drivers/gpu/drm/amd/amdgpu/amdgpu_device.c

   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 MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
  69 MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
  70 MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
  71 MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin");
  72 MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin");
  73 MODULE_FIRMWARE("amdgpu/navi10_gpu_info.bin");
  74
  75 #define AMDGPU_RESUME_MS                2000
  76
  77 static const char *amdgpu_asic_name[] = {
  78         "TAHITI",
  79         "PITCAIRN",
  80         "VERDE",
  81         "OLAND",
  82         "HAINAN",
  83         "BONAIRE",
  84         "KAVERI",
  85         "KABINI",
  86         "HAWAII",
  87         "MULLINS",
  88         "TOPAZ",
  89         "TONGA",
  90         "FIJI",
  91         "CARRIZO",
  92         "STONEY",
  93         "POLARIS10",
  94         "POLARIS11",
  95         "POLARIS12",
  96         "VEGAM",
  97         "VEGA10",
  98         "VEGA12",
  99         "VEGA20",
 100         "RAVEN",
 101         "NAVI10",
 102         "LAST",
 103 };
 104
 105 /**
 106  * DOC: pcie_replay_count
 107  *
 108  * The amdgpu driver provides a sysfs API for reporting the total number
 109  * of PCIe replays (NAKs)
 110  * The file pcie_replay_count is used for this and returns the total
 111  * number of replays as a sum of the NAKs generated and NAKs received
 112  */
 113
 114 static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev,
 115                 struct device_attribute *attr, char *buf)
 116 {
 117         struct drm_device *ddev = dev_get_drvdata(dev);
 118         struct amdgpu_device *adev = ddev->dev_private;
 119         uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev);
 120
 121         return snprintf(buf, PAGE_SIZE, "%llu\n", cnt);
 122 }
 123
 124 static DEVICE_ATTR(pcie_replay_count, S_IRUGO,
 125                 amdgpu_device_get_pcie_replay_count, NULL);
 126
 127 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
 128
 129 /**
 130  * amdgpu_device_is_px - Is the device is a dGPU with HG/PX power control
 131  *
 132  * @dev: drm_device pointer
 133  *
 134  * Returns true if the device is a dGPU with HG/PX power control,
 135  * otherwise return false.
 136  */
 137 bool amdgpu_device_is_px(struct drm_device *dev)
 138 {
 139         struct amdgpu_device *adev = dev->dev_private;
 140
 141         if (adev->flags & AMD_IS_PX)
 142                 return true;
 143         return false;
 144 }
 145
 146 /*
 147  * MMIO register access helper functions.
 148  */
 149 /**
 150  * amdgpu_mm_rreg - read a memory mapped IO register
 151  *
 152  * @adev: amdgpu_device pointer
 153  * @reg: dword aligned register offset
 154  * @acc_flags: access flags which require special behavior
 155  *
 156  * Returns the 32 bit value from the offset specified.
 157  */
 158 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
 159                         uint32_t acc_flags)
 160 {
 161         uint32_t ret;
 162
 163         if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
 164                 return amdgpu_virt_kiq_rreg(adev, reg);
 165
 166         if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
 167                 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
 168         else {
 169                 unsigned long flags;
 170
 171                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
 172                 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
 173                 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
 174                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
 175         }
 176         trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
 177         return ret;
 178 }
 179
 180 /*
 181  * MMIO register read with bytes helper functions
 182  * @offset:bytes offset from MMIO start
 183  *
 184 */
 185
 186 /**
 187  * amdgpu_mm_rreg8 - read a memory mapped IO register
 188  *
 189  * @adev: amdgpu_device pointer
 190  * @offset: byte aligned register offset
 191  *
 192  * Returns the 8 bit value from the offset specified.
 193  */
 194 uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) {
 195         if (offset < adev->rmmio_size)
 196                 return (readb(adev->rmmio + offset));
 197         BUG();
 198 }
 199
 200 /*
 201  * MMIO register write with bytes helper functions
 202  * @offset:bytes offset from MMIO start
 203  * @value: the value want to be written to the register
 204  *
 205 */
 206 /**
 207  * amdgpu_mm_wreg8 - read a memory mapped IO register
 208  *
 209  * @adev: amdgpu_device pointer
 210  * @offset: byte aligned register offset
 211  * @value: 8 bit value to write
 212  *
 213  * Writes the value specified to the offset specified.
 214  */
 215 void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) {
 216         if (offset < adev->rmmio_size)
 217                 writeb(value, adev->rmmio + offset);
 218         else
 219                 BUG();
 220 }
 221
 222 /**
 223  * amdgpu_mm_wreg - write to a memory mapped IO register
 224  *
 225  * @adev: amdgpu_device pointer
 226  * @reg: dword aligned register offset
 227  * @v: 32 bit value to write to the register
 228  * @acc_flags: access flags which require special behavior
 229  *
 230  * Writes the value specified to the offset specified.
 231  */
 232 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
 233                     uint32_t acc_flags)
 234 {
 235         trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
 236
 237         if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
 238                 adev->last_mm_index = v;
 239         }
 240
 241         if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
 242                 return amdgpu_virt_kiq_wreg(adev, reg, v);
 243
 244         if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
 245                 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
 246         else {
 247                 unsigned long flags;
 248
 249                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
 250                 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
 251                 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
 252                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
 253         }
 254
 255         if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
 256                 udelay(500);
 257         }
 258 }
 259
 260 /**
 261  * amdgpu_io_rreg - read an IO register
 262  *
 263  * @adev: amdgpu_device pointer
 264  * @reg: dword aligned register offset
 265  *
 266  * Returns the 32 bit value from the offset specified.
 267  */
 268 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
 269 {
 270         if ((reg * 4) < adev->rio_mem_size)
 271                 return ioread32(adev->rio_mem + (reg * 4));
 272         else {
 273                 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
 274                 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
 275         }
 276 }
 277
 278 /**
 279  * amdgpu_io_wreg - write to an IO register
 280  *
 281  * @adev: amdgpu_device pointer
 282  * @reg: dword aligned register offset
 283  * @v: 32 bit value to write to the register
 284  *
 285  * Writes the value specified to the offset specified.
 286  */
 287 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
 288 {
 289         if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
 290                 adev->last_mm_index = v;
 291         }
 292
 293         if ((reg * 4) < adev->rio_mem_size)
 294                 iowrite32(v, adev->rio_mem + (reg * 4));
 295         else {
 296                 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
 297                 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
 298         }
 299
 300         if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
 301                 udelay(500);
 302         }
 303 }
 304
 305 /**
 306  * amdgpu_mm_rdoorbell - read a doorbell dword
 307  *
 308  * @adev: amdgpu_device pointer
 309  * @index: doorbell index
 310  *
 311  * Returns the value in the doorbell aperture at the
 312  * requested doorbell index (CIK).
 313  */
 314 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
 315 {
 316         if (index < adev->doorbell.num_doorbells) {
 317                 return readl(adev->doorbell.ptr + index);
 318         } else {
 319                 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
 320                 return 0;
 321         }
 322 }
 323
 324 /**
 325  * amdgpu_mm_wdoorbell - write a doorbell dword
 326  *
 327  * @adev: amdgpu_device pointer
 328  * @index: doorbell index
 329  * @v: value to write
 330  *
 331  * Writes @v to the doorbell aperture at the
 332  * requested doorbell index (CIK).
 333  */
 334 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
 335 {
 336         if (index < adev->doorbell.num_doorbells) {
 337                 writel(v, adev->doorbell.ptr + index);
 338         } else {
 339                 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
 340         }
 341 }
 342
 343 /**
 344  * amdgpu_mm_rdoorbell64 - read a doorbell Qword
 345  *
 346  * @adev: amdgpu_device pointer
 347  * @index: doorbell index
 348  *
 349  * Returns the value in the doorbell aperture at the
 350  * requested doorbell index (VEGA10+).
 351  */
 352 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
 353 {
 354         if (index < adev->doorbell.num_doorbells) {
 355                 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
 356         } else {
 357                 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
 358                 return 0;
 359         }
 360 }
 361
 362 /**
 363  * amdgpu_mm_wdoorbell64 - write a doorbell Qword
 364  *
 365  * @adev: amdgpu_device pointer
 366  * @index: doorbell index
 367  * @v: value to write
 368  *
 369  * Writes @v to the doorbell aperture at the
 370  * requested doorbell index (VEGA10+).
 371  */
 372 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
 373 {
 374         if (index < adev->doorbell.num_doorbells) {
 375                 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
 376         } else {
 377                 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
 378         }
 379 }
 380
 381 /**
 382  * amdgpu_invalid_rreg - dummy reg read function
 383  *
 384  * @adev: amdgpu device pointer
 385  * @reg: offset of register
 386  *
 387  * Dummy register read function.  Used for register blocks
 388  * that certain asics don't have (all asics).
 389  * Returns the value in the register.
 390  */
 391 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
 392 {
 393         DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
 394         BUG();
 395         return 0;
 396 }
 397
 398 /**
 399  * amdgpu_invalid_wreg - dummy reg write function
 400  *
 401  * @adev: amdgpu device pointer
 402  * @reg: offset of register
 403  * @v: value to write to the register
 404  *
 405  * Dummy register read function.  Used for register blocks
 406  * that certain asics don't have (all asics).
 407  */
 408 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
 409 {
 410         DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
 411                   reg, v);
 412         BUG();
 413 }
 414
 415 /**
 416  * amdgpu_block_invalid_rreg - dummy reg read function
 417  *
 418  * @adev: amdgpu device pointer
 419  * @block: offset of instance
 420  * @reg: offset of register
 421  *
 422  * Dummy register read function.  Used for register blocks
 423  * that certain asics don't have (all asics).
 424  * Returns the value in the register.
 425  */
 426 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
 427                                           uint32_t block, uint32_t reg)
 428 {
 429         DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
 430                   reg, block);
 431         BUG();
 432         return 0;
 433 }
 434
 435 /**
 436  * amdgpu_block_invalid_wreg - dummy reg write function
 437  *
 438  * @adev: amdgpu device pointer
 439  * @block: offset of instance
 440  * @reg: offset of register
 441  * @v: value to write to the register
 442  *
 443  * Dummy register read function.  Used for register blocks
 444  * that certain asics don't have (all asics).
 445  */
 446 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
 447                                       uint32_t block,
 448                                       uint32_t reg, uint32_t v)
 449 {
 450         DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
 451                   reg, block, v);
 452         BUG();
 453 }
 454
 455 /**
 456  * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
 457  *
 458  * @adev: amdgpu device pointer
 459  *
 460  * Allocates a scratch page of VRAM for use by various things in the
 461  * driver.
 462  */
 463 static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
 464 {
 465         return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
 466                                        PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
 467                                        &adev->vram_scratch.robj,
 468                                        &adev->vram_scratch.gpu_addr,
 469                                        (void **)&adev->vram_scratch.ptr);
 470 }
 471
 472 /**
 473  * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
 474  *
 475  * @adev: amdgpu device pointer
 476  *
 477  * Frees the VRAM scratch page.
 478  */
 479 static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
 480 {
 481         amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
 482 }
 483
 484 /**
 485  * amdgpu_device_program_register_sequence - program an array of registers.
 486  *
 487  * @adev: amdgpu_device pointer
 488  * @registers: pointer to the register array
 489  * @array_size: size of the register array
 490  *
 491  * Programs an array or registers with and and or masks.
 492  * This is a helper for setting golden registers.
 493  */
 494 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
 495                                              const u32 *registers,
 496                                              const u32 array_size)
 497 {
 498         u32 tmp, reg, and_mask, or_mask;
 499         int i;
 500
 501         if (array_size % 3)
 502                 return;
 503
 504         for (i = 0; i < array_size; i +=3) {
 505                 reg = registers[i + 0];
 506                 and_mask = registers[i + 1];
 507                 or_mask = registers[i + 2];
 508
 509                 if (and_mask == 0xffffffff) {
 510                         tmp = or_mask;
 511                 } else {
 512                         tmp = RREG32(reg);
 513                         tmp &= ~and_mask;
 514                         if (adev->family >= AMDGPU_FAMILY_AI)
 515                                 tmp |= (or_mask & and_mask);
 516                         else
 517                                 tmp |= or_mask;
 518                 }
 519                 WREG32(reg, tmp);
 520         }
 521 }
 522
 523 /**
 524  * amdgpu_device_pci_config_reset - reset the GPU
 525  *
 526  * @adev: amdgpu_device pointer
 527  *
 528  * Resets the GPU using the pci config reset sequence.
 529  * Only applicable to asics prior to vega10.
 530  */
 531 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
 532 {
 533         pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
 534 }
 535
 536 /*
 537  * GPU doorbell aperture helpers function.
 538  */
 539 /**
 540  * amdgpu_device_doorbell_init - Init doorbell driver information.
 541  *
 542  * @adev: amdgpu_device pointer
 543  *
 544  * Init doorbell driver information (CIK)
 545  * Returns 0 on success, error on failure.
 546  */
 547 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
 548 {
 549
 550         /* No doorbell on SI hardware generation */
 551         if (adev->asic_type < CHIP_BONAIRE) {
 552                 adev->doorbell.base = 0;
 553                 adev->doorbell.size = 0;
 554                 adev->doorbell.num_doorbells = 0;
 555                 adev->doorbell.ptr = NULL;
 556                 return 0;
 557         }
 558
 559         if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
 560                 return -EINVAL;
 561
 562         amdgpu_asic_init_doorbell_index(adev);
 563
 564         /* doorbell bar mapping */
 565         adev->doorbell.base = pci_resource_start(adev->pdev, 2);
 566         adev->doorbell.size = pci_resource_len(adev->pdev, 2);
 567
 568         adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
 569                                              adev->doorbell_index.max_assignment+1);
 570         if (adev->doorbell.num_doorbells == 0)
 571                 return -EINVAL;
 572
 573         /* For Vega, reserve and map two pages on doorbell BAR since SDMA
 574          * paging queue doorbell use the second page. The
 575          * AMDGPU_DOORBELL64_MAX_ASSIGNMENT definition assumes all the
 576          * doorbells are in the first page. So with paging queue enabled,
 577          * the max num_doorbells should + 1 page (0x400 in dword)
 578          */
 579         if (adev->asic_type >= CHIP_VEGA10)
 580                 adev->doorbell.num_doorbells += 0x400;
 581
 582         adev->doorbell.ptr = ioremap(adev->doorbell.base,
 583                                      adev->doorbell.num_doorbells *
 584                                      sizeof(u32));
 585         if (adev->doorbell.ptr == NULL)
 586                 return -ENOMEM;
 587
 588         return 0;
 589 }
 590
 591 /**
 592  * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
 593  *
 594  * @adev: amdgpu_device pointer
 595  *
 596  * Tear down doorbell driver information (CIK)
 597  */
 598 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
 599 {
 600         iounmap(adev->doorbell.ptr);
 601         adev->doorbell.ptr = NULL;
 602 }
 603
 604
 605
 606 /*
 607  * amdgpu_device_wb_*()
 608  * Writeback is the method by which the GPU updates special pages in memory
 609  * with the status of certain GPU events (fences, ring pointers,etc.).
 610  */
 611
 612 /**
 613  * amdgpu_device_wb_fini - Disable Writeback and free memory
 614  *
 615  * @adev: amdgpu_device pointer
 616  *
 617  * Disables Writeback and frees the Writeback memory (all asics).
 618  * Used at driver shutdown.
