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/module.h> 31 #include <linux/console.h> 32 #include <linux/slab.h> 33 34 #include <drm/drm_atomic_helper.h> 35 #include <drm/drm_probe_helper.h> 36 #include <drm/amdgpu_drm.h> 37 #include <linux/vgaarb.h> 38 #include <linux/vga_switcheroo.h> 39 #include <linux/efi.h> 40 #include "amdgpu.h" 41 #include "amdgpu_trace.h" 42 #include "amdgpu_i2c.h" 43 #include "atom.h" 44 #include "amdgpu_atombios.h" 45 #include "amdgpu_atomfirmware.h" 46 #include "amd_pcie.h" 47 #ifdef CONFIG_DRM_AMDGPU_SI 48 #include "si.h" 49 #endif 50 #ifdef CONFIG_DRM_AMDGPU_CIK 51 #include "cik.h" 52 #endif 53 #include "vi.h" 54 #include "soc15.h" 55 #include "nv.h" 56 #include "bif/bif_4_1_d.h" 57 #include <linux/pci.h> 58 #include <linux/firmware.h> 59 #include "amdgpu_vf_error.h" 60 61 #include "amdgpu_amdkfd.h" 62 #include "amdgpu_pm.h" 63 64 #include "amdgpu_xgmi.h" 65 #include "amdgpu_ras.h" 66 #include "amdgpu_pmu.h" 67 68 #include <linux/suspend.h> 69 70 MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin"); 71 MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin"); 72 MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin"); 73 MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin"); 74 MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin"); 75 MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin"); 76 MODULE_FIRMWARE("amdgpu/renoir_gpu_info.bin"); 77 MODULE_FIRMWARE("amdgpu/navi10_gpu_info.bin"); 78 MODULE_FIRMWARE("amdgpu/navi14_gpu_info.bin"); 79 MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin"); 80 81 #define AMDGPU_RESUME_MS 2000 82 83 const char *amdgpu_asic_name[] = { 84 "TAHITI", 85 "PITCAIRN", 86 "VERDE", 87 "OLAND", 88 "HAINAN", 89 "BONAIRE", 90 "KAVERI", 91 "KABINI", 92 "HAWAII", 93 "MULLINS", 94 "TOPAZ", 95 "TONGA", 96 "FIJI", 97 "CARRIZO", 98 "STONEY", 99 "POLARIS10", 100 "POLARIS11", 101 "POLARIS12", 102 "VEGAM", 103 "VEGA10", 104 "VEGA12", 105 "VEGA20", 106 "RAVEN", 107 "ARCTURUS", 108 "RENOIR", 109 "NAVI10", 110 "NAVI14", 111 "NAVI12", 112 "LAST", 113 }; 114 115 /** 116 * DOC: pcie_replay_count 117 * 118 * The amdgpu driver provides a sysfs API for reporting the total number 119 * of PCIe replays (NAKs) 120 * The file pcie_replay_count is used for this and returns the total 121 * number of replays as a sum of the NAKs generated and NAKs received 122 */ 123 124 static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev, 125 struct device_attribute *attr, char *buf) 126 { 127 struct drm_device *ddev = dev_get_drvdata(dev); 128 struct amdgpu_device *adev = ddev->dev_private; 129 uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev); 130 131 return snprintf(buf, PAGE_SIZE, "%llu\n", cnt); 132 } 133 134 static DEVICE_ATTR(pcie_replay_count, S_IRUGO, 135 amdgpu_device_get_pcie_replay_count, NULL); 136 137 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev); 138 139 /** 140 * amdgpu_device_is_px - Is the device is a dGPU with HG/PX power control 141 * 142 * @dev: drm_device pointer 143 * 144 * Returns true if the device is a dGPU with HG/PX power control, 145 * otherwise return false. 146 */ 147 bool amdgpu_device_is_px(struct drm_device *dev) 148 { 149 struct amdgpu_device *adev = dev->dev_private; 150 151 if (adev->flags & AMD_IS_PX) 152 return true; 153 return false; 154 } 155 156 /** 157 * VRAM access helper functions. 158 * 159 * amdgpu_device_vram_access - read/write a buffer in vram 160 * 161 * @adev: amdgpu_device pointer 162 * @pos: offset of the buffer in vram 163 * @buf: virtual address of the buffer in system memory 164 * @size: read/write size, sizeof(@buf) must > @size 165 * @write: true - write to vram, otherwise - read from vram 166 */ 167 void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, 168 uint32_t *buf, size_t size, bool write) 169 { 170 uint64_t last; 171 unsigned long flags; 172 173 last = size - 4; 174 for (last += pos; pos <= last; pos += 4) { 175 spin_lock_irqsave(&adev->mmio_idx_lock, flags); 176 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)pos) | 0x80000000); 177 WREG32_NO_KIQ(mmMM_INDEX_HI, pos >> 31); 178 if (write) 179 WREG32_NO_KIQ(mmMM_DATA, *buf++); 180 else 181 *buf++ = RREG32_NO_KIQ(mmMM_DATA); 182 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags); 183 } 184 } 185 186 /* 187 * MMIO register access helper functions. 188 */ 189 /** 190 * amdgpu_mm_rreg - read a memory mapped IO register 191 * 192 * @adev: amdgpu_device pointer 193 * @reg: dword aligned register offset 194 * @acc_flags: access flags which require special behavior 195 * 196 * Returns the 32 bit value from the offset specified. 197 */ 198 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, 199 uint32_t acc_flags) 200 { 201 uint32_t ret; 202 203 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev)) 204 return amdgpu_virt_kiq_rreg(adev, reg); 205 206 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX)) 207 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4)); 208 else { 209 unsigned long flags; 210 211 spin_lock_irqsave(&adev->mmio_idx_lock, flags); 212 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4)); 213 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4)); 214 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags); 215 } 216 trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret); 217 return ret; 218 } 219 220 /* 221 * MMIO register read with bytes helper functions 222 * @offset:bytes offset from MMIO start 223 * 224 */ 225 226 /** 227 * amdgpu_mm_rreg8 - read a memory mapped IO register 228 * 229 * @adev: amdgpu_device pointer 230 * @offset: byte aligned register offset 231 * 232 * Returns the 8 bit value from the offset specified. 233 */ 234 uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) { 235 if (offset < adev->rmmio_size) 236 return (readb(adev->rmmio + offset)); 237 BUG(); 238 } 239 240 /* 241 * MMIO register write with bytes helper functions 242 * @offset:bytes offset from MMIO start 243 * @value: the value want to be written to the register 244 * 245 */ 246 /** 247 * amdgpu_mm_wreg8 - read a memory mapped IO register 248 * 249 * @adev: amdgpu_device pointer 250 * @offset: byte aligned register offset 251 * @value: 8 bit value to write 252 * 253 * Writes the value specified to the offset specified. 254 */ 255 void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) { 256 if (offset < adev->rmmio_size) 257 writeb(value, adev->rmmio + offset); 258 else 259 BUG(); 260 } 261 262 /** 263 * amdgpu_mm_wreg - write to a memory mapped IO register 264 * 265 * @adev: amdgpu_device pointer 266 * @reg: dword aligned register offset 267 * @v: 32 bit value to write to the register 268 * @acc_flags: access flags which require special behavior 269 * 270 * Writes the value specified to the offset specified. 271 */ 272 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v, 273 uint32_t acc_flags) 274 { 275 trace_amdgpu_mm_wreg(adev->pdev->device, reg, v); 276 277 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) { 278 adev->last_mm_index = v; 279 } 280 281 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev)) 282 return amdgpu_virt_kiq_wreg(adev, reg, v); 283 284 if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX)) 285 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4)); 286 else { 287 unsigned long flags; 288 289 spin_lock_irqsave(&adev->mmio_idx_lock, flags); 290 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4)); 291 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4)); 292 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags); 293 } 294 295 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) { 296 udelay(500); 297 } 298 } 299 300 /** 301 * amdgpu_io_rreg - read an IO register 302 * 303 * @adev: amdgpu_device pointer 304 * @reg: dword aligned register offset 305 * 306 * Returns the 32 bit value from the offset specified. 307 */ 308 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg) 309 { 310 if ((reg * 4) < adev->rio_mem_size) 311 return ioread32(adev->rio_mem + (reg * 4)); 312 else { 313 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4)); 314 return ioread32(adev->rio_mem + (mmMM_DATA * 4)); 315 } 316 } 317 318 /** 319 * amdgpu_io_wreg - write to an IO register 320 * 321 * @adev: amdgpu_device pointer 322 * @reg: dword aligned register offset 323 * @v: 32 bit value to write to the register 324 * 325 * Writes the value specified to the offset specified. 326 */ 327 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v) 328 { 329 if (adev->asic_type >= CHIP_VEGA10 && reg == 0) { 330 adev->last_mm_index = v; 331 } 332 333 if ((reg * 4) < adev->rio_mem_size) 334 iowrite32(v, adev->rio_mem + (reg * 4)); 335 else { 336 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4)); 337 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4)); 338 } 339 340 if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) { 341 udelay(500); 342 } 343 } 344 345 /** 346 * amdgpu_mm_rdoorbell - read a doorbell dword 347 * 348 * @adev: amdgpu_device pointer 349 * @index: doorbell index 350 * 351 * Returns the value in the doorbell aperture at the 352 * requested doorbell index (CIK). 353 */ 354 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index) 355 { 356 if (index < adev->doorbell.num_doorbells) { 357 return readl(adev->doorbell.ptr + index); 358 } else { 359 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index); 360 return 0; 361 } 362 } 363 364 /** 365 * amdgpu_mm_wdoorbell - write a doorbell dword 366 * 367 * @adev: amdgpu_device pointer 368 * @index: doorbell index 369 * @v: value to write 370 * 371 * Writes @v to the doorbell aperture at the 372 * requested doorbell index (CIK). 373 */ 374 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v) 375 { 376 if (index < adev->doorbell.num_doorbells) { 377 writel(v, adev->doorbell.ptr + index); 378 } else { 379 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index); 380 } 381 } 382 383 /** 384 * amdgpu_mm_rdoorbell64 - read a doorbell Qword 385 * 386 * @adev: amdgpu_device pointer 387 * @index: doorbell index 388 * 389 * Returns the value in the doorbell aperture at the 390 * requested doorbell index (VEGA10+). 391 */ 392 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index) 393 { 394 if (index < adev->doorbell.num_doorbells) { 395 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index)); 396 } else { 397 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index); 398 return 0; 399 } 400 } 401 402 /** 403 * amdgpu_mm_wdoorbell64 - write a doorbell Qword 404 * 405 * @adev: amdgpu_device pointer 406 * @index: doorbell index 407 * @v: value to write 408 * 409 * Writes @v to the doorbell aperture at the 410 * requested doorbell index (VEGA10+). 411 */ 412 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v) 413 { 414 if (index < adev->doorbell.num_doorbells) { 415 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v); 416 } else { 417 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index); 418 } 419 } 420 421 /** 422 * amdgpu_invalid_rreg - dummy reg read function 423 * 424 * @adev: amdgpu device pointer 425 * @reg: offset of register 426 * 427 * Dummy register read function. Used for register blocks 428 * that certain asics don't have (all asics). 429 * Returns the value in the register. 430 */ 431 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg) 432 { 433 DRM_ERROR("Invalid callback to read register 0x%04X\n", reg); 434 BUG(); 435 return 0; 436 } 437 438 /** 439 * amdgpu_invalid_wreg - dummy reg write function 440 * 441 * @adev: amdgpu device pointer 442 * @reg: offset of register 443 * @v: value to write to the register 444 * 445 * Dummy register read function. Used for register blocks 446 * that certain asics don't have (all asics). 447 */ 448 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v) 449 { 450 DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n", 451 reg, v); 452 BUG(); 453 } 454 455 /** 456 * amdgpu_invalid_rreg64 - dummy 64 bit reg read function 457 * 458 * @adev: amdgpu device pointer 459 * @reg: offset of register 460 * 461 * Dummy register read function. Used for register blocks 462 * that certain asics don't have (all asics). 463 * Returns the value in the register. 464 */ 465 static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg) 466 { 467 DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg); 468 BUG(); 469 return 0; 470 } 471 472 /** 473 * amdgpu_invalid_wreg64 - dummy reg write function 474 * 475 * @adev: amdgpu device pointer 476 * @reg: offset of register 477 * @v: value to write to the register 478 * 479 * Dummy register read function. Used for register blocks 480 * that certain asics don't have (all asics). 481 */ 482 static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v) 483 { 484 DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n", 485 reg, v); 486 BUG(); 487 } 488 489 /** 490 * amdgpu_block_invalid_rreg - dummy reg read function 491 * 492 * @adev: amdgpu device pointer 493 * @block: offset of instance 494 * @reg: offset of register 495 * 496 * Dummy register read function. Used for register blocks 497 * that certain asics don't have (all asics). 498 * Returns the value in the register. 499 */ 500 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev, 501 uint32_t block, uint32_t reg) 502 { 503 DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n", 504 reg, block); 505 BUG(); 506 return 0; 507 } 508 509 /** 510 * amdgpu_block_invalid_wreg - dummy reg write function 511 * 512 * @adev: amdgpu device pointer 513 * @block: offset of instance 514 * @reg: offset of register 515 * @v: value to write to the register 516 * 517 * Dummy register read function. Used for register blocks 518 * that certain asics don't have (all asics). 519 */ 520 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev, 521 uint32_t block, 522 uint32_t reg, uint32_t v) 523 { 524 DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n", 525 reg, block, v); 526 BUG(); 527 } 528 529 /** 530 * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page 531 * 532 * @adev: amdgpu device pointer 533 * 534 * Allocates a scratch page of VRAM for use by various things in the 535 * driver. 536 */ 537 static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev) 538 { 539 return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE, 540 PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM, 541 &adev->vram_scratch.robj, 542 &adev->vram_scratch.gpu_addr, 543 (void **)&adev->vram_scratch.ptr); 544 } 545 546 /** 547 * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page 548 * 549 * @adev: amdgpu device pointer 550 * 551 * Frees the VRAM scratch page. 552 */ 553 static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev) 554 { 555 amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL); 556 } 557 558 /** 559 * amdgpu_device_program_register_sequence - program an array of registers. 560 * 561 * @adev: amdgpu_device pointer 562 * @registers: pointer to the register array 563 * @array_size: size of the register array 564 * 565 * Programs an array or registers with and and or masks. 566 * This is a helper for setting golden registers. 567 */ 568 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev, 569 const u32 *registers, 570 const u32 array_size) 571 { 572 u32 tmp, reg, and_mask, or_mask; 573 int i; 574 575 if (array_size % 3) 576 return; 577 578 for (i = 0; i < array_size; i +=3) { 579 reg = registers[i + 0]; 580 and_mask = registers[i + 1]; 581 or_mask = registers[i + 2]; 582 583 if (and_mask == 0xffffffff) { 584 tmp = or_mask; 585 } else { 586 tmp = RREG32(reg); 587 tmp &= ~and_mask; 588 if (adev->family >= AMDGPU_FAMILY_AI) 589 tmp |= (or_mask & and_mask); 590 else 591 tmp |= or_mask; 592 } 593 WREG32(reg, tmp); 594 } 595 } 596 597 /** 598 * amdgpu_device_pci_config_reset - reset the GPU 599 * 600 * @adev: amdgpu_device pointer 601 * 602 * Resets the GPU using the pci config reset sequence. 