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/kthread.h>
  29 #include <linux/console.h>
  30 #include <linux/slab.h>
  31 #include <linux/debugfs.h>
  32 #include <drm/drmP.h>
  33 #include <drm/drm_crtc_helper.h>
  34 #include <drm/drm_atomic_helper.h>
  35 #include <drm/amdgpu_drm.h>
  36 #include <linux/vgaarb.h>
  37 #include <linux/vga_switcheroo.h>
  38 #include <linux/efi.h>
  39 #include "amdgpu.h"
  40 #include "amdgpu_trace.h"
  41 #include "amdgpu_i2c.h"
  42 #include "atom.h"
  43 #include "amdgpu_atombios.h"
  44 #include "amdgpu_atomfirmware.h"
  45 #include "amd_pcie.h"
  46 #ifdef CONFIG_DRM_AMDGPU_SI
  47 #include "si.h"
  48 #endif
  49 #ifdef CONFIG_DRM_AMDGPU_CIK
  50 #include "cik.h"
  51 #endif
  52 #include "vi.h"
  53 #include "soc15.h"
  54 #include "bif/bif_4_1_d.h"
  55 #include <linux/pci.h>
  56 #include <linux/firmware.h>
  57 #include "amdgpu_vf_error.h"
  58
  59 #include "amdgpu_amdkfd.h"
  60 #include "amdgpu_pm.h"
  61
  62 MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
  63 MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
  64
  65 #define AMDGPU_RESUME_MS                2000
  66
  67 static int amdgpu_debugfs_regs_init(struct amdgpu_device *adev);
  68 static void amdgpu_debugfs_regs_cleanup(struct amdgpu_device *adev);
  69 static int amdgpu_debugfs_test_ib_ring_init(struct amdgpu_device *adev);
  70 static int amdgpu_debugfs_vbios_dump_init(struct amdgpu_device *adev);
  71
  72 static const char *amdgpu_asic_name[] = {
  73         "TAHITI",
  74         "PITCAIRN",
  75         "VERDE",
  76         "OLAND",
  77         "HAINAN",
  78         "BONAIRE",
  79         "KAVERI",
  80         "KABINI",
  81         "HAWAII",
  82         "MULLINS",
  83         "TOPAZ",
  84         "TONGA",
  85         "FIJI",
  86         "CARRIZO",
  87         "STONEY",
  88         "POLARIS10",
  89         "POLARIS11",
  90         "POLARIS12",
  91         "VEGA10",
  92         "RAVEN",
  93         "LAST",
  94 };
  95
  96 bool amdgpu_device_is_px(struct drm_device *dev)
  97 {
  98         struct amdgpu_device *adev = dev->dev_private;
  99
 100         if (adev->flags & AMD_IS_PX)
 101                 return true;
 102         return false;
 103 }
 104
 105 /*
 106  * MMIO register access helper functions.
 107  */
 108 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
 109                         uint32_t acc_flags)
 110 {
 111         uint32_t ret;
 112
 113         if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
 114                 return amdgpu_virt_kiq_rreg(adev, reg);
 115
 116         if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
 117                 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
 118         else {
 119                 unsigned long flags;
 120
 121                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
 122                 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
 123                 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
 124                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
 125         }
 126         trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
 127         return ret;
 128 }
 129
 130 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
 131                     uint32_t acc_flags)
 132 {
 133         trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
 134
 135         if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
 136                 adev->last_mm_index = v;
 137         }
 138
 139         if (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev))
 140                 return amdgpu_virt_kiq_wreg(adev, reg, v);
 141
 142         if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
 143                 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
 144         else {
 145                 unsigned long flags;
 146
 147                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
 148                 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
 149                 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
 150                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
 151         }
 152
 153         if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
 154                 udelay(500);
 155         }
 156 }
 157
 158 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
 159 {
 160         if ((reg * 4) < adev->rio_mem_size)
 161                 return ioread32(adev->rio_mem + (reg * 4));
 162         else {
 163                 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
 164                 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
 165         }
 166 }
 167
 168 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
 169 {
 170         if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
 171                 adev->last_mm_index = v;
 172         }
 173
 174         if ((reg * 4) < adev->rio_mem_size)
 175                 iowrite32(v, adev->rio_mem + (reg * 4));
 176         else {
 177                 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
 178                 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
 179         }
 180
 181         if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
 182                 udelay(500);
 183         }
 184 }
 185
 186 /**
 187  * amdgpu_mm_rdoorbell - read a doorbell dword
 188  *
 189  * @adev: amdgpu_device pointer
 190  * @index: doorbell index
 191  *
 192  * Returns the value in the doorbell aperture at the
 193  * requested doorbell index (CIK).
 194  */
 195 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
 196 {
 197         if (index < adev->doorbell.num_doorbells) {
 198                 return readl(adev->doorbell.ptr + index);
 199         } else {
 200                 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
 201                 return 0;
 202         }
 203 }
 204
 205 /**
 206  * amdgpu_mm_wdoorbell - write a doorbell dword
 207  *
 208  * @adev: amdgpu_device pointer
 209  * @index: doorbell index
 210  * @v: value to write
 211  *
 212  * Writes @v to the doorbell aperture at the
 213  * requested doorbell index (CIK).
 214  */
 215 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
 216 {
 217         if (index < adev->doorbell.num_doorbells) {
 218                 writel(v, adev->doorbell.ptr + index);
 219         } else {
 220                 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
 221         }
 222 }
 223
 224 /**
 225  * amdgpu_mm_rdoorbell64 - read a doorbell Qword
 226  *
 227  * @adev: amdgpu_device pointer
 228  * @index: doorbell index
 229  *
 230  * Returns the value in the doorbell aperture at the
 231  * requested doorbell index (VEGA10+).
 232  */
 233 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
 234 {
 235         if (index < adev->doorbell.num_doorbells) {
 236                 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
 237         } else {
 238                 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
 239                 return 0;
 240         }
 241 }
 242
 243 /**
 244  * amdgpu_mm_wdoorbell64 - write a doorbell Qword
 245  *
 246  * @adev: amdgpu_device pointer
 247  * @index: doorbell index
 248  * @v: value to write
 249  *
 250  * Writes @v to the doorbell aperture at the
 251  * requested doorbell index (VEGA10+).
 252  */
 253 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
 254 {
 255         if (index < adev->doorbell.num_doorbells) {
 256                 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
 257         } else {
 258                 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
 259         }
 260 }
 261
 262 /**
 263  * amdgpu_invalid_rreg - dummy reg read function
 264  *
 265  * @adev: amdgpu device pointer
 266  * @reg: offset of register
 267  *
 268  * Dummy register read function.  Used for register blocks
 269  * that certain asics don't have (all asics).
 270  * Returns the value in the register.
 271  */
 272 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
 273 {
 274         DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
 275         BUG();
 276         return 0;
 277 }
 278
 279 /**
 280  * amdgpu_invalid_wreg - dummy reg write function
 281  *
 282  * @adev: amdgpu device pointer
 283  * @reg: offset of register
 284  * @v: value to write to the register
 285  *
 286  * Dummy register read function.  Used for register blocks
 287  * that certain asics don't have (all asics).
 288  */
 289 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
 290 {
 291         DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
 292                   reg, v);
 293         BUG();
 294 }
 295
 296 /**
 297  * amdgpu_block_invalid_rreg - dummy reg read function
 298  *
 299  * @adev: amdgpu device pointer
 300  * @block: offset of instance
 301  * @reg: offset of register
 302  *
 303  * Dummy register read function.  Used for register blocks
 304  * that certain asics don't have (all asics).
 305  * Returns the value in the register.
 306  */
 307 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
 308                                           uint32_t block, uint32_t reg)
 309 {
 310         DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
 311                   reg, block);
 312         BUG();
 313         return 0;
 314 }
 315
 316 /**
 317  * amdgpu_block_invalid_wreg - dummy reg write function
 318  *
 319  * @adev: amdgpu device pointer
 320  * @block: offset of instance
 321  * @reg: offset of register
 322  * @v: value to write to the register
 323  *
 324  * Dummy register read function.  Used for register blocks
 325  * that certain asics don't have (all asics).
 326  */
 327 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
 328                                       uint32_t block,
 329                                       uint32_t reg, uint32_t v)
 330 {
 331         DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
 332                   reg, block, v);
 333         BUG();
 334 }
 335
 336 static int amdgpu_vram_scratch_init(struct amdgpu_device *adev)
 337 {
 338         return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
 339                                        PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
 340                                        &adev->vram_scratch.robj,
 341                                        &adev->vram_scratch.gpu_addr,
 342                                        (void **)&adev->vram_scratch.ptr);
 343 }
 344
 345 static void amdgpu_vram_scratch_fini(struct amdgpu_device *adev)
 346 {
 347         amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
 348 }
 349
 350 /**
 351  * amdgpu_program_register_sequence - program an array of registers.
 352  *
 353  * @adev: amdgpu_device pointer
 354  * @registers: pointer to the register array
 355  * @array_size: size of the register array
 356  *
 357  * Programs an array or registers with and and or masks.
 358  * This is a helper for setting golden registers.
 359  */
 360 void amdgpu_program_register_sequence(struct amdgpu_device *adev,
 361                                       const u32 *registers,
 362                                       const u32 array_size)
 363 {
 364         u32 tmp, reg, and_mask, or_mask;
 365         int i;
 366
 367         if (array_size % 3)
 368                 return;
 369
 370         for (i = 0; i < array_size; i +=3) {
 371                 reg = registers[i + 0];
 372                 and_mask = registers[i + 1];
 373                 or_mask = registers[i + 2];
 374
 375                 if (and_mask == 0xffffffff) {
 376                         tmp = or_mask;
 377                 } else {
 378                         tmp = RREG32(reg);
 379                         tmp &= ~and_mask;
 380                         tmp |= or_mask;
 381                 }
 382                 WREG32(reg, tmp);
 383         }
 384 }
 385
 386 void amdgpu_pci_config_reset(struct amdgpu_device *adev)
 387 {
 388         pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
 389 }
 390
 391 /*
 392  * GPU doorbell aperture helpers function.
 393  */
 394 /**
 395  * amdgpu_doorbell_init - Init doorbell driver information.
 396  *
 397  * @adev: amdgpu_device pointer
 398  *
 399  * Init doorbell driver information (CIK)
 400  * Returns 0 on success, error on failure.
 401  */
 402 static int amdgpu_doorbell_init(struct amdgpu_device *adev)
 403 {
 404         /* No doorbell on SI hardware generation */
 405         if (adev->asic_type < CHIP_BONAIRE) {
 406                 adev->doorbell.base = 0;
 407                 adev->doorbell.size = 0;
 408                 adev->doorbell.num_doorbells = 0;
 409                 adev->doorbell.ptr = NULL;
 410                 return 0;
 411         }
 412
 413         /* doorbell bar mapping */
 414         adev->doorbell.base = pci_resource_start(adev->pdev, 2);
 415         adev->doorbell.size = pci_resource_len(adev->pdev, 2);
 416
 417         adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
 418                                              AMDGPU_DOORBELL_MAX_ASSIGNMENT+1);
 419         if (adev->doorbell.num_doorbells == 0)
 420                 return -EINVAL;
 421
 422         adev->doorbell.ptr = ioremap(adev->doorbell.base,
 423                                      adev->doorbell.num_doorbells *
 424                                      sizeof(u32));
 425         if (adev->doorbell.ptr == NULL)
 426                 return -ENOMEM;
 427
 428         return 0;
 429 }
 430
 431 /**
 432  * amdgpu_doorbell_fini - Tear down doorbell driver information.
 433  *
 434  * @adev: amdgpu_device pointer
 435  *
 436  * Tear down doorbell driver information (CIK)
 437  */
 438 static void amdgpu_doorbell_fini(struct amdgpu_device *adev)
 439 {
 440         iounmap(adev->doorbell.ptr);
 441         adev->doorbell.ptr = NULL;
 442 }
 443
 444 /**
 445  * amdgpu_doorbell_get_kfd_info - Report doorbell configuration required to
 446  *                                setup amdkfd
 447  *
 448  * @adev: amdgpu_device pointer
 449  * @aperture_base: output returning doorbell aperture base physical address
 450  * @aperture_size: output returning doorbell aperture size in bytes
 451  * @start_offset: output returning # of doorbell bytes reserved for amdgpu.
 452  *
 453  * amdgpu and amdkfd share the doorbell aperture. amdgpu sets it up,
 454  * takes doorbells required for its own rings and reports the setup to amdkfd.
 455  * amdgpu reserved doorbells are at the start of the doorbell aperture.
 456  */
 457 void amdgpu_doorbell_get_kfd_info(struct amdgpu_device *adev,
 458                                 phys_addr_t *aperture_base,
 459                                 size_t *aperture_size,
 460                                 size_t *start_offset)
 461 {
 462         /*
 463          * The first num_doorbells are used by amdgpu.
 464          * amdkfd takes whatever's left in the aperture.