 619  */
 620 static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
 621 {
 622         if (adev->wb.wb_obj) {
 623                 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
 624                                       &adev->wb.gpu_addr,
 625                                       (void **)&adev->wb.wb);
 626                 adev->wb.wb_obj = NULL;
 627         }
 628 }
 629
 630 /**
 631  * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
 632  *
 633  * @adev: amdgpu_device pointer
 634  *
 635  * Initializes writeback and allocates writeback memory (all asics).
 636  * Used at driver startup.
 637  * Returns 0 on success or an -error on failure.
 638  */
 639 static int amdgpu_device_wb_init(struct amdgpu_device *adev)
 640 {
 641         int r;
 642
 643         if (adev->wb.wb_obj == NULL) {
 644                 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
 645                 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
 646                                             PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
 647                                             &adev->wb.wb_obj, &adev->wb.gpu_addr,
 648                                             (void **)&adev->wb.wb);
 649                 if (r) {
 650                         dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
 651                         return r;
 652                 }
 653
 654                 adev->wb.num_wb = AMDGPU_MAX_WB;
 655                 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
 656
 657                 /* clear wb memory */
 658                 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
 659         }
 660
 661         return 0;
 662 }
 663
 664 /**
 665  * amdgpu_device_wb_get - Allocate a wb entry
 666  *
 667  * @adev: amdgpu_device pointer
 668  * @wb: wb index
 669  *
 670  * Allocate a wb slot for use by the driver (all asics).
 671  * Returns 0 on success or -EINVAL on failure.
 672  */
 673 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
 674 {
 675         unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
 676
 677         if (offset < adev->wb.num_wb) {
 678                 __set_bit(offset, adev->wb.used);
 679                 *wb = offset << 3; /* convert to dw offset */
 680                 return 0;
 681         } else {
 682                 return -EINVAL;
 683         }
 684 }
 685
 686 /**
 687  * amdgpu_device_wb_free - Free a wb entry
 688  *
 689  * @adev: amdgpu_device pointer
 690  * @wb: wb index
 691  *
 692  * Free a wb slot allocated for use by the driver (all asics)
 693  */
 694 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
 695 {
 696         wb >>= 3;
 697         if (wb < adev->wb.num_wb)
 698                 __clear_bit(wb, adev->wb.used);
 699 }
 700
 701 /**
 702  * amdgpu_device_resize_fb_bar - try to resize FB BAR
 703  *
 704  * @adev: amdgpu_device pointer
 705  *
 706  * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
 707  * to fail, but if any of the BARs is not accessible after the size we abort
 708  * driver loading by returning -ENODEV.
 709  */
 710 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
 711 {
 712         u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
 713         u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
 714         struct pci_bus *root;
 715         struct resource *res;
 716         unsigned i;
 717         u16 cmd;
 718         int r;
 719
 720         /* Bypass for VF */
 721         if (amdgpu_sriov_vf(adev))
 722                 return 0;
 723
 724         /* Check if the root BUS has 64bit memory resources */
 725         root = adev->pdev->bus;
 726         while (root->parent)
 727                 root = root->parent;
 728
 729         pci_bus_for_each_resource(root, res, i) {
 730                 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
 731                     res->start > 0x100000000ull)
 732                         break;
 733         }
 734
 735         /* Trying to resize is pointless without a root hub window above 4GB */
 736         if (!res)
 737                 return 0;
 738
 739         /* Disable memory decoding while we change the BAR addresses and size */
 740         pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
 741         pci_write_config_word(adev->pdev, PCI_COMMAND,
 742                               cmd & ~PCI_COMMAND_MEMORY);
 743
 744         /* Free the VRAM and doorbell BAR, we most likely need to move both. */
 745         amdgpu_device_doorbell_fini(adev);
 746         if (adev->asic_type >= CHIP_BONAIRE)
 747                 pci_release_resource(adev->pdev, 2);
 748
 749         pci_release_resource(adev->pdev, 0);
 750
 751         r = pci_resize_resource(adev->pdev, 0, rbar_size);
 752         if (r == -ENOSPC)
 753                 DRM_INFO("Not enough PCI address space for a large BAR.");
 754         else if (r && r != -ENOTSUPP)
 755                 DRM_ERROR("Problem resizing BAR0 (%d).", r);
 756
 757         pci_assign_unassigned_bus_resources(adev->pdev->bus);
 758
 759         /* When the doorbell or fb BAR isn't available we have no chance of
 760          * using the device.
 761          */
 762         r = amdgpu_device_doorbell_init(adev);
 763         if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
 764                 return -ENODEV;
 765
 766         pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
 767
 768         return 0;
 769 }
 770
 771 /*
 772  * GPU helpers function.
 773  */
 774 /**
 775  * amdgpu_device_need_post - check if the hw need post or not
 776  *
 777  * @adev: amdgpu_device pointer
 778  *
 779  * Check if the asic has been initialized (all asics) at driver startup
 780  * or post is needed if  hw reset is performed.
 781  * Returns true if need or false if not.
 782  */
 783 bool amdgpu_device_need_post(struct amdgpu_device *adev)
 784 {
 785         uint32_t reg;
 786
 787         if (amdgpu_sriov_vf(adev))
 788                 return false;
 789
 790         if (amdgpu_passthrough(adev)) {
 791                 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
 792                  * some old smc fw still need driver do vPost otherwise gpu hang, while
 793                  * those smc fw version above 22.15 doesn't have this flaw, so we force
 794                  * vpost executed for smc version below 22.15
 795                  */
 796                 if (adev->asic_type == CHIP_FIJI) {
 797                         int err;
 798                         uint32_t fw_ver;
 799                         err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
 800                         /* force vPost if error occured */
 801                         if (err)
 802                                 return true;
 803
 804                         fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
 805                         if (fw_ver < 0x00160e00)
 806                                 return true;
 807                 }
 808         }
 809
 810         if (adev->has_hw_reset) {
 811                 adev->has_hw_reset = false;
 812                 return true;
 813         }
 814
 815         /* bios scratch used on CIK+ */
 816         if (adev->asic_type >= CHIP_BONAIRE)
 817                 return amdgpu_atombios_scratch_need_asic_init(adev);
 818
 819         /* check MEM_SIZE for older asics */
 820         reg = amdgpu_asic_get_config_memsize(adev);
 821
 822         if ((reg != 0) && (reg != 0xffffffff))
 823                 return false;
 824
 825         return true;
 826 }
 827
 828 /* if we get transitioned to only one device, take VGA back */
 829 /**
 830  * amdgpu_device_vga_set_decode - enable/disable vga decode
 831  *
 832  * @cookie: amdgpu_device pointer
 833  * @state: enable/disable vga decode
 834  *
 835  * Enable/disable vga decode (all asics).
 836  * Returns VGA resource flags.
 837  */
 838 static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
 839 {
 840         struct amdgpu_device *adev = cookie;
 841         amdgpu_asic_set_vga_state(adev, state);
 842         if (state)
 843                 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
 844                        VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
 845         else
 846                 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
 847 }
 848
 849 /**
 850  * amdgpu_device_check_block_size - validate the vm block size
 851  *
 852  * @adev: amdgpu_device pointer
 853  *
 854  * Validates the vm block size specified via module parameter.
 855  * The vm block size defines number of bits in page table versus page directory,
 856  * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
 857  * page table and the remaining bits are in the page directory.
 858  */
 859 static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
 860 {
 861         /* defines number of bits in page table versus page directory,
 862          * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
 863          * page table and the remaining bits are in the page directory */
 864         if (amdgpu_vm_block_size == -1)
 865                 return;
 866
 867         if (amdgpu_vm_block_size < 9) {
 868                 dev_warn(adev->dev, "VM page table size (%d) too small\n",
 869                          amdgpu_vm_block_size);
 870                 amdgpu_vm_block_size = -1;
 871         }
 872 }
 873
 874 /**
 875  * amdgpu_device_check_vm_size - validate the vm size
 876  *
 877  * @adev: amdgpu_device pointer
 878  *
 879  * Validates the vm size in GB specified via module parameter.
 880  * The VM size is the size of the GPU virtual memory space in GB.
 881  */
 882 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
 883 {
 884         /* no need to check the default value */
 885         if (amdgpu_vm_size == -1)
 886                 return;
 887
 888         if (amdgpu_vm_size < 1) {
 889                 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
 890                          amdgpu_vm_size);
 891                 amdgpu_vm_size = -1;
 892         }
 893 }
 894
 895 static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
 896 {
 897         struct sysinfo si;
 898         bool is_os_64 = (sizeof(void *) == 8) ? true : false;
 899         uint64_t total_memory;
 900         uint64_t dram_size_seven_GB = 0x1B8000000;
 901         uint64_t dram_size_three_GB = 0xB8000000;
 902
 903         if (amdgpu_smu_memory_pool_size == 0)
 904                 return;
 905
 906         if (!is_os_64) {
 907                 DRM_WARN("Not 64-bit OS, feature not supported\n");
 908                 goto def_value;
 909         }
 910         si_meminfo(&si);
 911         total_memory = (uint64_t)si.totalram * si.mem_unit;
 912
 913         if ((amdgpu_smu_memory_pool_size == 1) ||
 914                 (amdgpu_smu_memory_pool_size == 2)) {
 915                 if (total_memory < dram_size_three_GB)
 916                         goto def_value1;
 917         } else if ((amdgpu_smu_memory_pool_size == 4) ||
 918                 (amdgpu_smu_memory_pool_size == 8)) {
 919                 if (total_memory < dram_size_seven_GB)
 920                         goto def_value1;
 921         } else {
 922                 DRM_WARN("Smu memory pool size not supported\n");
 923                 goto def_value;
 924         }
 925         adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
 926
 927         return;
 928
 929 def_value1:
 930         DRM_WARN("No enough system memory\n");
 931 def_value:
 932         adev->pm.smu_prv_buffer_size = 0;
 933 }
 934
 935 /**
 936  * amdgpu_device_check_arguments - validate module params
 937  *
 938  * @adev: amdgpu_device pointer
 939  *
 940  * Validates certain module parameters and updates
 941  * the associated values used by the driver (all asics).
 942  */
 943 static int amdgpu_device_check_arguments(struct amdgpu_device *adev)
 944 {
 945         int ret = 0;
 946
 947         if (amdgpu_sched_jobs < 4) {
 948                 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
 949                          amdgpu_sched_jobs);
 950                 amdgpu_sched_jobs = 4;
 951         } else if (!is_power_of_2(amdgpu_sched_jobs)){
 952                 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
 953                          amdgpu_sched_jobs);
 954                 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
 955         }
 956
 957         if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
 958                 /* gart size must be greater or equal to 32M */
 959                 dev_warn(adev->dev, "gart size (%d) too small\n",
 960                          amdgpu_gart_size);
 961                 amdgpu_gart_size = -1;
 962         }
 963
 964         if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
 965                 /* gtt size must be greater or equal to 32M */
 966                 dev_warn(adev->dev, "gtt size (%d) too small\n",
 967                                  amdgpu_gtt_size);
 968                 amdgpu_gtt_size = -1;
 969         }
 970
 971         /* valid range is between 4 and 9 inclusive */
 972         if (amdgpu_vm_fragment_size != -1 &&
 973             (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
 974                 dev_warn(adev->dev, "valid range is between 4 and 9\n");
 975                 amdgpu_vm_fragment_size = -1;
 976         }
 977
 978         amdgpu_device_check_smu_prv_buffer_size(adev);
 979
 980         amdgpu_device_check_vm_size(adev);
 981
 982         amdgpu_device_check_block_size(adev);
 983
 984         ret = amdgpu_device_get_job_timeout_settings(adev);
 985         if (ret) {
 986                 dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n");
 987                 return ret;
 988         }
 989
 990         adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
 991
 992         return ret;
 993 }
 994
 995 /**
 996  * amdgpu_switcheroo_set_state - set switcheroo state
 997  *
 998  * @pdev: pci dev pointer
 999  * @state: vga_switcheroo state
1000  *
1001  * Callback for the switcheroo driver.  Suspends or resumes the
1002  * the asics before or after it is powered up using ACPI methods.
1003  */
1004 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
1005 {
1006         struct drm_device *dev = pci_get_drvdata(pdev);
1007
1008         if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
1009                 return;
1010
1011         if (state == VGA_SWITCHEROO_ON) {
1012                 pr_info("amdgpu: switched on\n");
1013                 /* don't suspend or resume card normally */
1014                 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1015
1016                 amdgpu_device_resume(dev, true, true);
1017
1018                 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1019                 drm_kms_helper_poll_enable(dev);
1020         } else {
1021                 pr_info("amdgpu: switched off\n");
1022                 drm_kms_helper_poll_disable(dev);
1023                 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1024                 amdgpu_device_suspend(dev, true, true);
1025                 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1026         }
1027 }
1028
1029 /**
1030  * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1031  *
1032  * @pdev: pci dev pointer
1033  *
1034  * Callback for the switcheroo driver.  Check of the switcheroo
1035  * state can be changed.
1036  * Returns true if the state can be changed, false if not.
1037  */
1038 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1039 {
1040         struct drm_device *dev = pci_get_drvdata(pdev);
1041
1042         /*
1043         * FIXME: open_count is protected by drm_global_mutex but that would lead to
1044         * locking inversion with the driver load path. And the access here is
1045         * completely racy anyway. So don't bother with locking for now.
1046         */
1047         return dev->open_count == 0;
1048 }
1049
1050 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1051         .set_gpu_state = amdgpu_switcheroo_set_state,
1052         .reprobe = NULL,
1053         .can_switch = amdgpu_switcheroo_can_switch,
1054 };
1055
1056 /**
1057  * amdgpu_device_ip_set_clockgating_state - set the CG state
1058  *
1059  * @dev: amdgpu_device pointer
1060  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1061  * @state: clockgating state (gate or ungate)
1062  *
1063  * Sets the requested clockgating state for all instances of
1064  * the hardware IP specified.
1065  * Returns the error code from the last instance.
1066  */
1067 int amdgpu_device_ip_set_clockgating_state(void *dev,
1068                                            enum amd_ip_block_type block_type,
1069                                            enum amd_clockgating_state state)
1070 {
1071         struct amdgpu_device *adev = dev;
1072         int i, r = 0;
1073
1074         for (i = 0; i < adev->num_ip_blocks; i++) {
1075                 if (!adev->ip_blocks[i].status.valid)
1076                         continue;
1077                 if (adev->ip_blocks[i].version->type != block_type)
1078                         continue;
1079                 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1080                         continue;
1081                 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1082                         (void *)adev, state);
1083                 if (r)
1084                         DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1085                                   adev->ip_blocks[i].version->funcs->name, r);
1086         }
1087         return r;
1088 }
1089
1090 /**
1091  * amdgpu_device_ip_set_powergating_state - set the PG state
1092  *
1093  * @dev: amdgpu_device pointer
1094  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1095  * @state: powergating state (gate or ungate)
1096  *
1097  * Sets the requested powergating state for all instances of
1098  * the hardware IP specified.
1099  * Returns the error code from the last instance.
1100  */
1101 int amdgpu_device_ip_set_powergating_state(void *dev,
1102                                            enum amd_ip_block_type block_type,
1103                                            enum amd_powergating_state state)
1104 {
1105         struct amdgpu_device *adev = dev;
1106         int i, r = 0;
1107
1108         for (i = 0; i < adev->num_ip_blocks; i++) {
1109                 if (!adev->ip_blocks[i].status.valid)
1110                         continue;
1111                 if (adev->ip_blocks[i].version->type != block_type)
1112                         continue;
1113                 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1114                         continue;
1115                 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1116                         (void *)adev, state);
1117                 if (r)
1118                         DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1119                                   adev->ip_blocks[i].version->funcs->name, r);
1120         }
1121         return r;
1122 }
1123
1124 /**
1125  * amdgpu_device_ip_get_clockgating_state - get the CG state
1126  *
1127  * @adev: amdgpu_device pointer
1128  * @flags: clockgating feature flags
1129  *
1130  * Walks the list of IPs on the device and updates the clockgating
1131  * flags for each IP.
