) 2294 { 2295 if (amdgpu_sriov_vf(adev)) { 2296 if (adev->is_atom_fw) { 2297 if (amdgpu_atomfirmware_gpu_supports_virtualization(adev)) 2298 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS; 2299 } else { 2300 if (amdgpu_atombios_has_gpu_virtualization_table(adev)) 2301 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS; 2302 } 2303 2304 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS)) 2305 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0); 2306 } 2307 } 2308 2309 /** 2310 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic 2311 * 2312 * @asic_type: AMD asic type 2313 * 2314 * Check if there is DC (new modesetting infrastructre) support for an asic. 2315 * returns true if DC has support, false if not. 2316 */ 2317 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type) 2318 { 2319 switch (asic_type) { 2320 #if defined(CONFIG_DRM_AMD_DC) 2321 case CHIP_BONAIRE: 2322 case CHIP_KAVERI: 2323 case CHIP_KABINI: 2324 case CHIP_MULLINS: 2325 /* 2326 * We have systems in the wild with these ASICs that require 2327 * LVDS and VGA support which is not supported with DC. 2328 * 2329 * Fallback to the non-DC driver here by default so as not to 2330 * cause regressions. 2331 */ 2332 return amdgpu_dc > 0; 2333 case CHIP_HAWAII: 2334 case CHIP_CARRIZO: 2335 case CHIP_STONEY: 2336 case CHIP_POLARIS10: 2337 case CHIP_POLARIS11: 2338 case CHIP_POLARIS12: 2339 case CHIP_VEGAM: 2340 case CHIP_TONGA: 2341 case CHIP_FIJI: 2342 case CHIP_VEGA10: 2343 case CHIP_VEGA12: 2344 case CHIP_VEGA20: 2345 #if defined(CONFIG_DRM_AMD_DC_DCN1_0) 2346 case CHIP_RAVEN: 2347 #endif 2348 return amdgpu_dc != 0; 2349 #endif 2350 default: 2351 return false; 2352 } 2353 } 2354 2355 /** 2356 * amdgpu_device_has_dc_support - check if dc is supported 2357 * 2358 * @adev: amdgpu_device_pointer 2359 * 2360 * Returns true for supported, false for not supported 2361 */ 2362 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev) 2363 { 2364 if (amdgpu_sriov_vf(adev)) 2365 return false; 2366 2367 return amdgpu_device_asic_has_dc_support(adev->asic_type); 2368 } 2369 2370 /** 2371 * amdgpu_device_init - initialize the driver 2372 * 2373 * @adev: amdgpu_device pointer 2374 * @ddev: drm dev pointer 2375 * @pdev: pci dev pointer 2376 * @flags: driver flags 2377 * 2378 * Initializes the driver info and hw (all asics). 2379 * Returns 0 for success or an error on failure. 2380 * Called at driver startup. 2381 */ 2382 int amdgpu_device_init(struct amdgpu_device *adev, 2383 struct drm_device *ddev, 2384 struct pci_dev *pdev, 2385 uint32_t flags) 2386 { 2387 int r, i; 2388 bool runtime = false; 2389 u32 max_MBps; 2390 2391 adev->shutdown = false; 2392 #ifdef __linux__ 2393 adev->dev = &pdev->dev; 2394 #endif 2395 adev->ddev = ddev; 2396 adev->pdev = pdev; 2397 adev->flags = flags; 2398 adev->asic_type = flags & AMD_ASIC_MASK; 2399 adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT; 2400 if (amdgpu_emu_mode == 1) 2401 adev->usec_timeout *= 2; 2402 adev->gmc.gart_size = 512 * 1024 * 1024; 2403 adev->accel_working = false; 2404 adev->num_rings = 0; 2405 adev->mman.buffer_funcs = NULL; 2406 adev->mman.buffer_funcs_ring = NULL; 2407 adev->vm_manager.vm_pte_funcs = NULL; 2408 adev->vm_manager.vm_pte_num_rings = 0; 2409 adev->gmc.gmc_funcs = NULL; 2410 adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS); 2411 bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES); 2412 2413 adev->smc_rreg = &amdgpu_invalid_rreg; 2414 adev->smc_wreg = &amdgpu_invalid_wreg; 2415 adev->pcie_rreg = &amdgpu_invalid_rreg; 2416 adev->pcie_wreg = &amdgpu_invalid_wreg; 2417 adev->pciep_rreg = &amdgpu_invalid_rreg; 2418 adev->pciep_wreg = &amdgpu_invalid_wreg; 2419 adev->uvd_ctx_rreg = &amdgpu_invalid_rreg; 2420 adev->uvd_ctx_wreg = &amdgpu_invalid_wreg; 2421 adev->didt_rreg = &amdgpu_invalid_rreg; 2422 adev->didt_wreg = &amdgpu_invalid_wreg; 2423 adev->gc_cac_rreg = &amdgpu_invalid_rreg; 2424 adev->gc_cac_wreg = &amdgpu_invalid_wreg; 2425 adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg; 2426 adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg; 2427 2428 printf("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n", 2429 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device, 2430 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision); 2431 2432 /* mutex initialization are all done here so we 2433 * can recall function without having locking issues */ 2434 atomic_set(&adev->irq.ih.lock, 0); 2435 rw_init(&adev->firmware.mutex, "agfw"); 2436 rw_init(&adev->pm.mutex, "agpm"); 2437 rw_init(&adev->gfx.gpu_clock_mutex, "gfxclk"); 2438 rw_init(&adev->srbm_mutex, "srbm"); 2439 rw_init(&adev->gfx.pipe_reserve_mutex, "pipers"); 2440 rw_init(&adev->grbm_idx_mutex, "grbmidx"); 2441 rw_init(&adev->mn_lock, "agpumn"); 2442 rw_init(&adev->virt.vf_errors.lock, "vferr"); 2443 hash_init(adev->mn_hash); 2444 rw_init(&adev->lock_reset, "aglkrst"); 2445 2446 amdgpu_device_check_arguments(adev); 2447 2448 mtx_init(&adev->mmio_idx_lock, IPL_TTY); 2449 mtx_init(&adev->smc_idx_lock, IPL_TTY); 2450 mtx_init(&adev->pcie_idx_lock, IPL_TTY); 