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

   1 /*
   2  * Copyright 2008 Advanced Micro Devices, Inc.
   3  * Copyright 2008 Red Hat Inc.
   4  * Copyright 2009 Jerome Glisse.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  22  * OTHER DEALINGS IN THE SOFTWARE.
  23  *
  24  * Authors: Dave Airlie
  25  *          Alex Deucher
  26  *          Jerome Glisse
  27  */
  28 #include <linux/power_supply.h>
  29 #include <linux/kthread.h>
  30 #include <linux/module.h>
  31 #include <linux/console.h>
  32 #include <linux/slab.h>
  33
  34 #include <drm/drm_atomic_helper.h>
  35 #include <drm/drm_probe_helper.h>
  36 #include <drm/amdgpu_drm.h>
  37 #include <linux/vgaarb.h>
  38 #include <linux/vga_switcheroo.h>
  39 #include <linux/efi.h>
  40 #include "amdgpu.h"
  41 #include "amdgpu_trace.h"
  42 #include "amdgpu_i2c.h"
  43 #include "atom.h"
  44 #include "amdgpu_atombios.h"
  45 #include "amdgpu_atomfirmware.h"
  46 #include "amd_pcie.h"
  47 #ifdef CONFIG_DRM_AMDGPU_SI
  48 #include "si.h"
  49 #endif
  50 #ifdef CONFIG_DRM_AMDGPU_CIK
  51 #include "cik.h"
  52 #endif
  53 #include "vi.h"
  54 #include "soc15.h"
  55 #include "nv.h"
  56 #include "bif/bif_4_1_d.h"
  57 #include <linux/pci.h>
  58 #include <linux/firmware.h>
  59 #include "amdgpu_vf_error.h"
  60
  61 #include "amdgpu_amdkfd.h"
  62 #include "amdgpu_pm.h"
  63
  64 #include "amdgpu_xgmi.h"
  65 #include "amdgpu_ras.h"
  66 #include "amdgpu_pmu.h"
  67
  68 #include <linux/suspend.h>
  69 #include <drm/task_barrier.h>
  70
  71 MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
  72 MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
  73 MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
  74 MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin");
  75 MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin");
  76 MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin");
  77 MODULE_FIRMWARE("amdgpu/renoir_gpu_info.bin");
  78 MODULE_FIRMWARE("amdgpu/navi10_gpu_info.bin");
  79 MODULE_FIRMWARE("amdgpu/navi14_gpu_info.bin");
  80 MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin");
  81
  82 #define AMDGPU_RESUME_MS                2000
  83
  84 const char *amdgpu_asic_name[] = {
  85         "TAHITI",
  86         "PITCAIRN",
  87         "VERDE",
  88         "OLAND",
  89         "HAINAN",
  90         "BONAIRE",
  91         "KAVERI",
  92         "KABINI",
  93         "HAWAII",
  94         "MULLINS",
  95         "TOPAZ",
  96         "TONGA",
  97         "FIJI",
  98         "CARRIZO",
  99         "STONEY",
 100         "POLARIS10",
 101         "POLARIS11",
 102         "POLARIS12",
 103         "VEGAM",
 104         "VEGA10",
 105         "VEGA12",
 106         "VEGA20",
 107         "RAVEN",
 108         "ARCTURUS",
 109         "RENOIR",
 110         "NAVI10",
 111         "NAVI14",
 112         "NAVI12",
 113         "LAST",
 114 };
 115
 116 /**
 117  * DOC: pcie_replay_count
 118  *
 119  * The amdgpu driver provides a sysfs API for reporting the total number
 120  * of PCIe replays (NAKs)
 121  * The file pcie_replay_count is used for this and returns the total
 122  * number of replays as a sum of the NAKs generated and NAKs received
 123  */
 124
 125 static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev,
 126                 struct device_attribute *attr, char *buf)
 127 {
 128         struct drm_device *ddev = dev_get_drvdata(dev);
 129         struct amdgpu_device *adev = ddev->dev_private;
 130         uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev);
 131
 132         return snprintf(buf, PAGE_SIZE, "%llu\n", cnt);
 133 }
 134
 135 static DEVICE_ATTR(pcie_replay_count, S_IRUGO,
 136                 amdgpu_device_get_pcie_replay_count, NULL);
 137
 138 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
 139
 140 /**
 141  * amdgpu_device_supports_boco - Is the device a dGPU with HG/PX power control
 142  *
 143  * @dev: drm_device pointer
 144  *
 145  * Returns true if the device is a dGPU with HG/PX power control,
 146  * otherwise return false.
 147  */
 148 bool amdgpu_device_supports_boco(struct drm_device *dev)
 149 {
 150         struct amdgpu_device *adev = dev->dev_private;
 151
 152         if (adev->flags & AMD_IS_PX)
 153                 return true;
 154         return false;
 155 }
 156
 157 /**
 158  * amdgpu_device_supports_baco - Does the device support BACO
 159  *
 160  * @dev: drm_device pointer
 161  *
 162  * Returns true if the device supporte BACO,
 163  * otherwise return false.
 164  */
 165 bool amdgpu_device_supports_baco(struct drm_device *dev)
 166 {
 167         struct amdgpu_device *adev = dev->dev_private;
 168
 169         return amdgpu_asic_supports_baco(adev);
 170 }
 171
 172 /**
 173  * VRAM access helper functions.
 174  *
 175  * amdgpu_device_vram_access - read/write a buffer in vram
 176  *
 177  * @adev: amdgpu_device pointer
 178  * @pos: offset of the buffer in vram
 179  * @buf: virtual address of the buffer in system memory
 180  * @size: read/write size, sizeof(@buf) must > @size
 181  * @write: true - write to vram, otherwise - read from vram
 182  */
 183 void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos,
 184                                uint32_t *buf, size_t size, bool write)
 185 {
 186         unsigned long flags;
 187         uint32_t hi = ~0;
 188         uint64_t last;
 189
 190
 191 #ifdef CONFIG_64BIT
 192         last = min(pos + size, adev->gmc.visible_vram_size);
 193         if (last > pos) {
 194                 void __iomem *addr = adev->mman.aper_base_kaddr + pos;
 195                 size_t count = last - pos;
 196
 197                 if (write) {
 198                         memcpy_toio(addr, buf, count);
 199                         mb();
 200                         amdgpu_asic_flush_hdp(adev, NULL);
 201                 } else {
 202                         amdgpu_asic_invalidate_hdp(adev, NULL);
 203                         mb();
 204                         memcpy_fromio(buf, addr, count);
 205                 }
 206
 207                 if (count == size)
 208                         return;
 209
 210                 pos += count;
 211                 buf += count / 4;
 212                 size -= count;
 213         }
 214 #endif
 215
 216         spin_lock_irqsave(&adev->mmio_idx_lock, flags);
 217         for (last = pos + size; pos < last; pos += 4) {
 218                 uint32_t tmp = pos >> 31;
 219
 220                 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)pos) | 0x80000000);
 221                 if (tmp != hi) {
 222                         WREG32_NO_KIQ(mmMM_INDEX_HI, tmp);
 223                         hi = tmp;
 224                 }
 225                 if (write)
 226                         WREG32_NO_KIQ(mmMM_DATA, *buf++);
 227                 else
 228                         *buf++ = RREG32_NO_KIQ(mmMM_DATA);
 229         }
 230         spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
 231 }
 232
 233 /*
 234  * MMIO register access helper functions.
 235  */
 236 /**
 237  * amdgpu_mm_rreg - read a memory mapped IO register
 238  *
 239  * @adev: amdgpu_device pointer
 240  * @reg: dword aligned register offset
 241  * @acc_flags: access flags which require special behavior
 242  *
 243  * Returns the 32 bit value from the offset specified.
 244  */
 245 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
 246                         uint32_t acc_flags)
 247 {
 248         uint32_t ret;
 249
 250         if ((acc_flags & AMDGPU_REGS_KIQ) || (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev)))
 251                 return amdgpu_kiq_rreg(adev, reg);
 252
 253         if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
 254                 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
 255         else {
 256                 unsigned long flags;
 257
 258                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
 259                 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
 260                 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
 261                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
 262         }
 263         trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
 264         return ret;
 265 }
 266
 267 /*
 268  * MMIO register read with bytes helper functions
 269  * @offset:bytes offset from MMIO start
 270  *
 271 */
 272
 273 /**
 274  * amdgpu_mm_rreg8 - read a memory mapped IO register
 275  *
 276  * @adev: amdgpu_device pointer
 277  * @offset: byte aligned register offset
 278  *
 279  * Returns the 8 bit value from the offset specified.
 280  */
 281 uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset) {
 282         if (offset < adev->rmmio_size)
 283                 return (readb(adev->rmmio + offset));
 284         BUG();
 285 }
 286
 287 /*
 288  * MMIO register write with bytes helper functions
 289  * @offset:bytes offset from MMIO start
 290  * @value: the value want to be written to the register
 291  *
 292 */
 293 /**
 294  * amdgpu_mm_wreg8 - read a memory mapped IO register
 295  *
 296  * @adev: amdgpu_device pointer
 297  * @offset: byte aligned register offset
 298  * @value: 8 bit value to write
 299  *
 300  * Writes the value specified to the offset specified.