 465          */
 466         if (adev->doorbell.size > adev->doorbell.num_doorbells * sizeof(u32)) {
 467                 *aperture_base = adev->doorbell.base;
 468                 *aperture_size = adev->doorbell.size;
 469                 *start_offset = adev->doorbell.num_doorbells * sizeof(u32);
 470         } else {
 471                 *aperture_base = 0;
 472                 *aperture_size = 0;
 473                 *start_offset = 0;
 474         }
 475 }
 476
 477 /*
 478  * amdgpu_wb_*()
 479  * Writeback is the method by which the GPU updates special pages in memory
 480  * with the status of certain GPU events (fences, ring pointers,etc.).
 481  */
 482
 483 /**
 484  * amdgpu_wb_fini - Disable Writeback and free memory
 485  *
 486  * @adev: amdgpu_device pointer
 487  *
 488  * Disables Writeback and frees the Writeback memory (all asics).
 489  * Used at driver shutdown.
 490  */
 491 static void amdgpu_wb_fini(struct amdgpu_device *adev)
 492 {
 493         if (adev->wb.wb_obj) {
 494                 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
 495                                       &adev->wb.gpu_addr,
 496                                       (void **)&adev->wb.wb);
 497                 adev->wb.wb_obj = NULL;
 498         }
 499 }
 500
 501 /**
 502  * amdgpu_wb_init- Init Writeback driver info and allocate memory
 503  *
 504  * @adev: amdgpu_device pointer
 505  *
 506  * Initializes writeback and allocates writeback memory (all asics).
 507  * Used at driver startup.
 508  * Returns 0 on success or an -error on failure.
 509  */
 510 static int amdgpu_wb_init(struct amdgpu_device *adev)
 511 {
 512         int r;
 513
 514         if (adev->wb.wb_obj == NULL) {
 515                 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
 516                 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
 517                                             PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
 518                                             &adev->wb.wb_obj, &adev->wb.gpu_addr,
 519                                             (void **)&adev->wb.wb);
 520                 if (r) {
 521                         dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
 522                         return r;
 523                 }
 524
 525                 adev->wb.num_wb = AMDGPU_MAX_WB;
 526                 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
 527
 528                 /* clear wb memory */
 529                 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t));
 530         }
 531
 532         return 0;
 533 }
 534
 535 /**
 536  * amdgpu_wb_get - Allocate a wb entry
 537  *
 538  * @adev: amdgpu_device pointer
 539  * @wb: wb index
 540  *
 541  * Allocate a wb slot for use by the driver (all asics).
 542  * Returns 0 on success or -EINVAL on failure.
 543  */
 544 int amdgpu_wb_get(struct amdgpu_device *adev, u32 *wb)
 545 {
 546         unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
 547
 548         if (offset < adev->wb.num_wb) {
 549                 __set_bit(offset, adev->wb.used);
 550                 *wb = offset << 3; /* convert to dw offset */
 551                 return 0;
 552         } else {
 553                 return -EINVAL;
 554         }
 555 }
 556
 557 /**
 558  * amdgpu_wb_free - Free a wb entry
 559  *
 560  * @adev: amdgpu_device pointer
 561  * @wb: wb index
 562  *
 563  * Free a wb slot allocated for use by the driver (all asics)
 564  */
 565 void amdgpu_wb_free(struct amdgpu_device *adev, u32 wb)
 566 {
 567         if (wb < adev->wb.num_wb)
 568                 __clear_bit(wb >> 3, adev->wb.used);
 569 }
 570
 571 /**
 572  * amdgpu_vram_location - try to find VRAM location
 573  * @adev: amdgpu device structure holding all necessary informations
 574  * @mc: memory controller structure holding memory informations
 575  * @base: base address at which to put VRAM
 576  *
 577  * Function will try to place VRAM at base address provided
 578  * as parameter (which is so far either PCI aperture address or
 579  * for IGP TOM base address).
 580  *
 581  * If there is not enough space to fit the unvisible VRAM in the 32bits
 582  * address space then we limit the VRAM size to the aperture.
 583  *
 584  * Note: We don't explicitly enforce VRAM start to be aligned on VRAM size,
 585  * this shouldn't be a problem as we are using the PCI aperture as a reference.
 586  * Otherwise this would be needed for rv280, all r3xx, and all r4xx, but
 587  * not IGP.
 588  *
 589  * Note: we use mc_vram_size as on some board we need to program the mc to
 590  * cover the whole aperture even if VRAM size is inferior to aperture size
 591  * Novell bug 204882 + along with lots of ubuntu ones
 592  *
 593  * Note: when limiting vram it's safe to overwritte real_vram_size because
 594  * we are not in case where real_vram_size is inferior to mc_vram_size (ie
 595  * note afected by bogus hw of Novell bug 204882 + along with lots of ubuntu
 596  * ones)
 597  *
 598  * Note: IGP TOM addr should be the same as the aperture addr, we don't
 599  * explicitly check for that though.
 600  *
 601  * FIXME: when reducing VRAM size align new size on power of 2.
 602  */
 603 void amdgpu_vram_location(struct amdgpu_device *adev, struct amdgpu_mc *mc, u64 base)
 604 {
 605         uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
 606
 607         mc->vram_start = base;
 608         if (mc->mc_vram_size > (adev->mc.mc_mask - base + 1)) {
 609                 dev_warn(adev->dev, "limiting VRAM to PCI aperture size\n");
 610                 mc->real_vram_size = mc->aper_size;
 611                 mc->mc_vram_size = mc->aper_size;
 612         }
 613         mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
 614         if (limit && limit < mc->real_vram_size)
 615                 mc->real_vram_size = limit;
 616         dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
 617                         mc->mc_vram_size >> 20, mc->vram_start,
 618                         mc->vram_end, mc->real_vram_size >> 20);
 619 }
 620
 621 /**
 622  * amdgpu_gart_location - try to find GTT location
 623  * @adev: amdgpu device structure holding all necessary informations
 624  * @mc: memory controller structure holding memory informations
 625  *
 626  * Function will place try to place GTT before or after VRAM.
 627  *
 628  * If GTT size is bigger than space left then we ajust GTT size.
 629  * Thus function will never fails.
 630  *
 631  * FIXME: when reducing GTT size align new size on power of 2.
 632  */
 633 void amdgpu_gart_location(struct amdgpu_device *adev, struct amdgpu_mc *mc)
 634 {
 635         u64 size_af, size_bf;
 636
 637         size_af = adev->mc.mc_mask - mc->vram_end;
 638         size_bf = mc->vram_start;
 639         if (size_bf > size_af) {
 640                 if (mc->gart_size > size_bf) {
 641                         dev_warn(adev->dev, "limiting GTT\n");
 642                         mc->gart_size = size_bf;
 643                 }
 644                 mc->gart_start = 0;
 645         } else {
 646                 if (mc->gart_size > size_af) {
 647                         dev_warn(adev->dev, "limiting GTT\n");
 648                         mc->gart_size = size_af;
 649                 }
 650                 mc->gart_start = mc->vram_end + 1;
 651         }
 652         mc->gart_end = mc->gart_start + mc->gart_size - 1;
 653         dev_info(adev->dev, "GTT: %lluM 0x%016llX - 0x%016llX\n",
 654                         mc->gart_size >> 20, mc->gart_start, mc->gart_end);
 655 }
 656
 657 /*
 658  * Firmware Reservation functions
 659  */
 660 /**
 661  * amdgpu_fw_reserve_vram_fini - free fw reserved vram
 662  *
 663  * @adev: amdgpu_device pointer
 664  *
 665  * free fw reserved vram if it has been reserved.
 666  */
 667 void amdgpu_fw_reserve_vram_fini(struct amdgpu_device *adev)
 668 {
 669         amdgpu_bo_free_kernel(&adev->fw_vram_usage.reserved_bo,
 670                 NULL, &adev->fw_vram_usage.va);
 671 }
 672
 673 /**
 674  * amdgpu_fw_reserve_vram_init - create bo vram reservation from fw
 675  *
 676  * @adev: amdgpu_device pointer
 677  *
 678  * create bo vram reservation from fw.
 679  */
 680 int amdgpu_fw_reserve_vram_init(struct amdgpu_device *adev)
 681 {
 682         int r = 0;
 683         u64 gpu_addr;
 684         u64 vram_size = adev->mc.visible_vram_size;
 685
 686         adev->fw_vram_usage.va = NULL;
 687         adev->fw_vram_usage.reserved_bo = NULL;
 688
 689         if (adev->fw_vram_usage.size > 0 &&
 690                 adev->fw_vram_usage.size <= vram_size) {
 691
 692                 r = amdgpu_bo_create(adev, adev->fw_vram_usage.size,
 693                         PAGE_SIZE, true, 0,
 694                         AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
 695                         AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS, NULL, NULL, 0,
 696                         &adev->fw_vram_usage.reserved_bo);
 697                 if (r)
 698                         goto error_create;
 699
 700                 r = amdgpu_bo_reserve(adev->fw_vram_usage.reserved_bo, false);
 701                 if (r)
 702                         goto error_reserve;
 703                 r = amdgpu_bo_pin_restricted(adev->fw_vram_usage.reserved_bo,
 704                         AMDGPU_GEM_DOMAIN_VRAM,
 705                         adev->fw_vram_usage.start_offset,
 706                         (adev->fw_vram_usage.start_offset +
 707                         adev->fw_vram_usage.size), &gpu_addr);
 708                 if (r)
 709                         goto error_pin;
 710                 r = amdgpu_bo_kmap(adev->fw_vram_usage.reserved_bo,
 711                         &adev->fw_vram_usage.va);
 712                 if (r)
 713                         goto error_kmap;
 714
 715                 amdgpu_bo_unreserve(adev->fw_vram_usage.reserved_bo);
 716         }
 717         return r;
 718
 719 error_kmap:
 720         amdgpu_bo_unpin(adev->fw_vram_usage.reserved_bo);
 721 error_pin:
 722         amdgpu_bo_unreserve(adev->fw_vram_usage.reserved_bo);
 723 error_reserve:
 724         amdgpu_bo_unref(&adev->fw_vram_usage.reserved_bo);
 725 error_create:
 726         adev->fw_vram_usage.va = NULL;
 727         adev->fw_vram_usage.reserved_bo = NULL;
 728         return r;
 729 }
 730
 731
 732 /*
 733  * GPU helpers function.
 734  */
 735 /**
 736  * amdgpu_need_post - check if the hw need post or not
 737  *
 738  * @adev: amdgpu_device pointer
 739  *
 740  * Check if the asic has been initialized (all asics) at driver startup
 741  * or post is needed if  hw reset is performed.
 742  * Returns true if need or false if not.
 743  */
 744 bool amdgpu_need_post(struct amdgpu_device *adev)
 745 {
 746         uint32_t reg;
 747
 748         if (amdgpu_sriov_vf(adev))
 749                 return false;
 750
 751         if (amdgpu_passthrough(adev)) {
 752                 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
 753                  * some old smc fw still need driver do vPost otherwise gpu hang, while
 754                  * those smc fw version above 22.15 doesn't have this flaw, so we force
 755                  * vpost executed for smc version below 22.15
 756                  */
 757                 if (adev->asic_type == CHIP_FIJI) {
 758                         int err;
 759                         uint32_t fw_ver;
 760                         err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
 761                         /* force vPost if error occured */
 762                         if (err)
 763                                 return true;
 764
 765                         fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
 766                         if (fw_ver < 0x00160e00)
 767                                 return true;
 768                 }
 769         }
 770
 771         if (adev->has_hw_reset) {
 772                 adev->has_hw_reset = false;
 773                 return true;
 774         }
 775
 776         /* bios scratch used on CIK+ */
 777         if (adev->asic_type >= CHIP_BONAIRE)
 778                 return amdgpu_atombios_scratch_need_asic_init(adev);
 779
 780         /* check MEM_SIZE for older asics */
 781         reg = amdgpu_asic_get_config_memsize(adev);
 782
 783         if ((reg != 0) && (reg != 0xffffffff))
 784                 return false;
 785
 786         return true;
 787 }
 788
 789 /**
 790  * amdgpu_dummy_page_init - init dummy page used by the driver
 791  *
 792  * @adev: amdgpu_device pointer
 793  *
 794  * Allocate the dummy page used by the driver (all asics).
 795  * This dummy page is used by the driver as a filler for gart entries
 796  * when pages are taken out of the GART
 797  * Returns 0 on sucess, -ENOMEM on failure.
 798  */
 799 int amdgpu_dummy_page_init(struct amdgpu_device *adev)
 800 {
 801         if (adev->dummy_page.page)
 802                 return 0;
 803         adev->dummy_page.page = alloc_page(GFP_DMA32 | GFP_KERNEL | __GFP_ZERO);
 804         if (adev->dummy_page.page == NULL)
 805                 return -ENOMEM;
 806         adev->dummy_page.addr = pci_map_page(adev->pdev, adev->dummy_page.page,
 807                                         0, PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 808         if (pci_dma_mapping_error(adev->pdev, adev->dummy_page.addr)) {
 809                 dev_err(&adev->pdev->dev, "Failed to DMA MAP the dummy page\n");
 810                 __free_page(adev->dummy_page.page);
 811                 adev->dummy_page.page = NULL;
 812                 return -ENOMEM;
 813         }
 814         return 0;
 815 }
 816
 817 /**
 818  * amdgpu_dummy_page_fini - free dummy page used by the driver
 819  *
 820  * @adev: amdgpu_device pointer
 821  *
 822  * Frees the dummy page used by the driver (all asics).