2451 mtx_init(&adev->uvd_ctx_idx_lock, IPL_TTY); 2452 mtx_init(&adev->didt_idx_lock, IPL_TTY); 2453 mtx_init(&adev->gc_cac_idx_lock, IPL_TTY); 2454 mtx_init(&adev->se_cac_idx_lock, IPL_TTY); 2455 mtx_init(&adev->audio_endpt_idx_lock, IPL_TTY); 2456 mtx_init(&adev->mm_stats.lock, IPL_TTY); 2457 2458 INIT_LIST_HEAD(&adev->shadow_list); 2459 rw_init(&adev->shadow_list_lock, "sdwlst"); 2460 2461 INIT_LIST_HEAD(&adev->ring_lru_list); 2462 mtx_init(&adev->ring_lru_list_lock, IPL_TTY); 2463 2464 INIT_DELAYED_WORK(&adev->late_init_work, 2465 amdgpu_device_ip_late_init_func_handler); 2466 2467 adev->pm.ac_power = power_supply_is_system_supplied() > 0 ? true : false; 2468 2469 #ifdef __linux__ 2470 /* Registers mapping */ 2471 /* TODO: block userspace mapping of io register */ 2472 if (adev->asic_type >= CHIP_BONAIRE) { 2473 adev->rmmio_base = pci_resource_start(adev->pdev, 5); 2474 adev->rmmio_size = pci_resource_len(adev->pdev, 5); 2475 } else { 2476 adev->rmmio_base = pci_resource_start(adev->pdev, 2); 2477 adev->rmmio_size = pci_resource_len(adev->pdev, 2); 2478 } 2479 2480 adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size); 2481 if (adev->rmmio == NULL) { 2482 return -ENOMEM; 2483 } 2484 #endif 2485 DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base); 2486 DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size); 2487 2488 /* doorbell bar mapping */ 2489 amdgpu_device_doorbell_init(adev); 2490 2491 /* io port mapping */ 2492 #ifdef __linux__ 2493 for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) { 2494 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) { 2495 adev->rio_mem_size = pci_resource_len(adev->pdev, i); 2496 adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size); 2497 break; 2498 } 2499 } 2500 if (adev->rio_mem == NULL) 2501 DRM_INFO("PCI I/O BAR is not found.\n"); 2502 #endif 2503 2504 amdgpu_device_get_pcie_info(adev); 2505 2506 /* early init functions */ 2507 r = amdgpu_device_ip_early_init(adev); 2508 if (r) 2509 return r; 2510 2511 /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */ 2512 /* this will fail for cards that aren't VGA class devices, just 2513 * ignore it */ 2514 #ifdef notyet 2515 vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode); 2516 #endif 2517 2518 if (amdgpu_device_is_px(ddev)) 2519 runtime = true; 2520 #ifdef notyet 2521 if (!pci_is_thunderbolt_attached(adev->pdev)) 2522 vga_switcheroo_register_client(adev->pdev, 2523 &amdgpu_switcheroo_ops, runtime); 2524 if (runtime) 2525 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain); 2526 #endif 2527 2528 if (amdgpu_emu_mode == 1) { 2529 /* post the asic on emulation mode */ 2530 emu_soc_asic_init(adev); 2531 goto fence_driver_init; 2532 } 2533 2534 /* Read BIOS */ 2535 if (!amdgpu_get_bios(adev)) { 2536 r = -EINVAL; 2537 goto failed; 2538 } 2539 2540 r = amdgpu_atombios_init(adev); 2541 if (r) { 2542 dev_err(adev->dev, "amdgpu_atombios_init failed\n"); 2543 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0); 2544 goto failed; 2545 } 2546 2547 /* detect if we are with an SRIOV vbios */ 2548 amdgpu_device_detect_sriov_bios(adev); 2549 2550 /* Post card if necessary */ 2551 if (amdgpu_device_need_post(adev)) { 2552 if (!adev->bios) { 2553 dev_err(adev->dev, "no vBIOS found\n"); 2554 r = -EINVAL; 2555 goto failed; 2556 } 2557 DRM_INFO("GPU posting now...\n"); 2558 r = amdgpu_atom_asic_init(adev->mode_info.atom_context); 2559 if (r) { 2560 dev_err(adev->dev, "gpu post error!\n"); 2561 goto failed; 2562 } 2563 } 2564 2565 if (adev->is_atom_fw) { 2566 /* Initialize clocks */ 2567 r = amdgpu_atomfirmware_get_clock_info(adev); 2568 if (r) { 2569 dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n"); 2570 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0); 2571 goto failed; 2572 } 2573 } else { 2574 /* Initialize clocks */ 2575 r = amdgpu_atombios_get_clock_info(adev); 2576 if (r) { 2577 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n"); 2578 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0); 2579 goto failed; 2580 } 2581 /* init i2c buses */ 2582 if (!amdgpu_device_has_dc_support(adev)) 2583 amdgpu_atombios_i2c_init(adev); 2584 } 2585 2586 fence_driver_init: 2587 /* Fence driver */ 2588 r = amdgpu_fence_driver_init(adev); 2589 if (r) { 2590 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n"); 2591 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0); 2592 goto failed; 2593 } 2594 2595 /* init the mode config */ 2596 drm_mode_config_init(adev->ddev); 2597 2598 r = amdgpu_device_ip_init(adev); 2599 if (r) { 2600 /* failed in exclusive mode due to timeout */ 2601 if (amdgpu_sriov_vf(adev) && 2602 !amdgpu_sriov_runtime(adev) && 2603 amdgpu_virt_mmio_blocked(adev) && 2604 !amdgpu_virt_wait_reset(adev)) { 2605 dev_err(adev->dev, "VF exclusive mode timeout\n"); 2606 /* Don't send request since VF is inactive. */ 2607 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME; 2608 adev->virt.ops = NULL; 2609 r = -EAGAIN; 2610 goto failed; 2611 } 2612 dev_err(adev->dev, "amdgpu_device_ip_init failed\n"); 2613 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0); 2614 goto failed; 2615 } 2616 2617 adev->accel_working = true; 2618 2619 amdgpu_vm_check_compute_bug(adev); 2620 2621 /* Initialize the buffer migration limit. */ 2622 if (amdgpu_moverate >= 0) 2623 max_MBps = amdgpu_moverate; 2624 else 2625 max_MBps = 8; /* Allow 8 MB/s. */ 2626 /* Get a log2 for easy divisions. */ 2627 adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps)); 2628 2629 r = amdgpu_ib_pool_init(adev); 2630 if (r) { 2631 dev_err(adev->dev, "IB initialization failed (%d).