 301  */
 302 void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value) {
 303         if (offset < adev->rmmio_size)
 304                 writeb(value, adev->rmmio + offset);
 305         else
 306                 BUG();
 307 }
 308
 309 /**
 310  * amdgpu_mm_wreg - write to a memory mapped IO register
 311  *
 312  * @adev: amdgpu_device pointer
 313  * @reg: dword aligned register offset
 314  * @v: 32 bit value to write to the register
 315  * @acc_flags: access flags which require special behavior
 316  *
 317  * Writes the value specified to the offset specified.
 318  */
 319 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
 320                     uint32_t acc_flags)
 321 {
 322         trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
 323
 324         if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
 325                 adev->last_mm_index = v;
 326         }
 327
 328         if ((acc_flags & AMDGPU_REGS_KIQ) || (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev)))
 329                 return amdgpu_kiq_wreg(adev, reg, v);
 330
 331         if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
 332                 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
 333         else {
 334                 unsigned long flags;
 335
 336                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
 337                 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
 338                 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
 339                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
 340         }
 341
 342         if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
 343                 udelay(500);
 344         }
 345 }
 346
 347 /**
 348  * amdgpu_io_rreg - read an IO register
 349  *
 350  * @adev: amdgpu_device pointer
 351  * @reg: dword aligned register offset
 352  *
 353  * Returns the 32 bit value from the offset specified.
 354  */
 355 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
 356 {
 357         if ((reg * 4) < adev->rio_mem_size)
 358                 return ioread32(adev->rio_mem + (reg * 4));
 359         else {
 360                 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
 361                 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
 362         }
 363 }
 364
 365 /**
 366  * amdgpu_io_wreg - write to an IO register
 367  *
 368  * @adev: amdgpu_device pointer
 369  * @reg: dword aligned register offset
 370  * @v: 32 bit value to write to the register
 371  *
 372  * Writes the value specified to the offset specified.
 373  */
 374 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
 375 {
 376         if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
 377                 adev->last_mm_index = v;
 378         }
 379
 380         if ((reg * 4) < adev->rio_mem_size)
 381                 iowrite32(v, adev->rio_mem + (reg * 4));
 382         else {
 383                 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
 384                 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
 385         }
 386
 387         if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
 388                 udelay(500);
 389         }
 390 }
 391
 392 /**
 393  * amdgpu_mm_rdoorbell - read a doorbell dword
 394  *
 395  * @adev: amdgpu_device pointer
 396  * @index: doorbell index
 397  *
 398  * Returns the value in the doorbell aperture at the
 399  * requested doorbell index (CIK).
 400  */
 401 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
 402 {
 403         if (index < adev->doorbell.num_doorbells) {
 404                 return readl(adev->doorbell.ptr + index);
 405         } else {
 406                 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
 407                 return 0;
 408         }
 409 }
 410
 411 /**
 412  * amdgpu_mm_wdoorbell - write a doorbell dword
 413  *
 414  * @adev: amdgpu_device pointer
 415  * @index: doorbell index
 416  * @v: value to write
 417  *
 418  * Writes @v to the doorbell aperture at the
 419  * requested doorbell index (CIK).
 420  */
 421 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
 422 {
 423         if (index < adev->doorbell.num_doorbells) {
 424                 writel(v, adev->doorbell.ptr + index);
 425         } else {
 426                 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
 427         }
 428 }
 429
 430 /**
 431  * amdgpu_mm_rdoorbell64 - read a doorbell Qword
 432  *
 433  * @adev: amdgpu_device pointer
 434  * @index: doorbell index
 435  *
 436  * Returns the value in the doorbell aperture at the
 437  * requested doorbell index (VEGA10+).
 438  */
 439 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
 440 {
 441         if (index < adev->doorbell.num_doorbells) {
 442                 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
 443         } else {
 444                 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
 445                 return 0;
 446         }
 447 }
 448
 449 /**
 450  * amdgpu_mm_wdoorbell64 - write a doorbell Qword
 451  *
 452  * @adev: amdgpu_device pointer
 453  * @index: doorbell index
 454  * @v: value to write
 455  *
 456  * Writes @v to the doorbell aperture at the
 457  * requested doorbell index (VEGA10+).
 458  */
 459 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
 460 {
 461         if (index < adev->doorbell.num_doorbells) {
 462                 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
 463         } else {
 464                 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
 465         }
 466 }
 467
 468 /**
 469  * amdgpu_invalid_rreg - dummy reg read function
 470  *
 471  * @adev: amdgpu device pointer
 472  * @reg: offset of register
 473  *
 474  * Dummy register read function.  Used for register blocks
 475  * that certain asics don't have (all asics).
 476  * Returns the value in the register.
 477  */
 478 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
 479 {
 480         DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
 481         BUG();
 482         return 0;
 483 }
 484
 485 /**
 486  * amdgpu_invalid_wreg - dummy reg write function
 487  *
 488  * @adev: amdgpu device pointer
 489  * @reg: offset of register
 490  * @v: value to write to the register
 491  *
 492  * Dummy register read function.  Used for register blocks
 493  * that certain asics don't have (all asics).
 494  */
 495 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
 496 {
 497         DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
 498                   reg, v);
 499         BUG();
 500 }
 501
 502 /**
 503  * amdgpu_invalid_rreg64 - dummy 64 bit reg read function
 504  *
 505  * @adev: amdgpu device pointer
 506  * @reg: offset of register
 507  *
 508  * Dummy register read function.  Used for register blocks
 509  * that certain asics don't have (all asics).
 510  * Returns the value in the register.
 511  */
 512 static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg)
 513 {
 514         DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg);
 515         BUG();
 516         return 0;
 517 }
 518
 519 /**
 520  * amdgpu_invalid_wreg64 - dummy reg write function
 521  *
 522  * @adev: amdgpu device pointer
 523  * @reg: offset of register
 524  * @v: value to write to the register
 525  *
 526  * Dummy register read function.  Used for register blocks
 527  * that certain asics don't have (all asics).
 528  */
 529 static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v)
 530 {
 531         DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n",
 532                   reg, v);
 533         BUG();
 534 }
 535
 536 /**
 537  * amdgpu_block_invalid_rreg - dummy reg read function
 538  *
 539  * @adev: amdgpu device pointer
 540  * @block: offset of instance
 541  * @reg: offset of register
 542  *
 543  * Dummy register read function.  Used for register blocks
 544  * that certain asics don't have (all asics).
 545  * Returns the value in the register.
 546  */
 547 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
 548                                           uint32_t block, uint32_t reg)
 549 {
 550         DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
 551                   reg, block);
 552         BUG();
 553         return 0;
 554 }
 555
 556 /**
 557  * amdgpu_block_invalid_wreg - dummy reg write function
 558  *
 559  * @adev: amdgpu device pointer
 560  * @block: offset of instance
 561  * @reg: offset of register
 562  * @v: value to write to the register
 563  *
 564  * Dummy register read function.  Used for register blocks
 565  * that certain asics don't have (all asics).
 566  */
 567 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
 568                                       uint32_t block,
 569                                       uint32_t reg, uint32_t v)
 570 {
 571         DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
 572                   reg, block, v);
 573         BUG();
 574 }
 575
 576 /**
 577  * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
 578  *
 579  * @adev: amdgpu device pointer
 580  *
 581  * Allocates a scratch page of VRAM for use by various things in the
 582  * driver.
 583  */
 584 static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
 585 {
 586         return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
 587                                        PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
 588                                        &adev->vram_scratch.robj,
 589                                        &adev->vram_scratch.gpu_addr,
 590                                        (void **)&adev->vram_scratch.ptr);
 591 }
 592
 593 /**
 594  * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
 595  *
 596  * @adev: amdgpu device pointer
 597  *
 598  * Frees the VRAM scratch page.
 599  */
 600 static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
 601 {
 602         amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
 603 }
 604
 605 /**
 606  * amdgpu_device_program_register_sequence - program an array of registers.
 607  *
 608  * @adev: amdgpu_device pointer
 609  * @registers: pointer to the register array
 610  * @array_size: size of the register array
 611  *
 612  * Programs an array or registers with and and or masks.
 613  * This is a helper for setting golden registers.
 614  */
 615 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
 616                                              const u32 *registers,
 617                                              const u32 array_size)
 618 {
 619         u32 tmp, reg, and_mask, or_mask;
 620         int i;
 621
 622         if (array_size % 3)
 623                 return;
 624
 625         for (i = 0; i < array_size; i +=3) {
 626                 reg = registers[i + 0];
 627                 and_mask = registers[i + 1];
 628                 or_mask = registers[i + 2];
 629
 630                 if (and_mask == 0xffffffff) {
 631                         tmp = or_mask;
 632                 } else {
 633                         tmp = RREG32(reg);
 634                         tmp &= ~and_mask;
 635                         if (adev->family >= AMDGPU_FAMILY_AI)
 636                                 tmp |= (or_mask & and_mask);
 637                         else
 638                                 tmp |= or_mask;
 639                 }
 640                 WREG32(reg, tmp);
 641         }
 642 }
 643
 644 /**
 645  * amdgpu_device_pci_config_reset - reset the GPU
 646  *
 647  * @adev: amdgpu_device pointer
 648  *
 649  * Resets the GPU using the pci config reset sequence.