 823  */
 824 void amdgpu_dummy_page_fini(struct amdgpu_device *adev)
 825 {
 826         if (adev->dummy_page.page == NULL)
 827                 return;
 828         pci_unmap_page(adev->pdev, adev->dummy_page.addr,
 829                         PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 830         __free_page(adev->dummy_page.page);
 831         adev->dummy_page.page = NULL;
 832 }
 833
 834
 835 /* ATOM accessor methods */
 836 /*
 837  * ATOM is an interpreted byte code stored in tables in the vbios.  The
 838  * driver registers callbacks to access registers and the interpreter
 839  * in the driver parses the tables and executes then to program specific
 840  * actions (set display modes, asic init, etc.).  See amdgpu_atombios.c,
 841  * atombios.h, and atom.c
 842  */
 843
 844 /**
 845  * cail_pll_read - read PLL register
 846  *
 847  * @info: atom card_info pointer
 848  * @reg: PLL register offset
 849  *
 850  * Provides a PLL register accessor for the atom interpreter (r4xx+).
 851  * Returns the value of the PLL register.
 852  */
 853 static uint32_t cail_pll_read(struct card_info *info, uint32_t reg)
 854 {
 855         return 0;
 856 }
 857
 858 /**
 859  * cail_pll_write - write PLL register
 860  *
 861  * @info: atom card_info pointer
 862  * @reg: PLL register offset
 863  * @val: value to write to the pll register
 864  *
 865  * Provides a PLL register accessor for the atom interpreter (r4xx+).
 866  */
 867 static void cail_pll_write(struct card_info *info, uint32_t reg, uint32_t val)
 868 {
 869
 870 }
 871
 872 /**
 873  * cail_mc_read - read MC (Memory Controller) register
 874  *
 875  * @info: atom card_info pointer
 876  * @reg: MC register offset
 877  *
 878  * Provides an MC register accessor for the atom interpreter (r4xx+).
 879  * Returns the value of the MC register.
 880  */
 881 static uint32_t cail_mc_read(struct card_info *info, uint32_t reg)
 882 {
 883         return 0;
 884 }
 885
 886 /**
 887  * cail_mc_write - write MC (Memory Controller) register
 888  *
 889  * @info: atom card_info pointer
 890  * @reg: MC register offset
 891  * @val: value to write to the pll register
 892  *
 893  * Provides a MC register accessor for the atom interpreter (r4xx+).
 894  */
 895 static void cail_mc_write(struct card_info *info, uint32_t reg, uint32_t val)
 896 {
 897
 898 }
 899
 900 /**
 901  * cail_reg_write - write MMIO register
 902  *
 903  * @info: atom card_info pointer
 904  * @reg: MMIO register offset
 905  * @val: value to write to the pll register
 906  *
 907  * Provides a MMIO register accessor for the atom interpreter (r4xx+).
 908  */
 909 static void cail_reg_write(struct card_info *info, uint32_t reg, uint32_t val)
 910 {
 911         struct amdgpu_device *adev = info->dev->dev_private;
 912
 913         WREG32(reg, val);
 914 }
 915
 916 /**
 917  * cail_reg_read - read MMIO register
 918  *
 919  * @info: atom card_info pointer
 920  * @reg: MMIO register offset
 921  *
 922  * Provides an MMIO register accessor for the atom interpreter (r4xx+).
 923  * Returns the value of the MMIO register.
 924  */
 925 static uint32_t cail_reg_read(struct card_info *info, uint32_t reg)
 926 {
 927         struct amdgpu_device *adev = info->dev->dev_private;
 928         uint32_t r;
 929
 930         r = RREG32(reg);
 931         return r;
 932 }
 933
 934 /**
 935  * cail_ioreg_write - write IO register
 936  *
 937  * @info: atom card_info pointer
 938  * @reg: IO register offset
 939  * @val: value to write to the pll register
 940  *
 941  * Provides a IO register accessor for the atom interpreter (r4xx+).
 942  */
 943 static void cail_ioreg_write(struct card_info *info, uint32_t reg, uint32_t val)
 944 {
 945         struct amdgpu_device *adev = info->dev->dev_private;
 946
 947         WREG32_IO(reg, val);
 948 }
 949
 950 /**
 951  * cail_ioreg_read - read IO register
 952  *
 953  * @info: atom card_info pointer
 954  * @reg: IO register offset
 955  *
 956  * Provides an IO register accessor for the atom interpreter (r4xx+).
 957  * Returns the value of the IO register.
 958  */
 959 static uint32_t cail_ioreg_read(struct card_info *info, uint32_t reg)
 960 {
 961         struct amdgpu_device *adev = info->dev->dev_private;
 962         uint32_t r;
 963
 964         r = RREG32_IO(reg);
 965         return r;
 966 }
 967
 968 static ssize_t amdgpu_atombios_get_vbios_version(struct device *dev,
 969                                                  struct device_attribute *attr,
 970                                                  char *buf)
 971 {
 972         struct drm_device *ddev = dev_get_drvdata(dev);
 973         struct amdgpu_device *adev = ddev->dev_private;
 974         struct atom_context *ctx = adev->mode_info.atom_context;
 975
 976         return snprintf(buf, PAGE_SIZE, "%s\n", ctx->vbios_version);
 977 }
 978
 979 static DEVICE_ATTR(vbios_version, 0444, amdgpu_atombios_get_vbios_version,
 980                    NULL);
 981
 982 /**
 983  * amdgpu_atombios_fini - free the driver info and callbacks for atombios
 984  *
 985  * @adev: amdgpu_device pointer
 986  *
 987  * Frees the driver info and register access callbacks for the ATOM
 988  * interpreter (r4xx+).
 989  * Called at driver shutdown.
 990  */
 991 static void amdgpu_atombios_fini(struct amdgpu_device *adev)
 992 {
 993         if (adev->mode_info.atom_context) {
 994                 kfree(adev->mode_info.atom_context->scratch);
 995                 kfree(adev->mode_info.atom_context->iio);
 996         }
 997         kfree(adev->mode_info.atom_context);
 998         adev->mode_info.atom_context = NULL;
 999         kfree(adev->mode_info.atom_card_info);
1000         adev->mode_info.atom_card_info = NULL;
1001         device_remove_file(adev->dev, &dev_attr_vbios_version);
1002 }
1003
1004 /**
1005  * amdgpu_atombios_init - init the driver info and callbacks for atombios
1006  *
1007  * @adev: amdgpu_device pointer
1008  *
1009  * Initializes the driver info and register access callbacks for the
1010  * ATOM interpreter (r4xx+).
1011  * Returns 0 on sucess, -ENOMEM on failure.
1012  * Called at driver startup.
1013  */
1014 static int amdgpu_atombios_init(struct amdgpu_device *adev)
1015 {
1016         struct card_info *atom_card_info =
1017             kzalloc(sizeof(struct card_info), GFP_KERNEL);
1018         int ret;
1019
1020         if (!atom_card_info)
1021                 return -ENOMEM;
1022
1023         adev->mode_info.atom_card_info = atom_card_info;
1024         atom_card_info->dev = adev->ddev;
1025         atom_card_info->reg_read = cail_reg_read;
1026         atom_card_info->reg_write = cail_reg_write;
1027         /* needed for iio ops */
1028         if (adev->rio_mem) {
1029                 atom_card_info->ioreg_read = cail_ioreg_read;
1030                 atom_card_info->ioreg_write = cail_ioreg_write;
1031         } else {
1032                 DRM_INFO("PCI I/O BAR is not found. Using MMIO to access ATOM BIOS\n");
1033                 atom_card_info->ioreg_read = cail_reg_read;
1034                 atom_card_info->ioreg_write = cail_reg_write;
1035         }
1036         atom_card_info->mc_read = cail_mc_read;
1037         atom_card_info->mc_write = cail_mc_write;
1038         atom_card_info->pll_read = cail_pll_read;
1039         atom_card_info->pll_write = cail_pll_write;
1040
1041         adev->mode_info.atom_context = amdgpu_atom_parse(atom_card_info, adev->bios);
1042         if (!adev->mode_info.atom_context) {
1043                 amdgpu_atombios_fini(adev);
1044                 return -ENOMEM;
1045         }
1046
1047         mutex_init(&adev->mode_info.atom_context->mutex);
1048         if (adev->is_atom_fw) {
1049                 amdgpu_atomfirmware_scratch_regs_init(adev);
1050                 amdgpu_atomfirmware_allocate_fb_scratch(adev);
1051         } else {
1052                 amdgpu_atombios_scratch_regs_init(adev);
1053                 amdgpu_atombios_allocate_fb_scratch(adev);
1054         }
1055
1056         ret = device_create_file(adev->dev, &dev_attr_vbios_version);
1057         if (ret) {
1058                 DRM_ERROR("Failed to create device file for VBIOS version\n");
1059                 return ret;
1060         }
1061
1062         return 0;
1063 }
1064
1065 /* if we get transitioned to only one device, take VGA back */
1066 /**
1067  * amdgpu_vga_set_decode - enable/disable vga decode
1068  *
1069  * @cookie: amdgpu_device pointer
1070  * @state: enable/disable vga decode
1071  *
1072  * Enable/disable vga decode (all asics).
1073  * Returns VGA resource flags.
1074  */
1075 static unsigned int amdgpu_vga_set_decode(void *cookie, bool state)
1076 {
1077         struct amdgpu_device *adev = cookie;
1078         amdgpu_asic_set_vga_state(adev, state);
1079         if (state)
1080                 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
1081                        VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
1082         else
1083                 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
1084 }
1085
1086 static void amdgpu_check_block_size(struct amdgpu_device *adev)
1087 {
1088         /* defines number of bits in page table versus page directory,
1089          * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
1090          * page table and the remaining bits are in the page directory */
1091         if (amdgpu_vm_block_size == -1)
1092                 return;
1093
1094         if (amdgpu_vm_block_size < 9) {
1095                 dev_warn(adev->dev, "VM page table size (%d) too small\n",
1096                          amdgpu_vm_block_size);
1097                 goto def_value;
1098         }
1099
1100         if (amdgpu_vm_block_size > 24 ||
1101             (amdgpu_vm_size * 1024) < (1ull << amdgpu_vm_block_size)) {
1102                 dev_warn(adev->dev, "VM page table size (%d) too large\n",
1103                          amdgpu_vm_block_size);
1104                 goto def_value;
1105         }
1106
1107         return;
1108
1109 def_value:
1110         amdgpu_vm_block_size = -1;
1111 }
1112
1113 static void amdgpu_check_vm_size(struct amdgpu_device *adev)
1114 {
1115         /* no need to check the default value */
1116         if (amdgpu_vm_size == -1)
1117                 return;
1118
1119         if (!is_power_of_2(amdgpu_vm_size)) {
1120                 dev_warn(adev->dev, "VM size (%d) must be a power of 2\n",
1121                          amdgpu_vm_size);
1122                 goto def_value;
1123         }
1124
1125         if (amdgpu_vm_size < 1) {
1126                 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
1127                          amdgpu_vm_size);
1128                 goto def_value;
1129         }
1130
1131         /*
1132          * Max GPUVM size for Cayman, SI, CI VI are 40 bits.
1133          */
1134         if (amdgpu_vm_size > 1024) {
1135                 dev_warn(adev->dev, "VM size (%d) too large, max is 1TB\n",
1136                          amdgpu_vm_size);
1137                 goto def_value;
1138         }
1139
1140         return;
1141
1142 def_value:
1143         amdgpu_vm_size = -1;
1144 }
1145
1146 /**
1147  * amdgpu_check_arguments - validate module params
1148  *
1149  * @adev: amdgpu_device pointer
1150  *
1151  * Validates certain module parameters and updates
1152  * the associated values used by the driver (all asics).
1153  */
1154 static void amdgpu_check_arguments(struct amdgpu_device *adev)
1155 {
1156         if (amdgpu_sched_jobs < 4) {
1157                 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
1158                          amdgpu_sched_jobs);
1159                 amdgpu_sched_jobs = 4;
1160         } else if (!is_power_of_2(amdgpu_sched_jobs)){
1161                 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
1162                          amdgpu_sched_jobs);
1163                 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
1164         }
1165
1166         if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
1167                 /* gart size must be greater or equal to 32M */
1168                 dev_warn(adev->dev, "gart size (%d) too small\n",
1169                          amdgpu_gart_size);
1170                 amdgpu_gart_size = -1;
1171         }
1172
1173         if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
1174                 /* gtt size must be greater or equal to 32M */
1175                 dev_warn(adev->dev, "gtt size (%d) too small\n",
1176                                  amdgpu_gtt_size);
1177                 amdgpu_gtt_size = -1;
1178         }
1179
1180         /* valid range is between 4 and 9 inclusive */
1181         if (amdgpu_vm_fragment_size != -1 &&
1182             (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
1183                 dev_warn(adev->dev, "valid range is between 4 and 9\n");
1184                 amdgpu_vm_fragment_size = -1;
1185         }
1186
1187         amdgpu_check_vm_size(adev);
1188
1189         amdgpu_check_block_size(adev);
1190
1191         if (amdgpu_vram_page_split != -1 && (amdgpu_vram_page_split < 16 ||
1192             !is_power_of_2(amdgpu_vram_page_split))) {
1193                 dev_warn(adev->dev, "invalid VRAM page split (%d)\n",
1194                          amdgpu_vram_page_split);
1195                 amdgpu_vram_page_split = 1024;
1196         }
1197 }
1198
1199 /**
1200  * amdgpu_switcheroo_set_state - set switcheroo state
1201  *
1202  * @pdev: pci dev pointer
1203  * @state: vga_switcheroo state
1204  *
1205  * Callback for the switcheroo driver.  Suspends or resumes the
1206  * the asics before or after it is powered up using ACPI methods.