\n", r); 2632 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r); 2633 goto failed; 2634 } 2635 2636 amdgpu_fbdev_init(adev); 2637 2638 r = amdgpu_pm_sysfs_init(adev); 2639 if (r) 2640 DRM_ERROR("registering pm debugfs failed (%d).\n", r); 2641 2642 r = amdgpu_debugfs_gem_init(adev); 2643 if (r) 2644 DRM_ERROR("registering gem debugfs failed (%d).\n", r); 2645 2646 r = amdgpu_debugfs_regs_init(adev); 2647 if (r) 2648 DRM_ERROR("registering register debugfs failed (%d).\n", r); 2649 2650 r = amdgpu_debugfs_firmware_init(adev); 2651 if (r) 2652 DRM_ERROR("registering firmware debugfs failed (%d).\n", r); 2653 2654 r = amdgpu_debugfs_init(adev); 2655 if (r) 2656 DRM_ERROR("Creating debugfs files failed (%d).\n", r); 2657 2658 if ((amdgpu_testing & 1)) { 2659 if (adev->accel_working) 2660 amdgpu_test_moves(adev); 2661 else 2662 DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n"); 2663 } 2664 if (amdgpu_benchmarking) { 2665 if (adev->accel_working) 2666 amdgpu_benchmark(adev, amdgpu_benchmarking); 2667 else 2668 DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n"); 2669 } 2670 2671 /* enable clockgating, etc. after ib tests, etc. since some blocks require 2672 * explicit gating rather than handling it automatically. 2673 */ 2674 r = amdgpu_device_ip_late_init(adev); 2675 if (r) { 2676 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n"); 2677 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r); 2678 goto failed; 2679 } 2680 2681 return 0; 2682 2683 failed: 2684 amdgpu_vf_error_trans_all(adev); 2685 if (runtime) 2686 vga_switcheroo_fini_domain_pm_ops(adev->dev); 2687 2688 return r; 2689 } 2690 2691 /** 2692 * amdgpu_device_fini - tear down the driver 2693 * 2694 * @adev: amdgpu_device pointer 2695 * 2696 * Tear down the driver info (all asics). 2697 * Called at driver shutdown. 2698 */ 2699 void amdgpu_device_fini(struct amdgpu_device *adev) 2700 { 2701 int r; 2702 2703 DRM_INFO("amdgpu: finishing device.\n"); 2704 adev->shutdown = true; 2705 /* disable all interrupts */ 2706 amdgpu_irq_disable_all(adev); 2707 if (adev->mode_info.mode_config_initialized){ 2708 if (!amdgpu_device_has_dc_support(adev)) 2709 drm_crtc_force_disable_all(adev->ddev); 2710 else 2711 drm_atomic_helper_shutdown(adev->ddev); 2712 } 2713 amdgpu_ib_pool_fini(adev); 2714 amdgpu_fence_driver_fini(adev); 2715 amdgpu_pm_sysfs_fini(adev); 2716 amdgpu_fbdev_fini(adev); 2717 r = amdgpu_device_ip_fini(adev); 2718 if (adev->firmware.gpu_info_fw) { 2719 release_firmware(adev->firmware.gpu_info_fw); 2720 adev->firmware.gpu_info_fw = NULL; 2721 } 2722 adev->accel_working = false; 2723 cancel_delayed_work_sync(&adev->late_init_work); 2724 /* free i2c buses */ 2725 if (!amdgpu_device_has_dc_support(adev)) 2726 amdgpu_i2c_fini(adev); 2727 2728 if (amdgpu_emu_mode != 1) 2729 amdgpu_atombios_fini(adev); 2730 2731 kfree(adev->bios); 2732 adev->bios = NULL; 2733 if (!pci_is_thunderbolt_attached(adev->pdev)) 2734 vga_switcheroo_unregister_client(adev->pdev); 2735 if (adev->flags & AMD_IS_PX) 2736 vga_switcheroo_fini_domain_pm_ops(adev->dev); 2737 vga_client_register(adev->pdev, NULL, NULL, NULL); 2738 #ifdef __linux__ 2739 if (adev->rio_mem) 2740 pci_iounmap(adev->pdev, adev->rio_mem); 2741 adev->rio_mem = NULL; 2742 iounmap(adev->rmmio); 2743 adev->rmmio = NULL; 2744 #else 2745 if (adev->rio_mem_size > 0) 2746 bus_space_unmap(adev->rio_mem_bst, adev->rio_mem_bsh, 2747 adev->rio_mem_size); 2748 adev->rio_mem_size = 0; 2749 2750 if (adev->rmmio_size > 0) 2751 bus_space_unmap(adev->rmmio_bst, adev->rmmio_bsh, 2752 adev->rmmio_size); 2753 adev->rmmio_size = 0; 2754 #endif 2755 amdgpu_device_doorbell_fini(adev); 2756 amdgpu_debugfs_regs_cleanup(adev); 2757 } 2758 2759 2760 /* 2761 * Suspend & resume. 2762 */ 2763 /** 2764 * amdgpu_device_suspend - initiate device suspend 2765 * 2766 * @dev: drm dev pointer 2767 * @suspend: suspend state 2768 * @fbcon : notify the fbdev of suspend 2769 * 2770 * Puts the hw in the suspend state (all asics). 2771 * Returns 0 for success or an error on failure. 2772 * Called at driver suspend. 