 650  * Only applicable to asics prior to vega10.
 651  */
 652 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
 653 {
 654         pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
 655 }
 656
 657 /*
 658  * GPU doorbell aperture helpers function.
 659  */
 660 /**
 661  * amdgpu_device_doorbell_init - Init doorbell driver information.
 662  *
 663  * @adev: amdgpu_device pointer
 664  *
 665  * Init doorbell driver information (CIK)
 666  * Returns 0 on success, error on failure.
 667  */
 668 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
 669 {
 670
 671         /* No doorbell on SI hardware generation */
 672         if (adev->asic_type < CHIP_BONAIRE) {
 673                 adev->doorbell.base = 0;
 674                 adev->doorbell.size = 0;
 675                 adev->doorbell.num_doorbells = 0;
 676                 adev->doorbell.ptr = NULL;
 677                 return 0;
 678         }
 679
 680         if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
 681                 return -EINVAL;
 682
 683         amdgpu_asic_init_doorbell_index(adev);
 684
 685         /* doorbell bar mapping */
 686         adev->doorbell.base = pci_resource_start(adev->pdev, 2);
 687         adev->doorbell.size = pci_resource_len(adev->pdev, 2);
 688
 689         adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
 690                                              adev->doorbell_index.max_assignment+1);
 691         if (adev->doorbell.num_doorbells == 0)
 692                 return -EINVAL;
 693
 694         /* For Vega, reserve and map two pages on doorbell BAR since SDMA
 695          * paging queue doorbell use the second page. The
 696          * AMDGPU_DOORBELL64_MAX_ASSIGNMENT definition assumes all the
 697          * doorbells are in the first page. So with paging queue enabled,
 698          * the max num_doorbells should + 1 page (0x400 in dword)
 699          */
 700         if (adev->asic_type >= CHIP_VEGA10)
 701                 adev->doorbell.num_doorbells += 0x400;
 702
 703         adev->doorbell.ptr = ioremap(adev->doorbell.base,
 704                                      adev->doorbell.num_doorbells *
 705                                      sizeof(u32));
 706         if (adev->doorbell.ptr == NULL)
 707                 return -ENOMEM;
 708
 709         return 0;
 710 }
 711
 712 /**
 713  * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
 714  *
 715  * @adev: amdgpu_device pointer
 716  *
 717  * Tear down doorbell driver information (CIK)
 718  */
 719 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
 720 {
 721         iounmap(adev->doorbell.ptr);
 722         adev->doorbell.ptr = NULL;
 723 }
 724
 725
 726
 727 /*
 728  * amdgpu_device_wb_*()
 729  * Writeback is the method by which the GPU updates special pages in memory
 730  * with the status of certain GPU events (fences, ring pointers,etc.).
 731  */
 732
 733 /**
 734  * amdgpu_device_wb_fini - Disable Writeback and free memory
 735  *
 736  * @adev: amdgpu_device pointer
 737  *
 738  * Disables Writeback and frees the Writeback memory (all asics).
 739  * Used at driver shutdown.
 740  */
 741 static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
 742 {
 743         if (adev->wb.wb_obj) {
 744                 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
 745                                       &adev->wb.gpu_addr,
 746                                       (void **)&adev->wb.wb);
 747                 adev->wb.wb_obj = NULL;
 748         }
 749 }
 750
 751 /**
 752  * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
 753  *
 754  * @adev: amdgpu_device pointer
 755  *
 756  * Initializes writeback and allocates writeback memory (all asics).
 757  * Used at driver startup.
 758  * Returns 0 on success or an -error on failure.
 759  */
 760 static int amdgpu_device_wb_init(struct amdgpu_device *adev)
 761 {
 762         int r;
 763
 764         if (adev->wb.wb_obj == NULL) {
 765                 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
 766                 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
 767                                             PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
 768                                             &adev->wb.wb_obj, &adev->wb.gpu_addr,
 769                                             (void **)&adev->wb.wb);
 770                 if (r) {
 771                         dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
 772                         return r;
 773                 }
 774
 775                 adev->wb.num_wb = AMDGPU_MAX_WB;
 776                 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
 777
 778                 /* clear wb memory */
 779                 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
 780         }
 781
 782         return 0;
 783 }
 784
 785 /**
 786  * amdgpu_device_wb_get - Allocate a wb entry
 787  *
 788  * @adev: amdgpu_device pointer
 789  * @wb: wb index
 790  *
 791  * Allocate a wb slot for use by the driver (all asics).
 792  * Returns 0 on success or -EINVAL on failure.
 793  */
 794 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
 795 {
 796         unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
 797
 798         if (offset < adev->wb.num_wb) {
 799                 __set_bit(offset, adev->wb.used);
 800                 *wb = offset << 3; /* convert to dw offset */
 801                 return 0;
 802         } else {
 803                 return -EINVAL;
 804         }
 805 }
 806
 807 /**
 808  * amdgpu_device_wb_free - Free a wb entry
 809  *
 810  * @adev: amdgpu_device pointer
 811  * @wb: wb index
 812  *
 813  * Free a wb slot allocated for use by the driver (all asics)
 814  */
 815 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
 816 {
 817         wb >>= 3;
 818         if (wb < adev->wb.num_wb)
 819                 __clear_bit(wb, adev->wb.used);
 820 }
 821
 822 /**
 823  * amdgpu_device_resize_fb_bar - try to resize FB BAR
 824  *
 825  * @adev: amdgpu_device pointer
 826  *
 827  * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
 828  * to fail, but if any of the BARs is not accessible after the size we abort
 829  * driver loading by returning -ENODEV.
 830  */
 831 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
 832 {
 833         u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
 834         u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
 835         struct pci_bus *root;
 836         struct resource *res;
 837         unsigned i;
 838         u16 cmd;
 839         int r;
 840
 841         /* Bypass for VF */
 842         if (amdgpu_sriov_vf(adev))
 843                 return 0;
 844
 845         /* Check if the root BUS has 64bit memory resources */
 846         root = adev->pdev->bus;
 847         while (root->parent)
 848                 root = root->parent;
 849
 850         pci_bus_for_each_resource(root, res, i) {
 851                 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
 852                     res->start > 0x100000000ull)
 853                         break;
 854         }
 855
 856         /* Trying to resize is pointless without a root hub window above 4GB */
 857         if (!res)
 858                 return 0;
 859
 860         /* Disable memory decoding while we change the BAR addresses and size */
 861         pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
 862         pci_write_config_word(adev->pdev, PCI_COMMAND,
 863                               cmd & ~PCI_COMMAND_MEMORY);
 864
 865         /* Free the VRAM and doorbell BAR, we most likely need to move both. */
 866         amdgpu_device_doorbell_fini(adev);
 867         if (adev->asic_type >= CHIP_BONAIRE)
 868                 pci_release_resource(adev->pdev, 2);
 869
 870         pci_release_resource(adev->pdev, 0);
 871
 872         r = pci_resize_resource(adev->pdev, 0, rbar_size);
 873         if (r == -ENOSPC)
 874                 DRM_INFO("Not enough PCI address space for a large BAR.");
 875         else if (r && r != -ENOTSUPP)
 876                 DRM_ERROR("Problem resizing BAR0 (%d).", r);
 877
 878         pci_assign_unassigned_bus_resources(adev->pdev->bus);
 879
 880         /* When the doorbell or fb BAR isn't available we have no chance of
 881          * using the device.
 882          */
 883         r = amdgpu_device_doorbell_init(adev);
 884         if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
 885                 return -ENODEV;
 886
 887         pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
 888
 889         return 0;
 890 }
 891
 892 /*
 893  * GPU helpers function.
 894  */
 895 /**
 896  * amdgpu_device_need_post - check if the hw need post or not
 897  *
 898  * @adev: amdgpu_device pointer
 899  *
 900  * Check if the asic has been initialized (all asics) at driver startup
 901  * or post is needed if  hw reset is performed.
 902  * Returns true if need or false if not.
 903  */
 904 bool amdgpu_device_need_post(struct amdgpu_device *adev)
 905 {
 906         uint32_t reg;
 907
 908         if (amdgpu_sriov_vf(adev))
 909                 return false;
 910
 911         if (amdgpu_passthrough(adev)) {
 912                 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
 913                  * some old smc fw still need driver do vPost otherwise gpu hang, while
 914                  * those smc fw version above 22.15 doesn't have this flaw, so we force
 915                  * vpost executed for smc version below 22.15
 916                  */
 917                 if (adev->asic_type == CHIP_FIJI) {
 918                         int err;
 919                         uint32_t fw_ver;
 920                         err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
 921                         /* force vPost if error occured */
 922                         if (err)
 923                                 return true;
 924
 925                         fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
 926                         if (fw_ver < 0x00160e00)
 927                                 return true;
 928                 }
 929         }
 930
 931         if (adev->has_hw_reset) {
 932                 adev->has_hw_reset = false;
 933                 return true;
 934         }
 935
 936         /* bios scratch used on CIK+ */
 937         if (adev->asic_type >= CHIP_BONAIRE)
 938                 return amdgpu_atombios_scratch_need_asic_init(adev);
 939
 940         /* check MEM_SIZE for older asics */
 941         reg = amdgpu_asic_get_config_memsize(adev);
 942
 943         if ((reg != 0) && (reg != 0xffffffff))
 944                 return false;
 945
 946         return true;
 947 }
 948
 949 /* if we get transitioned to only one device, take VGA back */
 950 /**
 951  * amdgpu_device_vga_set_decode - enable/disable vga decode
 952  *
 953  * @cookie: amdgpu_device pointer
 954  * @state: enable/disable vga decode
 955  *
 956  * Enable/disable vga decode (all asics).