603 * Only applicable to asics prior to vega10. 604 */ 605 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev) 606 { 607 pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA); 608 } 609 610 /* 611 * GPU doorbell aperture helpers function. 612 */ 613 /** 614 * amdgpu_device_doorbell_init - Init doorbell driver information. 615 * 616 * @adev: amdgpu_device pointer 617 * 618 * Init doorbell driver information (CIK) 619 * Returns 0 on success, error on failure. 620 */ 621 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev) 622 { 623 624 /* No doorbell on SI hardware generation */ 625 if (adev->asic_type < CHIP_BONAIRE) { 626 adev->doorbell.base = 0; 627 adev->doorbell.size = 0; 628 adev->doorbell.num_doorbells = 0; 629 adev->doorbell.ptr = NULL; 630 return 0; 631 } 632 633 if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET) 634 return -EINVAL; 635 636 amdgpu_asic_init_doorbell_index(adev); 637 638 /* doorbell bar mapping */ 639 adev->doorbell.base = pci_resource_start(adev->pdev, 2); 640 adev->doorbell.size = pci_resource_len(adev->pdev, 2); 641 642 adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32), 643 adev->doorbell_index.max_assignment+1); 644 if (adev->doorbell.num_doorbells == 0) 645 return -EINVAL; 646 647 /* For Vega, reserve and map two pages on doorbell BAR since SDMA 648 * paging queue doorbell use the second page. The 649 * AMDGPU_DOORBELL64_MAX_ASSIGNMENT definition assumes all the 650 * doorbells are in the first page. So with paging queue enabled, 651 * the max num_doorbells should + 1 page (0x400 in dword) 652 */ 653 if (adev->asic_type >= CHIP_VEGA10) 654 adev->doorbell.num_doorbells += 0x400; 655 656 adev->doorbell.ptr = ioremap(adev->doorbell.base, 657 adev->doorbell.num_doorbells * 658 sizeof(u32)); 659 if (adev->doorbell.ptr == NULL) 660 return -ENOMEM; 661 662 return 0; 663 } 664 665 /** 666 * amdgpu_device_doorbell_fini - Tear down doorbell driver information. 667 * 668 * @adev: amdgpu_device pointer 669 * 670 * Tear down doorbell driver information (CIK) 671 */ 672 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev) 673 { 674 iounmap(adev->doorbell.ptr); 675 adev->doorbell.ptr = NULL; 676 } 677 678 679 680 /* 681 * amdgpu_device_wb_*() 682 * Writeback is the method by which the GPU updates special pages in memory 683 * with the status of certain GPU events (fences, ring pointers,etc.). 684 */ 685 686 /** 687 * amdgpu_device_wb_fini - Disable Writeback and free memory 688 * 689 * @adev: amdgpu_device pointer 690 * 691 * Disables Writeback and frees the Writeback memory (all asics). 692 * Used at driver shutdown. 693 */ 694 static void amdgpu_device_wb_fini(struct amdgpu_device *adev) 695 { 696 if (adev->wb.wb_obj) { 697 amdgpu_bo_free_kernel(&adev->wb.wb_obj, 698 &adev->wb.gpu_addr, 699 (void **)&adev->wb.wb); 700 adev->wb.wb_obj = NULL; 701 } 702 } 703 704 /** 705 * amdgpu_device_wb_init- Init Writeback driver info and allocate memory 706 * 707 * @adev: amdgpu_device pointer 708 * 709 * Initializes writeback and allocates writeback memory (all asics). 710 * Used at driver startup. 711 * Returns 0 on success or an -error on failure. 712 */ 713 static int amdgpu_device_wb_init(struct amdgpu_device *adev) 714 { 715 int r; 716 717 if (adev->wb.wb_obj == NULL) { 718 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */ 719 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8, 720 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT, 721 &adev->wb.wb_obj, &adev->wb.gpu_addr, 722 (void **)&adev->wb.wb); 723 if (r) { 724 dev_warn(adev->dev, "(%d) create WB bo failed\n", r); 725 return r; 726 } 727 728 adev->wb.num_wb = AMDGPU_MAX_WB; 729 memset(&adev->wb.used, 0, sizeof(adev->wb.used)); 730 731 /* clear wb memory */ 732 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8); 733 } 734 735 return 0; 736 } 737 738 /** 739 * amdgpu_device_wb_get - Allocate a wb entry 740 * 741 * @adev: amdgpu_device pointer 742 * @wb: wb index 743 * 744 * Allocate a wb slot for use by the driver (all asics). 745 * Returns 0 on success or -EINVAL on failure. 746 */ 747 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb) 748 { 749 unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb); 750 751 if (offset < adev->wb.num_wb) { 752 __set_bit(offset, adev->wb.used); 753 *wb = offset << 3; /* convert to dw offset */ 754 return 0; 755 } else { 756 return -EINVAL; 757 } 758 } 759 760 /** 761 * amdgpu_device_wb_free - Free a wb entry 762 * 763 * @adev: amdgpu_device pointer 764 * @wb: wb index 765 * 766 * Free a wb slot allocated for use by the driver (all asics) 767 */ 768 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb) 769 { 770 wb >>= 3; 771 if (wb < adev->wb.num_wb) 772 __clear_bit(wb, adev->wb.used); 773 } 774 775 /** 776 * amdgpu_device_resize_fb_bar - try to resize FB BAR 777 * 778 * @adev: amdgpu_device pointer 779 * 780 * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not 781 * to fail, but if any of the BARs is not accessible after the size we abort 782 * driver loading by returning -ENODEV. 783 */ 784 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev) 785 { 786 u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size); 787 u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1; 788 struct pci_bus *root; 789 struct resource *res; 790 unsigned i; 791 u16 cmd; 792 int r; 793 794 /* Bypass for VF */ 795 if (amdgpu_sriov_vf(adev)) 796 return 0; 797 798 /* Check if the root BUS has 64bit memory resources */ 799 root = adev->pdev->bus; 800 while (root->parent) 801 root = root->parent; 802 803 pci_bus_for_each_resource(root, res, i) { 804 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) && 805 res->start > 0x100000000ull) 806 break; 807 } 808 809 /* Trying to resize is pointless without a root hub window above 4GB */ 810 if (!res) 811 return 0; 812 813 /* Disable memory decoding while we change the BAR addresses and size */ 814 pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd); 815 pci_write_config_word(adev->pdev, PCI_COMMAND, 816 cmd & ~PCI_COMMAND_MEMORY); 817 818 /* Free the VRAM and doorbell BAR, we most likely need to move both. */ 819 amdgpu_device_doorbell_fini(adev); 820 if (adev->asic_type >= CHIP_BONAIRE) 821 pci_release_resource(adev->pdev, 2); 822 823 pci_release_resource(adev->pdev, 0); 824 825 r = pci_resize_resource(adev->pdev, 0, rbar_size); 826 if (r == -ENOSPC) 827 DRM_INFO("Not enough PCI address space for a large BAR."); 828 else if (r && r != -ENOTSUPP) 829 DRM_ERROR("Problem resizing BAR0 (%d).", r); 830 831 pci_assign_unassigned_bus_resources(adev->pdev->bus); 832 833 /* When the doorbell or fb BAR isn't available we have no chance of 834 * using the device. 835 */ 836 r = amdgpu_device_doorbell_init(adev); 837 if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET)) 838 return -ENODEV; 839 840 pci_write_config_word(adev->pdev, PCI_COMMAND, cmd); 841 842 return 0; 843 } 844 845 /* 846 * GPU helpers function. 847 */ 848 /** 849 * amdgpu_device_need_post - check if the hw need post or not 850 * 851 * @adev: amdgpu_device pointer 852 * 853 * Check if the asic has been initialized (all asics) at driver startup 854 * or post is needed if hw reset is performed. 855 * Returns true if need or false if not. 856 */ 857 bool amdgpu_device_need_post(struct amdgpu_device *adev) 858 { 859 uint32_t reg; 860 861 if (amdgpu_sriov_vf(adev)) 862 return false; 863 864 if (amdgpu_passthrough(adev)) { 865 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot 866 * some old smc fw still need driver do vPost otherwise gpu hang, while 867 * those smc fw version above 22.15 doesn't have this flaw, so we force 868 * vpost executed for smc version below 22.15 869 */ 870 if (adev->asic_type == CHIP_FIJI) { 871 int err; 872 uint32_t fw_ver; 873 err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev); 874 /* force vPost if error occured */ 875 if (err) 876 return true; 877 878 fw_ver = *((uint32_t *)adev->pm.fw->data + 69); 879 if (fw_ver < 0x00160e00) 880 return true; 881 } 882 } 883 884 if (adev->has_hw_reset) { 885 adev->has_hw_reset = false; 886 return true; 887 } 888 889 /* bios scratch used on CIK+ */ 890 if (adev->asic_type >= CHIP_BONAIRE) 891 return amdgpu_atombios_scratch_need_asic_init(adev); 892 893 /* check MEM_SIZE for older asics */ 894 reg = amdgpu_asic_get_config_memsize(adev); 895 896 if ((reg != 0) && (reg != 0xffffffff)) 897 return false; 898 899 return true; 900 } 901 902 /* if we get transitioned to only one device, take VGA back */ 903 /** 904 * amdgpu_device_vga_set_decode - enable/disable vga decode 905 * 906 * @cookie: amdgpu_device pointer 907 * @state: enable/disable vga decode 908 * 909 * Enable/disable vga decode (all asics). 910 * Returns VGA resource flags. 911 */ 912 static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state) 913 { 914 struct amdgpu_device *adev = cookie; 915 amdgpu_asic_set_vga_state(adev, state); 916 if (state) 917 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM | 918 VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM; 919 else 920 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM; 921 } 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 struct sysinfo si; 972 bool is_os_64 = (sizeof(void *) == 8) ? true : false; 973 uint64_t total_memory; 974 uint64_t dram_size_seven_GB = 0x1B8000000; 975 uint64_t dram_size_three_GB = 0xB8000000; 976 977 if (amdgpu_smu_memory_pool_size == 0) 978 return; 979 980 if (!is_os_64) { 981 DRM_WARN("Not 64-bit OS, feature not supported\n"); 982 goto def_value; 983 } 984 si_meminfo(&si); 985 total_memory = (uint64_t)si.totalram * si.mem_unit; 986 987 if ((amdgpu_smu_memory_pool_size == 1) || 988 (amdgpu_smu_memory_pool_size == 2)) { 989 if (total_memory < dram_size_three_GB) 990 goto def_value1; 991 } else if ((amdgpu_smu_memory_pool_size == 4) || 992 (amdgpu_smu_memory_pool_size == 8)) { 993 if (total_memory < dram_size_seven_GB) 994 goto def_value1; 995 } else { 996 DRM_WARN("Smu memory pool size not supported\n"); 997 goto def_value; 998 } 999 adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28; 1000 1001 return; 1002 1003 def_value1: 1004 DRM_WARN("No enough system memory\n"); 1005 def_value: 1006 adev->pm.smu_prv_buffer_size = 0; 1007 } 1008 1009 /** 1010 * amdgpu_device_check_arguments - validate module params 1011 * 1012 * @adev: amdgpu_device pointer 1013 * 1014 * Validates certain module parameters and updates 1015 * the associated values used by the driver (all asics). 1016 */ 1017 static int amdgpu_device_check_arguments(struct amdgpu_device *adev) 1018 { 1019 int ret = 0; 1020 1021 if (amdgpu_sched_jobs < 4) { 1022 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n", 1023 amdgpu_sched_jobs); 1024 amdgpu_sched_jobs = 4; 1025 } else if (!is_power_of_2(amdgpu_sched_jobs)){ 1026 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n", 1027 amdgpu_sched_jobs); 1028 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs); 1029 } 1030 1031 if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) { 1032 /* gart size must be greater or equal to 32M */ 1033 dev_warn(adev->dev, "gart size (%d) too small\n", 1034 amdgpu_gart_size); 1035 amdgpu_gart_size = -1; 1036 } 1037 1038 if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) { 1039 /* gtt size must be greater or equal to 32M */ 1040 dev_warn(adev->dev, "gtt size (%d) too small\n", 1041 amdgpu_gtt_size); 1042 amdgpu_gtt_size = -1; 1043 } 1044 1045 /* valid range is between 4 and 9 inclusive */ 1046 if (amdgpu_vm_fragment_size != -1 && 1047 (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) { 1048 dev_warn(adev->dev, "valid range is between 4 and 9\n"); 1049 amdgpu_vm_fragment_size = -1; 1050 } 1051 1052 amdgpu_device_check_smu_prv_buffer_size(adev); 1053 1054 amdgpu_device_check_vm_size(adev); 1055 1056 amdgpu_device_check_block_size(adev); 1057 1058 adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type); 1059 1060 return ret; 1061 } 1062 1063 /** 1064 * amdgpu_switcheroo_set_state - set switcheroo state 1065 * 1066 * @pdev: pci dev pointer 1067 * @state: vga_switcheroo state 1068 * 1069 * Callback for the switcheroo driver. Suspends or resumes the 1070 * the asics before or after it is powered up using ACPI methods. 1071 */ 1072 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state) 1073 { 1074 struct drm_device *dev = pci_get_drvdata(pdev); 1075 1076 if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF) 1077 return; 1078 1079 if (state == VGA_SWITCHEROO_ON) { 1080 pr_info("amdgpu: switched on\n"); 1081 /* don't suspend or resume card normally */ 1082 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING; 1083 1084 amdgpu_device_resume(dev, true, true); 1085 1086 dev->switch_power_state = DRM_SWITCH_POWER_ON; 1087 drm_kms_helper_poll_enable(dev); 1088 } else { 1089 pr_info("amdgpu: switched off\n"); 1090 drm_kms_helper_poll_disable(dev); 1091 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING; 1092 amdgpu_device_suspend(dev, true, true); 1093 dev->switch_power_state = DRM_SWITCH_POWER_OFF; 1094 } 1095 } 1096 1097 /** 1098 * amdgpu_switcheroo_can_switch - see if switcheroo state can change 1099 * 1100 * @pdev: pci dev pointer 1101 * 1102 * Callback for the switcheroo driver. Check of the switcheroo 1103 * state can be changed. 1104 * Returns true if the state can be changed, false if not. 1105 */ 1106 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev) 1107 { 1108 struct drm_device *dev = pci_get_drvdata(pdev); 1109 1110 /* 1111 * FIXME: open_count is protected by drm_global_mutex but that would lead to 1112 * locking inversion with the driver load path. And the access here is 1113 * completely racy anyway. So don't bother with locking for now. 1114 */ 1115 return dev->open_count == 0; 1116 } 1117 1118 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = { 1119 .set_gpu_state = amdgpu_switcheroo_set_state, 1120 .reprobe = NULL, 1121 .can_switch = amdgpu_switcheroo_can_switch, 1122 }; 1123 1124 /** 1125 * amdgpu_device_ip_set_clockgating_state - set the CG state 1126 * 1127 * @dev: amdgpu_device pointer 1128 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) 1129 * @state: clockgating state (gate or ungate) 1130 * 1131 * Sets the requested clockgating state for all instances of 1132 * the hardware IP specified. 1133 * Returns the error code from the last instance. 1134 */ 1135 int amdgpu_device_ip_set_clockgating_state(void *dev, 1136 enum amd_ip_block_type block_type, 1137 enum amd_clockgating_state state) 1138 { 1139 struct amdgpu_device *adev = dev; 1140 int i, r = 0; 1141 1142 for (i = 0; i < adev->num_ip_blocks; i++) { 1143 if (!adev->ip_blocks[i].status.valid) 1144 continue; 1145 if (adev->ip_blocks[i].version->type != block_type) 1146 continue; 1147 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state) 1148 continue; 1149 r = adev->ip_blocks[i].version->funcs->set_clockgating_state( 1150 (void *)adev, state); 1151 if (r) 1152 DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n", 1153 adev->ip_blocks[i].version->funcs->name, r); 1154 } 1155 return r; 1156 } 1157 1158 /** 1159 * amdgpu_device_ip_set_powergating_state - set the PG state 1160 * 1161 * @dev: amdgpu_device pointer 1162 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) 1163 * @state: powergating state (gate or ungate) 1164 * 1165 * Sets the requested powergating state for all instances of 1166 * the hardware IP specified. 1167 * Returns the error code from the last instance. 