1207  */
1208 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
1209 {
1210         struct drm_device *dev = pci_get_drvdata(pdev);
1211
1212         if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
1213                 return;
1214
1215         if (state == VGA_SWITCHEROO_ON) {
1216                 pr_info("amdgpu: switched on\n");
1217                 /* don't suspend or resume card normally */
1218                 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1219
1220                 amdgpu_device_resume(dev, true, true);
1221
1222                 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1223                 drm_kms_helper_poll_enable(dev);
1224         } else {
1225                 pr_info("amdgpu: switched off\n");
1226                 drm_kms_helper_poll_disable(dev);
1227                 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1228                 amdgpu_device_suspend(dev, true, true);
1229                 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1230         }
1231 }
1232
1233 /**
1234  * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1235  *
1236  * @pdev: pci dev pointer
1237  *
1238  * Callback for the switcheroo driver.  Check of the switcheroo
1239  * state can be changed.
1240  * Returns true if the state can be changed, false if not.
1241  */
1242 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1243 {
1244         struct drm_device *dev = pci_get_drvdata(pdev);
1245
1246         /*
1247         * FIXME: open_count is protected by drm_global_mutex but that would lead to
1248         * locking inversion with the driver load path. And the access here is
1249         * completely racy anyway. So don't bother with locking for now.
1250         */
1251         return dev->open_count == 0;
1252 }
1253
1254 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1255         .set_gpu_state = amdgpu_switcheroo_set_state,
1256         .reprobe = NULL,
1257         .can_switch = amdgpu_switcheroo_can_switch,
1258 };
1259
1260 int amdgpu_set_clockgating_state(struct amdgpu_device *adev,
1261                                   enum amd_ip_block_type block_type,
1262                                   enum amd_clockgating_state state)
1263 {
1264         int i, r = 0;
1265
1266         for (i = 0; i < adev->num_ip_blocks; i++) {
1267                 if (!adev->ip_blocks[i].status.valid)
1268                         continue;
1269                 if (adev->ip_blocks[i].version->type != block_type)
1270                         continue;
1271                 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1272                         continue;
1273                 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1274                         (void *)adev, state);
1275                 if (r)
1276                         DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1277                                   adev->ip_blocks[i].version->funcs->name, r);
1278         }
1279         return r;
1280 }
1281
1282 int amdgpu_set_powergating_state(struct amdgpu_device *adev,
1283                                   enum amd_ip_block_type block_type,
1284                                   enum amd_powergating_state state)
1285 {
1286         int i, r = 0;
1287
1288         for (i = 0; i < adev->num_ip_blocks; i++) {
1289                 if (!adev->ip_blocks[i].status.valid)
1290                         continue;
1291                 if (adev->ip_blocks[i].version->type != block_type)
1292                         continue;
1293                 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1294                         continue;
1295                 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1296                         (void *)adev, state);
1297                 if (r)
1298                         DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1299                                   adev->ip_blocks[i].version->funcs->name, r);
1300         }
1301         return r;
1302 }
1303
1304 void amdgpu_get_clockgating_state(struct amdgpu_device *adev, u32 *flags)
1305 {
1306         int i;
1307
1308         for (i = 0; i < adev->num_ip_blocks; i++) {
1309                 if (!adev->ip_blocks[i].status.valid)
1310                         continue;
1311                 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1312                         adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1313         }
1314 }
1315
1316 int amdgpu_wait_for_idle(struct amdgpu_device *adev,
1317                          enum amd_ip_block_type block_type)
1318 {
1319         int i, r;
1320
1321         for (i = 0; i < adev->num_ip_blocks; i++) {
1322                 if (!adev->ip_blocks[i].status.valid)
1323                         continue;
1324                 if (adev->ip_blocks[i].version->type == block_type) {
1325                         r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1326                         if (r)
1327                                 return r;
1328                         break;
1329                 }
1330         }
1331         return 0;
1332
1333 }
1334
1335 bool amdgpu_is_idle(struct amdgpu_device *adev,
1336                     enum amd_ip_block_type block_type)
1337 {
1338         int i;
1339
1340         for (i = 0; i < adev->num_ip_blocks; i++) {
1341                 if (!adev->ip_blocks[i].status.valid)
1342                         continue;
1343                 if (adev->ip_blocks[i].version->type == block_type)
1344                         return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1345         }
1346         return true;
1347
1348 }
1349
1350 struct amdgpu_ip_block * amdgpu_get_ip_block(struct amdgpu_device *adev,
1351                                              enum amd_ip_block_type type)
1352 {
1353         int i;
1354
1355         for (i = 0; i < adev->num_ip_blocks; i++)
1356                 if (adev->ip_blocks[i].version->type == type)
1357                         return &adev->ip_blocks[i];
1358
1359         return NULL;
1360 }
1361
1362 /**
1363  * amdgpu_ip_block_version_cmp
1364  *
1365  * @adev: amdgpu_device pointer
1366  * @type: enum amd_ip_block_type
1367  * @major: major version
1368  * @minor: minor version
1369  *
1370  * return 0 if equal or greater
1371  * return 1 if smaller or the ip_block doesn't exist
1372  */
1373 int amdgpu_ip_block_version_cmp(struct amdgpu_device *adev,
1374                                 enum amd_ip_block_type type,
1375                                 u32 major, u32 minor)
1376 {
1377         struct amdgpu_ip_block *ip_block = amdgpu_get_ip_block(adev, type);
1378
1379         if (ip_block && ((ip_block->version->major > major) ||
1380                         ((ip_block->version->major == major) &&
1381                         (ip_block->version->minor >= minor))))
1382                 return 0;
1383
1384         return 1;
1385 }
1386
1387 /**
1388  * amdgpu_ip_block_add
1389  *
1390  * @adev: amdgpu_device pointer
1391  * @ip_block_version: pointer to the IP to add
1392  *
1393  * Adds the IP block driver information to the collection of IPs
1394  * on the asic.
1395  */
1396 int amdgpu_ip_block_add(struct amdgpu_device *adev,
1397                         const struct amdgpu_ip_block_version *ip_block_version)
1398 {
1399         if (!ip_block_version)
1400                 return -EINVAL;
1401
1402         DRM_DEBUG("add ip block number %d <%s>\n", adev->num_ip_blocks,
1403                   ip_block_version->funcs->name);
1404
1405         adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1406
1407         return 0;
1408 }
1409
1410 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1411 {
1412         adev->enable_virtual_display = false;
1413
1414         if (amdgpu_virtual_display) {
1415                 struct drm_device *ddev = adev->ddev;
1416                 const char *pci_address_name = pci_name(ddev->pdev);
1417                 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1418
1419                 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1420                 pciaddstr_tmp = pciaddstr;
1421                 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1422                         pciaddname = strsep(&pciaddname_tmp, ",");
1423                         if (!strcmp("all", pciaddname)
1424                             || !strcmp(pci_address_name, pciaddname)) {
1425                                 long num_crtc;
1426                                 int res = -1;
1427
1428                                 adev->enable_virtual_display = true;
1429
1430                                 if (pciaddname_tmp)
1431                                         res = kstrtol(pciaddname_tmp, 10,
1432                                                       &num_crtc);
1433
1434                                 if (!res) {
1435                                         if (num_crtc < 1)
1436                                                 num_crtc = 1;
1437                                         if (num_crtc > 6)
1438                                                 num_crtc = 6;
1439                                         adev->mode_info.num_crtc = num_crtc;
1440                                 } else {
1441                                         adev->mode_info.num_crtc = 1;
1442                                 }
1443                                 break;
1444                         }
1445                 }
1446
1447                 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1448                          amdgpu_virtual_display, pci_address_name,
1449                          adev->enable_virtual_display, adev->mode_info.num_crtc);
1450
1451                 kfree(pciaddstr);
1452         }
1453 }
1454
1455 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1456 {
1457         const char *chip_name;
1458         char fw_name[30];
1459         int err;
1460         const struct gpu_info_firmware_header_v1_0 *hdr;
1461
1462         adev->firmware.gpu_info_fw = NULL;
1463
1464         switch (adev->asic_type) {
1465         case CHIP_TOPAZ:
1466         case CHIP_TONGA:
1467         case CHIP_FIJI:
1468         case CHIP_POLARIS11:
1469         case CHIP_POLARIS10:
1470         case CHIP_POLARIS12:
1471         case CHIP_CARRIZO:
1472         case CHIP_STONEY:
1473 #ifdef CONFIG_DRM_AMDGPU_SI
1474         case CHIP_VERDE:
1475         case CHIP_TAHITI:
1476         case CHIP_PITCAIRN:
1477         case CHIP_OLAND:
1478         case CHIP_HAINAN:
1479 #endif
1480 #ifdef CONFIG_DRM_AMDGPU_CIK
1481         case CHIP_BONAIRE:
1482         case CHIP_HAWAII:
1483         case CHIP_KAVERI:
1484         case CHIP_KABINI:
1485         case CHIP_MULLINS:
1486 #endif
1487         default:
1488                 return 0;
1489         case CHIP_VEGA10:
1490                 chip_name = "vega10";
1491                 break;
1492         case CHIP_RAVEN:
1493                 chip_name = "raven";
1494                 break;
1495         }
1496
1497         snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1498         err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1499         if (err) {
1500                 dev_err(adev->dev,
1501                         "Failed to load gpu_info firmware \"%s\"\n",
1502                         fw_name);
1503                 goto out;
1504         }
1505         err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1506         if (err) {
1507                 dev_err(adev->dev,
1508                         "Failed to validate gpu_info firmware \"%s\"\n",
1509                         fw_name);
1510                 goto out;
1511         }
1512
1513         hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1514         amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1515
1516         switch (hdr->version_major) {
1517         case 1:
1518         {
1519                 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1520                         (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1521                                                                 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1522
1523                 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1524                 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1525                 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1526                 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1527                 adev->gfx.config.max_texture_channel_caches =
1528                         le32_to_cpu(gpu_info_fw->gc_num_tccs);
1529                 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1530                 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1531                 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1532                 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1533                 adev->gfx.config.double_offchip_lds_buf =
1534                         le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1535                 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1536                 adev->gfx.cu_info.max_waves_per_simd =
1537                         le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1538                 adev->gfx.cu_info.max_scratch_slots_per_cu =
1539                         le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1540                 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
1541                 break;
1542         }
1543         default:
1544                 dev_err(adev->dev,
1545                         "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1546                 err = -EINVAL;
1547                 goto out;
1548         }
1549 out:
1550         return err;
1551 }
1552
1553 static int amdgpu_early_init(struct amdgpu_device *adev)
1554 {
1555         int i, r;
1556
1557         amdgpu_device_enable_virtual_display(adev);
1558
1559         switch (adev->asic_type) {
1560         case CHIP_TOPAZ:
1561         case CHIP_TONGA:
1562         case CHIP_FIJI:
1563         case CHIP_POLARIS11:
1564         case CHIP_POLARIS10:
1565         case CHIP_POLARIS12:
1566         case CHIP_CARRIZO:
1567         case CHIP_STONEY:
1568                 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1569                         adev->family = AMDGPU_FAMILY_CZ;
1570                 else
1571                         adev->family = AMDGPU_FAMILY_VI;
1572
1573                 r = vi_set_ip_blocks(adev);
1574                 if (r)
1575                         return r;
1576                 break;
1577 #ifdef CONFIG_DRM_AMDGPU_SI
1578         case CHIP_VERDE:
1579         case CHIP_TAHITI:
1580         case CHIP_PITCAIRN:
1581         case CHIP_OLAND:
1582         case CHIP_HAINAN:
1583                 adev->family = AMDGPU_FAMILY_SI;
1584                 r = si_set_ip_blocks(adev);
1585                 if (r)
1586                         return r;
1587                 break;
1588 #endif
1589 #ifdef CONFIG_DRM_AMDGPU_CIK
1590         case CHIP_BONAIRE:
1591         case CHIP_HAWAII:
1592         case CHIP_KAVERI:
1593         case CHIP_KABINI:
1594         case CHIP_MULLINS:
1595                 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1596                         adev->family = AMDGPU_FAMILY_CI;
1597                 else
1598                         adev->family = AMDGPU_FAMILY_KV;
1599
1600                 r = cik_set_ip_blocks(adev);
1601                 if (r)
1602                         return r;
1603                 break;
1604 #endif
1605         case  CHIP_VEGA10:
1606         case  CHIP_RAVEN:
1607                 if (adev->asic_type == CHIP_RAVEN)
1608                         adev->family = AMDGPU_FAMILY_RV;
1609                 else
1610                         adev->family = AMDGPU_FAMILY_AI;
1611
1612                 r = soc15_set_ip_blocks(adev);
1613                 if (r)
1614                         return r;
1615                 break;
1616         default:
1617                 /* FIXME: not supported yet */
1618                 return -EINVAL;
1619         }
1620
1621         r = amdgpu_device_parse_gpu_info_fw(adev);
1622         if (r)
1623                 return r;
1624
1625         if (amdgpu_sriov_vf(adev)) {
1626                 r = amdgpu_virt_request_full_gpu(adev, true);
1627                 if (r)
1628                         return r;
1629         }
1630
1631         for (i = 0; i < adev->num_ip_blocks; i++) {
1632                 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1633                         DRM_ERROR("disabled ip block: %d <%s>\n",
1634                                   i, adev->ip_blocks[i].version->funcs->name);
1635                         adev->ip_blocks[i].status.valid = false;
1636                 } else {
1637                         if (adev->ip_blocks[i].version->funcs->early_init) {
1638                                 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1639                                 if (r == -ENOENT) {
1640                                         adev->ip_blocks[i].status.valid = false;
1641                                 } else if (r) {
1642                                         DRM_ERROR("early_init of IP block <%s> failed %d\n",
1643                                                   adev->ip_blocks[i].version->funcs->name, r);
1644                                         return r;
1645                                 } else {
1646                                         adev->ip_blocks[i].status.valid = true;
1647                                 }
1648                         } else {
1649                                 adev->ip_blocks[i].status.valid = true;
1650                         }
1651                 }
1652         }
1653
1654         adev->cg_flags &= amdgpu_cg_mask;
1655         adev->pg_flags &= amdgpu_pg_mask;
1656
1657         return 0;
1658 }
1659
1660 static int amdgpu_init(struct amdgpu_device *adev)
1661 {
1662         int i, r;
1663
1664         for (i = 0; i < adev->num_ip_blocks; i++) {
1665                 if (!adev->ip_blocks[i].status.valid)
1666                         continue;
1667                 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1668                 if (r) {