 957  * Returns VGA resource flags.
 958  */
 959 static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
 960 {
 961         struct amdgpu_device *adev = cookie;
 962         amdgpu_asic_set_vga_state(adev, state);
 963         if (state)
 964                 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
 965                        VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
 966         else
 967                 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
 968 }
 969
 970 /**
 971  * amdgpu_device_check_block_size - validate the vm block size
 972  *
 973  * @adev: amdgpu_device pointer
 974  *
 975  * Validates the vm block size specified via module parameter.
 976  * The vm block size defines number of bits in page table versus page directory,
 977  * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
 978  * page table and the remaining bits are in the page directory.
 979  */
 980 static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
 981 {
 982         /* defines number of bits in page table versus page directory,
 983          * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
 984          * page table and the remaining bits are in the page directory */
 985         if (amdgpu_vm_block_size == -1)
 986                 return;
 987
 988         if (amdgpu_vm_block_size < 9) {
 989                 dev_warn(adev->dev, "VM page table size (%d) too small\n",
 990                          amdgpu_vm_block_size);
 991                 amdgpu_vm_block_size = -1;
 992         }
 993 }
 994
 995 /**
 996  * amdgpu_device_check_vm_size - validate the vm size
 997  *
 998  * @adev: amdgpu_device pointer
 999  *
1000  * Validates the vm size in GB specified via module parameter.
1001  * The VM size is the size of the GPU virtual memory space in GB.
1002  */
1003 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
1004 {
1005         /* no need to check the default value */
1006         if (amdgpu_vm_size == -1)
1007                 return;
1008
1009         if (amdgpu_vm_size < 1) {
1010                 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
1011                          amdgpu_vm_size);
1012                 amdgpu_vm_size = -1;
1013         }
1014 }
1015
1016 static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
1017 {
1018         struct sysinfo si;
1019         bool is_os_64 = (sizeof(void *) == 8);
1020         uint64_t total_memory;
1021         uint64_t dram_size_seven_GB = 0x1B8000000;
1022         uint64_t dram_size_three_GB = 0xB8000000;
1023
1024         if (amdgpu_smu_memory_pool_size == 0)
1025                 return;
1026
1027         if (!is_os_64) {
1028                 DRM_WARN("Not 64-bit OS, feature not supported\n");
1029                 goto def_value;
1030         }
1031         si_meminfo(&si);
1032         total_memory = (uint64_t)si.totalram * si.mem_unit;
1033
1034         if ((amdgpu_smu_memory_pool_size == 1) ||
1035                 (amdgpu_smu_memory_pool_size == 2)) {
1036                 if (total_memory < dram_size_three_GB)
1037                         goto def_value1;
1038         } else if ((amdgpu_smu_memory_pool_size == 4) ||
1039                 (amdgpu_smu_memory_pool_size == 8)) {
1040                 if (total_memory < dram_size_seven_GB)
1041                         goto def_value1;
1042         } else {
1043                 DRM_WARN("Smu memory pool size not supported\n");
1044                 goto def_value;
1045         }
1046         adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
1047
1048         return;
1049
1050 def_value1:
1051         DRM_WARN("No enough system memory\n");
1052 def_value:
1053         adev->pm.smu_prv_buffer_size = 0;
1054 }
1055
1056 /**
1057  * amdgpu_device_check_arguments - validate module params
1058  *
1059  * @adev: amdgpu_device pointer
1060  *
1061  * Validates certain module parameters and updates
1062  * the associated values used by the driver (all asics).
1063  */
1064 static int amdgpu_device_check_arguments(struct amdgpu_device *adev)
1065 {
1066         if (amdgpu_sched_jobs < 4) {
1067                 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
1068                          amdgpu_sched_jobs);
1069                 amdgpu_sched_jobs = 4;
1070         } else if (!is_power_of_2(amdgpu_sched_jobs)){
1071                 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
1072                          amdgpu_sched_jobs);
1073                 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
1074         }
1075
1076         if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
1077                 /* gart size must be greater or equal to 32M */
1078                 dev_warn(adev->dev, "gart size (%d) too small\n",
1079                          amdgpu_gart_size);
1080                 amdgpu_gart_size = -1;
1081         }
1082
1083         if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
1084                 /* gtt size must be greater or equal to 32M */
1085                 dev_warn(adev->dev, "gtt size (%d) too small\n",
1086                                  amdgpu_gtt_size);
1087                 amdgpu_gtt_size = -1;
1088         }
1089
1090         /* valid range is between 4 and 9 inclusive */
1091         if (amdgpu_vm_fragment_size != -1 &&
1092             (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
1093                 dev_warn(adev->dev, "valid range is between 4 and 9\n");
1094                 amdgpu_vm_fragment_size = -1;
1095         }
1096
1097         amdgpu_device_check_smu_prv_buffer_size(adev);
1098
1099         amdgpu_device_check_vm_size(adev);
1100
1101         amdgpu_device_check_block_size(adev);
1102
1103         adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
1104
1105         return 0;
1106 }
1107
1108 /**
1109  * amdgpu_switcheroo_set_state - set switcheroo state
1110  *
1111  * @pdev: pci dev pointer
1112  * @state: vga_switcheroo state
1113  *
1114  * Callback for the switcheroo driver.  Suspends or resumes the
1115  * the asics before or after it is powered up using ACPI methods.
1116  */
1117 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
1118 {
1119         struct drm_device *dev = pci_get_drvdata(pdev);
1120         int r;
1121
1122         if (amdgpu_device_supports_boco(dev) && state == VGA_SWITCHEROO_OFF)
1123                 return;
1124
1125         if (state == VGA_SWITCHEROO_ON) {
1126                 pr_info("amdgpu: switched on\n");
1127                 /* don't suspend or resume card normally */
1128                 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1129
1130                 pci_set_power_state(dev->pdev, PCI_D0);
1131                 pci_restore_state(dev->pdev);
1132                 r = pci_enable_device(dev->pdev);
1133                 if (r)
1134                         DRM_WARN("pci_enable_device failed (%d)\n", r);
1135                 amdgpu_device_resume(dev, true);
1136
1137                 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1138                 drm_kms_helper_poll_enable(dev);
1139         } else {
1140                 pr_info("amdgpu: switched off\n");
1141                 drm_kms_helper_poll_disable(dev);
1142                 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1143                 amdgpu_device_suspend(dev, true);
1144                 pci_save_state(dev->pdev);
1145                 /* Shut down the device */
1146                 pci_disable_device(dev->pdev);
1147                 pci_set_power_state(dev->pdev, PCI_D3cold);
1148                 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1149         }
1150 }
1151
1152 /**
1153  * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1154  *
1155  * @pdev: pci dev pointer
1156  *
1157  * Callback for the switcheroo driver.  Check of the switcheroo
1158  * state can be changed.
1159  * Returns true if the state can be changed, false if not.
1160  */
1161 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1162 {
1163         struct drm_device *dev = pci_get_drvdata(pdev);
1164
1165         /*
1166         * FIXME: open_count is protected by drm_global_mutex but that would lead to
1167         * locking inversion with the driver load path. And the access here is
1168         * completely racy anyway. So don't bother with locking for now.
1169         */
1170         return dev->open_count == 0;
1171 }
1172
1173 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1174         .set_gpu_state = amdgpu_switcheroo_set_state,
1175         .reprobe = NULL,
1176         .can_switch = amdgpu_switcheroo_can_switch,
1177 };
1178
1179 /**
1180  * amdgpu_device_ip_set_clockgating_state - set the CG state
1181  *
1182  * @dev: amdgpu_device pointer
1183  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1184  * @state: clockgating state (gate or ungate)
1185  *
1186  * Sets the requested clockgating state for all instances of
1187  * the hardware IP specified.
1188  * Returns the error code from the last instance.
1189  */
1190 int amdgpu_device_ip_set_clockgating_state(void *dev,
1191                                            enum amd_ip_block_type block_type,
1192                                            enum amd_clockgating_state state)
1193 {
1194         struct amdgpu_device *adev = dev;
1195         int i, r = 0;
1196
1197         for (i = 0; i < adev->num_ip_blocks; i++) {
1198                 if (!adev->ip_blocks[i].status.valid)
1199                         continue;
1200                 if (adev->ip_blocks[i].version->type != block_type)
1201                         continue;
1202                 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1203                         continue;
1204                 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1205                         (void *)adev, state);
1206                 if (r)
1207                         DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1208                                   adev->ip_blocks[i].version->funcs->name, r);
1209         }
1210         return r;
1211 }
1212
1213 /**
1214  * amdgpu_device_ip_set_powergating_state - set the PG state
1215  *
1216  * @dev: amdgpu_device pointer
1217  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1218  * @state: powergating state (gate or ungate)
1219  *
1220  * Sets the requested powergating state for all instances of
1221  * the hardware IP specified.