1168 */ 1169 int amdgpu_device_ip_set_powergating_state(void *dev, 1170 enum amd_ip_block_type block_type, 1171 enum amd_powergating_state state) 1172 { 1173 struct amdgpu_device *adev = dev; 1174 int i, r = 0; 1175 1176 for (i = 0; i < adev->num_ip_blocks; i++) { 1177 if (!adev->ip_blocks[i].status.valid) 1178 continue; 1179 if (adev->ip_blocks[i].version->type != block_type) 1180 continue; 1181 if (!adev->ip_blocks[i].version->funcs->set_powergating_state) 1182 continue; 1183 r = adev->ip_blocks[i].version->funcs->set_powergating_state( 1184 (void *)adev, state); 1185 if (r) 1186 DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n", 1187 adev->ip_blocks[i].version->funcs->name, r); 1188 } 1189 return r; 1190 } 1191 1192 /** 1193 * amdgpu_device_ip_get_clockgating_state - get the CG state 1194 * 1195 * @adev: amdgpu_device pointer 1196 * @flags: clockgating feature flags 1197 * 1198 * Walks the list of IPs on the device and updates the clockgating 1199 * flags for each IP. 1200 * Updates @flags with the feature flags for each hardware IP where 1201 * clockgating is enabled. 1202 */ 1203 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev, 1204 u32 *flags) 1205 { 1206 int i; 1207 1208 for (i = 0; i < adev->num_ip_blocks; i++) { 1209 if (!adev->ip_blocks[i].status.valid) 1210 continue; 1211 if (adev->ip_blocks[i].version->funcs->get_clockgating_state) 1212 adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags); 1213 } 1214 } 1215 1216 /** 1217 * amdgpu_device_ip_wait_for_idle - wait for idle 1218 * 1219 * @adev: amdgpu_device pointer 1220 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) 1221 * 1222 * Waits for the request hardware IP to be idle. 1223 * Returns 0 for success or a negative error code on failure. 1224 */ 1225 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev, 1226 enum amd_ip_block_type block_type) 1227 { 1228 int i, r; 1229 1230 for (i = 0; i < adev->num_ip_blocks; i++) { 1231 if (!adev->ip_blocks[i].status.valid) 1232 continue; 1233 if (adev->ip_blocks[i].version->type == block_type) { 1234 r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev); 1235 if (r) 1236 return r; 1237 break; 1238 } 1239 } 1240 return 0; 1241 1242 } 1243 1244 /** 1245 * amdgpu_device_ip_is_idle - is the hardware IP idle 1246 * 1247 * @adev: amdgpu_device pointer 1248 * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.) 1249 * 1250 * Check if the hardware IP is idle or not. 1251 * Returns true if it the IP is idle, false if not. 1252 */ 1253 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev, 1254 enum amd_ip_block_type block_type) 1255 { 1256 int i; 1257 1258 for (i = 0; i < adev->num_ip_blocks; i++) { 1259 if (!adev->ip_blocks[i].status.valid) 1260 continue; 1261 if (adev->ip_blocks[i].version->type == block_type) 1262 return adev->ip_blocks[i].version->funcs->is_idle((void *)adev); 1263 } 1264 return true; 1265 1266 } 1267 1268 /** 1269 * amdgpu_device_ip_get_ip_block - get a hw IP pointer 1270 * 1271 * @adev: amdgpu_device pointer 1272 * @type: Type of hardware IP (SMU, GFX, UVD, etc.) 1273 * 1274 * Returns a pointer to the hardware IP block structure 1275 * if it exists for the asic, otherwise NULL. 1276 */ 1277 struct amdgpu_ip_block * 1278 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev, 1279 enum amd_ip_block_type type) 1280 { 1281 int i; 1282 1283 for (i = 0; i < adev->num_ip_blocks; i++) 1284 if (adev->ip_blocks[i].version->type == type) 1285 return &adev->ip_blocks[i]; 1286 1287 return NULL; 1288 } 1289 1290 /** 1291 * amdgpu_device_ip_block_version_cmp 1292 * 1293 * @adev: amdgpu_device pointer 1294 * @type: enum amd_ip_block_type 1295 * @major: major version 1296 * @minor: minor version 1297 * 1298 * return 0 if equal or greater 1299 * return 1 if smaller or the ip_block doesn't exist 1300 */ 1301 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev, 1302 enum amd_ip_block_type type, 1303 u32 major, u32 minor) 1304 { 1305 struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type); 1306 1307 if (ip_block && ((ip_block->version->major > major) || 1308 ((ip_block->version->major == major) && 1309 (ip_block->version->minor >= minor)))) 1310 return 0; 1311 1312 return 1; 1313 } 1314 1315 /** 1316 * amdgpu_device_ip_block_add 1317 * 1318 * @adev: amdgpu_device pointer 1319 * @ip_block_version: pointer to the IP to add 1320 * 1321 * Adds the IP block driver information to the collection of IPs 1322 * on the asic. 1323 */ 1324 int amdgpu_device_ip_block_add(struct amdgpu_device *adev, 1325 const struct amdgpu_ip_block_version *ip_block_version) 1326 { 1327 if (!ip_block_version) 1328 return -EINVAL; 1329 1330 DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks, 1331 ip_block_version->funcs->name); 1332 1333 adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version; 1334 1335 return 0; 1336 } 1337 1338 /** 1339 * amdgpu_device_enable_virtual_display - enable virtual display feature 1340 * 1341 * @adev: amdgpu_device pointer 1342 * 1343 * Enabled the virtual display feature if the user has enabled it via 1344 * the module parameter virtual_display. This feature provides a virtual 1345 * display hardware on headless boards or in virtualized environments. 1346 * This function parses and validates the configuration string specified by 1347 * the user and configues the virtual display configuration (number of 1348 * virtual connectors, crtcs, etc.) specified. 1349 */ 1350 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev) 1351 { 1352 adev->enable_virtual_display = false; 1353 1354 if (amdgpu_virtual_display) { 1355 struct drm_device *ddev = adev->ddev; 1356 const char *pci_address_name = pci_name(ddev->pdev); 1357 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname; 1358 1359 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL); 1360 pciaddstr_tmp = pciaddstr; 1361 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) { 1362 pciaddname = strsep(&pciaddname_tmp, ","); 1363 if (!strcmp("all", pciaddname) 1364 || !strcmp(pci_address_name, pciaddname)) { 1365 long num_crtc; 1366 int res = -1; 1367 1368 adev->enable_virtual_display = true; 1369 1370 if (pciaddname_tmp) 1371 res = kstrtol(pciaddname_tmp, 10, 1372 &num_crtc); 1373 1374 if (!res) { 1375 if (num_crtc < 1) 1376 num_crtc = 1; 1377 if (num_crtc > 6) 1378 num_crtc = 6; 1379 adev->mode_info.num_crtc = num_crtc; 1380 } else { 1381 adev->mode_info.num_crtc = 1; 1382 } 1383 break; 1384 } 1385 } 1386 1387 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n", 1388 amdgpu_virtual_display, pci_address_name, 1389 adev->enable_virtual_display, adev->mode_info.num_crtc); 1390 1391 kfree(pciaddstr); 1392 } 1393 } 1394 1395 /** 1396 * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware 1397 * 1398 * @adev: amdgpu_device pointer 1399 * 1400 * Parses the asic configuration parameters specified in the gpu info 1401 * firmware and makes them availale to the driver for use in configuring 1402 * the asic. 1403 * Returns 0 on success, -EINVAL on failure. 1404 */ 1405 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev) 1406 { 1407 const char *chip_name; 1408 char fw_name[30]; 1409 int err; 1410 const struct gpu_info_firmware_header_v1_0 *hdr; 1411 1412 adev->firmware.gpu_info_fw = NULL; 1413 1414 switch (adev->asic_type) { 1415 case CHIP_TOPAZ: 1416 case CHIP_TONGA: 1417 case CHIP_FIJI: 1418 case CHIP_POLARIS10: 1419 case CHIP_POLARIS11: 1420 case CHIP_POLARIS12: 1421 case CHIP_VEGAM: 1422 case CHIP_CARRIZO: 1423 case CHIP_STONEY: 1424 #ifdef CONFIG_DRM_AMDGPU_SI 1425 case CHIP_VERDE: 1426 case CHIP_TAHITI: 1427 case CHIP_PITCAIRN: 1428 case CHIP_OLAND: 1429 case CHIP_HAINAN: 1430 #endif 1431 #ifdef CONFIG_DRM_AMDGPU_CIK 1432 case CHIP_BONAIRE: 1433 case CHIP_HAWAII: 1434 case CHIP_KAVERI: 1435 case CHIP_KABINI: 1436 case CHIP_MULLINS: 1437 #endif 1438 case CHIP_VEGA20: 1439 default: 1440 return 0; 1441 case CHIP_VEGA10: 1442 chip_name = "vega10"; 1443 break; 1444 case CHIP_VEGA12: 1445 chip_name = "vega12"; 1446 break; 1447 case CHIP_RAVEN: 1448 if (adev->rev_id >= 8) 1449 chip_name = "raven2"; 1450 else if (adev->pdev->device == 0x15d8) 1451 chip_name = "picasso"; 1452 else 1453 chip_name = "raven"; 1454 break; 1455 case CHIP_ARCTURUS: 1456 chip_name = "arcturus"; 1457 break; 1458 case CHIP_RENOIR: 1459 chip_name = "renoir"; 1460 break; 1461 case CHIP_NAVI10: 1462 chip_name = "navi10"; 1463 break; 1464 case CHIP_NAVI14: 1465 chip_name = "navi14"; 1466 break; 1467 case CHIP_NAVI12: 1468 chip_name = "navi12"; 1469 break; 1470 } 1471 1472 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name); 1473 err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev); 1474 if (err) { 1475 dev_err(adev->dev, 1476 "Failed to load gpu_info firmware \"%s\"\n", 1477 fw_name); 1478 goto out; 1479 } 1480 err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw); 1481 if (err) { 1482 dev_err(adev->dev, 1483 "Failed to validate gpu_info firmware \"%s\"\n", 1484 fw_name); 1485 goto out; 1486 } 1487 1488 hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data; 1489 amdgpu_ucode_print_gpu_info_hdr(&hdr->header); 1490 1491 switch (hdr->version_major) { 1492 case 1: 1493 { 1494 const struct gpu_info_firmware_v1_0 *gpu_info_fw = 1495 (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data + 1496 le32_to_cpu(hdr->header.ucode_array_offset_bytes)); 1497 1498 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10) 1499 goto parse_soc_bounding_box; 1500 1501 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se); 1502 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh); 1503 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se); 1504 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se); 1505 adev->gfx.config.max_texture_channel_caches = 1506 le32_to_cpu(gpu_info_fw->gc_num_tccs); 1507 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs); 1508 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds); 1509 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth); 1510 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth); 1511 adev->gfx.config.double_offchip_lds_buf = 1512 le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer); 1513 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size); 1514 adev->gfx.cu_info.max_waves_per_simd = 1515 le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd); 1516 adev->gfx.cu_info.max_scratch_slots_per_cu = 1517 le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu); 1518 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size); 1519 if (hdr->version_minor >= 1) { 1520 const struct gpu_info_firmware_v1_1 *gpu_info_fw = 1521 (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data + 1522 le32_to_cpu(hdr->header.ucode_array_offset_bytes)); 1523 adev->gfx.config.num_sc_per_sh = 1524 le32_to_cpu(gpu_info_fw->num_sc_per_sh); 1525 adev->gfx.config.num_packer_per_sc = 1526 le32_to_cpu(gpu_info_fw->num_packer_per_sc); 1527 } 1528 1529 parse_soc_bounding_box: 1530 /* 1531 * soc bounding box info is not integrated in disocovery table, 1532 * we always need to parse it from gpu info firmware. 1533 */ 1534 if (hdr->version_minor == 2) { 1535 const struct gpu_info_firmware_v1_2 *gpu_info_fw = 1536 (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data + 1537 le32_to_cpu(hdr->header.ucode_array_offset_bytes)); 1538 adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box; 1539 } 1540 break; 1541 } 1542 default: 1543 dev_err(adev->dev, 1544 "Unsupported gpu_info table %d\n", hdr->header.ucode_version); 1545 err = -EINVAL; 1546 goto out; 1547 } 1548 out: 1549 return err; 1550 } 1551 1552 /** 1553 * amdgpu_device_ip_early_init - run early init for hardware IPs 1554 * 1555 * @adev: amdgpu_device pointer 1556 * 1557 * Early initialization pass for hardware IPs. The hardware IPs that make 1558 * up each asic are discovered each IP's early_init callback is run. This 1559 * is the first stage in initializing the asic. 1560 * Returns 0 on success, negative error code on failure. 1561 */ 1562 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev) 1563 { 1564 int i, r; 1565 1566 amdgpu_device_enable_virtual_display(adev); 1567 1568 switch (adev->asic_type) { 1569 case CHIP_TOPAZ: 1570 case CHIP_TONGA: 1571 case CHIP_FIJI: 1572 case CHIP_POLARIS10: 1573 case CHIP_POLARIS11: 1574 case CHIP_POLARIS12: 1575 case CHIP_VEGAM: 1576 case CHIP_CARRIZO: 1577 case CHIP_STONEY: 1578 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY) 1579 adev->family = AMDGPU_FAMILY_CZ; 1580 else 1581 adev->family = AMDGPU_FAMILY_VI; 1582 1583 r = vi_set_ip_blocks(adev); 1584 if (r) 1585 return r; 1586 break; 1587 #ifdef CONFIG_DRM_AMDGPU_SI 1588 case CHIP_VERDE: 1589 case CHIP_TAHITI: 1590 case CHIP_PITCAIRN: 1591 case CHIP_OLAND: 1592 case CHIP_HAINAN: 1593 adev->family = AMDGPU_FAMILY_SI; 1594 r = si_set_ip_blocks(adev); 1595 if (r) 1596 return r; 1597 break; 1598 #endif 1599 #ifdef CONFIG_DRM_AMDGPU_CIK 1600 case CHIP_BONAIRE: 1601 case CHIP_HAWAII: 1602 case CHIP_KAVERI: 1603 case CHIP_KABINI: 1604 case CHIP_MULLINS: 1605 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII)) 1606 adev->family = AMDGPU_FAMILY_CI; 1607 else 1608 adev->family = AMDGPU_FAMILY_KV; 1609 1610 r = cik_set_ip_blocks(adev); 1611 if (r) 1612 return r; 1613 break; 1614 #endif 1615 case CHIP_VEGA10: 1616 case CHIP_VEGA12: 1617 case CHIP_VEGA20: 1618 case CHIP_RAVEN: 1619 case CHIP_ARCTURUS: 1620 case CHIP_RENOIR: 1621 if (adev->asic_type == CHIP_RAVEN || 1622 adev->asic_type == CHIP_RENOIR) 1623 adev->family = AMDGPU_FAMILY_RV; 1624 else 1625 adev->family = AMDGPU_FAMILY_AI; 1626 1627 r = soc15_set_ip_blocks(adev); 1628 if (r) 1629 return r; 1630 break; 1631 case CHIP_NAVI10: 1632 case CHIP_NAVI14: 1633 case CHIP_NAVI12: 1634 adev->family = AMDGPU_FAMILY_NV; 1635 1636 r = nv_set_ip_blocks(adev); 1637 if (r) 1638 return r; 1639 break; 1640 default: 1641 /* FIXME: not supported yet */ 1642 return -EINVAL; 1643 } 1644 1645 r = amdgpu_device_parse_gpu_info_fw(adev); 1646 if (r) 1647 return r; 1648 1649 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10) 1650 amdgpu_discovery_get_gfx_info(adev); 1651 1652 amdgpu_amdkfd_device_probe(adev); 1653 1654 if (amdgpu_sriov_vf(adev)) { 1655 r = amdgpu_virt_request_full_gpu(adev, true); 1656 if (r) 1657 return -EAGAIN; 1658 } 1659 1660 adev->pm.pp_feature = amdgpu_pp_feature_mask; 1661 if (amdgpu_sriov_vf(adev) || sched_policy == KFD_SCHED_POLICY_NO_HWS) 1662 adev->pm.pp_feature &= ~PP_GFXOFF_MASK; 1663 1664 for (i = 0; i < adev->num_ip_blocks; i++) { 1665 if ((amdgpu_ip_block_mask & (1 << i)) == 0) { 1666 DRM_ERROR("disabled ip block: %d <%s>\n", 1667 i, adev->ip_blocks[i].version->funcs->name); 1668 adev->ip_blocks[i].status.valid = false; 1669 } else { 1670 if (adev->ip_blocks[i].version->funcs->early_init) { 1671 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev); 1672 if (r == -ENOENT) { 1673 adev->ip_blocks[i].status.valid = false; 1674 } else if (r) { 1675 DRM_ERROR("early_init of IP block <%s> failed %d\n", 1676 adev->ip_blocks[i].version->funcs->name, r); 1677 return r; 1678 } else { 1679 adev->ip_blocks[i].status.valid = true; 1680 } 1681 } else { 1682 adev->ip_blocks[i].status.valid = true; 1683 } 1684 } 1685 /* get the vbios after the asic_funcs are set up */ 1686 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) { 1687 /* Read BIOS */ 1688 if (!amdgpu_get_bios(adev)) 1689 return -EINVAL; 1690 1691 r = amdgpu_atombios_init(adev); 1692 if (r) { 1693 dev_err(adev->dev, "amdgpu_atombios_init failed\n"); 1694 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0); 1695 return r; 1696 } 1697 } 1698 } 1699 1700 adev->cg_flags &= amdgpu_cg_mask; 1701 adev->pg_flags &= amdgpu_pg_mask; 1702 1703 return 0; 1704 } 1705 1706 static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev) 1707 { 1708 int i, r; 1709 1710 for (i = 0; i < adev->num_ip_blocks; i++) { 1711 if (!adev->ip_blocks[i].status.sw) 