2773 */ 2774 int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon) 2775 { 2776 struct amdgpu_device *adev; 2777 struct drm_crtc *crtc; 2778 struct drm_connector *connector; 2779 int r; 2780 2781 if (dev == NULL || dev->dev_private == NULL) { 2782 return -ENODEV; 2783 } 2784 2785 adev = dev->dev_private; 2786 if (adev->shutdown) 2787 return 0; 2788 2789 #ifdef notyet 2790 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF) 2791 return 0; 2792 #endif 2793 2794 drm_kms_helper_poll_disable(dev); 2795 2796 if (fbcon) 2797 amdgpu_fbdev_set_suspend(adev, 1); 2798 2799 if (!amdgpu_device_has_dc_support(adev)) { 2800 /* turn off display hw */ 2801 drm_modeset_lock_all(dev); 2802 list_for_each_entry(connector, &dev->mode_config.connector_list, head) { 2803 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF); 2804 } 2805 drm_modeset_unlock_all(dev); 2806 /* unpin the front buffers and cursors */ 2807 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) { 2808 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc); 2809 struct drm_framebuffer *fb = crtc->primary->fb; 2810 struct amdgpu_bo *robj; 2811 2812 if (amdgpu_crtc->cursor_bo) { 2813 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo); 2814 r = amdgpu_bo_reserve(aobj, true); 2815 if (r == 0) { 2816 amdgpu_bo_unpin(aobj); 2817 amdgpu_bo_unreserve(aobj); 2818 } 2819 } 2820 2821 if (fb == NULL || fb->obj[0] == NULL) { 2822 continue; 2823 } 2824 robj = gem_to_amdgpu_bo(fb->obj[0]); 2825 /* don't unpin kernel fb objects */ 2826 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) { 2827 r = amdgpu_bo_reserve(robj, true); 2828 if (r == 0) { 2829 amdgpu_bo_unpin(robj); 2830 amdgpu_bo_unreserve(robj); 2831 } 2832 } 2833 } 2834 } 2835 2836 amdgpu_amdkfd_suspend(adev); 2837 2838 r = amdgpu_device_ip_suspend_phase1(adev); 2839 2840 /* evict vram memory */ 2841 amdgpu_bo_evict_vram(adev); 2842 2843 amdgpu_fence_driver_suspend(adev); 2844 2845 r = amdgpu_device_ip_suspend_phase2(adev); 2846 2847 /* evict remaining vram memory 2848 * This second call to evict vram is to evict the gart page table 2849 * using the CPU. 2850 */ 2851 amdgpu_bo_evict_vram(adev); 2852 2853 pci_save_state(dev->pdev); 2854 if (suspend) { 2855 /* Shut down the device */ 2856 pci_disable_device(dev->pdev); 2857 pci_set_power_state(dev->pdev, PCI_D3hot); 2858 } else { 2859 r = amdgpu_asic_reset(adev); 2860 if (r) 2861 DRM_ERROR("amdgpu asic reset failed\n"); 2862 } 2863 2864 return 0; 2865 } 2866 2867 /** 2868 * amdgpu_device_resume - initiate device resume 2869 * 2870 * @dev: drm dev pointer 2871 * @resume: resume state 2872 * @fbcon : notify the fbdev of resume 2873 * 2874 * Bring the hw back to operating state (all asics). 2875 * Returns 0 for success or an error on failure. 2876 * Called at driver resume. 2877 */ 2878 int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon) 2879 { 2880 struct drm_connector *connector; 2881 struct amdgpu_device *adev = dev->dev_private; 2882 struct drm_crtc *crtc; 2883 int r = 0; 2884 2885 #ifdef notyet 2886 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF) 2887 return 0; 2888 #endif 2889 2890 if (resume) { 2891 pci_set_power_state(dev->pdev, PCI_D0); 2892 pci_restore_state(dev->pdev); 2893 r = pci_enable_device(dev->pdev); 2894 if (r) 2895 return r; 2896 } 2897 2898 /* post card */ 2899 if (amdgpu_device_need_post(adev)) { 2900 r = amdgpu_atom_asic_init(adev->mode_info.atom_context); 2901 if (r) 2902 DRM_ERROR("amdgpu asic init failed\n"); 2903 } 2904 2905 r = amdgpu_device_ip_resume(adev); 2906 if (r) { 2907 DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r); 2908 return r; 2909 } 2910 amdgpu_fence_driver_resume(adev); 2911 2912 2913 r = amdgpu_device_ip_late_init(adev); 2914 if (r) 2915 return r; 2916 2917 if (!amdgpu_device_has_dc_support(adev)) { 2918 /* pin cursors */ 2919 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) { 2920 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc); 2921 2922 if (amdgpu_crtc->cursor_bo) { 2923 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo); 2924 r = amdgpu_bo_reserve(aobj, true); 2925 if (r == 0) { 2926 r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM); 2927 if (r != 0) 2928 DRM_ERROR("Failed to pin cursor BO (%d)\n", r); 2929 amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj); 2930 amdgpu_bo_unreserve(aobj); 2931 } 2932 } 2933 } 2934 } 2935 r = amdgpu_amdkfd_resume(adev); 2936 if (r) 2937 return r; 2938 2939 /* Make sure IB tests flushed */ 2940 flush_delayed_work(&adev->late_init_work); 2941 2942 /* blat the mode back in */ 2943 if (fbcon) { 2944 if (!amdgpu_device_has_dc_support(adev)) { 2945 /* pre DCE11 */ 2946 drm_helper_resume_force_mode(dev); 2947 2948 /* turn on display hw */ 2949 drm_modeset_lock_all(dev); 2950 list_for_each_entry(connector, &dev->mode_config.connector_list, head) { 2951 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON); 2952 } 2953 drm_modeset_unlock_all(dev); 2954 } 2955 amdgpu_fbdev_set_suspend(adev, 0); 2956 } 2957 2958 drm_kms_helper_poll_enable(dev); 2959 2960 /* 2961 * Most of the connector probing functions try to acquire runtime pm 2962 * refs to ensure that the GPU is powered on when connector polling is 2963 * performed. Since we're calling this from a runtime PM callback, 2964 * trying to acquire rpm refs will cause us to deadlock. 