1132  * Updates @flags with the feature flags for each hardware IP where
1133  * clockgating is enabled.
1134  */
1135 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1136                                             u32 *flags)
1137 {
1138         int i;
1139
1140         for (i = 0; i < adev->num_ip_blocks; i++) {
1141                 if (!adev->ip_blocks[i].status.valid)
1142                         continue;
1143                 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1144                         adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1145         }
1146 }
1147
1148 /**
1149  * amdgpu_device_ip_wait_for_idle - wait for idle
1150  *
1151  * @adev: amdgpu_device pointer
1152  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1153  *
1154  * Waits for the request hardware IP to be idle.
1155  * Returns 0 for success or a negative error code on failure.
1156  */
1157 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1158                                    enum amd_ip_block_type block_type)
1159 {
1160         int i, r;
1161
1162         for (i = 0; i < adev->num_ip_blocks; i++) {
1163                 if (!adev->ip_blocks[i].status.valid)
1164                         continue;
1165                 if (adev->ip_blocks[i].version->type == block_type) {
1166                         r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1167                         if (r)
1168                                 return r;
1169                         break;
1170                 }
1171         }
1172         return 0;
1173
1174 }
1175
1176 /**
1177  * amdgpu_device_ip_is_idle - is the hardware IP idle
1178  *
1179  * @adev: amdgpu_device pointer
1180  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1181  *
1182  * Check if the hardware IP is idle or not.
1183  * Returns true if it the IP is idle, false if not.
1184  */
1185 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1186                               enum amd_ip_block_type block_type)
1187 {
1188         int i;
1189
1190         for (i = 0; i < adev->num_ip_blocks; i++) {
1191                 if (!adev->ip_blocks[i].status.valid)
1192                         continue;
1193                 if (adev->ip_blocks[i].version->type == block_type)
1194                         return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1195         }
1196         return true;
1197
1198 }
1199
1200 /**
1201  * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1202  *
1203  * @adev: amdgpu_device pointer
1204  * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
1205  *
1206  * Returns a pointer to the hardware IP block structure
1207  * if it exists for the asic, otherwise NULL.
1208  */
1209 struct amdgpu_ip_block *
1210 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1211                               enum amd_ip_block_type type)
1212 {
1213         int i;
1214
1215         for (i = 0; i < adev->num_ip_blocks; i++)
1216                 if (adev->ip_blocks[i].version->type == type)
1217                         return &adev->ip_blocks[i];
1218
1219         return NULL;
1220 }
1221
1222 /**
1223  * amdgpu_device_ip_block_version_cmp
1224  *
1225  * @adev: amdgpu_device pointer
1226  * @type: enum amd_ip_block_type
1227  * @major: major version
1228  * @minor: minor version
1229  *
1230  * return 0 if equal or greater
1231  * return 1 if smaller or the ip_block doesn't exist
1232  */
1233 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1234                                        enum amd_ip_block_type type,
1235                                        u32 major, u32 minor)
1236 {
1237         struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1238
1239         if (ip_block && ((ip_block->version->major > major) ||
1240                         ((ip_block->version->major == major) &&
1241                         (ip_block->version->minor >= minor))))
1242                 return 0;
1243
1244         return 1;
1245 }
1246
1247 /**
1248  * amdgpu_device_ip_block_add
1249  *
1250  * @adev: amdgpu_device pointer
1251  * @ip_block_version: pointer to the IP to add
1252  *
1253  * Adds the IP block driver information to the collection of IPs
1254  * on the asic.
1255  */
1256 int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1257                                const struct amdgpu_ip_block_version *ip_block_version)
1258 {
1259         if (!ip_block_version)
1260                 return -EINVAL;
1261
1262         DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1263                   ip_block_version->funcs->name);
1264
1265         adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1266
1267         return 0;
1268 }
1269
1270 /**
1271  * amdgpu_device_enable_virtual_display - enable virtual display feature
1272  *
1273  * @adev: amdgpu_device pointer
1274  *
1275  * Enabled the virtual display feature if the user has enabled it via
1276  * the module parameter virtual_display.  This feature provides a virtual
1277  * display hardware on headless boards or in virtualized environments.
1278  * This function parses and validates the configuration string specified by
1279  * the user and configues the virtual display configuration (number of
1280  * virtual connectors, crtcs, etc.) specified.
1281  */
1282 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1283 {
1284         adev->enable_virtual_display = false;
1285
1286         if (amdgpu_virtual_display) {
1287                 struct drm_device *ddev = adev->ddev;
1288                 const char *pci_address_name = pci_name(ddev->pdev);
1289                 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1290
1291                 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1292                 pciaddstr_tmp = pciaddstr;
1293                 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1294                         pciaddname = strsep(&pciaddname_tmp, ",");
1295                         if (!strcmp("all", pciaddname)
1296                             || !strcmp(pci_address_name, pciaddname)) {
1297                                 long num_crtc;
1298                                 int res = -1;
1299
1300                                 adev->enable_virtual_display = true;
1301
1302                                 if (pciaddname_tmp)
1303                                         res = kstrtol(pciaddname_tmp, 10,
1304                                                       &num_crtc);
1305
1306                                 if (!res) {
1307                                         if (num_crtc < 1)
1308                                                 num_crtc = 1;
1309                                         if (num_crtc > 6)
1310                                                 num_crtc = 6;
1311                                         adev->mode_info.num_crtc = num_crtc;
1312                                 } else {
1313                                         adev->mode_info.num_crtc = 1;
1314                                 }
1315                                 break;
1316                         }
1317                 }
1318
1319                 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1320                          amdgpu_virtual_display, pci_address_name,
1321                          adev->enable_virtual_display, adev->mode_info.num_crtc);
1322
1323                 kfree(pciaddstr);
1324         }
1325 }
1326
1327 /**
1328  * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1329  *
1330  * @adev: amdgpu_device pointer
1331  *
1332  * Parses the asic configuration parameters specified in the gpu info
1333  * firmware and makes them availale to the driver for use in configuring
1334  * the asic.
1335  * Returns 0 on success, -EINVAL on failure.
1336  */
1337 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1338 {
1339         const char *chip_name;
1340         char fw_name[30];
1341         int err;
1342         const struct gpu_info_firmware_header_v1_0 *hdr;
1343
1344         adev->firmware.gpu_info_fw = NULL;
1345
1346         switch (adev->asic_type) {
1347         case CHIP_TOPAZ:
1348         case CHIP_TONGA:
1349         case CHIP_FIJI:
1350         case CHIP_POLARIS10:
1351         case CHIP_POLARIS11:
1352         case CHIP_POLARIS12:
1353         case CHIP_VEGAM:
1354         case CHIP_CARRIZO:
1355         case CHIP_STONEY:
1356 #ifdef CONFIG_DRM_AMDGPU_SI
1357         case CHIP_VERDE:
1358         case CHIP_TAHITI:
1359         case CHIP_PITCAIRN:
1360         case CHIP_OLAND:
1361         case CHIP_HAINAN:
1362 #endif
1363 #ifdef CONFIG_DRM_AMDGPU_CIK
1364         case CHIP_BONAIRE:
1365         case CHIP_HAWAII:
1366         case CHIP_KAVERI:
1367         case CHIP_KABINI:
1368         case CHIP_MULLINS:
1369 #endif
1370         case CHIP_VEGA20:
1371         default:
1372                 return 0;
1373         case CHIP_VEGA10:
1374                 chip_name = "vega10";
1375                 break;
1376         case CHIP_VEGA12:
1377                 chip_name = "vega12";
1378                 break;
1379         case CHIP_RAVEN:
1380                 if (adev->rev_id >= 8)
1381                         chip_name = "raven2";
1382                 else if (adev->pdev->device == 0x15d8)
1383                         chip_name = "picasso";
1384                 else
1385                         chip_name = "raven";
1386                 break;
1387         case CHIP_NAVI10:
1388                 chip_name = "navi10";
1389                 break;
1390         }
1391
1392         snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1393         err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1394         if (err) {
1395                 dev_err(adev->dev,
1396                         "Failed to load gpu_info firmware \"%s\"\n",
1397                         fw_name);
1398                 goto out;
1399         }
1400         err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1401         if (err) {
1402                 dev_err(adev->dev,
1403                         "Failed to validate gpu_info firmware \"%s\"\n",
1404                         fw_name);
1405                 goto out;
1406         }
1407
1408         hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1409         amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1410
1411         switch (hdr->version_major) {
1412         case 1:
1413         {
1414                 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1415                         (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1416                                                                 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1417
1418                 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1419                 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1420                 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1421                 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1422                 adev->gfx.config.max_texture_channel_caches =
1423                         le32_to_cpu(gpu_info_fw->gc_num_tccs);
1424                 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1425                 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1426                 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1427                 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1428                 adev->gfx.config.double_offchip_lds_buf =
1429                         le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1430                 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1431                 adev->gfx.cu_info.max_waves_per_simd =
1432                         le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1433                 adev->gfx.cu_info.max_scratch_slots_per_cu =
1434                         le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1435                 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
1436                 if (hdr->version_minor >= 1) {
1437                         const struct gpu_info_firmware_v1_1 *gpu_info_fw =
1438                                 (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data +
1439                                                                         le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1440                         adev->gfx.config.num_sc_per_sh =
1441                                 le32_to_cpu(gpu_info_fw->num_sc_per_sh);
1442                         adev->gfx.config.num_packer_per_sc =
1443                                 le32_to_cpu(gpu_info_fw->num_packer_per_sc);
1444                 }
1445 #ifdef CONFIG_DRM_AMD_DC_DCN2_0
1446                 if (hdr->version_minor == 2) {
1447                         const struct gpu_info_firmware_v1_2 *gpu_info_fw =
1448                                 (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data +
1449                                                                         le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1450                         adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box;
1451                 }
1452 #endif
1453                 break;
1454         }
1455         default:
1456                 dev_err(adev->dev,
1457                         "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1458                 err = -EINVAL;
1459                 goto out;
1460         }
1461 out:
1462         return err;
1463 }
1464
1465 /**
1466  * amdgpu_device_ip_early_init - run early init for hardware IPs
1467  *
1468  * @adev: amdgpu_device pointer
1469  *
1470  * Early initialization pass for hardware IPs.  The hardware IPs that make
1471  * up each asic are discovered each IP's early_init callback is run.  This
1472  * is the first stage in initializing the asic.
1473  * Returns 0 on success, negative error code on failure.
1474  */
1475 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
1476 {
1477         int i, r;
1478
1479         amdgpu_device_enable_virtual_display(adev);
1480
1481         switch (adev->asic_type) {
1482         case CHIP_TOPAZ:
1483         case CHIP_TONGA:
1484         case CHIP_FIJI:
1485         case CHIP_POLARIS10:
1486         case CHIP_POLARIS11:
1487         case CHIP_POLARIS12:
1488         case CHIP_VEGAM:
1489         case CHIP_CARRIZO:
1490         case CHIP_STONEY:
1491                 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1492                         adev->family = AMDGPU_FAMILY_CZ;
1493                 else
1494                         adev->family = AMDGPU_FAMILY_VI;
1495
1496                 r = vi_set_ip_blocks(adev);
1497                 if (r)
1498                         return r;
1499                 break;
1500 #ifdef CONFIG_DRM_AMDGPU_SI
1501         case CHIP_VERDE:
1502         case CHIP_TAHITI:
1503         case CHIP_PITCAIRN:
1504         case CHIP_OLAND:
1505         case CHIP_HAINAN:
1506                 adev->family = AMDGPU_FAMILY_SI;
1507                 r = si_set_ip_blocks(adev);
1508                 if (r)
1509                         return r;
1510                 break;
1511 #endif
1512 #ifdef CONFIG_DRM_AMDGPU_CIK
1513         case CHIP_BONAIRE:
1514         case CHIP_HAWAII:
1515         case CHIP_KAVERI:
1516         case CHIP_KABINI:
1517         case CHIP_MULLINS:
1518                 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1519                         adev->family = AMDGPU_FAMILY_CI;
1520                 else
1521                         adev->family = AMDGPU_FAMILY_KV;
1522
1523                 r = cik_set_ip_blocks(adev);
1524                 if (r)
1525                         return r;
1526                 break;
1527 #endif
1528         case CHIP_VEGA10:
1529         case CHIP_VEGA12:
1530         case CHIP_VEGA20:
1531         case CHIP_RAVEN:
1532                 if (adev->asic_type == CHIP_RAVEN)
1533                         adev->family = AMDGPU_FAMILY_RV;
1534                 else
1535                         adev->family = AMDGPU_FAMILY_AI;
1536
1537                 r = soc15_set_ip_blocks(adev);
1538                 if (r)
1539                         return r;
1540                 break;
1541         case  CHIP_NAVI10:
1542                 adev->family = AMDGPU_FAMILY_NV;
1543
1544                 r = nv_set_ip_blocks(adev);
1545                 if (r)
1546                         return r;
1547                 break;
1548         default:
1549                 /* FIXME: not supported yet */
1550                 return -EINVAL;
1551         }
1552
1553         r = amdgpu_device_parse_gpu_info_fw(adev);
1554         if (r)
1555                 return r;
1556
1557         amdgpu_amdkfd_device_probe(adev);
1558
1559         if (amdgpu_sriov_vf(adev)) {
1560                 r = amdgpu_virt_request_full_gpu(adev, true);
1561                 if (r)
1562                         return -EAGAIN;
1563
1564                 /* query the reg access mode at the very beginning */
1565                 amdgpu_virt_init_reg_access_mode(adev);
1566         }
1567
1568         adev->pm.pp_feature = amdgpu_pp_feature_mask;
1569         if (amdgpu_sriov_vf(adev))
1570                 adev->pm.pp_feature &= ~PP_GFXOFF_MASK;
1571
1572         for (i = 0; i < adev->num_ip_blocks; i++) {
1573                 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1574                         DRM_ERROR("disabled ip block: %d <%s>\n",
1575                                   i, adev->ip_blocks[i].version->funcs->name);
1576                         adev->ip_blocks[i].status.valid = false;
1577                 } else {
1578                         if (adev->ip_blocks[i].version->funcs->early_init) {
1579                                 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1580                                 if (r == -ENOENT) {
1581                                         adev->ip_blocks[i].status.valid = false;
1582                                 } else if (r) {
1583                                         DRM_ERROR("early_init of IP block <%s> failed %d\n",
1584                                                   adev->ip_blocks[i].version->funcs->name, r);
1585                                         return r;
1586                                 } else {
1587                                         adev->ip_blocks[i].status.valid = true;
1588                                 }
1589                         } else {
1590                                 adev->ip_blocks[i].status.valid = true;
1591                         }
1592                 }
1593                 /* get the vbios after the asic_funcs are set up */
1594                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) {
1595                         /* Read BIOS */
1596                         if (!amdgpu_get_bios(adev))
1597                                 return -EINVAL;
1598
1599                         r = amdgpu_atombios_init(adev);
1600                         if (r) {
1601                                 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
1602                                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
1603                                 return r;
1604                         }
1605                 }
1606         }
1607
1608         adev->cg_flags &= amdgpu_cg_mask;
1609         adev->pg_flags &= amdgpu_pg_mask;
1610
1611         return 0;
1612 }
1613
1614 static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev)
1615 {
1616         int i, r;
1617
1618         for (i = 0; i < adev->num_ip_blocks; i++) {
1619                 if (!adev->ip_blocks[i].status.sw)
1620                         continue;
1621                 if (adev->ip_blocks[i].status.hw)
1622                         continue;
1623                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
1624                     (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) ||
1625                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
1626                         r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1627                         if (r) {
1628                                 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1629                                           adev->ip_blocks[i].version->funcs->name, r);
1630                                 return r;
1631                         }
1632                         adev->ip_blocks[i].status.hw = true;
1633                 }
1634         }
1635
1636         return 0;
1637 }
1638
1639 static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev)
1640 {
1641         int i, r;
1642
1643         for (i = 0; i < adev->num_ip_blocks; i++) {
1644                 if (!adev->ip_blocks[i].status.sw)
1645                         continue;
1646                 if (adev->ip_blocks[i].status.hw)
1647                         continue;
1648                 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1649                 if (r) {
1650                         DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1651                                   adev->ip_blocks[i].version->funcs->name, r);
1652                         return r;
1653                 }
1654                 adev->ip_blocks[i].status.hw = true;
1655         }
1656
1657         return 0;
1658 }
1659
1660 static int amdgpu_device_fw_loading(struct amdgpu_device *adev)
1661 {
1662         int r = 0;
1663         int i;
1664         uint32_t smu_version;
1665
1666         if (adev->asic_type >= CHIP_VEGA10) {
1667                 for (i = 0; i < adev->num_ip_blocks; i++) {
1668                         if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
1669                                 if (adev->in_gpu_reset || adev->in_suspend) {
1670                                         if (amdgpu_sriov_vf(adev) && adev->in_gpu_reset)
1671                                                 break; /* sriov gpu reset, psp need to do hw_init before IH because of hw limit */
1672                                         r = adev->ip_blocks[i].version->funcs->resume(adev);
1673                                         if (r) {
1674                                                 DRM_ERROR("resume of IP block <%s> failed %d\n",
1675                                                           adev->ip_blocks[i].version->funcs->name, r);
1676                                                 return r;
1677                                         }
1678                                 } else {
1679                                         r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1680                                         if (r) {
1681                                                 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1682                                                   adev->ip_blocks[i].version->funcs->name, r);
1683                                                 return r;
1684                                         }
1685                                 }
1686                                 adev->ip_blocks[i].status.hw = true;
1687                         }
1688                 }
1689         }
1690         r = amdgpu_pm_load_smu_firmware(adev, &smu_version);
1691
1692         return r;
1693 }
1694
1695 /**
1696  * amdgpu_device_ip_init - run init for hardware IPs
1697  *
1698  * @adev: amdgpu_device pointer
1699  *
1700  * Main initialization pass for hardware IPs.  The list of all the hardware
1701  * IPs that make up the asic is walked and the sw_init and hw_init callbacks
1702  * are run.  sw_init initializes the software state associated with each IP
1703  * and hw_init initializes the hardware associated with each IP.