1222  * Returns the error code from the last instance.
1223  */
1224 int amdgpu_device_ip_set_powergating_state(void *dev,
1225                                            enum amd_ip_block_type block_type,
1226                                            enum amd_powergating_state state)
1227 {
1228         struct amdgpu_device *adev = dev;
1229         int i, r = 0;
1230
1231         for (i = 0; i < adev->num_ip_blocks; i++) {
1232                 if (!adev->ip_blocks[i].status.valid)
1233                         continue;
1234                 if (adev->ip_blocks[i].version->type != block_type)
1235                         continue;
1236                 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1237                         continue;
1238                 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1239                         (void *)adev, state);
1240                 if (r)
1241                         DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1242                                   adev->ip_blocks[i].version->funcs->name, r);
1243         }
1244         return r;
1245 }
1246
1247 /**
1248  * amdgpu_device_ip_get_clockgating_state - get the CG state
1249  *
1250  * @adev: amdgpu_device pointer
1251  * @flags: clockgating feature flags
1252  *
1253  * Walks the list of IPs on the device and updates the clockgating
1254  * flags for each IP.
1255  * Updates @flags with the feature flags for each hardware IP where
1256  * clockgating is enabled.
1257  */
1258 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1259                                             u32 *flags)
1260 {
1261         int i;
1262
1263         for (i = 0; i < adev->num_ip_blocks; i++) {
1264                 if (!adev->ip_blocks[i].status.valid)
1265                         continue;
1266                 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1267                         adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1268         }
1269 }
1270
1271 /**
1272  * amdgpu_device_ip_wait_for_idle - wait for idle
1273  *
1274  * @adev: amdgpu_device pointer
1275  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1276  *
1277  * Waits for the request hardware IP to be idle.
1278  * Returns 0 for success or a negative error code on failure.
1279  */
1280 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1281                                    enum amd_ip_block_type block_type)
1282 {
1283         int i, r;
1284
1285         for (i = 0; i < adev->num_ip_blocks; i++) {
1286                 if (!adev->ip_blocks[i].status.valid)
1287                         continue;
1288                 if (adev->ip_blocks[i].version->type == block_type) {
1289                         r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1290                         if (r)
1291                                 return r;
1292                         break;
1293                 }
1294         }
1295         return 0;
1296
1297 }
1298
1299 /**
1300  * amdgpu_device_ip_is_idle - is the hardware IP idle
1301  *
1302  * @adev: amdgpu_device pointer
1303  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1304  *
1305  * Check if the hardware IP is idle or not.
1306  * Returns true if it the IP is idle, false if not.
1307  */
1308 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1309                               enum amd_ip_block_type block_type)
1310 {
1311         int i;
1312
1313         for (i = 0; i < adev->num_ip_blocks; i++) {
1314                 if (!adev->ip_blocks[i].status.valid)
1315                         continue;
1316                 if (adev->ip_blocks[i].version->type == block_type)
1317                         return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1318         }
1319         return true;
1320
1321 }
1322
1323 /**
1324  * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1325  *
1326  * @adev: amdgpu_device pointer
1327  * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
1328  *
1329  * Returns a pointer to the hardware IP block structure
1330  * if it exists for the asic, otherwise NULL.
1331  */
1332 struct amdgpu_ip_block *
1333 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1334                               enum amd_ip_block_type type)
1335 {
1336         int i;
1337
1338         for (i = 0; i < adev->num_ip_blocks; i++)
1339                 if (adev->ip_blocks[i].version->type == type)
1340                         return &adev->ip_blocks[i];
1341
1342         return NULL;
1343 }
1344
1345 /**
1346  * amdgpu_device_ip_block_version_cmp
1347  *
1348  * @adev: amdgpu_device pointer
1349  * @type: enum amd_ip_block_type
1350  * @major: major version
1351  * @minor: minor version
1352  *
1353  * return 0 if equal or greater
1354  * return 1 if smaller or the ip_block doesn't exist
1355  */
1356 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1357                                        enum amd_ip_block_type type,
1358                                        u32 major, u32 minor)
1359 {
1360         struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1361
1362         if (ip_block && ((ip_block->version->major > major) ||
1363                         ((ip_block->version->major == major) &&
1364                         (ip_block->version->minor >= minor))))
1365                 return 0;
1366
1367         return 1;
1368 }
1369
1370 /**
1371  * amdgpu_device_ip_block_add
1372  *
1373  * @adev: amdgpu_device pointer
1374  * @ip_block_version: pointer to the IP to add
1375  *
1376  * Adds the IP block driver information to the collection of IPs
1377  * on the asic.
1378  */
1379 int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1380                                const struct amdgpu_ip_block_version *ip_block_version)
1381 {
1382         if (!ip_block_version)
1383                 return -EINVAL;
1384
1385         DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1386                   ip_block_version->funcs->name);
1387
1388         adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1389
1390         return 0;
1391 }
1392
1393 /**
1394  * amdgpu_device_enable_virtual_display - enable virtual display feature
1395  *
1396  * @adev: amdgpu_device pointer
1397  *
1398  * Enabled the virtual display feature if the user has enabled it via
1399  * the module parameter virtual_display.  This feature provides a virtual
1400  * display hardware on headless boards or in virtualized environments.
1401  * This function parses and validates the configuration string specified by
1402  * the user and configues the virtual display configuration (number of
1403  * virtual connectors, crtcs, etc.) specified.
1404  */
1405 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1406 {
1407         adev->enable_virtual_display = false;
1408
1409         if (amdgpu_virtual_display) {
1410                 struct drm_device *ddev = adev->ddev;
1411                 const char *pci_address_name = pci_name(ddev->pdev);
1412                 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1413
1414                 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1415                 pciaddstr_tmp = pciaddstr;
1416                 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1417                         pciaddname = strsep(&pciaddname_tmp, ",");
1418                         if (!strcmp("all", pciaddname)
1419                             || !strcmp(pci_address_name, pciaddname)) {
1420                                 long num_crtc;
1421                                 int res = -1;
1422
1423                                 adev->enable_virtual_display = true;
1424
1425                                 if (pciaddname_tmp)
1426                                         res = kstrtol(pciaddname_tmp, 10,
1427                                                       &num_crtc);
1428
1429                                 if (!res) {
1430                                         if (num_crtc < 1)
1431                                                 num_crtc = 1;
1432                                         if (num_crtc > 6)
1433                                                 num_crtc = 6;
1434                                         adev->mode_info.num_crtc = num_crtc;
1435                                 } else {
1436                                         adev->mode_info.num_crtc = 1;
1437                                 }
1438                                 break;
1439                         }
1440                 }
1441
1442                 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1443                          amdgpu_virtual_display, pci_address_name,
1444                          adev->enable_virtual_display, adev->mode_info.num_crtc);
1445
1446                 kfree(pciaddstr);
1447         }
1448 }
1449
1450 /**
1451  * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1452  *
1453  * @adev: amdgpu_device pointer
1454  *
1455  * Parses the asic configuration parameters specified in the gpu info
1456  * firmware and makes them availale to the driver for use in configuring
1457  * the asic.
1458  * Returns 0 on success, -EINVAL on failure.