1712 continue; 1713 if (adev->ip_blocks[i].status.hw) 1714 continue; 1715 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || 1716 (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) || 1717 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) { 1718 r = adev->ip_blocks[i].version->funcs->hw_init(adev); 1719 if (r) { 1720 DRM_ERROR("hw_init of IP block <%s> failed %d\n", 1721 adev->ip_blocks[i].version->funcs->name, r); 1722 return r; 1723 } 1724 adev->ip_blocks[i].status.hw = true; 1725 } 1726 } 1727 1728 return 0; 1729 } 1730 1731 static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev) 1732 { 1733 int i, r; 1734 1735 for (i = 0; i < adev->num_ip_blocks; i++) { 1736 if (!adev->ip_blocks[i].status.sw) 1737 continue; 1738 if (adev->ip_blocks[i].status.hw) 1739 continue; 1740 r = adev->ip_blocks[i].version->funcs->hw_init(adev); 1741 if (r) { 1742 DRM_ERROR("hw_init of IP block <%s> failed %d\n", 1743 adev->ip_blocks[i].version->funcs->name, r); 1744 return r; 1745 } 1746 adev->ip_blocks[i].status.hw = true; 1747 } 1748 1749 return 0; 1750 } 1751 1752 static int amdgpu_device_fw_loading(struct amdgpu_device *adev) 1753 { 1754 int r = 0; 1755 int i; 1756 uint32_t smu_version; 1757 1758 if (adev->asic_type >= CHIP_VEGA10) { 1759 for (i = 0; i < adev->num_ip_blocks; i++) { 1760 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP) 1761 continue; 1762 1763 /* no need to do the fw loading again if already done*/ 1764 if (adev->ip_blocks[i].status.hw == true) 1765 break; 1766 1767 if (adev->in_gpu_reset || adev->in_suspend) { 1768 r = adev->ip_blocks[i].version->funcs->resume(adev); 1769 if (r) { 1770 DRM_ERROR("resume of IP block <%s> failed %d\n", 1771 adev->ip_blocks[i].version->funcs->name, r); 1772 return r; 1773 } 1774 } else { 1775 r = adev->ip_blocks[i].version->funcs->hw_init(adev); 1776 if (r) { 1777 DRM_ERROR("hw_init of IP block <%s> failed %d\n", 1778 adev->ip_blocks[i].version->funcs->name, r); 1779 return r; 1780 } 1781 } 1782 1783 adev->ip_blocks[i].status.hw = true; 1784 break; 1785 } 1786 } 1787 1788 r = amdgpu_pm_load_smu_firmware(adev, &smu_version); 1789 1790 return r; 1791 } 1792 1793 /** 1794 * amdgpu_device_ip_init - run init for hardware IPs 1795 * 1796 * @adev: amdgpu_device pointer 1797 * 1798 * Main initialization pass for hardware IPs. The list of all the hardware 1799 * IPs that make up the asic is walked and the sw_init and hw_init callbacks 1800 * are run. sw_init initializes the software state associated with each IP 1801 * and hw_init initializes the hardware associated with each IP. 1802 * Returns 0 on success, negative error code on failure. 1803 */ 1804 static int amdgpu_device_ip_init(struct amdgpu_device *adev) 1805 { 1806 int i, r; 1807 1808 r = amdgpu_ras_init(adev); 1809 if (r) 1810 return r; 1811 1812 for (i = 0; i < adev->num_ip_blocks; i++) { 1813 if (!adev->ip_blocks[i].status.valid) 1814 continue; 1815 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev); 1816 if (r) { 1817 DRM_ERROR("sw_init of IP block <%s> failed %d\n", 1818 adev->ip_blocks[i].version->funcs->name, r); 1819 goto init_failed; 1820 } 1821 adev->ip_blocks[i].status.sw = true; 1822 1823 /* need to do gmc hw init early so we can allocate gpu mem */ 1824 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) { 1825 r = amdgpu_device_vram_scratch_init(adev); 1826 if (r) { 1827 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r); 1828 goto init_failed; 1829 } 1830 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev); 1831 if (r) { 1832 DRM_ERROR("hw_init %d failed %d\n", i, r); 1833 goto init_failed; 1834 } 1835 r = amdgpu_device_wb_init(adev); 1836 if (r) { 1837 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r); 1838 goto init_failed; 1839 } 1840 adev->ip_blocks[i].status.hw = true; 1841 1842 /* right after GMC hw init, we create CSA */ 1843 if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) { 1844 r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj, 1845 AMDGPU_GEM_DOMAIN_VRAM, 1846 AMDGPU_CSA_SIZE); 1847 if (r) { 1848 DRM_ERROR("allocate CSA failed %d\n", r); 1849 goto init_failed; 1850 } 1851 } 1852 } 1853 } 1854 1855 r = amdgpu_ib_pool_init(adev); 1856 if (r) { 1857 dev_err(adev->dev, "IB initialization failed (%d).\n", r); 1858 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r); 1859 goto init_failed; 1860 } 1861 1862 r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/ 1863 if (r) 1864 goto init_failed; 1865 1866 r = amdgpu_device_ip_hw_init_phase1(adev); 1867 if (r) 1868 goto init_failed; 1869 1870 r = amdgpu_device_fw_loading(adev); 1871 if (r) 1872 goto init_failed; 1873 1874 r = amdgpu_device_ip_hw_init_phase2(adev); 1875 if (r) 1876 goto init_failed; 1877 1878 /* 1879 * retired pages will be loaded from eeprom and reserved here, 1880 * it should be called after amdgpu_device_ip_hw_init_phase2 since 1881 * for some ASICs the RAS EEPROM code relies on SMU fully functioning 1882 * for I2C communication which only true at this point. 1883 * recovery_init may fail, but it can free all resources allocated by 1884 * itself and its failure should not stop amdgpu init process. 1885 * 1886 * Note: theoretically, this should be called before all vram allocations 1887 * to protect retired page from abusing 1888 */ 1889 amdgpu_ras_recovery_init(adev); 1890 1891 if (adev->gmc.xgmi.num_physical_nodes > 1) 1892 amdgpu_xgmi_add_device(adev); 1893 amdgpu_amdkfd_device_init(adev); 1894 1895 init_failed: 1896 if (amdgpu_sriov_vf(adev)) { 1897 if (!r) 1898 amdgpu_virt_init_data_exchange(adev); 1899 amdgpu_virt_release_full_gpu(adev, true); 1900 } 1901 1902 return r; 1903 } 1904 1905 /** 1906 * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer 1907 * 1908 * @adev: amdgpu_device pointer 1909 * 1910 * Writes a reset magic value to the gart pointer in VRAM. The driver calls 1911 * this function before a GPU reset. If the value is retained after a 1912 * GPU reset, VRAM has not been lost. Some GPU resets may destry VRAM contents. 1913 */ 1914 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev) 1915 { 1916 memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM); 1917 } 1918 1919 /** 1920 * amdgpu_device_check_vram_lost - check if vram is valid 1921 * 1922 * @adev: amdgpu_device pointer 1923 * 1924 * Checks the reset magic value written to the gart pointer in VRAM. 1925 * The driver calls this after a GPU reset to see if the contents of 1926 * VRAM is lost or now. 1927 * returns true if vram is lost, false if not. 1928 */ 1929 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev) 1930 { 1931 return !!memcmp(adev->gart.ptr, adev->reset_magic, 1932 AMDGPU_RESET_MAGIC_NUM); 1933 } 1934 1935 /** 1936 * amdgpu_device_set_cg_state - set clockgating for amdgpu device 1937 * 1938 * @adev: amdgpu_device pointer 1939 * @state: clockgating state (gate or ungate) 1940 * 1941 * The list of all the hardware IPs that make up the asic is walked and the 1942 * set_clockgating_state callbacks are run. 1943 * Late initialization pass enabling clockgating for hardware IPs. 1944 * Fini or suspend, pass disabling clockgating for hardware IPs. 1945 * Returns 0 on success, negative error code on failure. 1946 */ 1947 1948 static int amdgpu_device_set_cg_state(struct amdgpu_device *adev, 1949 enum amd_clockgating_state state) 1950 { 1951 int i, j, r; 1952 1953 if (amdgpu_emu_mode == 1) 1954 return 0; 1955 1956 for (j = 0; j < adev->num_ip_blocks; j++) { 1957 i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1; 1958 if (!adev->ip_blocks[i].status.late_initialized) 1959 continue; 1960 /* skip CG for VCE/UVD, it's handled specially */ 1961 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD && 1962 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE && 1963 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN && 1964 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG && 1965 adev->ip_blocks[i].version->funcs->set_clockgating_state) { 1966 /* enable clockgating to save power */ 1967 r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev, 1968 state); 1969 if (r) { 1970 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n", 1971 adev->ip_blocks[i].version->funcs->name, r); 1972 return r; 1973 } 1974 } 1975 } 1976 1977 return 0; 1978 } 1979 1980 static int amdgpu_device_set_pg_state(struct amdgpu_device *adev, enum amd_powergating_state state) 1981 { 1982 int i, j, r; 1983 1984 if (amdgpu_emu_mode == 1) 1985 return 0; 1986 1987 for (j = 0; j < adev->num_ip_blocks; j++) { 1988 i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1; 1989 if (!adev->ip_blocks[i].status.late_initialized) 1990 continue; 1991 /* skip CG for VCE/UVD, it's handled specially */ 1992 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD && 1993 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE && 1994 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN && 1995 adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG && 1996 adev->ip_blocks[i].version->funcs->set_powergating_state) { 1997 /* enable powergating to save power */ 1998 r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev, 1999 state); 2000 if (r) { 2001 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n", 2002 adev->ip_blocks[i].version->funcs->name, r); 2003 return r; 2004 } 2005 } 2006 } 2007 return 0; 2008 } 2009 2010 static int amdgpu_device_enable_mgpu_fan_boost(void) 2011 { 2012 struct amdgpu_gpu_instance *gpu_ins; 2013 struct amdgpu_device *adev; 2014 int i, ret = 0; 2015 2016 mutex_lock(&mgpu_info.mutex); 2017 2018 /* 2019 * MGPU fan boost feature should be enabled 2020 * only when there are two or more dGPUs in 2021 * the system 2022 */ 2023 if (mgpu_info.num_dgpu < 2) 2024 goto out; 2025 2026 for (i = 0; i < mgpu_info.num_dgpu; i++) { 2027 gpu_ins = &(mgpu_info.gpu_ins[i]); 2028 adev = gpu_ins->adev; 2029 if (!(adev->flags & AMD_IS_APU) && 2030 !gpu_ins->mgpu_fan_enabled && 2031 adev->powerplay.pp_funcs && 2032 adev->powerplay.pp_funcs->enable_mgpu_fan_boost) { 2033 ret = amdgpu_dpm_enable_mgpu_fan_boost(adev); 2034 if (ret) 2035 break; 2036 2037 gpu_ins->mgpu_fan_enabled = 1; 2038 } 2039 } 2040 2041 out: 2042 mutex_unlock(&mgpu_info.mutex); 2043 2044 return ret; 2045 } 2046 2047 /** 2048 * amdgpu_device_ip_late_init - run late init for hardware IPs 2049 * 2050 * @adev: amdgpu_device pointer 2051 * 2052 * Late initialization pass for hardware IPs. The list of all the hardware 2053 * IPs that make up the asic is walked and the late_init callbacks are run. 2054 * late_init covers any special initialization that an IP requires 2055 * after all of the have been initialized or something that needs to happen 2056 * late in the init process. 2057 * Returns 0 on success, negative error code on failure. 2058 */ 2059 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev) 2060 { 2061 struct amdgpu_gpu_instance *gpu_instance; 2062 int i = 0, r; 2063 2064 for (i = 0; i < adev->num_ip_blocks; i++) { 2065 if (!adev->ip_blocks[i].status.hw) 2066 continue; 2067 if (adev->ip_blocks[i].version->funcs->late_init) { 2068 r = adev->ip_blocks[i].version->funcs->late_init((void *)adev); 2069 if (r) { 2070 DRM_ERROR("late_init of IP block <%s> failed %d\n", 2071 adev->ip_blocks[i].version->funcs->name, r); 2072 return r; 2073 } 2074 } 2075 adev->ip_blocks[i].status.late_initialized = true; 2076 } 2077 2078 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE); 2079 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE); 2080 2081 amdgpu_device_fill_reset_magic(adev); 2082 2083 r = amdgpu_device_enable_mgpu_fan_boost(); 2084 if (r) 2085 DRM_ERROR("enable mgpu fan boost failed (%d).\n", r); 2086 2087 2088 if (adev->gmc.xgmi.num_physical_nodes > 1) { 2089 mutex_lock(&mgpu_info.mutex); 2090 2091 /* 2092 * Reset device p-state to low as this was booted with high. 2093 * 2094 * This should be performed only after all devices from the same 2095 * hive get initialized. 2096 * 2097 * However, it's unknown how many device in the hive in advance. 2098 * As this is counted one by one during devices initializations. 2099 * 2100 * So, we wait for all XGMI interlinked devices initialized. 2101 * This may bring some delays as those devices may come from 2102 * different hives. But that should be OK. 2103 */ 2104 if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) { 2105 for (i = 0; i < mgpu_info.num_gpu; i++) { 2106 gpu_instance = &(mgpu_info.gpu_ins[i]); 2107 if (gpu_instance->adev->flags & AMD_IS_APU) 2108 continue; 2109 2110 r = amdgpu_xgmi_set_pstate(gpu_instance->adev, 0); 2111 if (r) { 2112 DRM_ERROR("pstate setting failed (%d).\n", r); 2113 break; 2114 } 2115 } 2116 } 2117 2118 mutex_unlock(&mgpu_info.mutex); 2119 } 2120 2121 return 0; 2122 } 2123 2124 /** 2125 * amdgpu_device_ip_fini - run fini for hardware IPs 2126 * 2127 * @adev: amdgpu_device pointer 2128 * 2129 * Main teardown pass for hardware IPs. The list of all the hardware 2130 * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks 2131 * are run. hw_fini tears down the hardware associated with each IP 2132 * and sw_fini tears down any software state associated with each IP. 2133 * Returns 0 on success, negative error code on failure. 2134 */ 2135 static int amdgpu_device_ip_fini(struct amdgpu_device *adev) 2136 { 2137 int i, r; 2138 2139 amdgpu_ras_pre_fini(adev); 2140 2141 if (adev->gmc.xgmi.num_physical_nodes > 1) 2142 amdgpu_xgmi_remove_device(adev); 2143 2144 amdgpu_amdkfd_device_fini(adev); 2145 2146 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE); 2147 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE); 2148 2149 /* need to disable SMC first */ 2150 for (i = 0; i < adev->num_ip_blocks; i++) { 2151 if (!adev->ip_blocks[i].status.hw) 2152 continue; 2153 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) { 2154 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev); 2155 /* XXX handle errors */ 2156 if (r) { 2157 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n", 2158 adev->ip_blocks[i].version->funcs->name, r); 2159 } 2160 adev->ip_blocks[i].status.hw = false; 2161 break; 2162 } 2163 } 2164 2165 for (i = adev->num_ip_blocks - 1; i >= 0; i--) { 2166 if (!adev->ip_blocks[i].status.hw) 2167 continue; 2168 2169 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev); 2170 /* XXX handle errors */ 2171 if (r) { 2172 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n", 2173 adev->ip_blocks[i].version->funcs->name, r); 2174 } 2175 2176 adev->ip_blocks[i].status.hw = false; 2177 } 2178 2179 2180 for (i = adev->num_ip_blocks - 1; i >= 0; i--) { 2181 if (!adev->ip_blocks[i].status.sw) 2182 continue; 2183 2184 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) { 2185 amdgpu_ucode_free_bo(adev); 2186 amdgpu_free_static_csa(&adev->virt.csa_obj); 2187 amdgpu_device_wb_fini(adev); 2188 amdgpu_device_vram_scratch_fini(adev); 2189 amdgpu_ib_pool_fini(adev); 2190 } 2191 2192 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev); 2193 /* XXX handle errors */ 2194 if (r) { 2195 DRM_DEBUG("sw_fini of IP block <%s> failed %d\n", 2196 adev->ip_blocks[i].version->funcs->name, r); 2197 } 2198 adev->ip_blocks[i].status.sw = false; 2199 adev->ip_blocks[i].status.valid = false; 2200 } 2201 2202 for (i = adev->num_ip_blocks - 1; i >= 0; i--) { 2203 if (!adev->ip_blocks[i].status.late_initialized) 2204 continue; 2205 if (adev->ip_blocks[i].version->funcs->late_fini) 2206 adev->ip_blocks[i].version->funcs->late_fini((void *)adev); 2207 adev->ip_blocks[i].status.late_initialized = false; 2208 } 2209 2210 amdgpu_ras_fini(adev); 2211 2212 if (amdgpu_sriov_vf(adev)) 2213 if (amdgpu_virt_release_full_gpu(adev, false)) 2214 DRM_ERROR("failed to release exclusive mode on fini\n"); 2215 2216 return 0; 2217 } 2218 2219 /** 2220 * amdgpu_device_delayed_init_work_handler - work handler for IB tests 2221 * 2222 * @work: work_struct. 