1704  * Returns 0 on success, negative error code on failure.
1705  */
1706 static int amdgpu_device_ip_init(struct amdgpu_device *adev)
1707 {
1708         int i, r;
1709
1710         r = amdgpu_ras_init(adev);
1711         if (r)
1712                 return r;
1713
1714         for (i = 0; i < adev->num_ip_blocks; i++) {
1715                 if (!adev->ip_blocks[i].status.valid)
1716                         continue;
1717                 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1718                 if (r) {
1719                         DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1720                                   adev->ip_blocks[i].version->funcs->name, r);
1721                         goto init_failed;
1722                 }
1723                 adev->ip_blocks[i].status.sw = true;
1724
1725                 /* need to do gmc hw init early so we can allocate gpu mem */
1726                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1727                         r = amdgpu_device_vram_scratch_init(adev);
1728                         if (r) {
1729                                 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1730                                 goto init_failed;
1731                         }
1732                         r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1733                         if (r) {
1734                                 DRM_ERROR("hw_init %d failed %d\n", i, r);
1735                                 goto init_failed;
1736                         }
1737                         r = amdgpu_device_wb_init(adev);
1738                         if (r) {
1739                                 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
1740                                 goto init_failed;
1741                         }
1742                         adev->ip_blocks[i].status.hw = true;
1743
1744                         /* right after GMC hw init, we create CSA */
1745                         if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
1746                                 r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj,
1747                                                                 AMDGPU_GEM_DOMAIN_VRAM,
1748                                                                 AMDGPU_CSA_SIZE);
1749                                 if (r) {
1750                                         DRM_ERROR("allocate CSA failed %d\n", r);
1751                                         goto init_failed;
1752                                 }
1753                         }
1754                 }
1755         }
1756
1757         r = amdgpu_ib_pool_init(adev);
1758         if (r) {
1759                 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
1760                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
1761                 goto init_failed;
1762         }
1763
1764         r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/
1765         if (r)
1766                 goto init_failed;
1767
1768         r = amdgpu_device_ip_hw_init_phase1(adev);
1769         if (r)
1770                 goto init_failed;
1771
1772         r = amdgpu_device_fw_loading(adev);
1773         if (r)
1774                 goto init_failed;
1775
1776         r = amdgpu_device_ip_hw_init_phase2(adev);
1777         if (r)
1778                 goto init_failed;
1779
1780         if (adev->gmc.xgmi.num_physical_nodes > 1)
1781                 amdgpu_xgmi_add_device(adev);
1782         amdgpu_amdkfd_device_init(adev);
1783
1784 init_failed:
1785         if (amdgpu_sriov_vf(adev)) {
1786                 if (!r)
1787                         amdgpu_virt_init_data_exchange(adev);
1788                 amdgpu_virt_release_full_gpu(adev, true);
1789         }
1790
1791         return r;
1792 }
1793
1794 /**
1795  * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
1796  *
1797  * @adev: amdgpu_device pointer
1798  *
1799  * Writes a reset magic value to the gart pointer in VRAM.  The driver calls
1800  * this function before a GPU reset.  If the value is retained after a
1801  * GPU reset, VRAM has not been lost.  Some GPU resets may destry VRAM contents.
1802  */
1803 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
1804 {
1805         memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1806 }
1807
1808 /**
1809  * amdgpu_device_check_vram_lost - check if vram is valid
1810  *
1811  * @adev: amdgpu_device pointer
1812  *
1813  * Checks the reset magic value written to the gart pointer in VRAM.
1814  * The driver calls this after a GPU reset to see if the contents of
1815  * VRAM is lost or now.
1816  * returns true if vram is lost, false if not.
1817  */
1818 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
1819 {
1820         return !!memcmp(adev->gart.ptr, adev->reset_magic,
1821                         AMDGPU_RESET_MAGIC_NUM);
1822 }
1823
1824 /**
1825  * amdgpu_device_set_cg_state - set clockgating for amdgpu device
1826  *
1827  * @adev: amdgpu_device pointer
1828  *
1829  * The list of all the hardware IPs that make up the asic is walked and the
1830  * set_clockgating_state callbacks are run.
1831  * Late initialization pass enabling clockgating for hardware IPs.
1832  * Fini or suspend, pass disabling clockgating for hardware IPs.
1833  * Returns 0 on success, negative error code on failure.
1834  */
1835
1836 static int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
1837                                                 enum amd_clockgating_state state)
1838 {
1839         int i, j, r;
1840
1841         if (amdgpu_emu_mode == 1)
1842                 return 0;
1843
1844         for (j = 0; j < adev->num_ip_blocks; j++) {
1845                 i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
1846                 if (!adev->ip_blocks[i].status.late_initialized)
1847                         continue;
1848                 /* skip CG for VCE/UVD, it's handled specially */
1849                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1850                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1851                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1852                     adev->ip_blocks[i].version->funcs->set_clockgating_state) {
1853                         /* enable clockgating to save power */
1854                         r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1855                                                                                      state);
1856                         if (r) {
1857                                 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
1858                                           adev->ip_blocks[i].version->funcs->name, r);
1859                                 return r;
1860                         }
1861                 }
1862         }
1863
1864         return 0;
1865 }
1866
1867 static int amdgpu_device_set_pg_state(struct amdgpu_device *adev, enum amd_powergating_state state)
1868 {
1869         int i, j, r;
1870
1871         if (amdgpu_emu_mode == 1)
1872                 return 0;
1873
1874         for (j = 0; j < adev->num_ip_blocks; j++) {
1875                 i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
1876                 if (!adev->ip_blocks[i].status.late_initialized)
1877                         continue;
1878                 /* skip CG for VCE/UVD, it's handled specially */
1879                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1880                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
1881                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
1882                     adev->ip_blocks[i].version->funcs->set_powergating_state) {
1883                         /* enable powergating to save power */
1884                         r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
1885                                                                                         state);
1886                         if (r) {
1887                                 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
1888                                           adev->ip_blocks[i].version->funcs->name, r);
1889                                 return r;
1890                         }
1891                 }
1892         }
1893         return 0;
1894 }
1895
1896 static int amdgpu_device_enable_mgpu_fan_boost(void)
1897 {
1898         struct amdgpu_gpu_instance *gpu_ins;
1899         struct amdgpu_device *adev;
1900         int i, ret = 0;
1901
1902         mutex_lock(&mgpu_info.mutex);
1903
1904         /*
1905          * MGPU fan boost feature should be enabled
1906          * only when there are two or more dGPUs in
1907          * the system
1908          */
1909         if (mgpu_info.num_dgpu < 2)
1910                 goto out;
1911
1912         for (i = 0; i < mgpu_info.num_dgpu; i++) {
1913                 gpu_ins = &(mgpu_info.gpu_ins[i]);
1914                 adev = gpu_ins->adev;
1915                 if (!(adev->flags & AMD_IS_APU) &&
1916                     !gpu_ins->mgpu_fan_enabled &&
1917                     adev->powerplay.pp_funcs &&
1918                     adev->powerplay.pp_funcs->enable_mgpu_fan_boost) {
1919                         ret = amdgpu_dpm_enable_mgpu_fan_boost(adev);
1920                         if (ret)
1921                                 break;
1922
1923                         gpu_ins->mgpu_fan_enabled = 1;
1924                 }
1925         }
1926
1927 out:
1928         mutex_unlock(&mgpu_info.mutex);
1929
1930         return ret;
1931 }
1932
1933 /**
1934  * amdgpu_device_ip_late_init - run late init for hardware IPs
1935  *
1936  * @adev: amdgpu_device pointer
1937  *
1938  * Late initialization pass for hardware IPs.  The list of all the hardware
1939  * IPs that make up the asic is walked and the late_init callbacks are run.
1940  * late_init covers any special initialization that an IP requires
1941  * after all of the have been initialized or something that needs to happen
1942  * late in the init process.
1943  * Returns 0 on success, negative error code on failure.
1944  */
1945 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
1946 {
1947         int i = 0, r;
1948
1949         for (i = 0; i < adev->num_ip_blocks; i++) {
1950                 if (!adev->ip_blocks[i].status.hw)
1951                         continue;
1952                 if (adev->ip_blocks[i].version->funcs->late_init) {
1953                         r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
1954                         if (r) {
1955                                 DRM_ERROR("late_init of IP block <%s> failed %d\n",
1956                                           adev->ip_blocks[i].version->funcs->name, r);
1957                                 return r;
1958                         }
1959                 }
1960                 adev->ip_blocks[i].status.late_initialized = true;
1961         }
1962
1963         amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
1964         amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
1965
1966         amdgpu_device_fill_reset_magic(adev);
1967
1968         r = amdgpu_device_enable_mgpu_fan_boost();
1969         if (r)
1970                 DRM_ERROR("enable mgpu fan boost failed (%d).\n", r);
1971
1972         /* set to low pstate by default */
1973         amdgpu_xgmi_set_pstate(adev, 0);
1974
1975         return 0;
1976 }
1977
1978 /**
1979  * amdgpu_device_ip_fini - run fini for hardware IPs
1980  *
1981  * @adev: amdgpu_device pointer
1982  *
1983  * Main teardown pass for hardware IPs.  The list of all the hardware
1984  * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
1985  * are run.  hw_fini tears down the hardware associated with each IP
1986  * and sw_fini tears down any software state associated with each IP.
1987  * Returns 0 on success, negative error code on failure.
1988  */
1989 static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
1990 {
1991         int i, r;
1992
1993         amdgpu_ras_pre_fini(adev);
1994
1995         if (adev->gmc.xgmi.num_physical_nodes > 1)
1996                 amdgpu_xgmi_remove_device(adev);
1997
1998         amdgpu_amdkfd_device_fini(adev);
1999
2000         amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2001         amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2002
2003         /* need to disable SMC first */
2004         for (i = 0; i < adev->num_ip_blocks; i++) {
2005                 if (!adev->ip_blocks[i].status.hw)
2006                         continue;
2007                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2008                         r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2009                         /* XXX handle errors */
2010                         if (r) {
2011                                 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2012                                           adev->ip_blocks[i].version->funcs->name, r);
2013                         }
2014                         adev->ip_blocks[i].status.hw = false;
2015                         break;
2016                 }
2017         }
2018
2019         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2020                 if (!adev->ip_blocks[i].status.hw)
2021                         continue;
2022
2023                 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2024                 /* XXX handle errors */
2025                 if (r) {
2026                         DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2027                                   adev->ip_blocks[i].version->funcs->name, r);
2028                 }
2029
2030                 adev->ip_blocks[i].status.hw = false;
2031         }
2032
2033
2034         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2035                 if (!adev->ip_blocks[i].status.sw)
2036                         continue;
2037
2038                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
2039                         amdgpu_ucode_free_bo(adev);
2040                         amdgpu_free_static_csa(&adev->virt.csa_obj);
2041                         amdgpu_device_wb_fini(adev);
2042                         amdgpu_device_vram_scratch_fini(adev);
2043                         amdgpu_ib_pool_fini(adev);
2044                 }
2045
2046                 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
2047                 /* XXX handle errors */
2048                 if (r) {
2049                         DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
2050                                   adev->ip_blocks[i].version->funcs->name, r);
2051                 }
2052                 adev->ip_blocks[i].status.sw = false;
2053                 adev->ip_blocks[i].status.valid = false;
2054         }
2055
2056         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2057                 if (!adev->ip_blocks[i].status.late_initialized)
2058                         continue;
2059                 if (adev->ip_blocks[i].version->funcs->late_fini)
2060                         adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
2061                 adev->ip_blocks[i].status.late_initialized = false;
2062         }
2063
2064         amdgpu_ras_fini(adev);
2065
2066         if (amdgpu_sriov_vf(adev))
2067                 if (amdgpu_virt_release_full_gpu(adev, false))
2068                         DRM_ERROR("failed to release exclusive mode on fini\n");
2069
2070         return 0;
2071 }
2072
2073 /**
2074  * amdgpu_device_delayed_init_work_handler - work handler for IB tests
2075  *
2076  * @work: work_struct.
2077  */
2078 static void amdgpu_device_delayed_init_work_handler(struct work_struct *work)
2079 {
2080         struct amdgpu_device *adev =
2081                 container_of(work, struct amdgpu_device, delayed_init_work.work);
2082         int r;
2083
2084         r = amdgpu_ib_ring_tests(adev);
2085         if (r)
2086                 DRM_ERROR("ib ring test failed (%d).\n", r);
2087 }
2088
2089 static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
2090 {
2091         struct amdgpu_device *adev =
2092                 container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
2093
2094         mutex_lock(&adev->gfx.gfx_off_mutex);
2095         if (!adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
2096                 if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
2097                         adev->gfx.gfx_off_state = true;
2098         }
2099         mutex_unlock(&adev->gfx.gfx_off_mutex);
2100 }
2101
2102 /**
2103  * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
2104  *
2105  * @adev: amdgpu_device pointer
2106  *
2107  * Main suspend function for hardware IPs.  The list of all the hardware
2108  * IPs that make up the asic is walked, clockgating is disabled and the
2109  * suspend callbacks are run.  suspend puts the hardware and software state
2110  * in each IP into a state suitable for suspend.