1459  */
1460 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1461 {
1462         const char *chip_name;
1463         char fw_name[30];
1464         int err;
1465         const struct gpu_info_firmware_header_v1_0 *hdr;
1466
1467         adev->firmware.gpu_info_fw = NULL;
1468
1469         switch (adev->asic_type) {
1470         case CHIP_TOPAZ:
1471         case CHIP_TONGA:
1472         case CHIP_FIJI:
1473         case CHIP_POLARIS10:
1474         case CHIP_POLARIS11:
1475         case CHIP_POLARIS12:
1476         case CHIP_VEGAM:
1477         case CHIP_CARRIZO:
1478         case CHIP_STONEY:
1479 #ifdef CONFIG_DRM_AMDGPU_SI
1480         case CHIP_VERDE:
1481         case CHIP_TAHITI:
1482         case CHIP_PITCAIRN:
1483         case CHIP_OLAND:
1484         case CHIP_HAINAN:
1485 #endif
1486 #ifdef CONFIG_DRM_AMDGPU_CIK
1487         case CHIP_BONAIRE:
1488         case CHIP_HAWAII:
1489         case CHIP_KAVERI:
1490         case CHIP_KABINI:
1491         case CHIP_MULLINS:
1492 #endif
1493         case CHIP_VEGA20:
1494         default:
1495                 return 0;
1496         case CHIP_VEGA10:
1497                 chip_name = "vega10";
1498                 break;
1499         case CHIP_VEGA12:
1500                 chip_name = "vega12";
1501                 break;
1502         case CHIP_RAVEN:
1503                 if (adev->rev_id >= 8)
1504                         chip_name = "raven2";
1505                 else if (adev->pdev->device == 0x15d8)
1506                         chip_name = "picasso";
1507                 else
1508                         chip_name = "raven";
1509                 break;
1510         case CHIP_ARCTURUS:
1511                 chip_name = "arcturus";
1512                 break;
1513         case CHIP_RENOIR:
1514                 chip_name = "renoir";
1515                 break;
1516         case CHIP_NAVI10:
1517                 chip_name = "navi10";
1518                 break;
1519         case CHIP_NAVI14:
1520                 chip_name = "navi14";
1521                 break;
1522         case CHIP_NAVI12:
1523                 chip_name = "navi12";
1524                 break;
1525         }
1526
1527         snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1528         err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1529         if (err) {
1530                 dev_err(adev->dev,
1531                         "Failed to load gpu_info firmware \"%s\"\n",
1532                         fw_name);
1533                 goto out;
1534         }
1535         err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1536         if (err) {
1537                 dev_err(adev->dev,
1538                         "Failed to validate gpu_info firmware \"%s\"\n",
1539                         fw_name);
1540                 goto out;
1541         }
1542
1543         hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1544         amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1545
1546         switch (hdr->version_major) {
1547         case 1:
1548         {
1549                 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1550                         (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1551                                                                 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1552
1553                 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10)
1554                         goto parse_soc_bounding_box;
1555
1556                 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1557                 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1558                 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1559                 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1560                 adev->gfx.config.max_texture_channel_caches =
1561                         le32_to_cpu(gpu_info_fw->gc_num_tccs);
1562                 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1563                 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1564                 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1565                 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1566                 adev->gfx.config.double_offchip_lds_buf =
1567                         le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1568                 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1569                 adev->gfx.cu_info.max_waves_per_simd =
1570                         le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1571                 adev->gfx.cu_info.max_scratch_slots_per_cu =
1572                         le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1573                 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
1574                 if (hdr->version_minor >= 1) {
1575                         const struct gpu_info_firmware_v1_1 *gpu_info_fw =
1576                                 (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data +
1577                                                                         le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1578                         adev->gfx.config.num_sc_per_sh =
1579                                 le32_to_cpu(gpu_info_fw->num_sc_per_sh);
1580                         adev->gfx.config.num_packer_per_sc =
1581                                 le32_to_cpu(gpu_info_fw->num_packer_per_sc);
1582                 }
1583
1584 parse_soc_bounding_box:
1585                 /*
1586                  * soc bounding box info is not integrated in disocovery table,
1587                  * we always need to parse it from gpu info firmware.
1588                  */
1589                 if (hdr->version_minor == 2) {
1590                         const struct gpu_info_firmware_v1_2 *gpu_info_fw =
1591                                 (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data +
1592                                                                         le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1593                         adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box;
1594                 }
1595                 break;
1596         }
1597         default:
1598                 dev_err(adev->dev,
1599                         "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1600                 err = -EINVAL;
1601                 goto out;
1602         }
1603 out:
1604         return err;
1605 }
1606
1607 /**
1608  * amdgpu_device_ip_early_init - run early init for hardware IPs
1609  *
1610  * @adev: amdgpu_device pointer
1611  *
1612  * Early initialization pass for hardware IPs.  The hardware IPs that make
1613  * up each asic are discovered each IP's early_init callback is run.  This
1614  * is the first stage in initializing the asic.
1615  * Returns 0 on success, negative error code on failure.
1616  */
1617 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
1618 {
1619         int i, r;
1620
1621         amdgpu_device_enable_virtual_display(adev);
1622
1623         switch (adev->asic_type) {
1624         case CHIP_TOPAZ:
1625         case CHIP_TONGA:
1626         case CHIP_FIJI:
1627         case CHIP_POLARIS10:
1628         case CHIP_POLARIS11:
1629         case CHIP_POLARIS12:
1630         case CHIP_VEGAM:
1631         case CHIP_CARRIZO:
1632         case CHIP_STONEY:
1633                 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1634                         adev->family = AMDGPU_FAMILY_CZ;
1635                 else
1636                         adev->family = AMDGPU_FAMILY_VI;
1637
1638                 r = vi_set_ip_blocks(adev);
1639                 if (r)
1640                         return r;
1641                 break;
1642 #ifdef CONFIG_DRM_AMDGPU_SI
1643         case CHIP_VERDE:
1644         case CHIP_TAHITI:
1645         case CHIP_PITCAIRN:
1646         case CHIP_OLAND:
1647         case CHIP_HAINAN:
1648                 adev->family = AMDGPU_FAMILY_SI;
1649                 r = si_set_ip_blocks(adev);
1650                 if (r)
1651                         return r;
1652                 break;
1653 #endif
1654 #ifdef CONFIG_DRM_AMDGPU_CIK
1655         case CHIP_BONAIRE:
1656         case CHIP_HAWAII:
1657         case CHIP_KAVERI:
1658         case CHIP_KABINI:
1659         case CHIP_MULLINS:
1660                 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1661                         adev->family = AMDGPU_FAMILY_CI;
1662                 else
1663                         adev->family = AMDGPU_FAMILY_KV;
1664
1665                 r = cik_set_ip_blocks(adev);
1666                 if (r)
1667                         return r;
1668                 break;
1669 #endif
1670         case CHIP_VEGA10:
1671         case CHIP_VEGA12:
1672         case CHIP_VEGA20:
1673         case CHIP_RAVEN:
1674         case CHIP_ARCTURUS:
1675         case CHIP_RENOIR:
1676                 if (adev->asic_type == CHIP_RAVEN ||
1677                     adev->asic_type == CHIP_RENOIR)
1678                         adev->family = AMDGPU_FAMILY_RV;
1679                 else
1680                         adev->family = AMDGPU_FAMILY_AI;
1681
1682                 r = soc15_set_ip_blocks(adev);
1683                 if (r)
1684                         return r;
1685                 break;
1686         case  CHIP_NAVI10:
1687         case  CHIP_NAVI14:
1688         case  CHIP_NAVI12:
1689                 adev->family = AMDGPU_FAMILY_NV;
1690
1691                 r = nv_set_ip_blocks(adev);
1692                 if (r)
1693                         return r;
1694                 break;
1695         default:
1696                 /* FIXME: not supported yet */
1697                 return -EINVAL;
1698         }
1699
1700         r = amdgpu_device_parse_gpu_info_fw(adev);
1701         if (r)
1702                 return r;
1703
1704         if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10)
1705                 amdgpu_discovery_get_gfx_info(adev);
1706
1707         amdgpu_amdkfd_device_probe(adev);
1708
1709         if (amdgpu_sriov_vf(adev)) {
1710                 r = amdgpu_virt_request_full_gpu(adev, true);
1711                 if (r)
1712                         return -EAGAIN;
1713         }
1714
1715         adev->pm.pp_feature = amdgpu_pp_feature_mask;
1716         if (amdgpu_sriov_vf(adev) || sched_policy == KFD_SCHED_POLICY_NO_HWS)
1717                 adev->pm.pp_feature &= ~PP_GFXOFF_MASK;
1718
1719         for (i = 0; i < adev->num_ip_blocks; i++) {
1720                 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1721                         DRM_ERROR("disabled ip block: %d <%s>\n",
1722                                   i, adev->ip_blocks[i].version->funcs->name);
1723                         adev->ip_blocks[i].status.valid = false;
1724                 } else {
1725                         if (adev->ip_blocks[i].version->funcs->early_init) {
1726                                 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1727                                 if (r == -ENOENT) {
1728                                         adev->ip_blocks[i].status.valid = false;
1729                                 } else if (r) {
1730                                         DRM_ERROR("early_init of IP block <%s> failed %d\n",
1731                                                   adev->ip_blocks[i].version->funcs->name, r);
1732                                         return r;
1733                                 } else {
1734                                         adev->ip_blocks[i].status.valid = true;
1735                                 }
1736                         } else {
1737                                 adev->ip_blocks[i].status.valid = true;
1738                         }
1739                 }
1740                 /* get the vbios after the asic_funcs are set up */
1741                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) {
1742                         /* Read BIOS */
1743                         if (!amdgpu_get_bios(adev))
1744                                 return -EINVAL;
1745
1746                         r = amdgpu_atombios_init(adev);
1747                         if (r) {
1748                                 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
1749                                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
1750                                 return r;
1751                         }
1752                 }
1753         }
1754
1755         adev->cg_flags &= amdgpu_cg_mask;
1756         adev->pg_flags &= amdgpu_pg_mask;
1757
1758         return 0;
1759 }
1760
1761 static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev)
1762 {
1763         int i, r;
1764
1765         for (i = 0; i < adev->num_ip_blocks; i++) {
1766                 if (!adev->ip_blocks[i].status.sw)
1767                         continue;
1768                 if (adev->ip_blocks[i].status.hw)
1769                         continue;
1770                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
1771                     (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) ||
1772                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
1773                         r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1774                         if (r) {
1775                                 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1776                                           adev->ip_blocks[i].version->funcs->name, r);
1777                                 return r;
1778                         }
1779                         adev->ip_blocks[i].status.hw = true;
1780                 }
1781         }
1782
1783         return 0;
1784 }
1785
1786 static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev)
1787 {
1788         int i, r;
1789
1790         for (i = 0; i < adev->num_ip_blocks; i++) {
1791                 if (!adev->ip_blocks[i].status.sw)
1792                         continue;
1793                 if (adev->ip_blocks[i].status.hw)
1794                         continue;
1795                 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1796                 if (r) {
1797                         DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1798                                   adev->ip_blocks[i].version->funcs->name, r);
1799                         return r;
1800                 }
1801                 adev->ip_blocks[i].status.hw = true;
1802         }
1803
1804         return 0;
1805 }
1806
1807 static int amdgpu_device_fw_loading(struct amdgpu_device *adev)
1808 {
1809         int r = 0;
1810         int i;
1811         uint32_t smu_version;
1812
1813         if (adev->asic_type >= CHIP_VEGA10) {
1814                 for (i = 0; i < adev->num_ip_blocks; i++) {
1815                         if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP)
1816                                 continue;
1817
1818                         /* no need to do the fw loading again if already done*/
1819                         if (adev->ip_blocks[i].status.hw == true)
1820                                 break;
1821
1822                         if (adev->in_gpu_reset || adev->in_suspend) {
1823                                 r = adev->ip_blocks[i].version->funcs->resume(adev);
1824                                 if (r) {
1825                                         DRM_ERROR("resume of IP block <%s> failed %d\n",
1826                                                           adev->ip_blocks[i].version->funcs->name, r);
1827                                         return r;
1828                                 }
1829                         } else {
1830                                 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1831                                 if (r) {
1832                                         DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1833                                                           adev->ip_blocks[i].version->funcs->name, r);
1834                                         return r;
1835                                 }
1836                         }
1837
1838                         adev->ip_blocks[i].status.hw = true;
1839                         break;
1840                 }
1841         }
1842
1843         if (!amdgpu_sriov_vf(adev) || adev->asic_type == CHIP_TONGA)
1844                 r = amdgpu_pm_load_smu_firmware(adev, &smu_version);
1845
1846         return r;
1847 }
1848
1849 /**
1850  * amdgpu_device_ip_init - run init for hardware IPs
1851  *
1852  * @adev: amdgpu_device pointer
1853  *
1854  * Main initialization pass for hardware IPs.  The list of all the hardware
1855  * IPs that make up the asic is walked and the sw_init and hw_init callbacks
1856  * are run.  sw_init initializes the software state associated with each IP
1857  * and hw_init initializes the hardware associated with each IP.