2223 */ 2224 static void amdgpu_device_delayed_init_work_handler(struct work_struct *work) 2225 { 2226 struct amdgpu_device *adev = 2227 container_of(work, struct amdgpu_device, delayed_init_work.work); 2228 int r; 2229 2230 r = amdgpu_ib_ring_tests(adev); 2231 if (r) 2232 DRM_ERROR("ib ring test failed (%d).\n", r); 2233 } 2234 2235 static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work) 2236 { 2237 struct amdgpu_device *adev = 2238 container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work); 2239 2240 mutex_lock(&adev->gfx.gfx_off_mutex); 2241 if (!adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) { 2242 if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true)) 2243 adev->gfx.gfx_off_state = true; 2244 } 2245 mutex_unlock(&adev->gfx.gfx_off_mutex); 2246 } 2247 2248 /** 2249 * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1) 2250 * 2251 * @adev: amdgpu_device pointer 2252 * 2253 * Main suspend function for hardware IPs. The list of all the hardware 2254 * IPs that make up the asic is walked, clockgating is disabled and the 2255 * suspend callbacks are run. suspend puts the hardware and software state 2256 * in each IP into a state suitable for suspend. 2257 * Returns 0 on success, negative error code on failure. 2258 */ 2259 static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev) 2260 { 2261 int i, r; 2262 2263 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE); 2264 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE); 2265 2266 for (i = adev->num_ip_blocks - 1; i >= 0; i--) { 2267 if (!adev->ip_blocks[i].status.valid) 2268 continue; 2269 /* displays are handled separately */ 2270 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) { 2271 /* XXX handle errors */ 2272 r = adev->ip_blocks[i].version->funcs->suspend(adev); 2273 /* XXX handle errors */ 2274 if (r) { 2275 DRM_ERROR("suspend of IP block <%s> failed %d\n", 2276 adev->ip_blocks[i].version->funcs->name, r); 2277 return r; 2278 } 2279 adev->ip_blocks[i].status.hw = false; 2280 } 2281 } 2282 2283 return 0; 2284 } 2285 2286 /** 2287 * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2) 2288 * 2289 * @adev: amdgpu_device pointer 2290 * 2291 * Main suspend function for hardware IPs. The list of all the hardware 2292 * IPs that make up the asic is walked, clockgating is disabled and the 2293 * suspend callbacks are run. suspend puts the hardware and software state 2294 * in each IP into a state suitable for suspend. 2295 * Returns 0 on success, negative error code on failure. 2296 */ 2297 static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev) 2298 { 2299 int i, r; 2300 2301 for (i = adev->num_ip_blocks - 1; i >= 0; i--) { 2302 if (!adev->ip_blocks[i].status.valid) 2303 continue; 2304 /* displays are handled in phase1 */ 2305 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) 2306 continue; 2307 /* PSP lost connection when err_event_athub occurs */ 2308 if (amdgpu_ras_intr_triggered() && 2309 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) { 2310 adev->ip_blocks[i].status.hw = false; 2311 continue; 2312 } 2313 /* XXX handle errors */ 2314 r = adev->ip_blocks[i].version->funcs->suspend(adev); 2315 /* XXX handle errors */ 2316 if (r) { 2317 DRM_ERROR("suspend of IP block <%s> failed %d\n", 2318 adev->ip_blocks[i].version->funcs->name, r); 2319 } 2320 adev->ip_blocks[i].status.hw = false; 2321 /* handle putting the SMC in the appropriate state */ 2322 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) { 2323 if (is_support_sw_smu(adev)) { 2324 r = smu_set_mp1_state(&adev->smu, adev->mp1_state); 2325 } else if (adev->powerplay.pp_funcs && 2326 adev->powerplay.pp_funcs->set_mp1_state) { 2327 r = adev->powerplay.pp_funcs->set_mp1_state( 2328 adev->powerplay.pp_handle, 2329 adev->mp1_state); 2330 } 2331 if (r) { 2332 DRM_ERROR("SMC failed to set mp1 state %d, %d\n", 2333 adev->mp1_state, r); 2334 return r; 2335 } 2336 } 2337 2338 adev->ip_blocks[i].status.hw = false; 2339 } 2340 2341 return 0; 2342 } 2343 2344 /** 2345 * amdgpu_device_ip_suspend - run suspend for hardware IPs 2346 * 2347 * @adev: amdgpu_device pointer 2348 * 2349 * Main suspend function for hardware IPs. The list of all the hardware 2350 * IPs that make up the asic is walked, clockgating is disabled and the 2351 * suspend callbacks are run. suspend puts the hardware and software state 2352 * in each IP into a state suitable for suspend. 2353 * Returns 0 on success, negative error code on failure. 2354 */ 2355 int amdgpu_device_ip_suspend(struct amdgpu_device *adev) 2356 { 2357 int r; 2358 2359 if (amdgpu_sriov_vf(adev)) 2360 amdgpu_virt_request_full_gpu(adev, false); 2361 2362 r = amdgpu_device_ip_suspend_phase1(adev); 2363 if (r) 2364 return r; 2365 r = amdgpu_device_ip_suspend_phase2(adev); 2366 2367 if (amdgpu_sriov_vf(adev)) 2368 amdgpu_virt_release_full_gpu(adev, false); 2369 2370 return r; 2371 } 2372 2373 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev) 2374 { 2375 int i, r; 2376 2377 static enum amd_ip_block_type ip_order[] = { 2378 AMD_IP_BLOCK_TYPE_GMC, 2379 AMD_IP_BLOCK_TYPE_COMMON, 2380 AMD_IP_BLOCK_TYPE_PSP, 2381 AMD_IP_BLOCK_TYPE_IH, 2382 }; 2383 2384 for (i = 0; i < ARRAY_SIZE(ip_order); i++) { 2385 int j; 2386 struct amdgpu_ip_block *block; 2387 2388 for (j = 0; j < adev->num_ip_blocks; j++) { 2389 block = &adev->ip_blocks[j]; 2390 2391 block->status.hw = false; 2392 if (block->version->type != ip_order[i] || 2393 !block->status.valid) 2394 continue; 2395 2396 r = block->version->funcs->hw_init(adev); 2397 DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded"); 2398 if (r) 2399 return r; 2400 block->status.hw = true; 2401 } 2402 } 2403 2404 return 0; 2405 } 2406 2407 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev) 2408 { 2409 int i, r; 2410 2411 static enum amd_ip_block_type ip_order[] = { 2412 AMD_IP_BLOCK_TYPE_SMC, 2413 AMD_IP_BLOCK_TYPE_DCE, 2414 AMD_IP_BLOCK_TYPE_GFX, 2415 AMD_IP_BLOCK_TYPE_SDMA, 2416 AMD_IP_BLOCK_TYPE_UVD, 2417 AMD_IP_BLOCK_TYPE_VCE 2418 }; 2419 2420 for (i = 0; i < ARRAY_SIZE(ip_order); i++) { 2421 int j; 2422 struct amdgpu_ip_block *block; 2423 2424 for (j = 0; j < adev->num_ip_blocks; j++) { 2425 block = &adev->ip_blocks[j]; 2426 2427 if (block->version->type != ip_order[i] || 2428 !block->status.valid || 2429 block->status.hw) 2430 continue; 2431 2432 r = block->version->funcs->hw_init(adev); 2433 DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded"); 2434 if (r) 2435 return r; 2436 block->status.hw = true; 2437 } 2438 } 2439 2440 return 0; 2441 } 2442 2443 /** 2444 * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs 2445 * 2446 * @adev: amdgpu_device pointer 2447 * 2448 * First resume function for hardware IPs. The list of all the hardware 2449 * IPs that make up the asic is walked and the resume callbacks are run for 2450 * COMMON, GMC, and IH. resume puts the hardware into a functional state 2451 * after a suspend and updates the software state as necessary. This 2452 * function is also used for restoring the GPU after a GPU reset. 2453 * Returns 0 on success, negative error code on failure. 2454 */ 2455 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev) 2456 { 2457 int i, r; 2458 2459 for (i = 0; i < adev->num_ip_blocks; i++) { 2460 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw) 2461 continue; 2462 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || 2463 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || 2464 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) { 2465 2466 r = adev->ip_blocks[i].version->funcs->resume(adev); 2467 if (r) { 2468 DRM_ERROR("resume of IP block <%s> failed %d\n", 2469 adev->ip_blocks[i].version->funcs->name, r); 2470 return r; 2471 } 2472 adev->ip_blocks[i].status.hw = true; 2473 } 2474 } 2475 2476 return 0; 2477 } 2478 2479 /** 2480 * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs 2481 * 2482 * @adev: amdgpu_device pointer 2483 * 2484 * First resume function for hardware IPs. The list of all the hardware 2485 * IPs that make up the asic is walked and the resume callbacks are run for 2486 * all blocks except COMMON, GMC, and IH. resume puts the hardware into a 2487 * functional state after a suspend and updates the software state as 2488 * necessary. This function is also used for restoring the GPU after a GPU 2489 * reset. 2490 * Returns 0 on success, negative error code on failure. 2491 */ 2492 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev) 2493 { 2494 int i, r; 2495 2496 for (i = 0; i < adev->num_ip_blocks; i++) { 2497 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw) 2498 continue; 2499 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON || 2500 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC || 2501 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH || 2502 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) 2503 continue; 2504 r = adev->ip_blocks[i].version->funcs->resume(adev); 2505 if (r) { 2506 DRM_ERROR("resume of IP block <%s> failed %d\n", 2507 adev->ip_blocks[i].version->funcs->name, r); 2508 return r; 2509 } 2510 adev->ip_blocks[i].status.hw = true; 2511 } 2512 2513 return 0; 2514 } 2515 2516 /** 2517 * amdgpu_device_ip_resume - run resume for hardware IPs 2518 * 2519 * @adev: amdgpu_device pointer 2520 * 2521 * Main resume function for hardware IPs. The hardware IPs 2522 * are split into two resume functions because they are 2523 * are also used in in recovering from a GPU reset and some additional 2524 * steps need to be take between them. In this case (S3/S4) they are 2525 * run sequentially. 2526 * Returns 0 on success, negative error code on failure. 2527 */ 2528 static int amdgpu_device_ip_resume(struct amdgpu_device *adev) 2529 { 2530 int r; 2531 2532 r = amdgpu_device_ip_resume_phase1(adev); 2533 if (r) 2534 return r; 2535 2536 r = amdgpu_device_fw_loading(adev); 2537 if (r) 2538 return r; 2539 2540 r = amdgpu_device_ip_resume_phase2(adev); 2541 2542 return r; 2543 } 2544 2545 /** 2546 * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV 2547 * 2548 * @adev: amdgpu_device pointer 2549 * 2550 * Query the VBIOS data tables to determine if the board supports SR-IOV. 2551 */ 2552 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev) 2553 { 2554 if (amdgpu_sriov_vf(adev)) { 2555 if (adev->is_atom_fw) { 2556 if (amdgpu_atomfirmware_gpu_supports_virtualization(adev)) 2557 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS; 2558 } else { 2559 if (amdgpu_atombios_has_gpu_virtualization_table(adev)) 2560 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS; 2561 } 2562 2563 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS)) 2564 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0); 2565 } 2566 } 2567 2568 /** 2569 * amdgpu_device_asic_has_dc_support - determine if DC supports the asic 2570 * 2571 * @asic_type: AMD asic type 2572 * 2573 * Check if there is DC (new modesetting infrastructre) support for an asic. 2574 * returns true if DC has support, false if not. 2575 */ 2576 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type) 2577 { 2578 switch (asic_type) { 2579 #if defined(CONFIG_DRM_AMD_DC) 2580 case CHIP_BONAIRE: 2581 case CHIP_KAVERI: 2582 case CHIP_KABINI: 2583 case CHIP_MULLINS: 2584 /* 2585 * We have systems in the wild with these ASICs that require 2586 * LVDS and VGA support which is not supported with DC. 2587 * 2588 * Fallback to the non-DC driver here by default so as not to 2589 * cause regressions. 2590 */ 2591 return amdgpu_dc > 0; 2592 case CHIP_HAWAII: 2593 case CHIP_CARRIZO: 2594 case CHIP_STONEY: 2595 case CHIP_POLARIS10: 2596 case CHIP_POLARIS11: 2597 case CHIP_POLARIS12: 2598 case CHIP_VEGAM: 2599 case CHIP_TONGA: 2600 case CHIP_FIJI: 2601 case CHIP_VEGA10: 2602 case CHIP_VEGA12: 2603 case CHIP_VEGA20: 2604 #if defined(CONFIG_DRM_AMD_DC_DCN) 2605 case CHIP_RAVEN: 2606 case CHIP_NAVI10: 2607 case CHIP_NAVI14: 2608 case CHIP_NAVI12: 2609 case CHIP_RENOIR: 2610 #endif 2611 return amdgpu_dc != 0; 2612 #endif 2613 default: 2614 return false; 2615 } 2616 } 2617 2618 /** 2619 * amdgpu_device_has_dc_support - check if dc is supported 2620 * 2621 * @adev: amdgpu_device_pointer 2622 * 2623 * Returns true for supported, false for not supported 2624 */ 2625 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev) 2626 { 2627 if (amdgpu_sriov_vf(adev)) 2628 return false; 2629 2630 return amdgpu_device_asic_has_dc_support(adev->asic_type); 2631 } 2632 2633 2634 static void amdgpu_device_xgmi_reset_func(struct work_struct *__work) 2635 { 2636 struct amdgpu_device *adev = 2637 container_of(__work, struct amdgpu_device, xgmi_reset_work); 2638 2639 adev->asic_reset_res = amdgpu_asic_reset(adev); 2640 if (adev->asic_reset_res) 2641 DRM_WARN("ASIC reset failed with error, %d for drm dev, %s", 2642 adev->asic_reset_res, adev->ddev->unique); 2643 } 2644 2645 static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev) 2646 { 2647 char *input = amdgpu_lockup_timeout; 2648 char *timeout_setting = NULL; 2649 int index = 0; 2650 long timeout; 2651 int ret = 0; 2652 2653 /* 2654 * By default timeout for non compute jobs is 10000. 2655 * And there is no timeout enforced on compute jobs. 2656 * In SR-IOV or passthrough mode, timeout for compute 2657 * jobs are 10000 by default. 2658 */ 2659 adev->gfx_timeout = msecs_to_jiffies(10000); 2660 adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout; 2661 if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev)) 2662 adev->compute_timeout = adev->gfx_timeout; 2663 else 2664 adev->compute_timeout = MAX_SCHEDULE_TIMEOUT; 2665 2666 if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) { 2667 while ((timeout_setting = strsep(&input, ",")) && 2668 strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) { 2669 ret = kstrtol(timeout_setting, 0, &timeout); 2670 if (ret) 2671 return ret; 2672 2673 if (timeout == 0) { 2674 index++; 2675 continue; 2676 } else if (timeout < 0) { 2677 timeout = MAX_SCHEDULE_TIMEOUT; 2678 } else { 2679 timeout = msecs_to_jiffies(timeout); 2680 } 2681 2682 switch (index++) { 2683 case 0: 2684 adev->gfx_timeout = timeout; 2685 break; 2686 case 1: 2687 adev->compute_timeout = timeout; 2688 break; 2689 case 2: 2690 adev->sdma_timeout = timeout; 2691 break; 2692 case 3: 2693 adev->video_timeout = timeout; 2694 break; 2695 default: 2696 break; 2697 } 2698 } 2699 /* 2700 * There is only one value specified and 2701 * it should apply to all non-compute jobs. 2702 */ 2703 if (index == 1) { 2704 adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout; 2705 if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev)) 2706 adev->compute_timeout = adev->gfx_timeout; 2707 } 2708 } 2709 2710 return ret; 2711 } 2712 2713 /** 2714 * amdgpu_device_init - initialize the driver 2715 * 2716 * @adev: amdgpu_device pointer 2717 * @ddev: drm dev pointer 2718 * @pdev: pci dev pointer 2719 * @flags: driver flags 2720 * 2721 * Initializes the driver info and hw (all asics). 2722 * Returns 0 for success or an error on failure. 2723 * Called at driver startup. 