1669                         DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1670                                   adev->ip_blocks[i].version->funcs->name, r);
1671                         return r;
1672                 }
1673                 adev->ip_blocks[i].status.sw = true;
1674                 /* need to do gmc hw init early so we can allocate gpu mem */
1675                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1676                         r = amdgpu_vram_scratch_init(adev);
1677                         if (r) {
1678                                 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1679                                 return r;
1680                         }
1681                         r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1682                         if (r) {
1683                                 DRM_ERROR("hw_init %d failed %d\n", i, r);
1684                                 return r;
1685                         }
1686                         r = amdgpu_wb_init(adev);
1687                         if (r) {
1688                                 DRM_ERROR("amdgpu_wb_init failed %d\n", r);
1689                                 return r;
1690                         }
1691                         adev->ip_blocks[i].status.hw = true;
1692
1693                         /* right after GMC hw init, we create CSA */
1694                         if (amdgpu_sriov_vf(adev)) {
1695                                 r = amdgpu_allocate_static_csa(adev);
1696                                 if (r) {
1697                                         DRM_ERROR("allocate CSA failed %d\n", r);
1698                                         return r;
1699                                 }
1700                         }
1701                 }
1702         }
1703
1704         for (i = 0; i < adev->num_ip_blocks; i++) {
1705                 if (!adev->ip_blocks[i].status.sw)
1706                         continue;
1707                 /* gmc hw init is done early */
1708                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC)
1709                         continue;
1710                 r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1711                 if (r) {
1712                         DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1713                                   adev->ip_blocks[i].version->funcs->name, r);
1714                         return r;
1715                 }
1716                 adev->ip_blocks[i].status.hw = true;
1717         }
1718
1719         return 0;
1720 }
1721
1722 static void amdgpu_fill_reset_magic(struct amdgpu_device *adev)
1723 {
1724         memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1725 }
1726
1727 static bool amdgpu_check_vram_lost(struct amdgpu_device *adev)
1728 {
1729         return !!memcmp(adev->gart.ptr, adev->reset_magic,
1730                         AMDGPU_RESET_MAGIC_NUM);
1731 }
1732
1733 static int amdgpu_late_set_cg_state(struct amdgpu_device *adev)
1734 {
1735         int i = 0, r;
1736
1737         for (i = 0; i < adev->num_ip_blocks; i++) {
1738                 if (!adev->ip_blocks[i].status.valid)
1739                         continue;
1740                 /* skip CG for VCE/UVD, it's handled specially */
1741                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1742                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE) {
1743                         /* enable clockgating to save power */
1744                         r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1745                                                                                      AMD_CG_STATE_GATE);
1746                         if (r) {
1747                                 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
1748                                           adev->ip_blocks[i].version->funcs->name, r);
1749                                 return r;
1750                         }
1751                 }
1752         }
1753         return 0;
1754 }
1755
1756 static int amdgpu_late_init(struct amdgpu_device *adev)
1757 {
1758         int i = 0, r;
1759
1760         for (i = 0; i < adev->num_ip_blocks; i++) {
1761                 if (!adev->ip_blocks[i].status.valid)
1762                         continue;
1763                 if (adev->ip_blocks[i].version->funcs->late_init) {
1764                         r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
1765                         if (r) {
1766                                 DRM_ERROR("late_init of IP block <%s> failed %d\n",
1767                                           adev->ip_blocks[i].version->funcs->name, r);
1768                                 return r;
1769                         }
1770                         adev->ip_blocks[i].status.late_initialized = true;
1771                 }
1772         }
1773
1774         mod_delayed_work(system_wq, &adev->late_init_work,
1775                         msecs_to_jiffies(AMDGPU_RESUME_MS));
1776
1777         amdgpu_fill_reset_magic(adev);
1778
1779         return 0;
1780 }
1781
1782 static int amdgpu_fini(struct amdgpu_device *adev)
1783 {
1784         int i, r;
1785
1786         /* need to disable SMC first */
1787         for (i = 0; i < adev->num_ip_blocks; i++) {
1788                 if (!adev->ip_blocks[i].status.hw)
1789                         continue;
1790                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
1791                         /* ungate blocks before hw fini so that we can shutdown the blocks safely */
1792                         r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1793                                                                                      AMD_CG_STATE_UNGATE);
1794                         if (r) {
1795                                 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1796                                           adev->ip_blocks[i].version->funcs->name, r);
1797                                 return r;
1798                         }
1799                         r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1800                         /* XXX handle errors */
1801                         if (r) {
1802                                 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1803                                           adev->ip_blocks[i].version->funcs->name, r);
1804                         }
1805                         adev->ip_blocks[i].status.hw = false;
1806                         break;
1807                 }
1808         }
1809
1810         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1811                 if (!adev->ip_blocks[i].status.hw)
1812                         continue;
1813                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1814                         amdgpu_wb_fini(adev);
1815                         amdgpu_vram_scratch_fini(adev);
1816                 }
1817
1818                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
1819                         adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE) {
1820                         /* ungate blocks before hw fini so that we can shutdown the blocks safely */
1821                         r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1822                                                                                      AMD_CG_STATE_UNGATE);
1823                         if (r) {
1824                                 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1825                                           adev->ip_blocks[i].version->funcs->name, r);
1826                                 return r;
1827                         }
1828                 }
1829
1830                 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
1831                 /* XXX handle errors */
1832                 if (r) {
1833                         DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1834                                   adev->ip_blocks[i].version->funcs->name, r);
1835                 }
1836
1837                 adev->ip_blocks[i].status.hw = false;
1838         }
1839
1840         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1841                 if (!adev->ip_blocks[i].status.sw)
1842                         continue;
1843                 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
1844                 /* XXX handle errors */
1845                 if (r) {
1846                         DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
1847                                   adev->ip_blocks[i].version->funcs->name, r);
1848                 }
1849                 adev->ip_blocks[i].status.sw = false;
1850                 adev->ip_blocks[i].status.valid = false;
1851         }
1852
1853         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1854                 if (!adev->ip_blocks[i].status.late_initialized)
1855                         continue;
1856                 if (adev->ip_blocks[i].version->funcs->late_fini)
1857                         adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
1858                 adev->ip_blocks[i].status.late_initialized = false;
1859         }
1860
1861         if (amdgpu_sriov_vf(adev))
1862                 amdgpu_virt_release_full_gpu(adev, false);
1863
1864         return 0;
1865 }
1866
1867 static void amdgpu_late_init_func_handler(struct work_struct *work)
1868 {
1869         struct amdgpu_device *adev =
1870                 container_of(work, struct amdgpu_device, late_init_work.work);
1871         amdgpu_late_set_cg_state(adev);
1872 }
1873
1874 int amdgpu_suspend(struct amdgpu_device *adev)
1875 {
1876         int i, r;
1877
1878         if (amdgpu_sriov_vf(adev))
1879                 amdgpu_virt_request_full_gpu(adev, false);
1880
1881         /* ungate SMC block first */
1882         r = amdgpu_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_SMC,
1883                                          AMD_CG_STATE_UNGATE);
1884         if (r) {
1885                 DRM_ERROR("set_clockgating_state(ungate) SMC failed %d\n",r);
1886         }
1887
1888         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1889                 if (!adev->ip_blocks[i].status.valid)
1890                         continue;
1891                 /* ungate blocks so that suspend can properly shut them down */
1892                 if (i != AMD_IP_BLOCK_TYPE_SMC) {
1893                         r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
1894                                                                                      AMD_CG_STATE_UNGATE);
1895                         if (r) {
1896                                 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1897                                           adev->ip_blocks[i].version->funcs->name, r);
1898                         }
1899                 }
1900                 /* XXX handle errors */
1901                 r = adev->ip_blocks[i].version->funcs->suspend(adev);
1902                 /* XXX handle errors */
1903                 if (r) {
1904                         DRM_ERROR("suspend of IP block <%s> failed %d\n",
1905                                   adev->ip_blocks[i].version->funcs->name, r);
1906                 }
1907         }
1908
1909         if (amdgpu_sriov_vf(adev))
1910                 amdgpu_virt_release_full_gpu(adev, false);
1911
1912         return 0;
1913 }
1914
1915 static int amdgpu_sriov_reinit_early(struct amdgpu_device *adev)
1916 {
1917         int i, r;
1918
1919         static enum amd_ip_block_type ip_order[] = {
1920                 AMD_IP_BLOCK_TYPE_GMC,
1921                 AMD_IP_BLOCK_TYPE_COMMON,
1922                 AMD_IP_BLOCK_TYPE_IH,
1923         };
1924
1925         for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
1926                 int j;
1927                 struct amdgpu_ip_block *block;
1928
1929                 for (j = 0; j < adev->num_ip_blocks; j++) {
1930                         block = &adev->ip_blocks[j];
1931
1932                         if (block->version->type != ip_order[i] ||
1933                                 !block->status.valid)
1934                                 continue;
1935
1936                         r = block->version->funcs->hw_init(adev);
1937                         DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
1938                 }
1939         }
1940
1941         return 0;
1942 }
1943
1944 static int amdgpu_sriov_reinit_late(struct amdgpu_device *adev)
1945 {
1946         int i, r;
1947
1948         static enum amd_ip_block_type ip_order[] = {
1949                 AMD_IP_BLOCK_TYPE_SMC,
1950                 AMD_IP_BLOCK_TYPE_PSP,
1951                 AMD_IP_BLOCK_TYPE_DCE,
1952                 AMD_IP_BLOCK_TYPE_GFX,
1953                 AMD_IP_BLOCK_TYPE_SDMA,
1954                 AMD_IP_BLOCK_TYPE_UVD,
1955                 AMD_IP_BLOCK_TYPE_VCE
1956         };
1957
1958         for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
1959                 int j;
1960                 struct amdgpu_ip_block *block;
1961
1962                 for (j = 0; j < adev->num_ip_blocks; j++) {
1963                         block = &adev->ip_blocks[j];
1964
1965                         if (block->version->type != ip_order[i] ||
1966                                 !block->status.valid)
1967                                 continue;
1968
1969                         r = block->version->funcs->hw_init(adev);
1970                         DRM_INFO("RE-INIT: %s %s\n", block->version->funcs->name, r?"failed":"successed");
1971                 }
1972         }
1973
1974         return 0;
1975 }
1976
1977 static int amdgpu_resume_phase1(struct amdgpu_device *adev)
1978 {
1979         int i, r;
1980
1981         for (i = 0; i < adev->num_ip_blocks; i++) {
1982                 if (!adev->ip_blocks[i].status.valid)
1983                         continue;
1984                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
1985                                 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
1986                                 adev->ip_blocks[i].version->type ==
1987                                 AMD_IP_BLOCK_TYPE_IH) {
1988                         r = adev->ip_blocks[i].version->funcs->resume(adev);
1989                         if (r) {
1990                                 DRM_ERROR("resume of IP block <%s> failed %d\n",
1991                                           adev->ip_blocks[i].version->funcs->name, r);
1992                                 return r;
1993                         }
1994                 }
1995         }
1996
1997         return 0;
1998 }
1999
2000 static int amdgpu_resume_phase2(struct amdgpu_device *adev)
2001 {
2002         int i, r;
2003
2004         for (i = 0; i < adev->num_ip_blocks; i++) {
2005                 if (!adev->ip_blocks[i].status.valid)
2006                         continue;
2007                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2008                                 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2009                                 adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH )
2010                         continue;
2011                 r = adev->ip_blocks[i].version->funcs->resume(adev);
2012                 if (r) {
2013                         DRM_ERROR("resume of IP block <%s> failed %d\n",
2014                                   adev->ip_blocks[i].version->funcs->name, r);
2015                         return r;
2016                 }
2017         }
2018
2019         return 0;
2020 }
2021
2022 static int amdgpu_resume(struct amdgpu_device *adev)
2023 {
2024         int r;
2025
2026         r = amdgpu_resume_phase1(adev);
2027         if (r)
2028                 return r;
2029         r = amdgpu_resume_phase2(adev);
2030
2031         return r;
2032 }
2033
2034 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
2035 {
2036         if (amdgpu_sriov_vf(adev)) {
2037                 if (adev->is_atom_fw) {
2038                         if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2039                                 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2040                 } else {
2041                         if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2042                                 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2043                 }
2044
2045                 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2046                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
2047         }
2048 }
2049
2050 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2051 {
2052         switch (asic_type) {
2053 #if defined(CONFIG_DRM_AMD_DC)
2054         case CHIP_BONAIRE:
2055         case CHIP_HAWAII:
2056         case CHIP_KAVERI:
2057         case CHIP_CARRIZO:
2058         case CHIP_STONEY:
2059         case CHIP_POLARIS11:
2060         case CHIP_POLARIS10:
2061         case CHIP_POLARIS12:
2062         case CHIP_TONGA:
2063         case CHIP_FIJI:
2064 #if defined(CONFIG_DRM_AMD_DC_PRE_VEGA)
2065                 return amdgpu_dc != 0;
2066 #endif
2067         case CHIP_KABINI:
2068         case CHIP_MULLINS:
2069                 return amdgpu_dc > 0;
2070         case CHIP_VEGA10:
2071 #if defined(CONFIG_DRM_AMD_DC_DCN1_0)
2072         case CHIP_RAVEN:
2073 #endif
2074                 return amdgpu_dc != 0;
2075 #endif
2076         default:
2077                 return false;
2078         }
2079 }
2080
2081 /**
2082  * amdgpu_device_has_dc_support - check if dc is supported
2083  *
2084  * @adev: amdgpu_device_pointer
2085  *
2086  * Returns true for supported, false for not supported
2087  */
2088 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2089 {
2090         if (amdgpu_sriov_vf(adev))
2091                 return false;
2092
2093         return amdgpu_device_asic_has_dc_support(adev->asic_type);
2094 }
2095
2096 /**
2097  * amdgpu_device_init - initialize the driver
2098  *
2099  * @adev: amdgpu_device pointer
2100  * @pdev: drm dev pointer
2101  * @pdev: pci dev pointer
2102  * @flags: driver flags
2103  *
2104  * Initializes the driver info and hw (all asics).