1858  * Returns 0 on success, negative error code on failure.
1859  */
1860 static int amdgpu_device_ip_init(struct amdgpu_device *adev)
1861 {
1862         int i, r;
1863
1864         r = amdgpu_ras_init(adev);
1865         if (r)
1866                 return r;
1867
1868         for (i = 0; i < adev->num_ip_blocks; i++) {
1869                 if (!adev->ip_blocks[i].status.valid)
1870                         continue;
1871                 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1872                 if (r) {
1873                         DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1874                                   adev->ip_blocks[i].version->funcs->name, r);
1875                         goto init_failed;
1876                 }
1877                 adev->ip_blocks[i].status.sw = true;
1878
1879                 /* need to do gmc hw init early so we can allocate gpu mem */
1880                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1881                         r = amdgpu_device_vram_scratch_init(adev);
1882                         if (r) {
1883                                 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1884                                 goto init_failed;
1885                         }
1886                         r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1887                         if (r) {
1888                                 DRM_ERROR("hw_init %d failed %d\n", i, r);
1889                                 goto init_failed;
1890                         }
1891                         r = amdgpu_device_wb_init(adev);
1892                         if (r) {
1893                                 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
1894                                 goto init_failed;
1895                         }
1896                         adev->ip_blocks[i].status.hw = true;
1897
1898                         /* right after GMC hw init, we create CSA */
1899                         if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
1900                                 r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj,
1901                                                                 AMDGPU_GEM_DOMAIN_VRAM,
1902                                                                 AMDGPU_CSA_SIZE);
1903                                 if (r) {
1904                                         DRM_ERROR("allocate CSA failed %d\n", r);
1905                                         goto init_failed;
1906                                 }
1907                         }
1908                 }
1909         }
1910
1911         if (amdgpu_sriov_vf(adev))
1912                 amdgpu_virt_init_data_exchange(adev);
1913
1914         r = amdgpu_ib_pool_init(adev);
1915         if (r) {
1916                 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
1917                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
1918                 goto init_failed;
1919         }
1920
1921         r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/
1922         if (r)
1923                 goto init_failed;
1924
1925         r = amdgpu_device_ip_hw_init_phase1(adev);
1926         if (r)
1927                 goto init_failed;
1928
1929         r = amdgpu_device_fw_loading(adev);
1930         if (r)
1931                 goto init_failed;
1932
1933         r = amdgpu_device_ip_hw_init_phase2(adev);
1934         if (r)
1935                 goto init_failed;
1936
1937         /*
1938          * retired pages will be loaded from eeprom and reserved here,
1939          * it should be called after amdgpu_device_ip_hw_init_phase2  since
1940          * for some ASICs the RAS EEPROM code relies on SMU fully functioning
1941          * for I2C communication which only true at this point.
1942          * recovery_init may fail, but it can free all resources allocated by
1943          * itself and its failure should not stop amdgpu init process.
1944          *
1945          * Note: theoretically, this should be called before all vram allocations
1946          * to protect retired page from abusing
1947          */
1948         amdgpu_ras_recovery_init(adev);
1949
1950         if (adev->gmc.xgmi.num_physical_nodes > 1)
1951                 amdgpu_xgmi_add_device(adev);
1952         amdgpu_amdkfd_device_init(adev);
1953
1954 init_failed:
1955         if (amdgpu_sriov_vf(adev))
1956                 amdgpu_virt_release_full_gpu(adev, true);
1957
1958         return r;
1959 }
1960
1961 /**
1962  * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
1963  *
1964  * @adev: amdgpu_device pointer
1965  *
1966  * Writes a reset magic value to the gart pointer in VRAM.  The driver calls
1967  * this function before a GPU reset.  If the value is retained after a
1968  * GPU reset, VRAM has not been lost.  Some GPU resets may destry VRAM contents.
1969  */
1970 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
1971 {
1972         memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1973 }
1974
1975 /**
1976  * amdgpu_device_check_vram_lost - check if vram is valid
1977  *
1978  * @adev: amdgpu_device pointer
1979  *
1980  * Checks the reset magic value written to the gart pointer in VRAM.
1981  * The driver calls this after a GPU reset to see if the contents of
1982  * VRAM is lost or now.
1983  * returns true if vram is lost, false if not.
1984  */
1985 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
1986 {
1987         return !!memcmp(adev->gart.ptr, adev->reset_magic,
1988                         AMDGPU_RESET_MAGIC_NUM);
1989 }
1990
1991 /**
1992  * amdgpu_device_set_cg_state - set clockgating for amdgpu device
1993  *
1994  * @adev: amdgpu_device pointer
1995  * @state: clockgating state (gate or ungate)
1996  *
1997  * The list of all the hardware IPs that make up the asic is walked and the
1998  * set_clockgating_state callbacks are run.
1999  * Late initialization pass enabling clockgating for hardware IPs.
2000  * Fini or suspend, pass disabling clockgating for hardware IPs.
2001  * Returns 0 on success, negative error code on failure.
2002  */
2003
2004 static int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
2005                                                 enum amd_clockgating_state state)
2006 {
2007         int i, j, r;
2008
2009         if (amdgpu_emu_mode == 1)
2010                 return 0;
2011
2012         for (j = 0; j < adev->num_ip_blocks; j++) {
2013                 i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2014                 if (!adev->ip_blocks[i].status.late_initialized)
2015                         continue;
2016                 /* skip CG for VCE/UVD, it's handled specially */
2017                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2018                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2019                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2020                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2021                     adev->ip_blocks[i].version->funcs->set_clockgating_state) {
2022                         /* enable clockgating to save power */
2023                         r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
2024                                                                                      state);
2025                         if (r) {
2026                                 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
2027                                           adev->ip_blocks[i].version->funcs->name, r);
2028                                 return r;
2029                         }
2030                 }
2031         }
2032
2033         return 0;
2034 }
2035
2036 static int amdgpu_device_set_pg_state(struct amdgpu_device *adev, enum amd_powergating_state state)
2037 {
2038         int i, j, r;
2039
2040         if (amdgpu_emu_mode == 1)
2041                 return 0;
2042
2043         for (j = 0; j < adev->num_ip_blocks; j++) {
2044                 i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2045                 if (!adev->ip_blocks[i].status.late_initialized)
2046                         continue;
2047                 /* skip CG for VCE/UVD, it's handled specially */
2048                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2049                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2050                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2051                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2052                     adev->ip_blocks[i].version->funcs->set_powergating_state) {
2053                         /* enable powergating to save power */
2054                         r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
2055                                                                                         state);
2056                         if (r) {
2057                                 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
2058                                           adev->ip_blocks[i].version->funcs->name, r);
2059                                 return r;
2060                         }
2061                 }
2062         }
2063         return 0;
2064 }
2065
2066 static int amdgpu_device_enable_mgpu_fan_boost(void)
2067 {
2068         struct amdgpu_gpu_instance *gpu_ins;
2069         struct amdgpu_device *adev;
2070         int i, ret = 0;
2071
2072         mutex_lock(&mgpu_info.mutex);
2073
2074         /*
2075          * MGPU fan boost feature should be enabled
2076          * only when there are two or more dGPUs in
2077          * the system
2078          */
2079         if (mgpu_info.num_dgpu < 2)
2080                 goto out;
2081
2082         for (i = 0; i < mgpu_info.num_dgpu; i++) {
2083                 gpu_ins = &(mgpu_info.gpu_ins[i]);
2084                 adev = gpu_ins->adev;
2085                 if (!(adev->flags & AMD_IS_APU) &&
2086                     !gpu_ins->mgpu_fan_enabled &&
2087                     adev->powerplay.pp_funcs &&
2088                     adev->powerplay.pp_funcs->enable_mgpu_fan_boost) {
2089                         ret = amdgpu_dpm_enable_mgpu_fan_boost(adev);
2090                         if (ret)
2091                                 break;
2092
2093                         gpu_ins->mgpu_fan_enabled = 1;
2094                 }
2095         }
2096
2097 out:
2098         mutex_unlock(&mgpu_info.mutex);
2099
2100         return ret;
2101 }
2102
2103 /**
2104  * amdgpu_device_ip_late_init - run late init for hardware IPs
2105  *
2106  * @adev: amdgpu_device pointer
2107  *
2108  * Late initialization pass for hardware IPs.  The list of all the hardware
2109  * IPs that make up the asic is walked and the late_init callbacks are run.