2965 * 2966 * Since we're guaranteed to be holding the rpm lock, it's safe to 2967 * temporarily disable the rpm helpers so this doesn't deadlock us. 2968 */ 2969 #if defined(CONFIG_PM) && defined(__linux__) 2970 dev->dev->power.disable_depth++; 2971 #endif 2972 if (!amdgpu_device_has_dc_support(adev)) 2973 drm_helper_hpd_irq_event(dev); 2974 else 2975 drm_kms_helper_hotplug_event(dev); 2976 #if defined(CONFIG_PM) && defined(__linux__) 2977 dev->dev->power.disable_depth--; 2978 #endif 2979 return 0; 2980 } 2981 2982 /** 2983 * amdgpu_device_ip_check_soft_reset - did soft reset succeed 2984 * 2985 * @adev: amdgpu_device pointer 2986 * 2987 * The list of all the hardware IPs that make up the asic is walked and 2988 * the check_soft_reset callbacks are run. check_soft_reset determines 2989 * if the asic is still hung or not. 2990 * Returns true if any of the IPs are still in a hung state, false if not. 2991 */ 2992 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev) 2993 { 2994 int i; 2995 bool asic_hang = false; 2996 2997 if (amdgpu_sriov_vf(adev)) 2998 return true; 2999 3000 if (amdgpu_asic_need_full_reset(adev)) 3001 return true; 3002 3003 for (i = 0; i < adev->num_ip_blocks; i++) { 3004 if (!adev->ip_blocks[i].status.valid) 3005 continue; 3006 if (adev->ip_blocks[i].version->funcs->check_soft_reset) 3007 adev->ip_blocks[i].status.hang = 3008 adev->ip_blocks[i].version->funcs->check_soft_reset(adev); 3009 if (adev->ip_blocks[i].status.hang) { 3010 DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name); 3011 asic_hang = true; 3012 } 3013 } 3014 return asic_hang; 3015 } 3016 3017 /** 3018 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset 3019 * 3020 * @adev: amdgpu_device pointer 3021 * 3022 * The list of all the hardware IPs that make up the asic is walked and the 3023 * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset 3024 * handles any IP specific hardware or software state changes that are 3025 * necessary for a soft reset to succeed. 3026 * Returns 0 on success, negative error code on failure. 3027 */ 3028 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev) 3029 { 3030 int i, r = 0; 3031 3032 for (i = 0; i < adev->num_ip_blocks; i++) { 3033 if (!adev->ip_blocks[i].status.valid) 3034 continue; 3035 if (adev->ip_blocks[i].status.hang && 3036 adev->ip_blocks[i].version->funcs->pre_soft_reset) { 3037 r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev); 3038 if (r) 3039 return r; 3040 } 3041 } 3042 3043 return 0; 3044 } 3045 3046 /** 3047 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed 3048 * 3049 * @adev: amdgpu_device pointer 3050 * 3051 * Some hardware IPs cannot be soft reset. If they are hung, a full gpu 3052 * reset is necessary to recover. 3053 * Returns true if a full asic reset is required, false if not. 3054 */ 3055 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev) 3056 { 3057 int i; 3058 3059 if (amdgpu_asic_need_full_reset(adev)) 3060 return true; 3061 3062 for (i = 0; i < adev->num_ip_blocks; i++) { 3063 if (!adev->ip_blocks[i].status.valid) 3064 continue; 3065 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) || 3066 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) || 3067 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) || 3068 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) || 3069 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) { 3070 if (adev->ip_blocks[i].status.hang) { 3071 DRM_INFO("Some block need full reset!\n"); 3072 return true; 3073 } 3074 } 3075 } 3076 return false; 3077 } 3078 3079 /** 3080 * amdgpu_device_ip_soft_reset - do a soft reset 3081 * 3082 * @adev: amdgpu_device pointer 3083 * 3084 * The list of all the hardware IPs that make up the asic is walked and the 3085 * soft_reset callbacks are run if the block is hung. soft_reset handles any 3086 * IP specific hardware or software state changes that are necessary to soft 3087 * reset the IP. 3088 * Returns 0 on success, negative error code on failure. 3089 */ 3090 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev) 3091 { 3092 int i, r = 0; 3093 3094 for (i = 0; i < adev->num_ip_blocks; i++) { 3095 if (!adev->ip_blocks[i].status.valid) 3096 continue; 3097 if (adev->ip_blocks[i].status.hang && 3098 adev->ip_blocks[i].version->funcs->soft_reset) { 3099 r = adev->ip_blocks[i].version->funcs->soft_reset(adev); 3100 if (r) 3101 return r; 3102 } 3103 } 3104 3105 return 0; 3106 } 3107 3108 /** 3109 * amdgpu_device_ip_post_soft_reset - clean up from soft reset 3110 * 3111 * @adev: amdgpu_device pointer 3112 * 3113 * The list of all the hardware IPs that make up the asic is walked and the 3114 * post_soft_reset callbacks are run if the asic was hung. post_soft_reset 3115 * handles any IP specific hardware or software state changes that are 3116 * necessary after the IP has been soft reset. 3117 * Returns 0 on success, negative error code on failure. 