2105  * Returns 0 for success or an error on failure.
2106  * Called at driver startup.
2107  */
2108 int amdgpu_device_init(struct amdgpu_device *adev,
2109                        struct drm_device *ddev,
2110                        struct pci_dev *pdev,
2111                        uint32_t flags)
2112 {
2113         int r, i;
2114         bool runtime = false;
2115         u32 max_MBps;
2116
2117         adev->shutdown = false;
2118         adev->dev = &pdev->dev;
2119         adev->ddev = ddev;
2120         adev->pdev = pdev;
2121         adev->flags = flags;
2122         adev->asic_type = flags & AMD_ASIC_MASK;
2123         adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
2124         adev->mc.gart_size = 512 * 1024 * 1024;
2125         adev->accel_working = false;
2126         adev->num_rings = 0;
2127         adev->mman.buffer_funcs = NULL;
2128         adev->mman.buffer_funcs_ring = NULL;
2129         adev->vm_manager.vm_pte_funcs = NULL;
2130         adev->vm_manager.vm_pte_num_rings = 0;
2131         adev->gart.gart_funcs = NULL;
2132         adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2133         bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
2134
2135         adev->smc_rreg = &amdgpu_invalid_rreg;
2136         adev->smc_wreg = &amdgpu_invalid_wreg;
2137         adev->pcie_rreg = &amdgpu_invalid_rreg;
2138         adev->pcie_wreg = &amdgpu_invalid_wreg;
2139         adev->pciep_rreg = &amdgpu_invalid_rreg;
2140         adev->pciep_wreg = &amdgpu_invalid_wreg;
2141         adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
2142         adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
2143         adev->didt_rreg = &amdgpu_invalid_rreg;
2144         adev->didt_wreg = &amdgpu_invalid_wreg;
2145         adev->gc_cac_rreg = &amdgpu_invalid_rreg;
2146         adev->gc_cac_wreg = &amdgpu_invalid_wreg;
2147         adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
2148         adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
2149
2150         DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
2151                  amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
2152                  pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
2153
2154         /* mutex initialization are all done here so we
2155          * can recall function without having locking issues */
2156         atomic_set(&adev->irq.ih.lock, 0);
2157         mutex_init(&adev->firmware.mutex);
2158         mutex_init(&adev->pm.mutex);
2159         mutex_init(&adev->gfx.gpu_clock_mutex);
2160         mutex_init(&adev->srbm_mutex);
2161         mutex_init(&adev->gfx.pipe_reserve_mutex);
2162         mutex_init(&adev->grbm_idx_mutex);
2163         mutex_init(&adev->mn_lock);
2164         mutex_init(&adev->virt.vf_errors.lock);
2165         hash_init(adev->mn_hash);
2166
2167         amdgpu_check_arguments(adev);
2168
2169         spin_lock_init(&adev->mmio_idx_lock);
2170         spin_lock_init(&adev->smc_idx_lock);
2171         spin_lock_init(&adev->pcie_idx_lock);
2172         spin_lock_init(&adev->uvd_ctx_idx_lock);
2173         spin_lock_init(&adev->didt_idx_lock);
2174         spin_lock_init(&adev->gc_cac_idx_lock);
2175         spin_lock_init(&adev->se_cac_idx_lock);
2176         spin_lock_init(&adev->audio_endpt_idx_lock);
2177         spin_lock_init(&adev->mm_stats.lock);
2178
2179         INIT_LIST_HEAD(&adev->shadow_list);
2180         mutex_init(&adev->shadow_list_lock);
2181
2182         INIT_LIST_HEAD(&adev->gtt_list);
2183         spin_lock_init(&adev->gtt_list_lock);
2184
2185         INIT_LIST_HEAD(&adev->ring_lru_list);
2186         spin_lock_init(&adev->ring_lru_list_lock);
2187
2188         INIT_DELAYED_WORK(&adev->late_init_work, amdgpu_late_init_func_handler);
2189
2190         /* Registers mapping */
2191         /* TODO: block userspace mapping of io register */
2192         if (adev->asic_type >= CHIP_BONAIRE) {
2193                 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
2194                 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
2195         } else {
2196                 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
2197                 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
2198         }
2199
2200         adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
2201         if (adev->rmmio == NULL) {
2202                 return -ENOMEM;
2203         }
2204         DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
2205         DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
2206
2207         /* doorbell bar mapping */
2208         amdgpu_doorbell_init(adev);
2209
2210         /* io port mapping */
2211         for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
2212                 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
2213                         adev->rio_mem_size = pci_resource_len(adev->pdev, i);
2214                         adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
2215                         break;
2216                 }
2217         }
2218         if (adev->rio_mem == NULL)
2219                 DRM_INFO("PCI I/O BAR is not found.\n");
2220
2221         /* early init functions */
2222         r = amdgpu_early_init(adev);
2223         if (r)
2224                 return r;
2225
2226         /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
2227         /* this will fail for cards that aren't VGA class devices, just
2228          * ignore it */
2229         vga_client_register(adev->pdev, adev, NULL, amdgpu_vga_set_decode);
2230
2231         if (amdgpu_runtime_pm == 1)
2232                 runtime = true;
2233         if (amdgpu_device_is_px(ddev))
2234                 runtime = true;
2235         if (!pci_is_thunderbolt_attached(adev->pdev))
2236                 vga_switcheroo_register_client(adev->pdev,
2237                                                &amdgpu_switcheroo_ops, runtime);
2238         if (runtime)
2239                 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
2240
2241         /* Read BIOS */
2242         if (!amdgpu_get_bios(adev)) {
2243                 r = -EINVAL;
2244                 goto failed;
2245         }
2246
2247         r = amdgpu_atombios_init(adev);
2248         if (r) {
2249                 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
2250                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
2251                 goto failed;
2252         }
2253
2254         /* detect if we are with an SRIOV vbios */
2255         amdgpu_device_detect_sriov_bios(adev);
2256
2257         /* Post card if necessary */
2258         if (amdgpu_need_post(adev)) {
2259                 if (!adev->bios) {
2260                         dev_err(adev->dev, "no vBIOS found\n");
2261                         r = -EINVAL;
2262                         goto failed;
2263                 }
2264                 DRM_INFO("GPU posting now...\n");
2265                 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2266                 if (r) {
2267                         dev_err(adev->dev, "gpu post error!\n");
2268                         goto failed;
2269                 }
2270         } else {
2271                 DRM_INFO("GPU post is not needed\n");
2272         }
2273
2274         if (adev->is_atom_fw) {
2275                 /* Initialize clocks */
2276                 r = amdgpu_atomfirmware_get_clock_info(adev);
2277                 if (r) {
2278                         dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
2279                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2280                         goto failed;
2281                 }
2282         } else {
2283                 /* Initialize clocks */
2284                 r = amdgpu_atombios_get_clock_info(adev);
2285                 if (r) {
2286                         dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
2287                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
2288                         goto failed;
2289                 }
2290                 /* init i2c buses */
2291                 if (!amdgpu_device_has_dc_support(adev))
2292                         amdgpu_atombios_i2c_init(adev);
2293         }
2294
2295         /* Fence driver */
2296         r = amdgpu_fence_driver_init(adev);
2297         if (r) {
2298                 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
2299                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
2300                 goto failed;
2301         }
2302
2303         /* init the mode config */
2304         drm_mode_config_init(adev->ddev);
2305
2306         r = amdgpu_init(adev);
2307         if (r) {
2308                 dev_err(adev->dev, "amdgpu_init failed\n");
2309                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
2310                 amdgpu_fini(adev);
2311                 goto failed;
2312         }
2313
2314         adev->accel_working = true;
2315
2316         amdgpu_vm_check_compute_bug(adev);
2317
2318         /* Initialize the buffer migration limit. */
2319         if (amdgpu_moverate >= 0)
2320                 max_MBps = amdgpu_moverate;
2321         else
2322                 max_MBps = 8; /* Allow 8 MB/s. */
2323         /* Get a log2 for easy divisions. */
2324         adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
2325
2326         r = amdgpu_ib_pool_init(adev);
2327         if (r) {
2328                 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
2329                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
2330                 goto failed;
2331         }
2332
2333         r = amdgpu_ib_ring_tests(adev);
2334         if (r)
2335                 DRM_ERROR("ib ring test failed (%d).\n", r);
2336
2337         if (amdgpu_sriov_vf(adev))
2338                 amdgpu_virt_init_data_exchange(adev);
2339
2340         amdgpu_fbdev_init(adev);
2341
2342         r = amdgpu_pm_sysfs_init(adev);
2343         if (r)
2344                 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
2345
2346         r = amdgpu_gem_debugfs_init(adev);
2347         if (r)
2348                 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
2349
2350         r = amdgpu_debugfs_regs_init(adev);
2351         if (r)
2352                 DRM_ERROR("registering register debugfs failed (%d).\n", r);
2353
2354         r = amdgpu_debugfs_test_ib_ring_init(adev);
2355         if (r)
2356                 DRM_ERROR("registering register test ib ring debugfs failed (%d).\n", r);
2357
2358         r = amdgpu_debugfs_firmware_init(adev);
2359         if (r)
2360                 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
2361
2362         r = amdgpu_debugfs_vbios_dump_init(adev);
2363         if (r)
2364                 DRM_ERROR("Creating vbios dump debugfs failed (%d).\n", r);
2365
2366         if ((amdgpu_testing & 1)) {
2367                 if (adev->accel_working)
2368                         amdgpu_test_moves(adev);
2369                 else
2370                         DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
2371         }
2372         if (amdgpu_benchmarking) {
2373                 if (adev->accel_working)
2374                         amdgpu_benchmark(adev, amdgpu_benchmarking);
2375                 else
2376                         DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
2377         }
2378
2379         /* enable clockgating, etc. after ib tests, etc. since some blocks require
2380          * explicit gating rather than handling it automatically.
2381          */
2382         r = amdgpu_late_init(adev);
2383         if (r) {
2384                 dev_err(adev->dev, "amdgpu_late_init failed\n");
2385                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
2386                 goto failed;
2387         }
2388
2389         return 0;
2390
2391 failed:
2392         amdgpu_vf_error_trans_all(adev);
2393         if (runtime)
2394                 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2395         return r;
2396 }
2397
2398 /**
2399  * amdgpu_device_fini - tear down the driver
2400  *
2401  * @adev: amdgpu_device pointer
2402  *
2403  * Tear down the driver info (all asics).
2404  * Called at driver shutdown.