2111  * Returns 0 on success, negative error code on failure.
2112  */
2113 static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
2114 {
2115         int i, r;
2116
2117         amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2118         amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2119
2120         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2121                 if (!adev->ip_blocks[i].status.valid)
2122                         continue;
2123                 /* displays are handled separately */
2124                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) {
2125                         /* XXX handle errors */
2126                         r = adev->ip_blocks[i].version->funcs->suspend(adev);
2127                         /* XXX handle errors */
2128                         if (r) {
2129                                 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2130                                           adev->ip_blocks[i].version->funcs->name, r);
2131                         }
2132                 }
2133         }
2134
2135         return 0;
2136 }
2137
2138 /**
2139  * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
2140  *
2141  * @adev: amdgpu_device pointer
2142  *
2143  * Main suspend function for hardware IPs.  The list of all the hardware
2144  * IPs that make up the asic is walked, clockgating is disabled and the
2145  * suspend callbacks are run.  suspend puts the hardware and software state
2146  * in each IP into a state suitable for suspend.
2147  * Returns 0 on success, negative error code on failure.
2148  */
2149 static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
2150 {
2151         int i, r;
2152
2153         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2154                 if (!adev->ip_blocks[i].status.valid)
2155                         continue;
2156                 /* displays are handled in phase1 */
2157                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
2158                         continue;
2159                 /* XXX handle errors */
2160                 r = adev->ip_blocks[i].version->funcs->suspend(adev);
2161                 /* XXX handle errors */
2162                 if (r) {
2163                         DRM_ERROR("suspend of IP block <%s> failed %d\n",
2164                                   adev->ip_blocks[i].version->funcs->name, r);
2165                 }
2166         }
2167
2168         return 0;
2169 }
2170
2171 /**
2172  * amdgpu_device_ip_suspend - run suspend for hardware IPs
2173  *
2174  * @adev: amdgpu_device pointer
2175  *
2176  * Main suspend function for hardware IPs.  The list of all the hardware
2177  * IPs that make up the asic is walked, clockgating is disabled and the
2178  * suspend callbacks are run.  suspend puts the hardware and software state
2179  * in each IP into a state suitable for suspend.
2180  * Returns 0 on success, negative error code on failure.
2181  */
2182 int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
2183 {
2184         int r;
2185
2186         if (amdgpu_sriov_vf(adev))
2187                 amdgpu_virt_request_full_gpu(adev, false);
2188
2189         r = amdgpu_device_ip_suspend_phase1(adev);
2190         if (r)
2191                 return r;
2192         r = amdgpu_device_ip_suspend_phase2(adev);
2193
2194         if (amdgpu_sriov_vf(adev))
2195                 amdgpu_virt_release_full_gpu(adev, false);
2196
2197         return r;
2198 }
2199
2200 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
2201 {
2202         int i, r;
2203
2204         static enum amd_ip_block_type ip_order[] = {
2205                 AMD_IP_BLOCK_TYPE_GMC,
2206                 AMD_IP_BLOCK_TYPE_COMMON,
2207                 AMD_IP_BLOCK_TYPE_PSP,
2208                 AMD_IP_BLOCK_TYPE_IH,
2209         };
2210
2211         for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2212                 int j;
2213                 struct amdgpu_ip_block *block;
2214
2215                 for (j = 0; j < adev->num_ip_blocks; j++) {
2216                         block = &adev->ip_blocks[j];
2217
2218                         if (block->version->type != ip_order[i] ||
2219                                 !block->status.valid)
2220                                 continue;
2221
2222                         r = block->version->funcs->hw_init(adev);
2223                         DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2224                         if (r)
2225                                 return r;
2226                 }
2227         }
2228
2229         return 0;
2230 }
2231
2232 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
2233 {
2234         int i, r;
2235
2236         static enum amd_ip_block_type ip_order[] = {
2237                 AMD_IP_BLOCK_TYPE_SMC,
2238                 AMD_IP_BLOCK_TYPE_DCE,
2239                 AMD_IP_BLOCK_TYPE_GFX,
2240                 AMD_IP_BLOCK_TYPE_SDMA,
2241                 AMD_IP_BLOCK_TYPE_UVD,
2242                 AMD_IP_BLOCK_TYPE_VCE
2243         };
2244
2245         for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2246                 int j;
2247                 struct amdgpu_ip_block *block;
2248
2249                 for (j = 0; j < adev->num_ip_blocks; j++) {
2250                         block = &adev->ip_blocks[j];
2251
2252                         if (block->version->type != ip_order[i] ||
2253                                 !block->status.valid)
2254                                 continue;
2255
2256                         r = block->version->funcs->hw_init(adev);
2257                         DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2258                         if (r)
2259                                 return r;
2260                 }
2261         }
2262
2263         return 0;
2264 }
2265
2266 /**
2267  * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
2268  *
2269  * @adev: amdgpu_device pointer
2270  *
2271  * First resume function for hardware IPs.  The list of all the hardware
2272  * IPs that make up the asic is walked and the resume callbacks are run for
2273  * COMMON, GMC, and IH.  resume puts the hardware into a functional state
2274  * after a suspend and updates the software state as necessary.  This
2275  * function is also used for restoring the GPU after a GPU reset.
2276  * Returns 0 on success, negative error code on failure.
2277  */
2278 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
2279 {
2280         int i, r;
2281
2282         for (i = 0; i < adev->num_ip_blocks; i++) {
2283                 if (!adev->ip_blocks[i].status.valid)
2284                         continue;
2285                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2286                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2287                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2288                         r = adev->ip_blocks[i].version->funcs->resume(adev);
2289                         if (r) {
2290                                 DRM_ERROR("resume of IP block <%s> failed %d\n",
2291                                           adev->ip_blocks[i].version->funcs->name, r);
2292                                 return r;
2293                         }
2294                 }
2295         }
2296
2297         return 0;
2298 }
2299
2300 /**
2301  * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
2302  *
2303  * @adev: amdgpu_device pointer
2304  *
2305  * First resume function for hardware IPs.  The list of all the hardware
2306  * IPs that make up the asic is walked and the resume callbacks are run for
2307  * all blocks except COMMON, GMC, and IH.  resume puts the hardware into a
2308  * functional state after a suspend and updates the software state as
2309  * necessary.  This function is also used for restoring the GPU after a GPU
2310  * reset.
2311  * Returns 0 on success, negative error code on failure.
2312  */
2313 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
2314 {
2315         int i, r;
2316
2317         for (i = 0; i < adev->num_ip_blocks; i++) {
2318                 if (!adev->ip_blocks[i].status.valid)
2319                         continue;
2320                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2321                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2322                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
2323                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
2324                         continue;
2325                 r = adev->ip_blocks[i].version->funcs->resume(adev);
2326                 if (r) {
2327                         DRM_ERROR("resume of IP block <%s> failed %d\n",
2328                                   adev->ip_blocks[i].version->funcs->name, r);
2329                         return r;
2330                 }
2331         }
2332
2333         return 0;
2334 }
2335
2336 /**
2337  * amdgpu_device_ip_resume - run resume for hardware IPs
2338  *
2339  * @adev: amdgpu_device pointer
2340  *
2341  * Main resume function for hardware IPs.  The hardware IPs
2342  * are split into two resume functions because they are
2343  * are also used in in recovering from a GPU reset and some additional
2344  * steps need to be take between them.  In this case (S3/S4) they are
2345  * run sequentially.
2346  * Returns 0 on success, negative error code on failure.
2347  */
2348 static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
2349 {
2350         int r;
2351
2352         r = amdgpu_device_ip_resume_phase1(adev);
2353         if (r)
2354                 return r;
2355
2356         r = amdgpu_device_fw_loading(adev);
2357         if (r)
2358                 return r;
2359
2360         r = amdgpu_device_ip_resume_phase2(adev);
2361
2362         return r;
2363 }
2364
2365 /**
2366  * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
2367  *
2368  * @adev: amdgpu_device pointer
2369  *
2370  * Query the VBIOS data tables to determine if the board supports SR-IOV.
2371  */
2372 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
2373 {
2374         if (amdgpu_sriov_vf(adev)) {
2375                 if (adev->is_atom_fw) {
2376                         if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2377                                 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2378                 } else {
2379                         if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2380                                 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2381                 }
2382
2383                 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2384                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
2385         }
2386 }
2387
2388 /**
2389  * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
2390  *
2391  * @asic_type: AMD asic type
2392  *
2393  * Check if there is DC (new modesetting infrastructre) support for an asic.
2394  * returns true if DC has support, false if not.
2395  */
2396 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2397 {
2398         switch (asic_type) {
2399 #if defined(CONFIG_DRM_AMD_DC)
2400         case CHIP_BONAIRE:
2401         case CHIP_KAVERI:
2402         case CHIP_KABINI:
2403         case CHIP_MULLINS:
2404                 /*
2405                  * We have systems in the wild with these ASICs that require
2406                  * LVDS and VGA support which is not supported with DC.
2407                  *
2408                  * Fallback to the non-DC driver here by default so as not to
2409                  * cause regressions.
2410                  */
2411                 return amdgpu_dc > 0;
2412         case CHIP_HAWAII:
2413         case CHIP_CARRIZO:
2414         case CHIP_STONEY:
2415         case CHIP_POLARIS10:
2416         case CHIP_POLARIS11:
2417         case CHIP_POLARIS12:
2418         case CHIP_VEGAM:
2419         case CHIP_TONGA:
2420         case CHIP_FIJI:
2421         case CHIP_VEGA10:
2422         case CHIP_VEGA12:
2423         case CHIP_VEGA20:
2424 #if defined(CONFIG_DRM_AMD_DC_DCN1_0)
2425         case CHIP_RAVEN:
2426 #endif
2427 #if defined(CONFIG_DRM_AMD_DC_DCN2_0)
2428         case CHIP_NAVI10:
2429 #endif
2430                 return amdgpu_dc != 0;
2431 #endif
2432         default:
2433                 return false;
2434         }
2435 }
2436
2437 /**
2438  * amdgpu_device_has_dc_support - check if dc is supported
2439  *
2440  * @adev: amdgpu_device_pointer
2441  *
2442  * Returns true for supported, false for not supported
2443  */
2444 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2445 {
2446         if (amdgpu_sriov_vf(adev))
2447                 return false;
2448
2449         return amdgpu_device_asic_has_dc_support(adev->asic_type);
2450 }
2451
2452
2453 static void amdgpu_device_xgmi_reset_func(struct work_struct *__work)
2454 {
2455         struct amdgpu_device *adev =
2456                 container_of(__work, struct amdgpu_device, xgmi_reset_work);
2457
2458         adev->asic_reset_res =  amdgpu_asic_reset(adev);
2459         if (adev->asic_reset_res)
2460                 DRM_WARN("ASIC reset failed with error, %d for drm dev, %s",
2461                          adev->asic_reset_res, adev->ddev->unique);
2462 }
2463
2464
2465 /**
2466  * amdgpu_device_init - initialize the driver
2467  *
2468  * @adev: amdgpu_device pointer
2469  * @ddev: drm dev pointer
2470  * @pdev: pci dev pointer
2471  * @flags: driver flags
2472  *
2473  * Initializes the driver info and hw (all asics).
2474  * Returns 0 for success or an error on failure.
2475  * Called at driver startup.
2476  */
2477 int amdgpu_device_init(struct amdgpu_device *adev,
2478                        struct drm_device *ddev,
2479                        struct pci_dev *pdev,
2480                        uint32_t flags)
2481 {
2482         int r, i;
2483         bool runtime = false;
2484         u32 max_MBps;
2485
2486         adev->shutdown = false;
2487         adev->dev = &pdev->dev;
2488         adev->ddev = ddev;
2489         adev->pdev = pdev;
2490         adev->flags = flags;
2491         adev->asic_type = flags & AMD_ASIC_MASK;
2492         adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
2493         if (amdgpu_emu_mode == 1)
2494                 adev->usec_timeout *= 2;
2495         adev->gmc.gart_size = 512 * 1024 * 1024;
2496         adev->accel_working = false;
2497         adev->num_rings = 0;
2498         adev->mman.buffer_funcs = NULL;
2499         adev->mman.buffer_funcs_ring = NULL;
2500         adev->vm_manager.vm_pte_funcs = NULL;
2501         adev->vm_manager.vm_pte_num_rqs = 0;
2502         adev->gmc.gmc_funcs = NULL;
2503         adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2504         bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
2505
2506         adev->smc_rreg = &amdgpu_invalid_rreg;
2507         adev->smc_wreg = &amdgpu_invalid_wreg;
2508         adev->pcie_rreg = &amdgpu_invalid_rreg;
2509         adev->pcie_wreg = &amdgpu_invalid_wreg;
2510         adev->pciep_rreg = &amdgpu_invalid_rreg;
2511         adev->pciep_wreg = &amdgpu_invalid_wreg;
2512         adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
2513         adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
2514         adev->didt_rreg = &amdgpu_invalid_rreg;
2515         adev->didt_wreg = &amdgpu_invalid_wreg;
2516         adev->gc_cac_rreg = &amdgpu_invalid_rreg;
2517         adev->gc_cac_wreg = &amdgpu_invalid_wreg;
2518         adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
2519         adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
2520
2521         DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
2522                  amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
2523                  pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
2524
2525         /* mutex initialization are all done here so we
2526          * can recall function without having locking issues */
2527         atomic_set(&adev->irq.ih.lock, 0);
2528         mutex_init(&adev->firmware.mutex);
2529         mutex_init(&adev->pm.mutex);
2530         mutex_init(&adev->gfx.gpu_clock_mutex);
2531         mutex_init(&adev->srbm_mutex);
2532         mutex_init(&adev->gfx.pipe_reserve_mutex);
2533         mutex_init(&adev->gfx.gfx_off_mutex);
2534         mutex_init(&adev->grbm_idx_mutex);
2535         mutex_init(&adev->mn_lock);
2536         mutex_init(&adev->virt.vf_errors.lock);
2537         hash_init(adev->mn_hash);
2538         mutex_init(&adev->lock_reset);
2539         mutex_init(&adev->virt.dpm_mutex);
2540         mutex_init(&adev->psp.mutex);
2541
2542         r = amdgpu_device_check_arguments(adev);
2543         if (r)
2544                 return r;
2545
2546         spin_lock_init(&adev->mmio_idx_lock);
2547         spin_lock_init(&adev->smc_idx_lock);
2548         spin_lock_init(&adev->pcie_idx_lock);
2549         spin_lock_init(&adev->uvd_ctx_idx_lock);
2550         spin_lock_init(&adev->didt_idx_lock);
2551         spin_lock_init(&adev->gc_cac_idx_lock);
2552         spin_lock_init(&adev->se_cac_idx_lock);
2553         spin_lock_init(&adev->audio_endpt_idx_lock);
2554         spin_lock_init(&adev->mm_stats.lock);
2555
2556         INIT_LIST_HEAD(&adev->shadow_list);
2557         mutex_init(&adev->shadow_list_lock);
2558
2559         INIT_LIST_HEAD(&adev->ring_lru_list);
2560         spin_lock_init(&adev->ring_lru_list_lock);
2561
2562         INIT_DELAYED_WORK(&adev->delayed_init_work,
2563                           amdgpu_device_delayed_init_work_handler);
2564         INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
2565                           amdgpu_device_delay_enable_gfx_off);
2566
2567         INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func);
2568
2569         adev->gfx.gfx_off_req_count = 1;
2570         adev->pm.ac_power = power_supply_is_system_supplied() > 0 ? true : false;
2571
2572         /* Registers mapping */
2573         /* TODO: block userspace mapping of io register */
2574         if (adev->asic_type >= CHIP_BONAIRE) {
2575                 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
2576                 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
2577         } else {
2578                 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
2579                 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
2580         }
2581
2582         adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
2583         if (adev->rmmio == NULL) {
2584                 return -ENOMEM;
2585         }
2586         DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
2587         DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
2588
2589         /* io port mapping */
2590         for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
2591                 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
2592                         adev->rio_mem_size = pci_resource_len(adev->pdev, i);
2593                         adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
2594                         break;
2595                 }
2596         }
2597         if (adev->rio_mem == NULL)
2598                 DRM_INFO("PCI I/O BAR is not found.\n");
2599
2600         /* enable PCIE atomic ops */
2601         r = pci_enable_atomic_ops_to_root(adev->pdev,
2602                                           PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
2603                                           PCI_EXP_DEVCAP2_ATOMIC_COMP64);
2604         if (r) {
2605                 adev->have_atomics_support = false;
2606                 DRM_INFO("PCIE atomic ops is not supported\n");
2607         } else {
2608                 adev->have_atomics_support = true;
2609         }
2610
2611         amdgpu_device_get_pcie_info(adev);
2612
2613         if (amdgpu_mcbp)
2614                 DRM_INFO("MCBP is enabled\n");
2615
2616         if (amdgpu_mes && adev->asic_type >= CHIP_NAVI10)
2617                 adev->enable_mes = true;
2618
2619         if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10) {
2620                 r = amdgpu_discovery_init(adev);
2621                 if (r) {
2622                         dev_err(adev->dev, "amdgpu_discovery_init failed\n");
2623                         return r;
2624                 }
2625         }
2626
2627         /* early init functions */
2628         r = amdgpu_device_ip_early_init(adev);
2629         if (r)
2630                 return r;
2631
2632         /* doorbell bar mapping and doorbell index init*/
2633         amdgpu_device_doorbell_init(adev);
2634
2635         /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
2636         /* this will fail for cards that aren't VGA class devices, just
2637          * ignore it */
2638         vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
2639
2640         if (amdgpu_device_is_px(ddev))
2641                 runtime = true;
2642         if (!pci_is_thunderbolt_attached(adev->pdev))
2643                 vga_switcheroo_register_client(adev->pdev,
2644                                                &amdgpu_switcheroo_ops, runtime);
2645         if (runtime)
2646                 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
2647
2648         if (amdgpu_emu_mode == 1) {
2649                 /* post the asic on emulation mode */
2650                 emu_soc_asic_init(adev);
2651                 goto fence_driver_init;
2652         }
2653
2654         /* detect if we are with an SRIOV vbios */
2655         amdgpu_device_detect_sriov_bios(adev);
2656
2657         /* check if we need to reset the asic
2658          *  E.g., driver was not cleanly unloaded previously, etc.