3118 */ 3119 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev) 3120 { 3121 int i, r = 0; 3122 3123 for (i = 0; i < adev->num_ip_blocks; i++) { 3124 if (!adev->ip_blocks[i].status.valid) 3125 continue; 3126 if (adev->ip_blocks[i].status.hang && 3127 adev->ip_blocks[i].version->funcs->post_soft_reset) 3128 r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev); 3129 if (r) 3130 return r; 3131 } 3132 3133 return 0; 3134 } 3135 3136 /** 3137 * amdgpu_device_recover_vram_from_shadow - restore shadowed VRAM buffers 3138 * 3139 * @adev: amdgpu_device pointer 3140 * @ring: amdgpu_ring for the engine handling the buffer operations 3141 * @bo: amdgpu_bo buffer whose shadow is being restored 3142 * @fence: dma_fence associated with the operation 3143 * 3144 * Restores the VRAM buffer contents from the shadow in GTT. Used to 3145 * restore things like GPUVM page tables after a GPU reset where 3146 * the contents of VRAM might be lost. 3147 * Returns 0 on success, negative error code on failure. 3148 */ 3149 static int amdgpu_device_recover_vram_from_shadow(struct amdgpu_device *adev, 3150 struct amdgpu_ring *ring, 3151 struct amdgpu_bo *bo, 3152 struct dma_fence **fence) 3153 { 3154 uint32_t domain; 3155 int r; 3156 3157 if (!bo->shadow) 3158 return 0; 3159 3160 r = amdgpu_bo_reserve(bo, true); 3161 if (r) 3162 return r; 3163 domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type); 3164 /* if bo has been evicted, then no need to recover */ 3165 if (domain == AMDGPU_GEM_DOMAIN_VRAM) { 3166 r = amdgpu_bo_validate(bo->shadow); 3167 if (r) { 3168 DRM_ERROR("bo validate failed!\n"); 3169 goto err; 3170 } 3171 3172 r = amdgpu_bo_restore_from_shadow(adev, ring, bo, 3173 NULL, fence, true); 3174 if (r) { 3175 DRM_ERROR("recover page table failed!\n"); 3176 goto err; 3177 } 3178 } 3179 err: 3180 amdgpu_bo_unreserve(bo); 3181 return r; 3182 } 3183 3184 /** 3185 * amdgpu_device_handle_vram_lost - Handle the loss of VRAM contents 3186 * 3187 * @adev: amdgpu_device pointer 3188 * 3189 * Restores the contents of VRAM buffers from the shadows in GTT. Used to 3190 * restore things like GPUVM page tables after a GPU reset where 3191 * the contents of VRAM might be lost. 3192 * Returns 0 on success, 1 on failure. 3193 */ 3194 static int amdgpu_device_handle_vram_lost(struct amdgpu_device *adev) 3195 { 3196 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring; 3197 struct amdgpu_bo *bo, *tmp; 3198 struct dma_fence *fence = NULL, *next = NULL; 3199 long r = 1; 3200 int i = 0; 3201 long tmo; 3202 3203 if (amdgpu_sriov_runtime(adev)) 3204 tmo = msecs_to_jiffies(8000); 3205 else 3206 tmo = msecs_to_jiffies(100); 3207 3208 DRM_INFO("recover vram bo from shadow start\n"); 3209 mutex_lock(&adev->shadow_list_lock); 3210 list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) { 3211 next = NULL; 3212 amdgpu_device_recover_vram_from_shadow(adev, ring, bo, &next); 3213 if (fence) { 3214 r = dma_fence_wait_timeout(fence, false, tmo); 3215 if (r == 0) 3216 pr_err("wait fence %p[%d] timeout\n", fence, i); 3217 else if (r < 0) 3218 pr_err("wait fence %p[%d] interrupted\n", fence, i); 3219 if (r < 1) { 3220 dma_fence_put(fence); 3221 fence = next; 3222 break; 3223 } 3224 i++; 3225 } 3226 3227 dma_fence_put(fence); 3228 fence = next; 3229 } 3230 mutex_unlock(&adev->shadow_list_lock); 3231 3232 if (fence) { 3233 r = dma_fence_wait_timeout(fence, false, tmo); 3234 if (r == 0) 3235 pr_err("wait fence %p[%d] timeout\n", fence, i); 3236 else if (r < 0) 3237 pr_err("wait fence %p[%d] interrupted\n", fence, i); 3238 3239 } 3240 dma_fence_put(fence); 3241 3242 if (r > 0) 3243 DRM_INFO("recover vram bo from shadow done\n"); 3244 else 3245 DRM_ERROR("recover vram bo from shadow failed\n"); 3246 3247 return (r > 0) ? 0 : 1; 3248 } 3249 3250 /** 3251 * amdgpu_device_reset - reset ASIC/GPU for bare-metal or passthrough 3252 * 3253 * @adev: amdgpu device pointer 3254 * 3255 * attempt to do soft-reset or full-reset and reinitialize Asic 3256 * return 0 means succeeded otherwise failed 3257 */ 3258 static int amdgpu_device_reset(struct amdgpu_device *adev) 3259 { 3260 bool need_full_reset, vram_lost = 0; 3261 int r; 3262 3263 need_full_reset = amdgpu_device_ip_need_full_reset(adev); 3264 3265 if (!need_full_reset) { 3266 amdgpu_device_ip_pre_soft_reset(adev); 3267 r = amdgpu_device_ip_soft_reset(adev); 3268 amdgpu_device_ip_post_soft_reset(adev); 3269 if (r || amdgpu_device_ip_check_soft_reset(adev)) { 3270 DRM_INFO("soft reset failed, will fallback to full reset!\n"); 3271 need_full_reset = true; 3272 } 3273 } 3274 3275 if (need_full_reset) { 3276 r = amdgpu_device_ip_suspend(adev); 3277 3278 retry: 3279 r = amdgpu_asic_reset(adev); 3280 /* post card */ 3281 amdgpu_atom_asic_init(adev->mode_info.atom_context); 3282 3283 if (!r) { 3284 dev_info(adev->dev, "GPU reset succeeded, trying to resume\n"); 3285 r = amdgpu_device_ip_resume_phase1(adev); 3286 if (r) 3287 goto out; 3288 3289 vram_lost = amdgpu_device_check_vram_lost(adev); 3290 if (vram_lost) { 3291 DRM_ERROR("VRAM is lost!\n"); 3292 atomic_inc(&adev->vram_lost_counter); 3293 } 3294 3295 r = amdgpu_gtt_mgr_recover( 3296 &adev->mman.bdev.man[TTM_PL_TT]); 3297 if (r) 3298 goto out; 3299 3300 r = amdgpu_device_ip_resume_phase2(adev); 3301 if (r) 3302 goto out; 3303 3304 if (vram_lost) 3305 amdgpu_device_fill_reset_magic(adev); 3306 } 3307 } 3308 3309 out: 3310 if (!r) { 3311 amdgpu_irq_gpu_reset_resume_helper(adev); 3312 r = amdgpu_ib_ring_tests(adev); 3313 if (r) { 3314 dev_err(adev->dev, "ib ring test failed (%d).