2724 */ 2725 int amdgpu_device_init(struct amdgpu_device *adev, 2726 struct drm_device *ddev, 2727 struct pci_dev *pdev, 2728 uint32_t flags) 2729 { 2730 int r, i; 2731 bool runtime = false; 2732 u32 max_MBps; 2733 2734 adev->shutdown = false; 2735 adev->dev = &pdev->dev; 2736 adev->ddev = ddev; 2737 adev->pdev = pdev; 2738 adev->flags = flags; 2739 2740 if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST) 2741 adev->asic_type = amdgpu_force_asic_type; 2742 else 2743 adev->asic_type = flags & AMD_ASIC_MASK; 2744 2745 adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT; 2746 if (amdgpu_emu_mode == 1) 2747 adev->usec_timeout *= 2; 2748 adev->gmc.gart_size = 512 * 1024 * 1024; 2749 adev->accel_working = false; 2750 adev->num_rings = 0; 2751 adev->mman.buffer_funcs = NULL; 2752 adev->mman.buffer_funcs_ring = NULL; 2753 adev->vm_manager.vm_pte_funcs = NULL; 2754 adev->vm_manager.vm_pte_num_rqs = 0; 2755 adev->gmc.gmc_funcs = NULL; 2756 adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS); 2757 bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES); 2758 2759 adev->smc_rreg = &amdgpu_invalid_rreg; 2760 adev->smc_wreg = &amdgpu_invalid_wreg; 2761 adev->pcie_rreg = &amdgpu_invalid_rreg; 2762 adev->pcie_wreg = &amdgpu_invalid_wreg; 2763 adev->pciep_rreg = &amdgpu_invalid_rreg; 2764 adev->pciep_wreg = &amdgpu_invalid_wreg; 2765 adev->pcie_rreg64 = &amdgpu_invalid_rreg64; 2766 adev->pcie_wreg64 = &amdgpu_invalid_wreg64; 2767 adev->uvd_ctx_rreg = &amdgpu_invalid_rreg; 2768 adev->uvd_ctx_wreg = &amdgpu_invalid_wreg; 2769 adev->didt_rreg = &amdgpu_invalid_rreg; 2770 adev->didt_wreg = &amdgpu_invalid_wreg; 2771 adev->gc_cac_rreg = &amdgpu_invalid_rreg; 2772 adev->gc_cac_wreg = &amdgpu_invalid_wreg; 2773 adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg; 2774 adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg; 2775 2776 DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n", 2777 amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device, 2778 pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision); 2779 2780 /* mutex initialization are all done here so we 2781 * can recall function without having locking issues */ 2782 atomic_set(&adev->irq.ih.lock, 0); 2783 mutex_init(&adev->firmware.mutex); 2784 mutex_init(&adev->pm.mutex); 2785 mutex_init(&adev->gfx.gpu_clock_mutex); 2786 mutex_init(&adev->srbm_mutex); 2787 mutex_init(&adev->gfx.pipe_reserve_mutex); 2788 mutex_init(&adev->gfx.gfx_off_mutex); 2789 mutex_init(&adev->grbm_idx_mutex); 2790 mutex_init(&adev->mn_lock); 2791 mutex_init(&adev->virt.vf_errors.lock); 2792 hash_init(adev->mn_hash); 2793 mutex_init(&adev->lock_reset); 2794 mutex_init(&adev->virt.dpm_mutex); 2795 mutex_init(&adev->psp.mutex); 2796 2797 r = amdgpu_device_check_arguments(adev); 2798 if (r) 2799 return r; 2800 2801 spin_lock_init(&adev->mmio_idx_lock); 2802 spin_lock_init(&adev->smc_idx_lock); 2803 spin_lock_init(&adev->pcie_idx_lock); 2804 spin_lock_init(&adev->uvd_ctx_idx_lock); 2805 spin_lock_init(&adev->didt_idx_lock); 2806 spin_lock_init(&adev->gc_cac_idx_lock); 2807 spin_lock_init(&adev->se_cac_idx_lock); 2808 spin_lock_init(&adev->audio_endpt_idx_lock); 2809 spin_lock_init(&adev->mm_stats.lock); 2810 2811 INIT_LIST_HEAD(&adev->shadow_list); 2812 mutex_init(&adev->shadow_list_lock); 2813 2814 INIT_LIST_HEAD(&adev->ring_lru_list); 2815 spin_lock_init(&adev->ring_lru_list_lock); 2816 2817 INIT_DELAYED_WORK(&adev->delayed_init_work, 2818 amdgpu_device_delayed_init_work_handler); 2819 INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work, 2820 amdgpu_device_delay_enable_gfx_off); 2821 2822 INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func); 2823 2824 adev->gfx.gfx_off_req_count = 1; 2825 adev->pm.ac_power = power_supply_is_system_supplied() > 0 ? true : false; 2826 2827 /* Registers mapping */ 2828 /* TODO: block userspace mapping of io register */ 2829 if (adev->asic_type >= CHIP_BONAIRE) { 2830 adev->rmmio_base = pci_resource_start(adev->pdev, 5); 2831 adev->rmmio_size = pci_resource_len(adev->pdev, 5); 2832 } else { 2833 adev->rmmio_base = pci_resource_start(adev->pdev, 2); 2834 adev->rmmio_size = pci_resource_len(adev->pdev, 2); 2835 } 2836 2837 adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size); 2838 if (adev->rmmio == NULL) { 2839 return -ENOMEM; 2840 } 2841 DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base); 2842 DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size); 2843 2844 /* io port mapping */ 2845 for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) { 2846 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) { 2847 adev->rio_mem_size = pci_resource_len(adev->pdev, i); 2848 adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size); 2849 break; 2850 } 2851 } 2852 if (adev->rio_mem == NULL) 2853 DRM_INFO("PCI I/O BAR is not found.\n"); 2854 2855 /* enable PCIE atomic ops */ 2856 r = pci_enable_atomic_ops_to_root(adev->pdev, 2857 PCI_EXP_DEVCAP2_ATOMIC_COMP32 | 2858 PCI_EXP_DEVCAP2_ATOMIC_COMP64); 2859 if (r) { 2860 adev->have_atomics_support = false; 2861 DRM_INFO("PCIE atomic ops is not supported\n"); 2862 } else { 2863 adev->have_atomics_support = true; 2864 } 2865 2866 amdgpu_device_get_pcie_info(adev); 2867 2868 if (amdgpu_mcbp) 2869 DRM_INFO("MCBP is enabled\n"); 2870 2871 if (amdgpu_mes && adev->asic_type >= CHIP_NAVI10) 2872 adev->enable_mes = true; 2873 2874 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10) { 2875 r = amdgpu_discovery_init(adev); 2876 if (r) { 2877 dev_err(adev->dev, "amdgpu_discovery_init failed\n"); 2878 return r; 2879 } 2880 } 2881 2882 /* early init functions */ 2883 r = amdgpu_device_ip_early_init(adev); 2884 if (r) 2885 return r; 2886 2887 r = amdgpu_device_get_job_timeout_settings(adev); 2888 if (r) { 2889 dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n"); 2890 return r; 2891 } 2892 2893 /* doorbell bar mapping and doorbell index init*/ 2894 amdgpu_device_doorbell_init(adev); 2895 2896 /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */ 2897 /* this will fail for cards that aren't VGA class devices, just 2898 * ignore it */ 2899 vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode); 2900 2901 if (amdgpu_device_is_px(ddev)) 2902 runtime = true; 2903 if (!pci_is_thunderbolt_attached(adev->pdev)) 2904 vga_switcheroo_register_client(adev->pdev, 2905 &amdgpu_switcheroo_ops, runtime); 2906 if (runtime) 2907 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain); 2908 2909 if (amdgpu_emu_mode == 1) { 2910 /* post the asic on emulation mode */ 2911 emu_soc_asic_init(adev); 2912 goto fence_driver_init; 2913 } 2914 2915 /* detect if we are with an SRIOV vbios */ 2916 amdgpu_device_detect_sriov_bios(adev); 2917 2918 /* check if we need to reset the asic 2919 * E.g., driver was not cleanly unloaded previously, etc. 2920 */ 2921 if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) { 2922 r = amdgpu_asic_reset(adev); 2923 if (r) { 2924 dev_err(adev->dev, "asic reset on init failed\n"); 2925 goto failed; 2926 } 2927 } 2928 2929 /* Post card if necessary */ 2930 if (amdgpu_device_need_post(adev)) { 2931 if (!adev->bios) { 2932 dev_err(adev->dev, "no vBIOS found\n"); 2933 r = -EINVAL; 2934 goto failed; 2935 } 2936 DRM_INFO("GPU posting now...\n"); 2937 r = amdgpu_atom_asic_init(adev->mode_info.atom_context); 2938 if (r) { 2939 dev_err(adev->dev, "gpu post error!\n"); 2940 goto failed; 2941 } 2942 } 2943 2944 if (adev->is_atom_fw) { 2945 /* Initialize clocks */ 2946 r = amdgpu_atomfirmware_get_clock_info(adev); 2947 if (r) { 2948 dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n"); 2949 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0); 2950 goto failed; 2951 } 2952 } else { 2953 /* Initialize clocks */ 2954 r = amdgpu_atombios_get_clock_info(adev); 2955 if (r) { 2956 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n"); 2957 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0); 2958 goto failed; 2959 } 2960 /* init i2c buses */ 2961 if (!amdgpu_device_has_dc_support(adev)) 2962 amdgpu_atombios_i2c_init(adev); 2963 } 2964 2965 fence_driver_init: 2966 /* Fence driver */ 2967 r = amdgpu_fence_driver_init(adev); 2968 if (r) { 2969 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n"); 2970 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0); 2971 goto failed; 2972 } 2973 2974 /* init the mode config */ 2975 drm_mode_config_init(adev->ddev); 2976 2977 r = amdgpu_device_ip_init(adev); 2978 if (r) { 2979 /* failed in exclusive mode due to timeout */ 2980 if (amdgpu_sriov_vf(adev) && 2981 !amdgpu_sriov_runtime(adev) && 2982 amdgpu_virt_mmio_blocked(adev) && 2983 !amdgpu_virt_wait_reset(adev)) { 2984 dev_err(adev->dev, "VF exclusive mode timeout\n"); 2985 /* Don't send request since VF is inactive. */ 2986 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME; 2987 adev->virt.ops = NULL; 2988 r = -EAGAIN; 2989 goto failed; 2990 } 2991 dev_err(adev->dev, "amdgpu_device_ip_init failed\n"); 2992 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0); 2993 if (amdgpu_virt_request_full_gpu(adev, false)) 2994 amdgpu_virt_release_full_gpu(adev, false); 2995 goto failed; 2996 } 2997 2998 adev->accel_working = true; 2999 3000 amdgpu_vm_check_compute_bug(adev); 3001 3002 /* Initialize the buffer migration limit. */ 3003 if (amdgpu_moverate >= 0) 3004 max_MBps = amdgpu_moverate; 3005 else 3006 max_MBps = 8; /* Allow 8 MB/s. */ 3007 /* Get a log2 for easy divisions. */ 3008 adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps)); 3009 3010 amdgpu_fbdev_init(adev); 3011 3012 if (amdgpu_sriov_vf(adev) && amdgim_is_hwperf(adev)) 3013 amdgpu_pm_virt_sysfs_init(adev); 3014 3015 r = amdgpu_pm_sysfs_init(adev); 3016 if (r) 3017 DRM_ERROR("registering pm debugfs failed (%d).\n", r); 3018 3019 r = amdgpu_ucode_sysfs_init(adev); 3020 if (r) 3021 DRM_ERROR("Creating firmware sysfs failed (%d).\n", r); 3022 3023 r = amdgpu_debugfs_gem_init(adev); 3024 if (r) 3025 DRM_ERROR("registering gem debugfs failed (%d).\n", r); 3026 3027 r = amdgpu_debugfs_regs_init(adev); 3028 if (r) 3029 DRM_ERROR("registering register debugfs failed (%d).\n", r); 3030 3031 r = amdgpu_debugfs_firmware_init(adev); 3032 if (r) 3033 DRM_ERROR("registering firmware debugfs failed (%d).\n", r); 3034 3035 r = amdgpu_debugfs_init(adev); 3036 if (r) 3037 DRM_ERROR("Creating debugfs files failed (%d).\n", r); 3038 3039 if ((amdgpu_testing & 1)) { 3040 if (adev->accel_working) 3041 amdgpu_test_moves(adev); 3042 else 3043 DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n"); 3044 } 3045 if (amdgpu_benchmarking) { 3046 if (adev->accel_working) 3047 amdgpu_benchmark(adev, amdgpu_benchmarking); 3048 else 3049 DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n"); 3050 } 3051 3052 /* 3053 * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost. 3054 * Otherwise the mgpu fan boost feature will be skipped due to the 3055 * gpu instance is counted less. 3056 */ 3057 amdgpu_register_gpu_instance(adev); 3058 3059 /* enable clockgating, etc. after ib tests, etc. since some blocks require 3060 * explicit gating rather than handling it automatically. 3061 */ 3062 r = amdgpu_device_ip_late_init(adev); 3063 if (r) { 3064 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n"); 3065 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r); 3066 goto failed; 3067 } 3068 3069 /* must succeed. */ 3070 amdgpu_ras_resume(adev); 3071 3072 queue_delayed_work(system_wq, &adev->delayed_init_work, 3073 msecs_to_jiffies(AMDGPU_RESUME_MS)); 3074 3075 r = device_create_file(adev->dev, &dev_attr_pcie_replay_count); 3076 if (r) { 3077 dev_err(adev->dev, "Could not create pcie_replay_count"); 3078 return r; 3079 } 3080 3081 if (IS_ENABLED(CONFIG_PERF_EVENTS)) 3082 r = amdgpu_pmu_init(adev); 3083 if (r) 3084 dev_err(adev->dev, "amdgpu_pmu_init failed\n"); 3085 3086 return 0; 3087 3088 failed: 3089 amdgpu_vf_error_trans_all(adev); 3090 if (runtime) 3091 vga_switcheroo_fini_domain_pm_ops(adev->dev); 3092 3093 return r; 3094 } 3095 3096 /** 3097 * amdgpu_device_fini - tear down the driver 3098 * 3099 * @adev: amdgpu_device pointer 3100 * 3101 * Tear down the driver info (all asics). 3102 * Called at driver shutdown. 3103 */ 3104 void amdgpu_device_fini(struct amdgpu_device *adev) 3105 { 3106 int r; 3107 3108 DRM_INFO("amdgpu: finishing device.\n"); 3109 flush_delayed_work(&adev->delayed_init_work); 3110 adev->shutdown = true; 3111 3112 /* disable all interrupts */ 3113 amdgpu_irq_disable_all(adev); 3114 if (adev->mode_info.mode_config_initialized){ 3115 if (!amdgpu_device_has_dc_support(adev)) 3116 drm_helper_force_disable_all(adev->ddev); 3117 else 3118 drm_atomic_helper_shutdown(adev->ddev); 3119 } 3120 amdgpu_fence_driver_fini(adev); 3121 amdgpu_pm_sysfs_fini(adev); 3122 amdgpu_fbdev_fini(adev); 3123 r = amdgpu_device_ip_fini(adev); 3124 if (adev->firmware.gpu_info_fw) { 3125 release_firmware(adev->firmware.gpu_info_fw); 3126 adev->firmware.gpu_info_fw = NULL; 3127 } 3128 adev->accel_working = false; 3129 /* free i2c buses */ 3130 if (!amdgpu_device_has_dc_support(adev)) 3131 amdgpu_i2c_fini(adev); 3132 3133 if (amdgpu_emu_mode != 1) 3134 amdgpu_atombios_fini(adev); 3135 3136 kfree(adev->bios); 3137 adev->bios = NULL; 3138 if (!pci_is_thunderbolt_attached(adev->pdev)) 3139 vga_switcheroo_unregister_client(adev->pdev); 3140 if (adev->flags & AMD_IS_PX) 3141 vga_switcheroo_fini_domain_pm_ops(adev->dev); 3142 vga_client_register(adev->pdev, NULL, NULL, NULL); 3143 if (adev->rio_mem) 3144 pci_iounmap(adev->pdev, adev->rio_mem); 3145 adev->rio_mem = NULL; 3146 iounmap(adev->rmmio); 3147 adev->rmmio = NULL; 3148 amdgpu_device_doorbell_fini(adev); 3149 if (amdgpu_sriov_vf(adev) && amdgim_is_hwperf(adev)) 3150 amdgpu_pm_virt_sysfs_fini(adev); 3151 3152 amdgpu_debugfs_regs_cleanup(adev); 3153 device_remove_file(adev->dev, &dev_attr_pcie_replay_count); 3154 amdgpu_ucode_sysfs_fini(adev); 3155 if (IS_ENABLED(CONFIG_PERF_EVENTS)) 3156 amdgpu_pmu_fini(adev); 3157 amdgpu_debugfs_preempt_cleanup(adev); 3158 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10) 3159 amdgpu_discovery_fini(adev); 3160 } 3161 3162 3163 /* 3164 * Suspend & resume. 3165 */ 3166 /** 3167 * amdgpu_device_suspend - initiate device suspend 3168 * 3169 * @dev: drm dev pointer 3170 * @suspend: suspend state 3171 * @fbcon : notify the fbdev of suspend 3172 * 3173 * Puts the hw in the suspend state (all asics). 3174 * Returns 0 for success or an error on failure. 3175 * Called at driver suspend. 3176 */ 3177 int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon) 3178 { 3179 struct amdgpu_device *adev; 3180 struct drm_crtc *crtc; 3181 struct drm_connector *connector; 3182 struct drm_connector_list_iter iter; 3183 int r; 3184 3185 if (dev == NULL || dev->dev_private == NULL) { 3186 return -ENODEV; 3187 } 3188 3189 adev = dev->dev_private; 3190 3191 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF) 3192 return 0; 3193 3194 adev->in_suspend = true; 3195 drm_kms_helper_poll_disable(dev); 3196 3197 if (fbcon) 3198 amdgpu_fbdev_set_suspend(adev, 1); 3199 3200 cancel_delayed_work_sync(&adev->delayed_init_work); 3201 3202 if (!amdgpu_device_has_dc_support(adev)) { 3203 /* turn off display hw */ 3204 drm_modeset_lock_all(dev); 3205 drm_connector_list_iter_begin(dev, &iter); 3206 drm_for_each_connector_iter(connector, &iter) 3207 drm_helper_connector_dpms(connector, 3208 DRM_MODE_DPMS_OFF); 3209 drm_connector_list_iter_end(&iter); 3210 drm_modeset_unlock_all(dev); 3211 /* unpin the front buffers and cursors */ 3212 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) { 3213 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc); 3214 struct drm_framebuffer *fb = crtc->primary->fb; 3215 struct amdgpu_bo *robj; 3216 3217 if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) { 3218 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo); 3219 r = amdgpu_bo_reserve(aobj, true); 3220 if (r == 0) { 3221 amdgpu_bo_unpin(aobj); 3222 amdgpu_bo_unreserve(aobj); 3223 } 3224 } 3225 3226 if (fb == NULL || fb->obj[0] == NULL) { 3227 continue; 3228 } 3229 robj = gem_to_amdgpu_bo(fb->obj[0]); 3230 /* don't unpin kernel fb objects */ 3231 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) { 3232 r = amdgpu_bo_reserve(robj, true); 3233 if (r == 0) { 3234 amdgpu_bo_unpin(robj); 3235 amdgpu_bo_unreserve(robj); 3236 } 3237 } 3238 } 3239 } 3240 3241 amdgpu_amdkfd_suspend(adev); 3242 3243 amdgpu_ras_suspend(adev); 3244 3245 r = amdgpu_device_ip_suspend_phase1(adev); 3246 3247 /* evict vram memory */ 3248 amdgpu_bo_evict_vram(adev); 3249 3250 amdgpu_fence_driver_suspend(adev); 3251 3252 r = amdgpu_device_ip_suspend_phase2(adev); 3253 3254 /* evict remaining vram memory 3255 * This second call to evict vram is to evict the gart page table 3256 * using the CPU. 3257 */ 3258 amdgpu_bo_evict_vram(adev); 3259 3260 if (suspend) { 3261 pci_save_state(dev->pdev); 3262 /* Shut down the device */ 3263 pci_disable_device(dev->pdev); 3264 pci_set_power_state(dev->pdev, PCI_D3hot); 3265 } 3266 3267 return 0; 3268 } 3269 3270 /** 3271 * amdgpu_device_resume - initiate device resume 3272 * 3273 * @dev: drm dev pointer 3274 * @resume: resume state 3275 * @fbcon : notify the fbdev of resume 3276 * 3277 * Bring the hw back to operating state (all asics). 