2659          */
2660         if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) {
2661                 r = amdgpu_asic_reset(adev);
2662                 if (r) {
2663                         dev_err(adev->dev, "asic reset on init failed\n");
2664                         goto failed;
2665                 }
2666         }
2667
2668         /* Post card if necessary */
2669         if (amdgpu_device_need_post(adev)) {
2670                 if (!adev->bios) {
2671                         dev_err(adev->dev, "no vBIOS found\n");
2672                         r = -EINVAL;
2673                         goto failed;
2674                 }
2675                 DRM_INFO("GPU posting now...\n");
2676                 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2677                 if (r) {
2678                         dev_err(adev->dev, "gpu post error!\n");
2679                         goto failed;
2680                 }
2681         }
2682
2683         if (adev->is_atom_fw) {
2684                 /* Initialize clocks */
2685                 r = amdgpu_atomfirmware_get_clock_info(adev);
2686                 if (r) {
2687                         dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
2688                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2689                         goto failed;
2690                 }
2691         } else {
2692                 /* Initialize clocks */
2693                 r = amdgpu_atombios_get_clock_info(adev);
2694                 if (r) {
2695                         dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
2696                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2697                         goto failed;
2698                 }
2699                 /* init i2c buses */
2700                 if (!amdgpu_device_has_dc_support(adev))
2701                         amdgpu_atombios_i2c_init(adev);
2702         }
2703
2704 fence_driver_init:
2705         /* Fence driver */
2706         r = amdgpu_fence_driver_init(adev);
2707         if (r) {
2708                 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
2709                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
2710                 goto failed;
2711         }
2712
2713         /* init the mode config */
2714         drm_mode_config_init(adev->ddev);
2715
2716         r = amdgpu_device_ip_init(adev);
2717         if (r) {
2718                 /* failed in exclusive mode due to timeout */
2719                 if (amdgpu_sriov_vf(adev) &&
2720                     !amdgpu_sriov_runtime(adev) &&
2721                     amdgpu_virt_mmio_blocked(adev) &&
2722                     !amdgpu_virt_wait_reset(adev)) {
2723                         dev_err(adev->dev, "VF exclusive mode timeout\n");
2724                         /* Don't send request since VF is inactive. */
2725                         adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
2726                         adev->virt.ops = NULL;
2727                         r = -EAGAIN;
2728                         goto failed;
2729                 }
2730                 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
2731                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
2732                 if (amdgpu_virt_request_full_gpu(adev, false))
2733                         amdgpu_virt_release_full_gpu(adev, false);
2734                 goto failed;
2735         }
2736
2737         adev->accel_working = true;
2738
2739         amdgpu_vm_check_compute_bug(adev);
2740
2741         /* Initialize the buffer migration limit. */
2742         if (amdgpu_moverate >= 0)
2743                 max_MBps = amdgpu_moverate;
2744         else
2745                 max_MBps = 8; /* Allow 8 MB/s. */
2746         /* Get a log2 for easy divisions. */
2747         adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
2748
2749         amdgpu_fbdev_init(adev);
2750
2751         if (amdgpu_sriov_vf(adev) && amdgim_is_hwperf(adev))
2752                 amdgpu_pm_virt_sysfs_init(adev);
2753
2754         r = amdgpu_pm_sysfs_init(adev);
2755         if (r)
2756                 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
2757
2758         r = amdgpu_ucode_sysfs_init(adev);
2759         if (r)
2760                 DRM_ERROR("Creating firmware sysfs failed (%d).\n", r);
2761
2762         r = amdgpu_debugfs_gem_init(adev);
2763         if (r)
2764                 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
2765
2766         r = amdgpu_debugfs_regs_init(adev);
2767         if (r)
2768                 DRM_ERROR("registering register debugfs failed (%d).\n", r);
2769
2770         r = amdgpu_debugfs_firmware_init(adev);
2771         if (r)
2772                 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
2773
2774         r = amdgpu_debugfs_init(adev);
2775         if (r)
2776                 DRM_ERROR("Creating debugfs files failed (%d).\n", r);
2777
2778         if ((amdgpu_testing & 1)) {
2779                 if (adev->accel_working)
2780                         amdgpu_test_moves(adev);
2781                 else
2782                         DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
2783         }
2784         if (amdgpu_benchmarking) {
2785                 if (adev->accel_working)
2786                         amdgpu_benchmark(adev, amdgpu_benchmarking);
2787                 else
2788                         DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
2789         }
2790
2791         /* enable clockgating, etc. after ib tests, etc. since some blocks require
2792          * explicit gating rather than handling it automatically.
2793          */
2794         r = amdgpu_device_ip_late_init(adev);
2795         if (r) {
2796                 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
2797                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
2798                 goto failed;
2799         }
2800
2801         /* must succeed. */
2802         amdgpu_ras_resume(adev);
2803
2804         queue_delayed_work(system_wq, &adev->delayed_init_work,
2805                            msecs_to_jiffies(AMDGPU_RESUME_MS));
2806
2807         r = device_create_file(adev->dev, &dev_attr_pcie_replay_count);
2808         if (r) {
2809                 dev_err(adev->dev, "Could not create pcie_replay_count");
2810                 return r;
2811         }
2812
2813         if (IS_ENABLED(CONFIG_PERF_EVENTS))
2814                 r = amdgpu_pmu_init(adev);
2815         if (r)
2816                 dev_err(adev->dev, "amdgpu_pmu_init failed\n");
2817
2818         return 0;
2819
2820 failed:
2821         amdgpu_vf_error_trans_all(adev);
2822         if (runtime)
2823                 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2824
2825         return r;
2826 }
2827
2828 /**
2829  * amdgpu_device_fini - tear down the driver
2830  *
2831  * @adev: amdgpu_device pointer
2832  *
2833  * Tear down the driver info (all asics).
2834  * Called at driver shutdown.
2835  */
2836 void amdgpu_device_fini(struct amdgpu_device *adev)
2837 {
2838         int r;
2839
2840         DRM_INFO("amdgpu: finishing device.\n");
2841         adev->shutdown = true;
2842         /* disable all interrupts */
2843         amdgpu_irq_disable_all(adev);
2844         if (adev->mode_info.mode_config_initialized){
2845                 if (!amdgpu_device_has_dc_support(adev))
2846                         drm_helper_force_disable_all(adev->ddev);
2847                 else
2848                         drm_atomic_helper_shutdown(adev->ddev);
2849         }
2850         amdgpu_fence_driver_fini(adev);
2851         amdgpu_pm_sysfs_fini(adev);
2852         amdgpu_fbdev_fini(adev);
2853         r = amdgpu_device_ip_fini(adev);
2854         if (adev->firmware.gpu_info_fw) {
2855                 release_firmware(adev->firmware.gpu_info_fw);
2856                 adev->firmware.gpu_info_fw = NULL;
2857         }
2858         adev->accel_working = false;
2859         cancel_delayed_work_sync(&adev->delayed_init_work);
2860         /* free i2c buses */
2861         if (!amdgpu_device_has_dc_support(adev))
2862                 amdgpu_i2c_fini(adev);
2863
2864         if (amdgpu_emu_mode != 1)
2865                 amdgpu_atombios_fini(adev);
2866
2867         kfree(adev->bios);
2868         adev->bios = NULL;
2869         if (!pci_is_thunderbolt_attached(adev->pdev))
2870                 vga_switcheroo_unregister_client(adev->pdev);
2871         if (adev->flags & AMD_IS_PX)
2872                 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2873         vga_client_register(adev->pdev, NULL, NULL, NULL);
2874         if (adev->rio_mem)
2875                 pci_iounmap(adev->pdev, adev->rio_mem);
2876         adev->rio_mem = NULL;
2877         iounmap(adev->rmmio);
2878         adev->rmmio = NULL;
2879         amdgpu_device_doorbell_fini(adev);
2880         if (amdgpu_sriov_vf(adev) && amdgim_is_hwperf(adev))
2881                 amdgpu_pm_virt_sysfs_fini(adev);
2882
2883         amdgpu_debugfs_regs_cleanup(adev);
2884         device_remove_file(adev->dev, &dev_attr_pcie_replay_count);
2885         amdgpu_ucode_sysfs_fini(adev);
2886         if (IS_ENABLED(CONFIG_PERF_EVENTS))
2887                 amdgpu_pmu_fini(adev);
2888         amdgpu_debugfs_preempt_cleanup(adev);
2889         if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10)
2890                 amdgpu_discovery_fini(adev);
2891 }
2892
2893
2894 /*
2895  * Suspend & resume.
2896  */
2897 /**
2898  * amdgpu_device_suspend - initiate device suspend
2899  *
2900  * @dev: drm dev pointer
2901  * @suspend: suspend state
2902  * @fbcon : notify the fbdev of suspend
2903  *
2904  * Puts the hw in the suspend state (all asics).
2905  * Returns 0 for success or an error on failure.
2906  * Called at driver suspend.
2907  */
2908 int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
2909 {
2910         struct amdgpu_device *adev;
2911         struct drm_crtc *crtc;
2912         struct drm_connector *connector;
2913         int r;
2914
2915         if (dev == NULL || dev->dev_private == NULL) {
2916                 return -ENODEV;
2917         }
2918
2919         adev = dev->dev_private;
2920
2921         if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2922                 return 0;
2923
2924         adev->in_suspend = true;
2925         drm_kms_helper_poll_disable(dev);
2926
2927         if (fbcon)
2928                 amdgpu_fbdev_set_suspend(adev, 1);
2929
2930         cancel_delayed_work_sync(&adev->delayed_init_work);
2931
2932         if (!amdgpu_device_has_dc_support(adev)) {
2933                 /* turn off display hw */
2934                 drm_modeset_lock_all(dev);
2935                 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2936                         drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
2937                 }
2938                 drm_modeset_unlock_all(dev);
2939                         /* unpin the front buffers and cursors */
2940                 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2941                         struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2942                         struct drm_framebuffer *fb = crtc->primary->fb;
2943                         struct amdgpu_bo *robj;
2944
2945                         if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) {
2946                                 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2947                                 r = amdgpu_bo_reserve(aobj, true);
2948                                 if (r == 0) {
2949                                         amdgpu_bo_unpin(aobj);
2950                                         amdgpu_bo_unreserve(aobj);
2951                                 }
2952                         }
2953
2954                         if (fb == NULL || fb->obj[0] == NULL) {
2955                                 continue;
2956                         }
2957                         robj = gem_to_amdgpu_bo(fb->obj[0]);
2958                         /* don't unpin kernel fb objects */
2959                         if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
2960                                 r = amdgpu_bo_reserve(robj, true);
2961                                 if (r == 0) {
2962                                         amdgpu_bo_unpin(robj);
2963                                         amdgpu_bo_unreserve(robj);
2964                                 }
2965                         }
2966                 }
2967         }
2968
2969         amdgpu_amdkfd_suspend(adev);
2970
2971         amdgpu_ras_suspend(adev);
2972
2973         r = amdgpu_device_ip_suspend_phase1(adev);
2974
2975         /* evict vram memory */
2976         amdgpu_bo_evict_vram(adev);
2977
2978         amdgpu_fence_driver_suspend(adev);
2979
2980         r = amdgpu_device_ip_suspend_phase2(adev);
2981
2982         /* evict remaining vram memory
2983          * This second call to evict vram is to evict the gart page table
2984          * using the CPU.
2985          */
2986         amdgpu_bo_evict_vram(adev);
2987
2988         pci_save_state(dev->pdev);
2989         if (suspend) {
2990                 /* Shut down the device */
2991                 pci_disable_device(dev->pdev);
2992                 pci_set_power_state(dev->pdev, PCI_D3hot);
2993         } else {
2994                 r = amdgpu_asic_reset(adev);
2995                 if (r)
2996                         DRM_ERROR("amdgpu asic reset failed\n");
2997         }
2998
2999         return 0;
3000 }
3001
3002 /**
3003  * amdgpu_device_resume - initiate device resume
3004  *
3005  * @dev: drm dev pointer
3006  * @resume: resume state
3007  * @fbcon : notify the fbdev of resume
3008  *
3009  * Bring the hw back to operating state (all asics).
3010  * Returns 0 for success or an error on failure.
3011  * Called at driver resume.