\n", r); 3315 r = amdgpu_device_ip_suspend(adev); 3316 need_full_reset = true; 3317 goto retry; 3318 } 3319 } 3320 3321 if (!r && ((need_full_reset && !(adev->flags & AMD_IS_APU)) || vram_lost)) 3322 r = amdgpu_device_handle_vram_lost(adev); 3323 3324 return r; 3325 } 3326 3327 /** 3328 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf 3329 * 3330 * @adev: amdgpu device pointer 3331 * @from_hypervisor: request from hypervisor 3332 * 3333 * do VF FLR and reinitialize Asic 3334 * return 0 means succeeded otherwise failed 3335 */ 3336 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, 3337 bool from_hypervisor) 3338 { 3339 int r; 3340 3341 if (from_hypervisor) 3342 r = amdgpu_virt_request_full_gpu(adev, true); 3343 else 3344 r = amdgpu_virt_reset_gpu(adev); 3345 if (r) 3346 return r; 3347 3348 /* Resume IP prior to SMC */ 3349 r = amdgpu_device_ip_reinit_early_sriov(adev); 3350 if (r) 3351 goto error; 3352 3353 /* we need recover gart prior to run SMC/CP/SDMA resume */ 3354 amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]); 3355 3356 /* now we are okay to resume SMC/CP/SDMA */ 3357 r = amdgpu_device_ip_reinit_late_sriov(adev); 3358 if (r) 3359 goto error; 3360 3361 amdgpu_irq_gpu_reset_resume_helper(adev); 3362 r = amdgpu_ib_ring_tests(adev); 3363 3364 error: 3365 amdgpu_virt_init_data_exchange(adev); 3366 amdgpu_virt_release_full_gpu(adev, true); 3367 if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) { 3368 atomic_inc(&adev->vram_lost_counter); 3369 r = amdgpu_device_handle_vram_lost(adev); 3370 } 3371 3372 return r; 3373 } 3374 3375 /** 3376 * amdgpu_device_gpu_recover - reset the asic and recover scheduler 3377 * 3378 * @adev: amdgpu device pointer 3379 * @job: which job trigger hang 3380 * @force: forces reset regardless of amdgpu_gpu_recovery 3381 * 3382 * Attempt to reset the GPU if it has hung (all asics). 3383 * Returns 0 for success or an error on failure. 3384 */ 3385 int amdgpu_device_gpu_recover(struct amdgpu_device *adev, 3386 struct amdgpu_job *job, bool force) 3387 { 3388 int i, r, resched; 3389 3390 if (!force && !amdgpu_device_ip_check_soft_reset(adev)) { 3391 DRM_INFO("No hardware hang detected. Did some blocks stall?\n"); 3392 return 0; 3393 } 3394 3395 if (!force && (amdgpu_gpu_recovery == 0 || 3396 (amdgpu_gpu_recovery == -1 && !amdgpu_sriov_vf(adev)))) { 3397 DRM_INFO("GPU recovery disabled.\n"); 3398 return 0; 3399 } 3400 3401 dev_info(adev->dev, "GPU reset begin!\n"); 3402 3403 mutex_lock(&adev->lock_reset); 3404 atomic_inc(&adev->gpu_reset_counter); 3405 adev->in_gpu_reset = 1; 3406 3407 /* Block kfd */ 3408 amdgpu_amdkfd_pre_reset(adev); 3409 3410 /* block TTM */ 3411 resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev); 3412 3413 /* block all schedulers and reset given job's ring */ 3414 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { 3415 struct amdgpu_ring *ring = adev->rings[i]; 3416 3417 if (!ring || !ring->sched.thread) 3418 continue; 3419 3420 kthread_park(ring->sched.thread); 3421 3422 if (job && job->base.sched == &ring->sched) 3423 continue; 3424 3425 drm_sched_hw_job_reset(&ring->sched, job ? &job->base : NULL); 3426 3427 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */ 3428 amdgpu_fence_driver_force_completion(ring); 3429 } 3430 3431 if (amdgpu_sriov_vf(adev)) 3432 r = amdgpu_device_reset_sriov(adev, job ? false : true); 3433 else 3434 r = amdgpu_device_reset(adev); 3435 3436 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { 3437 struct amdgpu_ring *ring = adev->rings[i]; 3438 3439 if (!ring || !ring->sched.thread) 3440 continue; 3441 3442 /* only need recovery sched of the given job's ring 3443 * or all rings (in the case @job is NULL) 3444 * after above amdgpu_reset accomplished 3445 */ 3446 if ((!job || job->base.sched == &ring->sched) && !r) 3447 drm_sched_job_recovery(&ring->sched); 3448 3449 kthread_unpark(ring->sched.thread); 3450 } 3451 3452 if (!amdgpu_device_has_dc_support(adev)) { 3453 drm_helper_resume_force_mode(adev->ddev); 3454 } 3455 3456 ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched); 3457 3458 if (r) { 3459 /* bad news, how to tell it to userspace ? */ 3460 dev_info(adev->dev, "GPU reset(%d) failed\n", atomic_read(&adev->gpu_reset_counter)); 3461 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r); 3462 } else { 3463 dev_info(adev->dev, "GPU reset(%d) succeeded!\n",atomic_read(&adev->gpu_reset_counter)); 3464 } 3465 3466 /*unlock kfd */ 3467 amdgpu_amdkfd_post_reset(adev); 3468 amdgpu_vf_error_trans_all(adev); 3469 adev->in_gpu_reset = 0; 3470 mutex_unlock(&adev->lock_reset); 3471 return r; 3472 } 3473 3474 /** 3475 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot 3476 * 3477 * @adev: amdgpu_device pointer 3478 * 3479 * Fetchs and stores in the driver the PCIE capabilities (gen speed 3480 * and lanes) of the slot the device is in. Handles APUs and 3481 * virtualized environments where PCIE config space may not be available. 