3278 * Returns 0 for success or an error on failure. 3279 * Called at driver resume. 3280 */ 3281 int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon) 3282 { 3283 struct drm_connector *connector; 3284 struct drm_connector_list_iter iter; 3285 struct amdgpu_device *adev = dev->dev_private; 3286 struct drm_crtc *crtc; 3287 int r = 0; 3288 3289 if (dev->switch_power_state == DRM_SWITCH_POWER_OFF) 3290 return 0; 3291 3292 if (resume) { 3293 pci_set_power_state(dev->pdev, PCI_D0); 3294 pci_restore_state(dev->pdev); 3295 r = pci_enable_device(dev->pdev); 3296 if (r) 3297 return r; 3298 } 3299 3300 /* post card */ 3301 if (amdgpu_device_need_post(adev)) { 3302 r = amdgpu_atom_asic_init(adev->mode_info.atom_context); 3303 if (r) 3304 DRM_ERROR("amdgpu asic init failed\n"); 3305 } 3306 3307 r = amdgpu_device_ip_resume(adev); 3308 if (r) { 3309 DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r); 3310 return r; 3311 } 3312 amdgpu_fence_driver_resume(adev); 3313 3314 3315 r = amdgpu_device_ip_late_init(adev); 3316 if (r) 3317 return r; 3318 3319 queue_delayed_work(system_wq, &adev->delayed_init_work, 3320 msecs_to_jiffies(AMDGPU_RESUME_MS)); 3321 3322 if (!amdgpu_device_has_dc_support(adev)) { 3323 /* pin cursors */ 3324 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) { 3325 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc); 3326 3327 if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) { 3328 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo); 3329 r = amdgpu_bo_reserve(aobj, true); 3330 if (r == 0) { 3331 r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM); 3332 if (r != 0) 3333 DRM_ERROR("Failed to pin cursor BO (%d)\n", r); 3334 amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj); 3335 amdgpu_bo_unreserve(aobj); 3336 } 3337 } 3338 } 3339 } 3340 r = amdgpu_amdkfd_resume(adev); 3341 if (r) 3342 return r; 3343 3344 /* Make sure IB tests flushed */ 3345 flush_delayed_work(&adev->delayed_init_work); 3346 3347 /* blat the mode back in */ 3348 if (fbcon) { 3349 if (!amdgpu_device_has_dc_support(adev)) { 3350 /* pre DCE11 */ 3351 drm_helper_resume_force_mode(dev); 3352 3353 /* turn on display hw */ 3354 drm_modeset_lock_all(dev); 3355 3356 drm_connector_list_iter_begin(dev, &iter); 3357 drm_for_each_connector_iter(connector, &iter) 3358 drm_helper_connector_dpms(connector, 3359 DRM_MODE_DPMS_ON); 3360 drm_connector_list_iter_end(&iter); 3361 3362 drm_modeset_unlock_all(dev); 3363 } 3364 amdgpu_fbdev_set_suspend(adev, 0); 3365 } 3366 3367 drm_kms_helper_poll_enable(dev); 3368 3369 amdgpu_ras_resume(adev); 3370 3371 /* 3372 * Most of the connector probing functions try to acquire runtime pm 3373 * refs to ensure that the GPU is powered on when connector polling is 3374 * performed. Since we're calling this from a runtime PM callback, 3375 * trying to acquire rpm refs will cause us to deadlock. 3376 * 3377 * Since we're guaranteed to be holding the rpm lock, it's safe to 3378 * temporarily disable the rpm helpers so this doesn't deadlock us. 3379 */ 3380 #ifdef CONFIG_PM 3381 dev->dev->power.disable_depth++; 3382 #endif 3383 if (!amdgpu_device_has_dc_support(adev)) 3384 drm_helper_hpd_irq_event(dev); 3385 else 3386 drm_kms_helper_hotplug_event(dev); 3387 #ifdef CONFIG_PM 3388 dev->dev->power.disable_depth--; 3389 #endif 3390 adev->in_suspend = false; 3391 3392 return 0; 3393 } 3394 3395 /** 3396 * amdgpu_device_ip_check_soft_reset - did soft reset succeed 3397 * 3398 * @adev: amdgpu_device pointer 3399 * 3400 * The list of all the hardware IPs that make up the asic is walked and 3401 * the check_soft_reset callbacks are run. check_soft_reset determines 3402 * if the asic is still hung or not. 3403 * Returns true if any of the IPs are still in a hung state, false if not. 3404 */ 3405 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev) 3406 { 3407 int i; 3408 bool asic_hang = false; 3409 3410 if (amdgpu_sriov_vf(adev)) 3411 return true; 3412 3413 if (amdgpu_asic_need_full_reset(adev)) 3414 return true; 3415 3416 for (i = 0; i < adev->num_ip_blocks; i++) { 3417 if (!adev->ip_blocks[i].status.valid) 3418 continue; 3419 if (adev->ip_blocks[i].version->funcs->check_soft_reset) 3420 adev->ip_blocks[i].status.hang = 3421 adev->ip_blocks[i].version->funcs->check_soft_reset(adev); 3422 if (adev->ip_blocks[i].status.hang) { 3423 DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name); 3424 asic_hang = true; 3425 } 3426 } 3427 return asic_hang; 3428 } 3429 3430 /** 3431 * amdgpu_device_ip_pre_soft_reset - prepare for soft reset 3432 * 3433 * @adev: amdgpu_device pointer 3434 * 3435 * The list of all the hardware IPs that make up the asic is walked and the 3436 * pre_soft_reset callbacks are run if the block is hung. pre_soft_reset 3437 * handles any IP specific hardware or software state changes that are 3438 * necessary for a soft reset to succeed. 3439 * Returns 0 on success, negative error code on failure. 3440 */ 3441 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev) 3442 { 3443 int i, r = 0; 3444 3445 for (i = 0; i < adev->num_ip_blocks; i++) { 3446 if (!adev->ip_blocks[i].status.valid) 3447 continue; 3448 if (adev->ip_blocks[i].status.hang && 3449 adev->ip_blocks[i].version->funcs->pre_soft_reset) { 3450 r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev); 3451 if (r) 3452 return r; 3453 } 3454 } 3455 3456 return 0; 3457 } 3458 3459 /** 3460 * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed 3461 * 3462 * @adev: amdgpu_device pointer 3463 * 3464 * Some hardware IPs cannot be soft reset. If they are hung, a full gpu 3465 * reset is necessary to recover. 3466 * Returns true if a full asic reset is required, false if not. 3467 */ 3468 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev) 3469 { 3470 int i; 3471 3472 if (amdgpu_asic_need_full_reset(adev)) 3473 return true; 3474 3475 for (i = 0; i < adev->num_ip_blocks; i++) { 3476 if (!adev->ip_blocks[i].status.valid) 3477 continue; 3478 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) || 3479 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) || 3480 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) || 3481 (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) || 3482 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) { 3483 if (adev->ip_blocks[i].status.hang) { 3484 DRM_INFO("Some block need full reset!\n"); 3485 return true; 3486 } 3487 } 3488 } 3489 return false; 3490 } 3491 3492 /** 3493 * amdgpu_device_ip_soft_reset - do a soft reset 3494 * 3495 * @adev: amdgpu_device pointer 3496 * 3497 * The list of all the hardware IPs that make up the asic is walked and the 3498 * soft_reset callbacks are run if the block is hung. soft_reset handles any 3499 * IP specific hardware or software state changes that are necessary to soft 3500 * reset the IP. 3501 * Returns 0 on success, negative error code on failure. 3502 */ 3503 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev) 3504 { 3505 int i, r = 0; 3506 3507 for (i = 0; i < adev->num_ip_blocks; i++) { 3508 if (!adev->ip_blocks[i].status.valid) 3509 continue; 3510 if (adev->ip_blocks[i].status.hang && 3511 adev->ip_blocks[i].version->funcs->soft_reset) { 3512 r = adev->ip_blocks[i].version->funcs->soft_reset(adev); 3513 if (r) 3514 return r; 3515 } 3516 } 3517 3518 return 0; 3519 } 3520 3521 /** 3522 * amdgpu_device_ip_post_soft_reset - clean up from soft reset 3523 * 3524 * @adev: amdgpu_device pointer 3525 * 3526 * The list of all the hardware IPs that make up the asic is walked and the 3527 * post_soft_reset callbacks are run if the asic was hung. post_soft_reset 3528 * handles any IP specific hardware or software state changes that are 3529 * necessary after the IP has been soft reset. 3530 * Returns 0 on success, negative error code on failure. 3531 */ 3532 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev) 3533 { 3534 int i, r = 0; 3535 3536 for (i = 0; i < adev->num_ip_blocks; i++) { 3537 if (!adev->ip_blocks[i].status.valid) 3538 continue; 3539 if (adev->ip_blocks[i].status.hang && 3540 adev->ip_blocks[i].version->funcs->post_soft_reset) 3541 r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev); 3542 if (r) 3543 return r; 3544 } 3545 3546 return 0; 3547 } 3548 3549 /** 3550 * amdgpu_device_recover_vram - Recover some VRAM contents 3551 * 3552 * @adev: amdgpu_device pointer 3553 * 3554 * Restores the contents of VRAM buffers from the shadows in GTT. Used to 3555 * restore things like GPUVM page tables after a GPU reset where 3556 * the contents of VRAM might be lost. 3557 * 3558 * Returns: 3559 * 0 on success, negative error code on failure. 3560 */ 3561 static int amdgpu_device_recover_vram(struct amdgpu_device *adev) 3562 { 3563 struct dma_fence *fence = NULL, *next = NULL; 3564 struct amdgpu_bo *shadow; 3565 long r = 1, tmo; 3566 3567 if (amdgpu_sriov_runtime(adev)) 3568 tmo = msecs_to_jiffies(8000); 3569 else 3570 tmo = msecs_to_jiffies(100); 3571 3572 DRM_INFO("recover vram bo from shadow start\n"); 3573 mutex_lock(&adev->shadow_list_lock); 3574 list_for_each_entry(shadow, &adev->shadow_list, shadow_list) { 3575 3576 /* No need to recover an evicted BO */ 3577 if (shadow->tbo.mem.mem_type != TTM_PL_TT || 3578 shadow->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET || 3579 shadow->parent->tbo.mem.mem_type != TTM_PL_VRAM) 3580 continue; 3581 3582 r = amdgpu_bo_restore_shadow(shadow, &next); 3583 if (r) 3584 break; 3585 3586 if (fence) { 3587 tmo = dma_fence_wait_timeout(fence, false, tmo); 3588 dma_fence_put(fence); 3589 fence = next; 3590 if (tmo == 0) { 3591 r = -ETIMEDOUT; 3592 break; 3593 } else if (tmo < 0) { 3594 r = tmo; 3595 break; 3596 } 3597 } else { 3598 fence = next; 3599 } 3600 } 3601 mutex_unlock(&adev->shadow_list_lock); 3602 3603 if (fence) 3604 tmo = dma_fence_wait_timeout(fence, false, tmo); 3605 dma_fence_put(fence); 3606 3607 if (r < 0 || tmo <= 0) { 3608 DRM_ERROR("recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo); 3609 return -EIO; 3610 } 3611 3612 DRM_INFO("recover vram bo from shadow done\n"); 3613 return 0; 3614 } 3615 3616 3617 /** 3618 * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf 3619 * 3620 * @adev: amdgpu device pointer 3621 * @from_hypervisor: request from hypervisor 3622 * 3623 * do VF FLR and reinitialize Asic 3624 * return 0 means succeeded otherwise failed 3625 */ 3626 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev, 3627 bool from_hypervisor) 3628 { 3629 int r; 3630 3631 if (from_hypervisor) 3632 r = amdgpu_virt_request_full_gpu(adev, true); 3633 else 3634 r = amdgpu_virt_reset_gpu(adev); 3635 if (r) 3636 return r; 3637 3638 amdgpu_amdkfd_pre_reset(adev); 3639 3640 /* Resume IP prior to SMC */ 3641 r = amdgpu_device_ip_reinit_early_sriov(adev); 3642 if (r) 3643 goto error; 3644 3645 /* we need recover gart prior to run SMC/CP/SDMA resume */ 3646 amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]); 3647 3648 r = amdgpu_device_fw_loading(adev); 3649 if (r) 3650 return r; 3651 3652 /* now we are okay to resume SMC/CP/SDMA */ 3653 r = amdgpu_device_ip_reinit_late_sriov(adev); 3654 if (r) 3655 goto error; 3656 3657 amdgpu_irq_gpu_reset_resume_helper(adev); 3658 r = amdgpu_ib_ring_tests(adev); 3659 amdgpu_amdkfd_post_reset(adev); 3660 3661 error: 3662 amdgpu_virt_init_data_exchange(adev); 3663 amdgpu_virt_release_full_gpu(adev, true); 3664 if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) { 3665 amdgpu_inc_vram_lost(adev); 3666 r = amdgpu_device_recover_vram(adev); 3667 } 3668 3669 return r; 3670 } 3671 3672 /** 3673 * amdgpu_device_should_recover_gpu - check if we should try GPU recovery 3674 * 3675 * @adev: amdgpu device pointer 3676 * 3677 * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover 3678 * a hung GPU. 3679 */ 3680 bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev) 3681 { 3682 if (!amdgpu_device_ip_check_soft_reset(adev)) { 3683 DRM_INFO("Timeout, but no hardware hang detected.\n"); 3684 return false; 3685 } 3686 3687 if (amdgpu_gpu_recovery == 0) 3688 goto disabled; 3689 3690 if (amdgpu_sriov_vf(adev)) 3691 return true; 3692 3693 if (amdgpu_gpu_recovery == -1) { 3694 switch (adev->asic_type) { 3695 case CHIP_BONAIRE: 3696 case CHIP_HAWAII: 3697 case CHIP_TOPAZ: 3698 case CHIP_TONGA: 3699 case CHIP_FIJI: 3700 case CHIP_POLARIS10: 3701 case CHIP_POLARIS11: 3702 case CHIP_POLARIS12: 3703 case CHIP_VEGAM: 3704 case CHIP_VEGA20: 3705 case CHIP_VEGA10: 3706 case CHIP_VEGA12: 3707 case CHIP_RAVEN: 3708 break; 3709 default: 3710 goto disabled; 3711 } 3712 } 3713 3714 return true; 3715 3716 disabled: 3717 DRM_INFO("GPU recovery disabled.\n"); 3718 return false; 3719 } 3720 3721 3722 static int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev, 3723 struct amdgpu_job *job, 3724 bool *need_full_reset_arg) 3725 { 3726 int i, r = 0; 3727 bool need_full_reset = *need_full_reset_arg; 3728 3729 /* block all schedulers and reset given job's ring */ 3730 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { 3731 struct amdgpu_ring *ring = adev->rings[i]; 3732 3733 if (!ring || !ring->sched.thread) 3734 continue; 3735 3736 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */ 3737 amdgpu_fence_driver_force_completion(ring); 3738 } 3739 3740 if(job) 3741 drm_sched_increase_karma(&job->base); 3742 3743 /* Don't suspend on bare metal if we are not going to HW reset the ASIC */ 3744 if (!amdgpu_sriov_vf(adev)) { 3745 3746 if (!need_full_reset) 3747 need_full_reset = amdgpu_device_ip_need_full_reset(adev); 3748 3749 if (!need_full_reset) { 3750 amdgpu_device_ip_pre_soft_reset(adev); 3751 r = amdgpu_device_ip_soft_reset(adev); 3752 amdgpu_device_ip_post_soft_reset(adev); 3753 if (r || amdgpu_device_ip_check_soft_reset(adev)) { 3754 DRM_INFO("soft reset failed, will fallback to full reset!\n"); 3755 need_full_reset = true; 3756 } 3757 } 3758 3759 if (need_full_reset) 3760 r = amdgpu_device_ip_suspend(adev); 3761 3762 *need_full_reset_arg = need_full_reset; 3763 } 3764 3765 return r; 3766 } 3767 3768 static int amdgpu_do_asic_reset(struct amdgpu_hive_info *hive, 3769 struct list_head *device_list_handle, 3770 bool *need_full_reset_arg) 3771 { 3772 struct amdgpu_device *tmp_adev = NULL; 3773 bool need_full_reset = *need_full_reset_arg, vram_lost = false; 3774 int r = 0; 3775 3776 /* 3777 * ASIC reset has to be done on all HGMI hive nodes ASAP 3778 * to allow proper links negotiation in FW (within 1 sec) 3779 */ 3780 if (need_full_reset) { 3781 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) { 3782 /* For XGMI run all resets in parallel to speed up the process */ 3783 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) { 3784 if (!queue_work(system_highpri_wq, &tmp_adev->xgmi_reset_work)) 3785 r = -EALREADY; 3786 } else 3787 r = amdgpu_asic_reset(tmp_adev); 3788 3789 if (r) { 3790 DRM_ERROR("ASIC reset failed with error, %d for drm dev, %s", 3791 r, tmp_adev->ddev->unique); 3792 break; 3793 } 3794 } 3795 3796 /* For XGMI wait for all PSP resets to complete before proceed */ 3797 if (!r) { 3798 list_for_each_entry(tmp_adev, device_list_handle, 3799 gmc.xgmi.head) { 3800 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) { 3801 flush_work(&tmp_adev->xgmi_reset_work); 3802 r = tmp_adev->asic_reset_res; 3803 if (r) 3804 break; 3805 } 3806 } 3807 } 3808 } 3809 3810 3811 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) { 3812 if (need_full_reset) { 3813 /* post card */ 3814 if (amdgpu_atom_asic_init(tmp_adev->mode_info.atom_context)) 3815 DRM_WARN("asic atom init failed!"); 3816 3817 if (!r) { 3818 dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n"); 3819 r = amdgpu_device_ip_resume_phase1(tmp_adev); 3820 if (r) 3821 goto out; 3822 3823 vram_lost = amdgpu_device_check_vram_lost(tmp_adev); 3824 if (vram_lost) { 3825 DRM_INFO("VRAM is lost due to GPU reset!