3012  */
3013 int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
3014 {
3015         struct drm_connector *connector;
3016         struct amdgpu_device *adev = dev->dev_private;
3017         struct drm_crtc *crtc;
3018         int r = 0;
3019
3020         if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3021                 return 0;
3022
3023         if (resume) {
3024                 pci_set_power_state(dev->pdev, PCI_D0);
3025                 pci_restore_state(dev->pdev);
3026                 r = pci_enable_device(dev->pdev);
3027                 if (r)
3028                         return r;
3029         }
3030
3031         /* post card */
3032         if (amdgpu_device_need_post(adev)) {
3033                 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
3034                 if (r)
3035                         DRM_ERROR("amdgpu asic init failed\n");
3036         }
3037
3038         r = amdgpu_device_ip_resume(adev);
3039         if (r) {
3040                 DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
3041                 return r;
3042         }
3043         amdgpu_fence_driver_resume(adev);
3044
3045
3046         r = amdgpu_device_ip_late_init(adev);
3047         if (r)
3048                 return r;
3049
3050         queue_delayed_work(system_wq, &adev->delayed_init_work,
3051                            msecs_to_jiffies(AMDGPU_RESUME_MS));
3052
3053         if (!amdgpu_device_has_dc_support(adev)) {
3054                 /* pin cursors */
3055                 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
3056                         struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
3057
3058                         if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) {
3059                                 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
3060                                 r = amdgpu_bo_reserve(aobj, true);
3061                                 if (r == 0) {
3062                                         r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM);
3063                                         if (r != 0)
3064                                                 DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
3065                                         amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj);
3066                                         amdgpu_bo_unreserve(aobj);
3067                                 }
3068                         }
3069                 }
3070         }
3071         r = amdgpu_amdkfd_resume(adev);
3072         if (r)
3073                 return r;
3074
3075         /* Make sure IB tests flushed */
3076         flush_delayed_work(&adev->delayed_init_work);
3077
3078         /* blat the mode back in */
3079         if (fbcon) {
3080                 if (!amdgpu_device_has_dc_support(adev)) {
3081                         /* pre DCE11 */
3082                         drm_helper_resume_force_mode(dev);
3083
3084                         /* turn on display hw */
3085                         drm_modeset_lock_all(dev);
3086                         list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
3087                                 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON);
3088                         }
3089                         drm_modeset_unlock_all(dev);
3090                 }
3091                 amdgpu_fbdev_set_suspend(adev, 0);
3092         }
3093
3094         drm_kms_helper_poll_enable(dev);
3095
3096         amdgpu_ras_resume(adev);
3097
3098         /*
3099          * Most of the connector probing functions try to acquire runtime pm
3100          * refs to ensure that the GPU is powered on when connector polling is
3101          * performed. Since we're calling this from a runtime PM callback,
3102          * trying to acquire rpm refs will cause us to deadlock.
3103          *
3104          * Since we're guaranteed to be holding the rpm lock, it's safe to
3105          * temporarily disable the rpm helpers so this doesn't deadlock us.
3106          */
3107 #ifdef CONFIG_PM
3108         dev->dev->power.disable_depth++;
3109 #endif
3110         if (!amdgpu_device_has_dc_support(adev))
3111                 drm_helper_hpd_irq_event(dev);
3112         else
3113                 drm_kms_helper_hotplug_event(dev);
3114 #ifdef CONFIG_PM
3115         dev->dev->power.disable_depth--;
3116 #endif
3117         adev->in_suspend = false;
3118
3119         return 0;
3120 }
3121
3122 /**
3123  * amdgpu_device_ip_check_soft_reset - did soft reset succeed
3124  *
3125  * @adev: amdgpu_device pointer
3126  *
3127  * The list of all the hardware IPs that make up the asic is walked and
3128  * the check_soft_reset callbacks are run.  check_soft_reset determines
3129  * if the asic is still hung or not.
3130  * Returns true if any of the IPs are still in a hung state, false if not.
3131  */
3132 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
3133 {
3134         int i;
3135         bool asic_hang = false;
3136
3137         if (amdgpu_sriov_vf(adev))
3138                 return true;
3139
3140         if (amdgpu_asic_need_full_reset(adev))
3141                 return true;
3142
3143         for (i = 0; i < adev->num_ip_blocks; i++) {
3144                 if (!adev->ip_blocks[i].status.valid)
3145                         continue;
3146                 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
3147                         adev->ip_blocks[i].status.hang =
3148                                 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
3149                 if (adev->ip_blocks[i].status.hang) {
3150                         DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
3151                         asic_hang = true;
3152                 }
3153         }
3154         return asic_hang;
3155 }
3156
3157 /**
3158  * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
3159  *
3160  * @adev: amdgpu_device pointer
3161  *
3162  * The list of all the hardware IPs that make up the asic is walked and the
3163  * pre_soft_reset callbacks are run if the block is hung.  pre_soft_reset
3164  * handles any IP specific hardware or software state changes that are
3165  * necessary for a soft reset to succeed.
3166  * Returns 0 on success, negative error code on failure.
3167  */
3168 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
3169 {
3170         int i, r = 0;
3171
3172         for (i = 0; i < adev->num_ip_blocks; i++) {
3173                 if (!adev->ip_blocks[i].status.valid)
3174                         continue;
3175                 if (adev->ip_blocks[i].status.hang &&
3176                     adev->ip_blocks[i].version->funcs->pre_soft_reset) {
3177                         r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
3178                         if (r)
3179                                 return r;
3180                 }
3181         }
3182
3183         return 0;
3184 }
3185
3186 /**
3187  * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
3188  *
3189  * @adev: amdgpu_device pointer
3190  *
3191  * Some hardware IPs cannot be soft reset.  If they are hung, a full gpu
3192  * reset is necessary to recover.
3193  * Returns true if a full asic reset is required, false if not.
3194  */
3195 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
3196 {
3197         int i;
3198
3199         if (amdgpu_asic_need_full_reset(adev))
3200                 return true;
3201
3202         for (i = 0; i < adev->num_ip_blocks; i++) {
3203                 if (!adev->ip_blocks[i].status.valid)
3204                         continue;
3205                 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
3206                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
3207                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
3208                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
3209                      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
3210                         if (adev->ip_blocks[i].status.hang) {
3211                                 DRM_INFO("Some block need full reset!\n");
3212                                 return true;
3213                         }
3214                 }
3215         }
3216         return false;
3217 }
3218
3219 /**
3220  * amdgpu_device_ip_soft_reset - do a soft reset
3221  *
3222  * @adev: amdgpu_device pointer
3223  *
3224  * The list of all the hardware IPs that make up the asic is walked and the
3225  * soft_reset callbacks are run if the block is hung.  soft_reset handles any
3226  * IP specific hardware or software state changes that are necessary to soft
3227  * reset the IP.
3228  * Returns 0 on success, negative error code on failure.
3229  */
3230 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
3231 {
3232         int i, r = 0;
3233
3234         for (i = 0; i < adev->num_ip_blocks; i++) {
3235                 if (!adev->ip_blocks[i].status.valid)
3236                         continue;
3237                 if (adev->ip_blocks[i].status.hang &&
3238                     adev->ip_blocks[i].version->funcs->soft_reset) {
3239                         r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
3240                         if (r)
3241                                 return r;
3242                 }
3243         }
3244
3245         return 0;
3246 }
3247
3248 /**
3249  * amdgpu_device_ip_post_soft_reset - clean up from soft reset
3250  *
3251  * @adev: amdgpu_device pointer
3252  *
3253  * The list of all the hardware IPs that make up the asic is walked and the
3254  * post_soft_reset callbacks are run if the asic was hung.  post_soft_reset
3255  * handles any IP specific hardware or software state changes that are
3256  * necessary after the IP has been soft reset.
3257  * Returns 0 on success, negative error code on failure.
3258  */
3259 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
3260 {
3261         int i, r = 0;
3262
3263         for (i = 0; i < adev->num_ip_blocks; i++) {
3264                 if (!adev->ip_blocks[i].status.valid)
3265                         continue;
3266                 if (adev->ip_blocks[i].status.hang &&
3267                     adev->ip_blocks[i].version->funcs->post_soft_reset)
3268                         r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
3269                 if (r)
3270                         return r;
3271         }
3272
3273         return 0;
3274 }
3275
3276 /**
3277  * amdgpu_device_recover_vram - Recover some VRAM contents
3278  *
3279  * @adev: amdgpu_device pointer
3280  *
3281  * Restores the contents of VRAM buffers from the shadows in GTT.  Used to
3282  * restore things like GPUVM page tables after a GPU reset where
3283  * the contents of VRAM might be lost.
3284  *
3285  * Returns:
3286  * 0 on success, negative error code on failure.
3287  */
3288 static int amdgpu_device_recover_vram(struct amdgpu_device *adev)
3289 {
3290         struct dma_fence *fence = NULL, *next = NULL;
3291         struct amdgpu_bo *shadow;
3292         long r = 1, tmo;
3293
3294         if (amdgpu_sriov_runtime(adev))
3295                 tmo = msecs_to_jiffies(8000);
3296         else
3297                 tmo = msecs_to_jiffies(100);
3298
3299         DRM_INFO("recover vram bo from shadow start\n");
3300         mutex_lock(&adev->shadow_list_lock);
3301         list_for_each_entry(shadow, &adev->shadow_list, shadow_list) {
3302
3303                 /* No need to recover an evicted BO */
3304                 if (shadow->tbo.mem.mem_type != TTM_PL_TT ||
3305                     shadow->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET ||
3306                     shadow->parent->tbo.mem.mem_type != TTM_PL_VRAM)
3307                         continue;
3308
3309                 r = amdgpu_bo_restore_shadow(shadow, &next);
3310                 if (r)
3311                         break;
3312
3313                 if (fence) {
3314                         tmo = dma_fence_wait_timeout(fence, false, tmo);
3315                         dma_fence_put(fence);
3316                         fence = next;
3317                         if (tmo == 0) {
3318                                 r = -ETIMEDOUT;
3319                                 break;
3320                         } else if (tmo < 0) {
3321                                 r = tmo;
3322                                 break;
3323                         }
3324                 } else {
3325                         fence = next;
3326                 }
3327         }
3328         mutex_unlock(&adev->shadow_list_lock);
3329
3330         if (fence)
3331                 tmo = dma_fence_wait_timeout(fence, false, tmo);
3332         dma_fence_put(fence);
3333
3334         if (r < 0 || tmo <= 0) {
3335                 DRM_ERROR("recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo);
3336                 return -EIO;
3337         }
3338
3339         DRM_INFO("recover vram bo from shadow done\n");
3340         return 0;
3341 }
3342
3343
3344 /**
3345  * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
3346  *
3347  * @adev: amdgpu device pointer
3348  * @from_hypervisor: request from hypervisor
3349  *
3350  * do VF FLR and reinitialize Asic
3351  * return 0 means succeeded otherwise failed
3352  */
3353 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
3354                                      bool from_hypervisor)
3355 {
3356         int r;
3357
3358         if (from_hypervisor)
3359                 r = amdgpu_virt_request_full_gpu(adev, true);
3360         else
3361                 r = amdgpu_virt_reset_gpu(adev);
3362         if (r)
3363                 return r;
3364
3365         amdgpu_amdkfd_pre_reset(adev);
3366
3367         /* Resume IP prior to SMC */
3368         r = amdgpu_device_ip_reinit_early_sriov(adev);
3369         if (r)
3370                 goto error;
3371
3372         /* we need recover gart prior to run SMC/CP/SDMA resume */
3373         amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
3374
3375         r = amdgpu_device_fw_loading(adev);
3376         if (r)
3377                 return r;
3378
3379         /* now we are okay to resume SMC/CP/SDMA */
3380         r = amdgpu_device_ip_reinit_late_sriov(adev);
3381         if (r)
3382                 goto error;
3383
3384         amdgpu_irq_gpu_reset_resume_helper(adev);
3385         r = amdgpu_ib_ring_tests(adev);
3386         amdgpu_amdkfd_post_reset(adev);
3387
3388 error:
3389         amdgpu_virt_init_data_exchange(adev);
3390         amdgpu_virt_release_full_gpu(adev, true);
3391         if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
3392                 atomic_inc(&adev->vram_lost_counter);
3393                 r = amdgpu_device_recover_vram(adev);
3394         }
3395
3396         return r;
3397 }
3398
3399 /**
3400  * amdgpu_device_should_recover_gpu - check if we should try GPU recovery
3401  *
3402  * @adev: amdgpu device pointer
3403  *
3404  * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover
3405  * a hung GPU.
3406  */
3407 bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev)
3408 {
3409         if (!amdgpu_device_ip_check_soft_reset(adev)) {
3410                 DRM_INFO("Timeout, but no hardware hang detected.\n");
3411                 return false;
3412         }
3413
3414         if (amdgpu_gpu_recovery == 0)
3415                 goto disabled;
3416
3417         if (amdgpu_sriov_vf(adev))
3418                 return true;
3419
3420         if (amdgpu_gpu_recovery == -1) {
3421                 switch (adev->asic_type) {
3422                 case CHIP_BONAIRE:
3423                 case CHIP_HAWAII:
3424                 case CHIP_TOPAZ:
3425                 case CHIP_TONGA:
3426                 case CHIP_FIJI:
3427                 case CHIP_POLARIS10:
3428                 case CHIP_POLARIS11:
3429                 case CHIP_POLARIS12:
3430                 case CHIP_VEGAM:
3431                 case CHIP_VEGA20:
3432                 case CHIP_VEGA10:
3433                 case CHIP_VEGA12:
3434                         break;
3435                 default:
3436                         goto disabled;
3437                 }
3438         }
3439
3440         return true;
3441
3442 disabled:
3443                 DRM_INFO("GPU recovery disabled.\n");
3444                 return false;
3445 }
3446
3447
3448 static int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
3449                                         struct amdgpu_job *job,
3450                                         bool *need_full_reset_arg)
3451 {
3452         int i, r = 0;
3453         bool need_full_reset  = *need_full_reset_arg;
3454
3455         /* block all schedulers and reset given job's ring */
3456         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3457                 struct amdgpu_ring *ring = adev->rings[i];
3458
3459                 if (!ring || !ring->sched.thread)
3460                         continue;
3461
3462                 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
3463                 amdgpu_fence_driver_force_completion(ring);
3464         }
3465
3466         if(job)
3467                 drm_sched_increase_karma(&job->base);
3468
3469         /* Don't suspend on bare metal if we are not going to HW reset the ASIC */
3470         if (!amdgpu_sriov_vf(adev)) {
3471
3472                 if (!need_full_reset)
3473                         need_full_reset = amdgpu_device_ip_need_full_reset(adev);
3474
3475                 if (!need_full_reset) {
3476                         amdgpu_device_ip_pre_soft_reset(adev);
3477                         r = amdgpu_device_ip_soft_reset(adev);
3478                         amdgpu_device_ip_post_soft_reset(adev);
3479                         if (r || amdgpu_device_ip_check_soft_reset(adev)) {
3480                                 DRM_INFO("soft reset failed, will fallback to full reset!\n");
3481                                 need_full_reset = true;
3482                         }
3483                 }
3484
3485                 if (need_full_reset)
3486                         r = amdgpu_device_ip_suspend(adev);
3487
3488                 *need_full_reset_arg = need_full_reset;
3489         }
3490
3491         return r;
3492 }
3493
3494 static int amdgpu_do_asic_reset(struct amdgpu_hive_info *hive,
3495                                struct list_head *device_list_handle,
3496                                bool *need_full_reset_arg)
3497 {
3498         struct amdgpu_device *tmp_adev = NULL;
3499         bool need_full_reset = *need_full_reset_arg, vram_lost = false;
3500         int r = 0;
3501
3502         /*
3503          * ASIC reset has to be done on all HGMI hive nodes ASAP
3504          * to allow proper links negotiation in FW (within 1 sec)
3505          */
3506         if (need_full_reset) {
3507                 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3508                         /* For XGMI run all resets in parallel to speed up the process */
3509                         if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
3510                                 if (!queue_work(system_highpri_wq, &tmp_adev->xgmi_reset_work))
3511                                         r = -EALREADY;
3512                         } else
3513                                 r = amdgpu_asic_reset(tmp_adev);
3514
3515                         if (r) {
3516                                 DRM_ERROR("ASIC reset failed with error, %d for drm dev, %s",
3517                                          r, tmp_adev->ddev->unique);
3518                                 break;
3519                         }
3520                 }
3521
3522                 /* For XGMI wait for all PSP resets to complete before proceed */
3523                 if (!r) {
3524                         list_for_each_entry(tmp_adev, device_list_handle,
3525                                             gmc.xgmi.head) {
3526                                 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
3527                                         flush_work(&tmp_adev->xgmi_reset_work);
3528                                         r = tmp_adev->asic_reset_res;
3529                                         if (r)
3530                                                 break;
3531                                 }
3532                         }
3533
3534                         list_for_each_entry(tmp_adev, device_list_handle,
3535                                         gmc.xgmi.head) {
3536                                 amdgpu_ras_reserve_bad_pages(tmp_adev);
3537                         }
3538                 }
3539         }
3540
3541
3542         list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3543                 if (need_full_reset) {
3544                         /* post card */
3545                         if (amdgpu_atom_asic_init(tmp_adev->mode_info.atom_context))
3546                                 DRM_WARN("asic atom init failed!");
3547
3548                         if (!r) {
3549                                 dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n");
3550                                 r = amdgpu_device_ip_resume_phase1(tmp_adev);
3551                                 if (r)
3552                                         goto out;
3553
3554                                 vram_lost = amdgpu_device_check_vram_lost(tmp_adev);
3555                                 if (vram_lost) {
3556                                         DRM_INFO("VRAM is lost due to GPU reset!\n");
3557                                         atomic_inc(&tmp_adev->vram_lost_counter);
3558                                 }
3559
3560                                 r = amdgpu_gtt_mgr_recover(
3561                                         &tmp_adev->mman.bdev.man[TTM_PL_TT]);
3562                                 if (r)
3563                                         goto out;
3564
3565                                 r = amdgpu_device_fw_loading(tmp_adev);
3566                                 if (r)
3567                                         return r;
3568
3569                                 r = amdgpu_device_ip_resume_phase2(tmp_adev);
3570                                 if (r)
3571                                         goto out;
3572
3573                                 if (vram_lost)
3574                                         amdgpu_device_fill_reset_magic(tmp_adev);
3575
3576                                 /*
3577                                  * Add this ASIC as tracked as reset was already
3578                                  * complete successfully.