2405  */
2406 void amdgpu_device_fini(struct amdgpu_device *adev)
2407 {
2408         int r;
2409
2410         DRM_INFO("amdgpu: finishing device.\n");
2411         adev->shutdown = true;
2412         if (adev->mode_info.mode_config_initialized)
2413                 drm_crtc_force_disable_all(adev->ddev);
2414         /* evict vram memory */
2415         amdgpu_bo_evict_vram(adev);
2416         amdgpu_ib_pool_fini(adev);
2417         amdgpu_fw_reserve_vram_fini(adev);
2418         amdgpu_fence_driver_fini(adev);
2419         amdgpu_fbdev_fini(adev);
2420         r = amdgpu_fini(adev);
2421         if (adev->firmware.gpu_info_fw) {
2422                 release_firmware(adev->firmware.gpu_info_fw);
2423                 adev->firmware.gpu_info_fw = NULL;
2424         }
2425         adev->accel_working = false;
2426         cancel_delayed_work_sync(&adev->late_init_work);
2427         /* free i2c buses */
2428         if (!amdgpu_device_has_dc_support(adev))
2429                 amdgpu_i2c_fini(adev);
2430         amdgpu_atombios_fini(adev);
2431         kfree(adev->bios);
2432         adev->bios = NULL;
2433         if (!pci_is_thunderbolt_attached(adev->pdev))
2434                 vga_switcheroo_unregister_client(adev->pdev);
2435         if (adev->flags & AMD_IS_PX)
2436                 vga_switcheroo_fini_domain_pm_ops(adev->dev);
2437         vga_client_register(adev->pdev, NULL, NULL, NULL);
2438         if (adev->rio_mem)
2439                 pci_iounmap(adev->pdev, adev->rio_mem);
2440         adev->rio_mem = NULL;
2441         iounmap(adev->rmmio);
2442         adev->rmmio = NULL;
2443         amdgpu_doorbell_fini(adev);
2444         amdgpu_pm_sysfs_fini(adev);
2445         amdgpu_debugfs_regs_cleanup(adev);
2446 }
2447
2448
2449 /*
2450  * Suspend & resume.
2451  */
2452 /**
2453  * amdgpu_device_suspend - initiate device suspend
2454  *
2455  * @pdev: drm dev pointer
2456  * @state: suspend state
2457  *
2458  * Puts the hw in the suspend state (all asics).
2459  * Returns 0 for success or an error on failure.
2460  * Called at driver suspend.
2461  */
2462 int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
2463 {
2464         struct amdgpu_device *adev;
2465         struct drm_crtc *crtc;
2466         struct drm_connector *connector;
2467         int r;
2468
2469         if (dev == NULL || dev->dev_private == NULL) {
2470                 return -ENODEV;
2471         }
2472
2473         adev = dev->dev_private;
2474
2475         if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2476                 return 0;
2477
2478         drm_kms_helper_poll_disable(dev);
2479
2480         if (!amdgpu_device_has_dc_support(adev)) {
2481                 /* turn off display hw */
2482                 drm_modeset_lock_all(dev);
2483                 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2484                         drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
2485                 }
2486                 drm_modeset_unlock_all(dev);
2487         }
2488
2489         amdgpu_amdkfd_suspend(adev);
2490
2491         /* unpin the front buffers and cursors */
2492         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2493                 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2494                 struct amdgpu_framebuffer *rfb = to_amdgpu_framebuffer(crtc->primary->fb);
2495                 struct amdgpu_bo *robj;
2496
2497                 if (amdgpu_crtc->cursor_bo) {
2498                         struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2499                         r = amdgpu_bo_reserve(aobj, true);
2500                         if (r == 0) {
2501                                 amdgpu_bo_unpin(aobj);
2502                                 amdgpu_bo_unreserve(aobj);
2503                         }
2504                 }
2505
2506                 if (rfb == NULL || rfb->obj == NULL) {
2507                         continue;
2508                 }
2509                 robj = gem_to_amdgpu_bo(rfb->obj);
2510                 /* don't unpin kernel fb objects */
2511                 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
2512                         r = amdgpu_bo_reserve(robj, true);
2513                         if (r == 0) {
2514                                 amdgpu_bo_unpin(robj);
2515                                 amdgpu_bo_unreserve(robj);
2516                         }
2517                 }
2518         }
2519         /* evict vram memory */
2520         amdgpu_bo_evict_vram(adev);
2521
2522         amdgpu_fence_driver_suspend(adev);
2523
2524         r = amdgpu_suspend(adev);
2525
2526         /* evict remaining vram memory
2527          * This second call to evict vram is to evict the gart page table
2528          * using the CPU.
2529          */
2530         amdgpu_bo_evict_vram(adev);
2531
2532         amdgpu_atombios_scratch_regs_save(adev);
2533         pci_save_state(dev->pdev);
2534         if (suspend) {
2535                 /* Shut down the device */
2536                 pci_disable_device(dev->pdev);
2537                 pci_set_power_state(dev->pdev, PCI_D3hot);
2538         } else {
2539                 r = amdgpu_asic_reset(adev);
2540                 if (r)
2541                         DRM_ERROR("amdgpu asic reset failed\n");
2542         }
2543
2544         if (fbcon) {
2545                 console_lock();
2546                 amdgpu_fbdev_set_suspend(adev, 1);
2547                 console_unlock();
2548         }
2549         return 0;
2550 }
2551
2552 /**
2553  * amdgpu_device_resume - initiate device resume
2554  *
2555  * @pdev: drm dev pointer
2556  *
2557  * Bring the hw back to operating state (all asics).
2558  * Returns 0 for success or an error on failure.
2559  * Called at driver resume.
2560  */
2561 int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
2562 {
2563         struct drm_connector *connector;
2564         struct amdgpu_device *adev = dev->dev_private;
2565         struct drm_crtc *crtc;
2566         int r = 0;
2567
2568         if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
2569                 return 0;
2570
2571         if (fbcon)
2572                 console_lock();
2573
2574         if (resume) {
2575                 pci_set_power_state(dev->pdev, PCI_D0);
2576                 pci_restore_state(dev->pdev);
2577                 r = pci_enable_device(dev->pdev);
2578                 if (r)
2579                         goto unlock;
2580         }
2581         amdgpu_atombios_scratch_regs_restore(adev);
2582
2583         /* post card */
2584         if (amdgpu_need_post(adev)) {
2585                 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2586                 if (r)
2587                         DRM_ERROR("amdgpu asic init failed\n");
2588         }
2589
2590         r = amdgpu_resume(adev);
2591         if (r) {
2592                 DRM_ERROR("amdgpu_resume failed (%d).\n", r);
2593                 goto unlock;
2594         }
2595         amdgpu_fence_driver_resume(adev);
2596
2597         if (resume) {
2598                 r = amdgpu_ib_ring_tests(adev);
2599                 if (r)
2600                         DRM_ERROR("ib ring test failed (%d).\n", r);
2601         }
2602
2603         r = amdgpu_late_init(adev);
2604         if (r)
2605                 goto unlock;
2606
2607         /* pin cursors */
2608         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2609                 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2610
2611                 if (amdgpu_crtc->cursor_bo) {
2612                         struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2613                         r = amdgpu_bo_reserve(aobj, true);
2614                         if (r == 0) {
2615                                 r = amdgpu_bo_pin(aobj,
2616                                                   AMDGPU_GEM_DOMAIN_VRAM,
2617                                                   &amdgpu_crtc->cursor_addr);
2618                                 if (r != 0)
2619                                         DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
2620                                 amdgpu_bo_unreserve(aobj);
2621                         }
2622                 }
2623         }
2624         r = amdgpu_amdkfd_resume(adev);
2625         if (r)
2626                 return r;
2627
2628         /* blat the mode back in */
2629         if (fbcon) {
2630                 if (!amdgpu_device_has_dc_support(adev)) {
2631                         /* pre DCE11 */
2632                         drm_helper_resume_force_mode(dev);
2633
2634                         /* turn on display hw */
2635                         drm_modeset_lock_all(dev);
2636                         list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2637                                 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON);
2638                         }
2639                         drm_modeset_unlock_all(dev);
2640                 } else {
2641                         /*
2642                          * There is no equivalent atomic helper to turn on
2643                          * display, so we defined our own function for this,
2644                          * once suspend resume is supported by the atomic
2645                          * framework this will be reworked
2646                          */
2647                         amdgpu_dm_display_resume(adev);
2648                 }
2649         }
2650
2651         drm_kms_helper_poll_enable(dev);
2652
2653         /*
2654          * Most of the connector probing functions try to acquire runtime pm
2655          * refs to ensure that the GPU is powered on when connector polling is
2656          * performed. Since we're calling this from a runtime PM callback,
2657          * trying to acquire rpm refs will cause us to deadlock.
2658          *
2659          * Since we're guaranteed to be holding the rpm lock, it's safe to
2660          * temporarily disable the rpm helpers so this doesn't deadlock us.
2661          */
2662 #ifdef CONFIG_PM
2663         dev->dev->power.disable_depth++;
2664 #endif
2665         if (!amdgpu_device_has_dc_support(adev))
2666                 drm_helper_hpd_irq_event(dev);
2667         else
2668                 drm_kms_helper_hotplug_event(dev);
2669 #ifdef CONFIG_PM
2670         dev->dev->power.disable_depth--;
2671 #endif
2672
2673         if (fbcon)
2674                 amdgpu_fbdev_set_suspend(adev, 0);
2675
2676 unlock:
2677         if (fbcon)
2678                 console_unlock();
2679
2680         return r;
2681 }
2682
2683 static bool amdgpu_check_soft_reset(struct amdgpu_device *adev)
2684 {
2685         int i;
2686         bool asic_hang = false;
2687
2688         if (amdgpu_sriov_vf(adev))
2689                 return true;
2690
2691         for (i = 0; i < adev->num_ip_blocks; i++) {
2692                 if (!adev->ip_blocks[i].status.valid)
2693                         continue;
2694                 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
2695                         adev->ip_blocks[i].status.hang =
2696                                 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
2697                 if (adev->ip_blocks[i].status.hang) {
2698                         DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
2699                         asic_hang = true;
2700                 }
2701         }
2702         return asic_hang;
2703 }
2704
2705 static int amdgpu_pre_soft_reset(struct amdgpu_device *adev)
2706 {
2707         int i, r = 0;
2708
2709         for (i = 0; i < adev->num_ip_blocks; i++) {
2710                 if (!adev->ip_blocks[i].status.valid)
2711                         continue;
2712                 if (adev->ip_blocks[i].status.hang &&
2713                     adev->ip_blocks[i].version->funcs->pre_soft_reset) {
2714                         r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
2715                         if (r)
2716                                 return r;
2717                 }
2718         }
2719
2720         return 0;
2721 }
2722
2723 static bool amdgpu_need_full_reset(struct amdgpu_device *adev)
2724 {
2725         int i;
2726
2727         for (i = 0; i < adev->num_ip_blocks; i++) {
2728                 if (!adev->ip_blocks[i].status.valid)
2729                         continue;
2730                 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
2731                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
2732                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
2733                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
2734                      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
2735                         if (adev->ip_blocks[i].status.hang) {
2736                                 DRM_INFO("Some block need full reset!\n");
2737                                 return true;
2738                         }
2739                 }
2740         }
2741         return false;
2742 }
2743
2744 static int amdgpu_soft_reset(struct amdgpu_device *adev)
2745 {
2746         int i, r = 0;
2747
2748         for (i = 0; i < adev->num_ip_blocks; i++) {
2749                 if (!adev->ip_blocks[i].status.valid)
2750                         continue;
2751                 if (adev->ip_blocks[i].status.hang &&
2752                     adev->ip_blocks[i].version->funcs->soft_reset) {
2753                         r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
2754                         if (r)
2755                                 return r;
2756                 }
2757         }
2758
2759         return 0;
2760 }
2761
2762 static int amdgpu_post_soft_reset(struct amdgpu_device *adev)
2763 {
2764         int i, r = 0;
2765
2766         for (i = 0; i < adev->num_ip_blocks; i++) {
2767                 if (!adev->ip_blocks[i].status.valid)
2768                         continue;
2769                 if (adev->ip_blocks[i].status.hang &&
2770                     adev->ip_blocks[i].version->funcs->post_soft_reset)
2771                         r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
2772                 if (r)
2773                         return r;
2774         }
2775
2776         return 0;
2777 }
2778
2779 bool amdgpu_need_backup(struct amdgpu_device *adev)
2780 {
2781         if (adev->flags & AMD_IS_APU)
2782                 return false;
2783
2784         return amdgpu_lockup_timeout > 0 ? true : false;
2785 }
2786
2787 static int amdgpu_recover_vram_from_shadow(struct amdgpu_device *adev,
2788                                            struct amdgpu_ring *ring,
2789                                            struct amdgpu_bo *bo,
2790                                            struct dma_fence **fence)
2791 {
2792         uint32_t domain;
2793         int r;
2794
2795         if (!bo->shadow)
2796                 return 0;
2797
2798         r = amdgpu_bo_reserve(bo, true);
2799         if (r)
2800                 return r;
2801         domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
2802         /* if bo has been evicted, then no need to recover */
2803         if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
2804                 r = amdgpu_bo_validate(bo->shadow);
2805                 if (r) {
2806                         DRM_ERROR("bo validate failed!\n");
2807                         goto err;
2808                 }
2809
2810                 r = amdgpu_bo_restore_from_shadow(adev, ring, bo,
2811                                                  NULL, fence, true);
2812                 if (r) {
2813                         DRM_ERROR("recover page table failed!\n");
2814                         goto err;
2815                 }
2816         }
2817 err:
2818         amdgpu_bo_unreserve(bo);
2819         return r;
2820 }
2821
2822 /**
2823  * amdgpu_sriov_gpu_reset - reset the asic
2824  *
2825  * @adev: amdgpu device pointer
2826  * @job: which job trigger hang
2827  *
2828  * Attempt the reset the GPU if it has hung (all asics).