\n"); 3826 amdgpu_inc_vram_lost(tmp_adev); 3827 } 3828 3829 r = amdgpu_gtt_mgr_recover( 3830 &tmp_adev->mman.bdev.man[TTM_PL_TT]); 3831 if (r) 3832 goto out; 3833 3834 r = amdgpu_device_fw_loading(tmp_adev); 3835 if (r) 3836 return r; 3837 3838 r = amdgpu_device_ip_resume_phase2(tmp_adev); 3839 if (r) 3840 goto out; 3841 3842 if (vram_lost) 3843 amdgpu_device_fill_reset_magic(tmp_adev); 3844 3845 /* 3846 * Add this ASIC as tracked as reset was already 3847 * complete successfully. 3848 */ 3849 amdgpu_register_gpu_instance(tmp_adev); 3850 3851 r = amdgpu_device_ip_late_init(tmp_adev); 3852 if (r) 3853 goto out; 3854 3855 /* must succeed. */ 3856 amdgpu_ras_resume(tmp_adev); 3857 3858 /* Update PSP FW topology after reset */ 3859 if (hive && tmp_adev->gmc.xgmi.num_physical_nodes > 1) 3860 r = amdgpu_xgmi_update_topology(hive, tmp_adev); 3861 } 3862 } 3863 3864 3865 out: 3866 if (!r) { 3867 amdgpu_irq_gpu_reset_resume_helper(tmp_adev); 3868 r = amdgpu_ib_ring_tests(tmp_adev); 3869 if (r) { 3870 dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r); 3871 r = amdgpu_device_ip_suspend(tmp_adev); 3872 need_full_reset = true; 3873 r = -EAGAIN; 3874 goto end; 3875 } 3876 } 3877 3878 if (!r) 3879 r = amdgpu_device_recover_vram(tmp_adev); 3880 else 3881 tmp_adev->asic_reset_res = r; 3882 } 3883 3884 end: 3885 *need_full_reset_arg = need_full_reset; 3886 return r; 3887 } 3888 3889 static bool amdgpu_device_lock_adev(struct amdgpu_device *adev, bool trylock) 3890 { 3891 if (trylock) { 3892 if (!mutex_trylock(&adev->lock_reset)) 3893 return false; 3894 } else 3895 mutex_lock(&adev->lock_reset); 3896 3897 atomic_inc(&adev->gpu_reset_counter); 3898 adev->in_gpu_reset = 1; 3899 switch (amdgpu_asic_reset_method(adev)) { 3900 case AMD_RESET_METHOD_MODE1: 3901 adev->mp1_state = PP_MP1_STATE_SHUTDOWN; 3902 break; 3903 case AMD_RESET_METHOD_MODE2: 3904 adev->mp1_state = PP_MP1_STATE_RESET; 3905 break; 3906 default: 3907 adev->mp1_state = PP_MP1_STATE_NONE; 3908 break; 3909 } 3910 3911 return true; 3912 } 3913 3914 static void amdgpu_device_unlock_adev(struct amdgpu_device *adev) 3915 { 3916 amdgpu_vf_error_trans_all(adev); 3917 adev->mp1_state = PP_MP1_STATE_NONE; 3918 adev->in_gpu_reset = 0; 3919 mutex_unlock(&adev->lock_reset); 3920 } 3921 3922 /** 3923 * amdgpu_device_gpu_recover - reset the asic and recover scheduler 3924 * 3925 * @adev: amdgpu device pointer 3926 * @job: which job trigger hang 3927 * 3928 * Attempt to reset the GPU if it has hung (all asics). 3929 * Attempt to do soft-reset or full-reset and reinitialize Asic 3930 * Returns 0 for success or an error on failure. 3931 */ 3932 3933 int amdgpu_device_gpu_recover(struct amdgpu_device *adev, 3934 struct amdgpu_job *job) 3935 { 3936 struct list_head device_list, *device_list_handle = NULL; 3937 bool need_full_reset, job_signaled; 3938 struct amdgpu_hive_info *hive = NULL; 3939 struct amdgpu_device *tmp_adev = NULL; 3940 int i, r = 0; 3941 bool in_ras_intr = amdgpu_ras_intr_triggered(); 3942 3943 /* 3944 * Flush RAM to disk so that after reboot 3945 * the user can read log and see why the system rebooted. 3946 */ 3947 if (in_ras_intr && amdgpu_ras_get_context(adev)->reboot) { 3948 3949 DRM_WARN("Emergency reboot."); 3950 3951 ksys_sync_helper(); 3952 emergency_restart(); 3953 } 3954 3955 need_full_reset = job_signaled = false; 3956 INIT_LIST_HEAD(&device_list); 3957 3958 dev_info(adev->dev, "GPU %s begin!\n", in_ras_intr ? "jobs stop":"reset"); 3959 3960 cancel_delayed_work_sync(&adev->delayed_init_work); 3961 3962 hive = amdgpu_get_xgmi_hive(adev, false); 3963 3964 /* 3965 * Here we trylock to avoid chain of resets executing from 3966 * either trigger by jobs on different adevs in XGMI hive or jobs on 3967 * different schedulers for same device while this TO handler is running. 3968 * We always reset all schedulers for device and all devices for XGMI 3969 * hive so that should take care of them too. 3970 */ 3971 3972 if (hive && !mutex_trylock(&hive->reset_lock)) { 3973 DRM_INFO("Bailing on TDR for s_job:%llx, hive: %llx as another already in progress", 3974 job ? job->base.id : -1, hive->hive_id); 3975 return 0; 3976 } 3977 3978 /* Start with adev pre asic reset first for soft reset check.*/ 3979 if (!amdgpu_device_lock_adev(adev, !hive)) { 3980 DRM_INFO("Bailing on TDR for s_job:%llx, as another already in progress", 3981 job ? job->base.id : -1); 3982 return 0; 3983 } 3984 3985 /* Block kfd: SRIOV would do it separately */ 3986 if (!amdgpu_sriov_vf(adev)) 3987 amdgpu_amdkfd_pre_reset(adev); 3988 3989 /* Build list of devices to reset */ 3990 if (adev->gmc.xgmi.num_physical_nodes > 1) { 3991 if (!hive) { 3992 /*unlock kfd: SRIOV would do it separately */ 3993 if (!amdgpu_sriov_vf(adev)) 3994 amdgpu_amdkfd_post_reset(adev); 3995 amdgpu_device_unlock_adev(adev); 3996 return -ENODEV; 3997 } 3998 3999 /* 4000 * In case we are in XGMI hive mode device reset is done for all the 4001 * nodes in the hive to retrain all XGMI links and hence the reset 4002 * sequence is executed in loop on all nodes. 4003 */ 4004 device_list_handle = &hive->device_list; 4005 } else { 4006 list_add_tail(&adev->gmc.xgmi.head, &device_list); 4007 device_list_handle = &device_list; 4008 } 4009 4010 /* block all schedulers and reset given job's ring */ 4011 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) { 4012 if (tmp_adev != adev) { 4013 amdgpu_device_lock_adev(tmp_adev, false); 4014 if (!amdgpu_sriov_vf(tmp_adev)) 4015 amdgpu_amdkfd_pre_reset(tmp_adev); 4016 } 4017 4018 /* 4019 * Mark these ASICs to be reseted as untracked first 4020 * And add them back after reset completed 4021 */ 4022 amdgpu_unregister_gpu_instance(tmp_adev); 4023 4024 /* disable ras on ALL IPs */ 4025 if (!in_ras_intr && amdgpu_device_ip_need_full_reset(tmp_adev)) 4026 amdgpu_ras_suspend(tmp_adev); 4027 4028 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { 4029 struct amdgpu_ring *ring = tmp_adev->rings[i]; 4030 4031 if (!ring || !ring->sched.thread) 4032 continue; 4033 4034 drm_sched_stop(&ring->sched, job ? &job->base : NULL); 4035 4036 if (in_ras_intr) 4037 amdgpu_job_stop_all_jobs_on_sched(&ring->sched); 4038 } 4039 } 4040 4041 4042 if (in_ras_intr) 4043 goto skip_sched_resume; 4044 4045 /* 4046 * Must check guilty signal here since after this point all old 4047 * HW fences are force signaled. 4048 * 4049 * job->base holds a reference to parent fence 4050 */ 4051 if (job && job->base.s_fence->parent && 4052 dma_fence_is_signaled(job->base.s_fence->parent)) 4053 job_signaled = true; 4054 4055 if (job_signaled) { 4056 dev_info(adev->dev, "Guilty job already signaled, skipping HW reset"); 4057 goto skip_hw_reset; 4058 } 4059 4060 4061 /* Guilty job will be freed after this*/ 4062 r = amdgpu_device_pre_asic_reset(adev, job, &need_full_reset); 4063 if (r) { 4064 /*TODO Should we stop ?*/ 4065 DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ", 4066 r, adev->ddev->unique); 4067 adev->asic_reset_res = r; 4068 } 4069 4070 retry: /* Rest of adevs pre asic reset from XGMI hive. */ 4071 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) { 4072 4073 if (tmp_adev == adev) 4074 continue; 4075 4076 r = amdgpu_device_pre_asic_reset(tmp_adev, 4077 NULL, 4078 &need_full_reset); 4079 /*TODO Should we stop ?*/ 4080 if (r) { 4081 DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ", 4082 r, tmp_adev->ddev->unique); 4083 tmp_adev->asic_reset_res = r; 4084 } 4085 } 4086 4087 /* Actual ASIC resets if needed.*/ 4088 /* TODO Implement XGMI hive reset logic for SRIOV */ 4089 if (amdgpu_sriov_vf(adev)) { 4090 r = amdgpu_device_reset_sriov(adev, job ? false : true); 4091 if (r) 4092 adev->asic_reset_res = r; 4093 } else { 4094 r = amdgpu_do_asic_reset(hive, device_list_handle, &need_full_reset); 4095 if (r && r == -EAGAIN) 4096 goto retry; 4097 } 4098 4099 skip_hw_reset: 4100 4101 /* Post ASIC reset for all devs .*/ 4102 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) { 4103 4104 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { 4105 struct amdgpu_ring *ring = tmp_adev->rings[i]; 4106 4107 if (!ring || !ring->sched.thread) 4108 continue; 4109 4110 /* No point to resubmit jobs if we didn't HW reset*/ 4111 if (!tmp_adev->asic_reset_res && !job_signaled) 4112 drm_sched_resubmit_jobs(&ring->sched); 4113 4114 drm_sched_start(&ring->sched, !tmp_adev->asic_reset_res); 4115 } 4116 4117 if (!amdgpu_device_has_dc_support(tmp_adev) && !job_signaled) { 4118 drm_helper_resume_force_mode(tmp_adev->ddev); 4119 } 4120 4121 tmp_adev->asic_reset_res = 0; 4122 4123 if (r) { 4124 /* bad news, how to tell it to userspace ? */ 4125 dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&tmp_adev->gpu_reset_counter)); 4126 amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r); 4127 } else { 4128 dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&tmp_adev->gpu_reset_counter)); 4129 } 4130 } 4131 4132 skip_sched_resume: 4133 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) { 4134 /*unlock kfd: SRIOV would do it separately */ 4135 if (!in_ras_intr && !amdgpu_sriov_vf(tmp_adev)) 4136 amdgpu_amdkfd_post_reset(tmp_adev); 4137 amdgpu_device_unlock_adev(tmp_adev); 4138 } 4139 4140 if (hive) 4141 mutex_unlock(&hive->reset_lock); 4142 4143 if (r) 4144 dev_info(adev->dev, "GPU reset end with ret = %d\n", r); 4145 return r; 4146 } 4147 4148 /** 4149 * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot 4150 * 4151 * @adev: amdgpu_device pointer 4152 * 4153 * Fetchs and stores in the driver the PCIE capabilities (gen speed 4154 * and lanes) of the slot the device is in. Handles APUs and 4155 * virtualized environments where PCIE config space may not be available. 4156 */ 4157 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev) 4158 { 4159 struct pci_dev *pdev; 4160 enum pci_bus_speed speed_cap, platform_speed_cap; 4161 enum pcie_link_width platform_link_width; 4162 4163 if (amdgpu_pcie_gen_cap) 4164 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap; 4165 4166 if (amdgpu_pcie_lane_cap) 4167 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap; 4168 4169 /* covers APUs as well */ 4170 if (pci_is_root_bus(adev->pdev->bus)) { 4171 if (adev->pm.pcie_gen_mask == 0) 4172 adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK; 4173 if (adev->pm.pcie_mlw_mask == 0) 4174 adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK; 4175 return; 4176 } 4177 4178 if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask) 4179 return; 4180 4181 pcie_bandwidth_available(adev->pdev, NULL, 4182 &platform_speed_cap, &platform_link_width); 4183 4184 if (adev->pm.pcie_gen_mask == 0) { 4185 /* asic caps */ 4186 pdev = adev->pdev; 4187 speed_cap = pcie_get_speed_cap(pdev); 4188 if (speed_cap == PCI_SPEED_UNKNOWN) { 4189 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | 4190 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | 4191 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3); 4192 } else { 4193 if (speed_cap == PCIE_SPEED_16_0GT) 4194 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | 4195 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | 4196 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 | 4197 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4); 4198 else if (speed_cap == PCIE_SPEED_8_0GT) 4199 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | 4200 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 | 4201 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3); 4202 else if (speed_cap == PCIE_SPEED_5_0GT) 4203 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 | 4204 CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2); 4205 else 4206 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1; 4207 } 4208 /* platform caps */ 4209 if (platform_speed_cap == PCI_SPEED_UNKNOWN) { 4210 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | 4211 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2); 4212 } else { 4213 if (platform_speed_cap == PCIE_SPEED_16_0GT) 4214 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | 4215 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 | 4216 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 | 4217 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4); 4218 else if (platform_speed_cap == PCIE_SPEED_8_0GT) 4219 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | 4220 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 | 4221 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3); 4222 else if (platform_speed_cap == PCIE_SPEED_5_0GT) 4223 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 | 4224 CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2); 4225 else 4226 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1; 4227 4228 } 4229 } 4230 if (adev->pm.pcie_mlw_mask == 0) { 4231 if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) { 4232 adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK; 4233 } else { 4234 switch (platform_link_width) { 4235 case PCIE_LNK_X32: 4236 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 | 4237 CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 | 4238 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 | 4239 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | 4240 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | 4241 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | 4242 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); 4243 break; 4244 case PCIE_LNK_X16: 4245 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 | 4246 CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 | 4247 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | 4248 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | 4249 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | 4250 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); 4251 break; 4252 case PCIE_LNK_X12: 4253 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 | 4254 CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | 4255 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | 4256 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | 4257 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); 4258 break; 4259 case PCIE_LNK_X8: 4260 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 | 4261 CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | 4262 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | 4263 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); 4264 break; 4265 case PCIE_LNK_X4: 4266 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 | 4267 CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | 4268 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); 4269 break; 4270 case PCIE_LNK_X2: 4271 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 | 4272 CAIL_PCIE_LINK_WIDTH_SUPPORT_X1); 4273 break; 4274 case PCIE_LNK_X1: 4275 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1; 4276 break; 4277 default: 4278 break; 4279 } 4280 } 4281 } 4282 } 4283 4284