3482 */ 3483 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev) 3484 { 3485 struct pci_dev *pdev; 3486 enum pci_bus_speed speed_cap; 3487 enum pcie_link_width link_width; 3488 3489 if (amdgpu_pcie_gen_cap) 3490 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap; 3491 3492 if (amdgpu_pcie_lane_cap) 3493 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap; 3494 3495 /* covers APUs as well */ 3496 if (pci_is_root_bus(adev->pdev->bus)) { 3497 if (adev->pm.pcie_gen_mask == 0) 3498 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK; 3499 if (adev->pm.pcie_mlw_mask == 0) 3500 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK; 3501 return; 3502 } 3503 3504 if (adev->pm.pcie_gen_mask == 0) { 3505 /* asic caps */ 3506 pdev = adev->pdev; 3507 speed_cap = pcie_get_speed_cap(pdev); 3508 if (speed_cap == PCI_SPEED_UNKNOWN) { 3509 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | 3510 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | 3511 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3); 3512 } else { 3513 if (speed_cap == PCIE_SPEED_16_0GT) 3514 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | 3515 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | 3516 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 | 3517 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4); 3518 else if (speed_cap == PCIE_SPEED_8_0GT) 3519 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | 3520 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | 3521 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3); 3522 else if (speed_cap == PCIE_SPEED_5_0GT) 3523 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | 3524 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2); 3525 else 3526 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1; 3527 } 3528 /* platform caps */ 3529 pdev = adev->ddev->pdev->bus->self; 3530 speed_cap = pcie_get_speed_cap(pdev); 3531 if (speed_cap == PCI_SPEED_UNKNOWN) { 3532 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | 3533 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2); 3534 } else { 3535 if (speed_cap == PCIE_SPEED_16_0GT) 3536 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | 3537 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 | 3538 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 | 3539 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4); 3540 else if (speed_cap == PCIE_SPEED_8_0GT) 3541 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | 3542 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 | 3543 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3); 3544 else if (speed_cap == PCIE_SPEED_5_0GT) 3545 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | 3546 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2); 3547 else 3548 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1; 3549 3550 } 3551 } 3552 if (adev->pm.pcie_mlw_mask == 0) { 3553 pdev = adev->ddev->pdev->bus->self; 3554 link_width = pcie_get_width_cap(pdev); 3555 if (link_width == PCIE_LNK_WIDTH_UNKNOWN) { 3556 adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK; 3557 } else { 3558 switch (link_width) { 3559 case PCIE_LNK_X32: 3560 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 | 3561 CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 | 3562 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 | 3563 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | 3564 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | 3565 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | 3566 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); 3567 break; 3568 case PCIE_LNK_X16: 3569 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 | 3570 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 | 3571 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | 3572 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | 3573 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | 3574 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); 3575 break; 3576 case PCIE_LNK_X12: 3577 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 | 3578 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | 3579 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | 3580 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | 3581 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); 3582 break; 3583 case PCIE_LNK_X8: 3584 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | 3585 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | 3586 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | 3587 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); 3588 break; 3589 case PCIE_LNK_X4: 3590 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | 3591 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | 3592 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); 3593 break; 3594 case PCIE_LNK_X2: 3595 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | 3596 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); 3597 break; 3598 case PCIE_LNK_X1: 3599 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1; 3600 break; 3601 default: 3602 break; 3603 } 3604 } 3605 } 3606 } 3607 3608