2829  * for SRIOV case.
2830  * Returns 0 for success or an error on failure.
2831  */
2832 int amdgpu_sriov_gpu_reset(struct amdgpu_device *adev, struct amdgpu_job *job)
2833 {
2834         int i, j, r = 0;
2835         int resched;
2836         struct amdgpu_bo *bo, *tmp;
2837         struct amdgpu_ring *ring;
2838         struct dma_fence *fence = NULL, *next = NULL;
2839
2840         mutex_lock(&adev->virt.lock_reset);
2841         atomic_inc(&adev->gpu_reset_counter);
2842         adev->in_sriov_reset = true;
2843
2844         /* block TTM */
2845         resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
2846
2847         /* we start from the ring trigger GPU hang */
2848         j = job ? job->ring->idx : 0;
2849
2850         /* block scheduler */
2851         for (i = j; i < j + AMDGPU_MAX_RINGS; ++i) {
2852                 ring = adev->rings[i % AMDGPU_MAX_RINGS];
2853                 if (!ring || !ring->sched.thread)
2854                         continue;
2855
2856                 kthread_park(ring->sched.thread);
2857
2858                 if (job && j != i)
2859                         continue;
2860
2861                 /* here give the last chance to check if job removed from mirror-list
2862                  * since we already pay some time on kthread_park */
2863                 if (job && list_empty(&job->base.node)) {
2864                         kthread_unpark(ring->sched.thread);
2865                         goto give_up_reset;
2866                 }
2867
2868                 if (amd_sched_invalidate_job(&job->base, amdgpu_job_hang_limit))
2869                         amd_sched_job_kickout(&job->base);
2870
2871                 /* only do job_reset on the hang ring if @job not NULL */
2872                 amd_sched_hw_job_reset(&ring->sched);
2873
2874                 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
2875                 amdgpu_fence_driver_force_completion_ring(ring);
2876         }
2877
2878         /* request to take full control of GPU before re-initialization  */
2879         if (job)
2880                 amdgpu_virt_reset_gpu(adev);
2881         else
2882                 amdgpu_virt_request_full_gpu(adev, true);
2883
2884
2885         /* Resume IP prior to SMC */
2886         amdgpu_sriov_reinit_early(adev);
2887
2888         /* we need recover gart prior to run SMC/CP/SDMA resume */
2889         amdgpu_ttm_recover_gart(adev);
2890
2891         /* now we are okay to resume SMC/CP/SDMA */
2892         amdgpu_sriov_reinit_late(adev);
2893
2894         amdgpu_irq_gpu_reset_resume_helper(adev);
2895
2896         if (amdgpu_ib_ring_tests(adev))
2897                 dev_err(adev->dev, "[GPU_RESET] ib ring test failed (%d).\n", r);
2898
2899         /* release full control of GPU after ib test */
2900         amdgpu_virt_release_full_gpu(adev, true);
2901
2902         DRM_INFO("recover vram bo from shadow\n");
2903
2904         ring = adev->mman.buffer_funcs_ring;
2905         mutex_lock(&adev->shadow_list_lock);
2906         list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
2907                 next = NULL;
2908                 amdgpu_recover_vram_from_shadow(adev, ring, bo, &next);
2909                 if (fence) {
2910                         r = dma_fence_wait(fence, false);
2911                         if (r) {
2912                                 WARN(r, "recovery from shadow isn't completed\n");
2913                                 break;
2914                         }
2915                 }
2916
2917                 dma_fence_put(fence);
2918                 fence = next;
2919         }
2920         mutex_unlock(&adev->shadow_list_lock);
2921
2922         if (fence) {
2923                 r = dma_fence_wait(fence, false);
2924                 if (r)
2925                         WARN(r, "recovery from shadow isn't completed\n");
2926         }
2927         dma_fence_put(fence);
2928
2929         for (i = j; i < j + AMDGPU_MAX_RINGS; ++i) {
2930                 ring = adev->rings[i % AMDGPU_MAX_RINGS];
2931                 if (!ring || !ring->sched.thread)
2932                         continue;
2933
2934                 if (job && j != i) {
2935                         kthread_unpark(ring->sched.thread);
2936                         continue;
2937                 }
2938
2939                 amd_sched_job_recovery(&ring->sched);
2940                 kthread_unpark(ring->sched.thread);
2941         }
2942
2943         drm_helper_resume_force_mode(adev->ddev);
2944 give_up_reset:
2945         ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
2946         if (r) {
2947                 /* bad news, how to tell it to userspace ? */
2948                 dev_info(adev->dev, "GPU reset failed\n");
2949         } else {
2950                 dev_info(adev->dev, "GPU reset successed!\n");
2951         }
2952
2953         adev->in_sriov_reset = false;
2954         mutex_unlock(&adev->virt.lock_reset);
2955         return r;
2956 }
2957
2958 /**
2959  * amdgpu_gpu_reset - reset the asic
2960  *
2961  * @adev: amdgpu device pointer
2962  *
2963  * Attempt the reset the GPU if it has hung (all asics).
2964  * Returns 0 for success or an error on failure.
2965  */
2966 int amdgpu_gpu_reset(struct amdgpu_device *adev)
2967 {
2968         struct drm_atomic_state *state = NULL;
2969         int i, r;
2970         int resched;
2971         bool need_full_reset, vram_lost = false;
2972
2973         if (!amdgpu_check_soft_reset(adev)) {
2974                 DRM_INFO("No hardware hang detected. Did some blocks stall?\n");
2975                 return 0;
2976         }
2977
2978         atomic_inc(&adev->gpu_reset_counter);
2979
2980         /* block TTM */
2981         resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
2982         /* store modesetting */
2983         if (amdgpu_device_has_dc_support(adev))
2984                 state = drm_atomic_helper_suspend(adev->ddev);
2985
2986         /* block scheduler */
2987         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2988                 struct amdgpu_ring *ring = adev->rings[i];
2989
2990                 if (!ring || !ring->sched.thread)
2991                         continue;
2992                 kthread_park(ring->sched.thread);
2993                 amd_sched_hw_job_reset(&ring->sched);
2994         }
2995         /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
2996         amdgpu_fence_driver_force_completion(adev);
2997
2998         need_full_reset = amdgpu_need_full_reset(adev);
2999
3000         if (!need_full_reset) {
3001                 amdgpu_pre_soft_reset(adev);
3002                 r = amdgpu_soft_reset(adev);
3003                 amdgpu_post_soft_reset(adev);
3004                 if (r || amdgpu_check_soft_reset(adev)) {
3005                         DRM_INFO("soft reset failed, will fallback to full reset!\n");
3006                         need_full_reset = true;
3007                 }
3008         }
3009
3010         if (need_full_reset) {
3011                 r = amdgpu_suspend(adev);
3012
3013 retry:
3014                 amdgpu_atombios_scratch_regs_save(adev);
3015                 r = amdgpu_asic_reset(adev);
3016                 amdgpu_atombios_scratch_regs_restore(adev);
3017                 /* post card */
3018                 amdgpu_atom_asic_init(adev->mode_info.atom_context);
3019
3020                 if (!r) {
3021                         dev_info(adev->dev, "GPU reset succeeded, trying to resume\n");
3022                         r = amdgpu_resume_phase1(adev);
3023                         if (r)
3024                                 goto out;
3025                         vram_lost = amdgpu_check_vram_lost(adev);
3026                         if (vram_lost) {
3027                                 DRM_ERROR("VRAM is lost!\n");
3028                                 atomic_inc(&adev->vram_lost_counter);
3029                         }
3030                         r = amdgpu_ttm_recover_gart(adev);
3031                         if (r)
3032                                 goto out;
3033                         r = amdgpu_resume_phase2(adev);
3034                         if (r)
3035                                 goto out;
3036                         if (vram_lost)
3037                                 amdgpu_fill_reset_magic(adev);
3038                 }
3039         }
3040 out:
3041         if (!r) {
3042                 amdgpu_irq_gpu_reset_resume_helper(adev);
3043                 r = amdgpu_ib_ring_tests(adev);
3044                 if (r) {
3045                         dev_err(adev->dev, "ib ring test failed (%d).\n", r);
3046                         r = amdgpu_suspend(adev);
3047                         need_full_reset = true;
3048                         goto retry;
3049                 }
3050                 /**
3051                  * recovery vm page tables, since we cannot depend on VRAM is
3052                  * consistent after gpu full reset.
3053                  */
3054                 if (need_full_reset && amdgpu_need_backup(adev)) {
3055                         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
3056                         struct amdgpu_bo *bo, *tmp;
3057                         struct dma_fence *fence = NULL, *next = NULL;
3058
3059                         DRM_INFO("recover vram bo from shadow\n");
3060                         mutex_lock(&adev->shadow_list_lock);
3061                         list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
3062                                 next = NULL;
3063                                 amdgpu_recover_vram_from_shadow(adev, ring, bo, &next);
3064                                 if (fence) {
3065                                         r = dma_fence_wait(fence, false);
3066                                         if (r) {
3067                                                 WARN(r, "recovery from shadow isn't completed\n");
3068                                                 break;
3069                                         }
3070                                 }
3071
3072                                 dma_fence_put(fence);
3073                                 fence = next;
3074                         }
3075                         mutex_unlock(&adev->shadow_list_lock);
3076                         if (fence) {
3077                                 r = dma_fence_wait(fence, false);
3078                                 if (r)
3079                                         WARN(r, "recovery from shadow isn't completed\n");
3080                         }
3081                         dma_fence_put(fence);
3082                 }
3083                 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3084                         struct amdgpu_ring *ring = adev->rings[i];
3085
3086                         if (!ring || !ring->sched.thread)
3087                                 continue;
3088
3089                         amd_sched_job_recovery(&ring->sched);
3090                         kthread_unpark(ring->sched.thread);
3091                 }
3092         } else {
3093                 dev_err(adev->dev, "asic resume failed (%d).\n", r);
3094                 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3095                         if (adev->rings[i] && adev->rings[i]->sched.thread) {
3096                                 kthread_unpark(adev->rings[i]->sched.thread);
3097                         }
3098                 }
3099         }
3100
3101         if (amdgpu_device_has_dc_support(adev)) {
3102                 r = drm_atomic_helper_resume(adev->ddev, state);
3103                 amdgpu_dm_display_resume(adev);
3104         } else
3105                 drm_helper_resume_force_mode(adev->ddev);
3106
3107         ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
3108         if (r) {
3109                 /* bad news, how to tell it to userspace ? */
3110                 dev_info(adev->dev, "GPU reset failed\n");
3111         }
3112         else {
3113                 dev_info(adev->dev, "GPU reset successed!\n");
3114         }
3115
3116         amdgpu_vf_error_trans_all(adev);
3117         return r;
3118 }
3119
3120 void amdgpu_get_pcie_info(struct amdgpu_device *adev)
3121 {
3122         u32 mask;
3123         int ret;
3124
3125         if (amdgpu_pcie_gen_cap)
3126                 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
3127
3128         if (amdgpu_pcie_lane_cap)
3129                 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
3130
3131         /* covers APUs as well */
3132         if (pci_is_root_bus(adev->pdev->bus)) {
3133                 if (adev->pm.pcie_gen_mask == 0)
3134                         adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
3135                 if (adev->pm.pcie_mlw_mask == 0)
3136                         adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
3137                 return;
3138         }
3139
3140         if (adev->pm.pcie_gen_mask == 0) {
3141                 ret = drm_pcie_get_speed_cap_mask(adev->ddev, &mask);
3142                 if (!ret) {
3143                         adev->pm.pcie_gen_mask = (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
3144                                                   CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
3145                                                   CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
3146
3147                         if (mask & DRM_PCIE_SPEED_25)
3148                                 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
3149                         if (mask & DRM_PCIE_SPEED_50)
3150                                 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2;
3151                         if (mask & DRM_PCIE_SPEED_80)
3152                                 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3;
3153                 } else {
3154                         adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
3155                 }
3156         }
3157         if (adev->pm.pcie_mlw_mask == 0) {
3158                 ret = drm_pcie_get_max_link_width(adev->ddev, &mask);
3159                 if (!ret) {
3160                         switch (mask) {
3161                         case 32:
3162                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
3163                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3164                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3165                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3166                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3167                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3168                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3169                                 break;
3170                         case 16:
3171                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
3172                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3173                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
3174                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
3175                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
3176                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
3177                                 break;
3178                         case 12:
3179                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
3180                                                           CAIL_PCIE_LINK_WID