3579                                  */
3580                                 amdgpu_register_gpu_instance(tmp_adev);
3581
3582                                 r = amdgpu_device_ip_late_init(tmp_adev);
3583                                 if (r)
3584                                         goto out;
3585
3586                                 /* must succeed. */
3587                                 amdgpu_ras_resume(tmp_adev);
3588
3589                                 /* Update PSP FW topology after reset */
3590                                 if (hive && tmp_adev->gmc.xgmi.num_physical_nodes > 1)
3591                                         r = amdgpu_xgmi_update_topology(hive, tmp_adev);
3592                         }
3593                 }
3594
3595
3596 out:
3597                 if (!r) {
3598                         amdgpu_irq_gpu_reset_resume_helper(tmp_adev);
3599                         r = amdgpu_ib_ring_tests(tmp_adev);
3600                         if (r) {
3601                                 dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r);
3602                                 r = amdgpu_device_ip_suspend(tmp_adev);
3603                                 need_full_reset = true;
3604                                 r = -EAGAIN;
3605                                 goto end;
3606                         }
3607                 }
3608
3609                 if (!r)
3610                         r = amdgpu_device_recover_vram(tmp_adev);
3611                 else
3612                         tmp_adev->asic_reset_res = r;
3613         }
3614
3615 end:
3616         *need_full_reset_arg = need_full_reset;
3617         return r;
3618 }
3619
3620 static bool amdgpu_device_lock_adev(struct amdgpu_device *adev, bool trylock)
3621 {
3622         if (trylock) {
3623                 if (!mutex_trylock(&adev->lock_reset))
3624                         return false;
3625         } else
3626                 mutex_lock(&adev->lock_reset);
3627
3628         atomic_inc(&adev->gpu_reset_counter);
3629         adev->in_gpu_reset = 1;
3630         /* Block kfd: SRIOV would do it separately */
3631         if (!amdgpu_sriov_vf(adev))
3632                 amdgpu_amdkfd_pre_reset(adev);
3633
3634         return true;
3635 }
3636
3637 static void amdgpu_device_unlock_adev(struct amdgpu_device *adev)
3638 {
3639         /*unlock kfd: SRIOV would do it separately */
3640         if (!amdgpu_sriov_vf(adev))
3641                 amdgpu_amdkfd_post_reset(adev);
3642         amdgpu_vf_error_trans_all(adev);
3643         adev->in_gpu_reset = 0;
3644         mutex_unlock(&adev->lock_reset);
3645 }
3646
3647
3648 /**
3649  * amdgpu_device_gpu_recover - reset the asic and recover scheduler
3650  *
3651  * @adev: amdgpu device pointer
3652  * @job: which job trigger hang
3653  *
3654  * Attempt to reset the GPU if it has hung (all asics).
3655  * Attempt to do soft-reset or full-reset and reinitialize Asic
3656  * Returns 0 for success or an error on failure.
3657  */
3658
3659 int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
3660                               struct amdgpu_job *job)
3661 {
3662         struct list_head device_list, *device_list_handle =  NULL;
3663         bool need_full_reset, job_signaled;
3664         struct amdgpu_hive_info *hive = NULL;
3665         struct amdgpu_device *tmp_adev = NULL;
3666         int i, r = 0;
3667
3668         need_full_reset = job_signaled = false;
3669         INIT_LIST_HEAD(&device_list);
3670
3671         dev_info(adev->dev, "GPU reset begin!\n");
3672
3673         cancel_delayed_work_sync(&adev->delayed_init_work);
3674
3675         hive = amdgpu_get_xgmi_hive(adev, false);
3676
3677         /*
3678          * Here we trylock to avoid chain of resets executing from
3679          * either trigger by jobs on different adevs in XGMI hive or jobs on
3680          * different schedulers for same device while this TO handler is running.
3681          * We always reset all schedulers for device and all devices for XGMI
3682          * hive so that should take care of them too.
3683          */
3684
3685         if (hive && !mutex_trylock(&hive->reset_lock)) {
3686                 DRM_INFO("Bailing on TDR for s_job:%llx, hive: %llx as another already in progress",
3687                          job->base.id, hive->hive_id);
3688                 return 0;
3689         }
3690
3691         /* Start with adev pre asic reset first for soft reset check.*/
3692         if (!amdgpu_device_lock_adev(adev, !hive)) {
3693                 DRM_INFO("Bailing on TDR for s_job:%llx, as another already in progress",
3694                                          job->base.id);
3695                 return 0;
3696         }
3697
3698         /* Build list of devices to reset */
3699         if  (adev->gmc.xgmi.num_physical_nodes > 1) {
3700                 if (!hive) {
3701                         amdgpu_device_unlock_adev(adev);
3702                         return -ENODEV;
3703                 }
3704
3705                 /*
3706                  * In case we are in XGMI hive mode device reset is done for all the
3707                  * nodes in the hive to retrain all XGMI links and hence the reset
3708                  * sequence is executed in loop on all nodes.
3709                  */
3710                 device_list_handle = &hive->device_list;
3711         } else {
3712                 list_add_tail(&adev->gmc.xgmi.head, &device_list);
3713                 device_list_handle = &device_list;
3714         }
3715
3716         /*
3717          * Mark these ASICs to be reseted as untracked first
3718          * And add them back after reset completed
3719          */
3720         list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head)
3721                 amdgpu_unregister_gpu_instance(tmp_adev);
3722
3723         /* block all schedulers and reset given job's ring */
3724         list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3725                 /* disable ras on ALL IPs */
3726                 if (amdgpu_device_ip_need_full_reset(tmp_adev))
3727                         amdgpu_ras_suspend(tmp_adev);
3728
3729                 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3730                         struct amdgpu_ring *ring = tmp_adev->rings[i];
3731
3732                         if (!ring || !ring->sched.thread)
3733                                 continue;
3734
3735                         drm_sched_stop(&ring->sched, &job->base);
3736                 }
3737         }
3738
3739
3740         /*
3741          * Must check guilty signal here since after this point all old
3742          * HW fences are force signaled.
3743          *
3744          * job->base holds a reference to parent fence
3745          */
3746         if (job && job->base.s_fence->parent &&
3747             dma_fence_is_signaled(job->base.s_fence->parent))
3748                 job_signaled = true;
3749
3750         if (!amdgpu_device_ip_need_full_reset(adev))
3751                 device_list_handle = &device_list;
3752
3753         if (job_signaled) {
3754                 dev_info(adev->dev, "Guilty job already signaled, skipping HW reset");
3755                 goto skip_hw_reset;
3756         }
3757
3758
3759         /* Guilty job will be freed after this*/
3760         r = amdgpu_device_pre_asic_reset(adev,
3761                                          job,
3762                                          &need_full_reset);
3763         if (r) {
3764                 /*TODO Should we stop ?*/
3765                 DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ",
3766                           r, adev->ddev->unique);
3767                 adev->asic_reset_res = r;
3768         }
3769
3770 retry:  /* Rest of adevs pre asic reset from XGMI hive. */
3771         list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3772
3773                 if (tmp_adev == adev)
3774                         continue;
3775
3776                 amdgpu_device_lock_adev(tmp_adev, false);
3777                 r = amdgpu_device_pre_asic_reset(tmp_adev,
3778                                                  NULL,
3779                                                  &need_full_reset);
3780                 /*TODO Should we stop ?*/
3781                 if (r) {
3782                         DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ",
3783                                   r, tmp_adev->ddev->unique);
3784                         tmp_adev->asic_reset_res = r;
3785                 }
3786         }
3787
3788         /* Actual ASIC resets if needed.*/
3789         /* TODO Implement XGMI hive reset logic for SRIOV */
3790         if (amdgpu_sriov_vf(adev)) {
3791                 r = amdgpu_device_reset_sriov(adev, job ? false : true);
3792                 if (r)
3793                         adev->asic_reset_res = r;
3794         } else {
3795                 r  = amdgpu_do_asic_reset(hive, device_list_handle, &need_full_reset);
3796                 if (r && r == -EAGAIN)
3797                         goto retry;
3798         }
3799
3800 skip_hw_reset:
3801
3802         /* Post ASIC reset for all devs .*/
3803         list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3804                 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3805                         struct amdgpu_ring *ring = tmp_adev->rings[i];
3806
3807                         if (!ring || !ring->sched.thread)
3808                                 continue;
3809
3810                         /* No point to resubmit jobs if we didn't HW reset*/
3811                         if (!tmp_adev->asic_reset_res && !job_signaled)
3812                                 drm_sched_resubmit_jobs(&ring->sched);
3813
3814                         drm_sched_start(&ring->sched, !tmp_adev->asic_reset_res);
3815                 }
3816
3817                 if (!amdgpu_device_has_dc_support(tmp_adev) && !job_signaled) {
3818                         drm_helper_resume_force_mode(tmp_adev->ddev);
3819                 }
3820
3821                 tmp_adev->asic_reset_res = 0;
3822
3823                 if (r) {
3824                         /* bad news, how to tell it to userspace ? */
3825                         dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&adev->gpu_reset_counter));
3826                         amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
3827                 } else {
3828                         dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&adev->gpu_reset_counter));
3829                 }
3830
3831                 amdgpu_device_unlock_adev(tmp_adev);
3832         }
3833
3834         if (hive)
3835                 mutex_unlock(&hive->reset_lock);
3836
3837         if (r)
3838                 dev_info(adev->dev, "GPU reset end with ret = %d\n", r);
3839         return r;
3840 }
3841
3842 /**
3843  * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
3844  *
3845  * @adev: amdgpu_device pointer
3846  *
3847  * Fetchs and stores in the driver the PCIE capabilities (gen speed
3848  * and lanes) of the slot the device is in. Handles APUs and
3849  * virtualized environments where PCIE config space may not be available.
3850  */
3851 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
3852 {
3853         struct pci_dev *pdev;
3854         enum pci_bus_speed speed_cap, platform_speed_cap;
3855         enum pcie_link_width platform_link_width;
3856
3857         if (amdgpu_pcie_gen_cap)
3858                 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
3859
3860         if (amdgpu_pcie_lane_cap)
3861                 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
3862
3863         /* covers APUs as well */
3864         if (pci_is_root_bus(adev->pdev->bus)) {
3865                 if (adev->pm.pcie_gen_mask == 0)
3866                         adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
3867                 if (adev->pm.pcie_mlw_mask == 0)
3868                         adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
3869                 return;
3870         }
3871
3872         if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask)
3873                 return;
3874
3875         pcie_bandwidth_available(adev->pdev, NULL,
3876                                  &platform_speed_cap, &platform_link_width);
3877
3878         if (adev->pm.pcie_gen_mask == 0) {
3879                 /* asic caps */
3880                 pdev = adev->pdev;
3881                 speed_cap = pcie_get_speed_cap(pdev);
3882                 if (speed_cap == PCI_SPEED_UNKNOWN) {
3883                         adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3884                                                   CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3885                                                   CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
3886                 } else {
3887                         if (speed_cap == PCIE_SPEED_16_0GT)
3888                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3889                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3890                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
3891                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
3892                         else if (speed_cap == PCIE_SPEED_8_0GT)
3893                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3894                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3895                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
3896                         else if (speed_cap == PCIE_SPEED_5_0GT)
3897                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3898                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
3899                         else
3900                                 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
3901                 }
3902                 /* platform caps */
3903                 if (platform_speed_cap == PCI_SPEED_UNKNOWN) {
3904                         adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3905                                                    CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
3906                 } else {
3907                         if (platform_speed_cap == PCIE_SPEED_16_0GT)
3908                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3909                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3910                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
3911                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
3912                         else if (platform_speed_cap == PCIE_SPEED_8_0GT)
3913                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3914                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3915                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
3916                         else if (platform_speed_cap == PCIE_SPEED_5_0GT)
3917                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3918                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
3919                         else
3920                                 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
3921
3922                 }
3923         }
3924         if (adev->pm.pcie_mlw_mask == 0) {
3925                 if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) {
3926                         adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
3927                 } else {
3928                         switch (platform_link_width) {
3929                         case PCIE_LNK_X32:
3930                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
3931                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3932                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3933                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3934                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3935                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3936                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3937                                 break;
3938                         case PCIE_LNK_X16:
3939                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3940                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3941                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3942                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3943                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3944                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3945                                 break;
3946                         case PCIE_LNK_X12:
3947                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3948                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3949                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3950                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3951                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3952                                 break;
3953                         case PCIE_LNK_X8:
3954                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3955                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3956                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3957                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3958                                 break;
3959                         case PCIE_LNK_X4:
3960                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3961                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3962                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3963                                 break;
3964                         case PCIE_LNK_X2:
3965                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3966                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3967                                 break;
3968                         case PCIE_LNK_X1:
3969                                 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
3970                                 break;
3971                         default:
3972                                 break;
3973                         }
3974                 }
3975         }
3976 }
3977