2110  * late_init covers any special initialization that an IP requires
2111  * after all of the have been initialized or something that needs to happen
2112  * late in the init process.
2113  * Returns 0 on success, negative error code on failure.
2114  */
2115 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
2116 {
2117         struct amdgpu_gpu_instance *gpu_instance;
2118         int i = 0, r;
2119
2120         for (i = 0; i < adev->num_ip_blocks; i++) {
2121                 if (!adev->ip_blocks[i].status.hw)
2122                         continue;
2123                 if (adev->ip_blocks[i].version->funcs->late_init) {
2124                         r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
2125                         if (r) {
2126                                 DRM_ERROR("late_init of IP block <%s> failed %d\n",
2127                                           adev->ip_blocks[i].version->funcs->name, r);
2128                                 return r;
2129                         }
2130                 }
2131                 adev->ip_blocks[i].status.late_initialized = true;
2132         }
2133
2134         amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
2135         amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
2136
2137         amdgpu_device_fill_reset_magic(adev);
2138
2139         r = amdgpu_device_enable_mgpu_fan_boost();
2140         if (r)
2141                 DRM_ERROR("enable mgpu fan boost failed (%d).\n", r);
2142
2143
2144         if (adev->gmc.xgmi.num_physical_nodes > 1) {
2145                 mutex_lock(&mgpu_info.mutex);
2146
2147                 /*
2148                  * Reset device p-state to low as this was booted with high.
2149                  *
2150                  * This should be performed only after all devices from the same
2151                  * hive get initialized.
2152                  *
2153                  * However, it's unknown how many device in the hive in advance.
2154                  * As this is counted one by one during devices initializations.
2155                  *
2156                  * So, we wait for all XGMI interlinked devices initialized.
2157                  * This may bring some delays as those devices may come from
2158                  * different hives. But that should be OK.
2159                  */
2160                 if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) {
2161                         for (i = 0; i < mgpu_info.num_gpu; i++) {
2162                                 gpu_instance = &(mgpu_info.gpu_ins[i]);
2163                                 if (gpu_instance->adev->flags & AMD_IS_APU)
2164                                         continue;
2165
2166                                 r = amdgpu_xgmi_set_pstate(gpu_instance->adev, 0);
2167                                 if (r) {
2168                                         DRM_ERROR("pstate setting failed (%d).\n", r);
2169                                         break;
2170                                 }
2171                         }
2172                 }
2173
2174                 mutex_unlock(&mgpu_info.mutex);
2175         }
2176
2177         return 0;
2178 }
2179
2180 /**
2181  * amdgpu_device_ip_fini - run fini for hardware IPs
2182  *
2183  * @adev: amdgpu_device pointer
2184  *
2185  * Main teardown pass for hardware IPs.  The list of all the hardware
2186  * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
2187  * are run.  hw_fini tears down the hardware associated with each IP
2188  * and sw_fini tears down any software state associated with each IP.
2189  * Returns 0 on success, negative error code on failure.
2190  */
2191 static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
2192 {
2193         int i, r;
2194
2195         amdgpu_ras_pre_fini(adev);
2196
2197         if (adev->gmc.xgmi.num_physical_nodes > 1)
2198                 amdgpu_xgmi_remove_device(adev);
2199
2200         amdgpu_amdkfd_device_fini(adev);
2201
2202         amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2203         amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2204
2205         /* need to disable SMC first */
2206         for (i = 0; i < adev->num_ip_blocks; i++) {
2207                 if (!adev->ip_blocks[i].status.hw)
2208                         continue;
2209                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2210                         r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2211                         /* XXX handle errors */
2212                         if (r) {
2213                                 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2214                                           adev->ip_blocks[i].version->funcs->name, r);
2215                         }
2216                         adev->ip_blocks[i].status.hw = false;
2217                         break;
2218                 }
2219         }
2220
2221         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2222                 if (!adev->ip_blocks[i].status.hw)
2223                         continue;
2224
2225                 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2226                 /* XXX handle errors */
2227                 if (r) {
2228                         DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2229                                   adev->ip_blocks[i].version->funcs->name, r);
2230                 }
2231
2232                 adev->ip_blocks[i].status.hw = false;
2233         }
2234
2235
2236         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2237                 if (!adev->ip_blocks[i].status.sw)
2238                         continue;
2239
2240                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
2241                         amdgpu_ucode_free_bo(adev);
2242                         amdgpu_free_static_csa(&adev->virt.csa_obj);
2243                         amdgpu_device_wb_fini(adev);
2244                         amdgpu_device_vram_scratch_fini(adev);
2245                         amdgpu_ib_pool_fini(adev);
2246                 }
2247
2248                 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
2249                 /* XXX handle errors */
2250                 if (r) {
2251                         DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
2252                                   adev->ip_blocks[i].version->funcs->name, r);
2253                 }
2254                 adev->ip_blocks[i].status.sw = false;
2255                 adev->ip_blocks[i].status.valid = false;
2256         }
2257
2258         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2259                 if (!adev->ip_blocks[i].status.late_initialized)
2260                         continue;
2261                 if (adev->ip_blocks[i].version->funcs->late_fini)
2262                         adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
2263                 adev->ip_blocks[i].status.late_initialized = false;
2264         }
2265
2266         amdgpu_ras_fini(adev);
2267
2268         if (amdgpu_sriov_vf(adev))
2269                 if (amdgpu_virt_release_full_gpu(adev, false))
2270                         DRM_ERROR("failed to release exclusive mode on fini\n");
2271
2272         return 0;
2273 }
2274
2275 /**
2276  * amdgpu_device_delayed_init_work_handler - work handler for IB tests
2277  *
2278  * @work: work_struct.
2279  */
2280 static void amdgpu_device_delayed_init_work_handler(struct work_struct *work)
2281 {
2282         struct amdgpu_device *adev =
2283                 container_of(work, struct amdgpu_device, delayed_init_work.work);
2284         int r;
2285
2286         r = amdgpu_ib_ring_tests(adev);
2287         if (r)
2288                 DRM_ERROR("ib ring test failed (%d).\n", r);
2289 }
2290
2291 static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
2292 {
2293         struct amdgpu_device *adev =
2294                 container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
2295
2296         mutex_lock(&adev->gfx.gfx_off_mutex);
2297         if (!adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
2298                 if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
2299                         adev->gfx.gfx_off_state = true;
2300         }
2301         mutex_unlock(&adev->gfx.gfx_off_mutex);
2302 }
2303
2304 /**
2305  * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
2306  *
2307  * @adev: amdgpu_device pointer
2308  *
2309  * Main suspend function for hardware IPs.  The list of all the hardware
2310  * IPs that make up the asic is walked, clockgating is disabled and the
2311  * suspend callbacks are run.  suspend puts the hardware and software state
2312  * in each IP into a state suitable for suspend.
2313  * Returns 0 on success, negative error code on failure.
2314  */
2315 static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
2316 {
2317         int i, r;
2318
2319         amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2320         amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2321
2322         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2323                 if (!adev->ip_blocks[i].status.valid)
2324                         continue;
2325                 /* displays are handled separately */
2326                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) {
2327                         /* XXX handle errors */
2328                         r = adev->ip_blocks[i].version->funcs->suspend(adev);
2329                         /* XXX handle errors */
2330                         if (r) {
2331                                 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2332                                           adev->ip_blocks[i].version->funcs->name, r);
2333                                 return r;
2334                         }
2335                         adev->ip_blocks[i].status.hw = false;
2336                 }
2337         }
2338
2339         return 0;
2340 }
2341
2342 /**
2343  * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
2344  *
2345  * @adev: amdgpu_device pointer
2346  *
2347  * Main suspend function for hardware IPs.  The list of all the hardware
2348  * IPs that make up the asic is walked, clockgating is disabled and the
2349  * suspend callbacks are run.  suspend puts the hardware and software state
2350  * in each IP into a state suitable for suspend.
2351  * Returns 0 on success, negative error code on failure.
2352  */
2353 static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
2354 {
2355         int i, r;
2356
2357         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2358                 if (!adev->ip_blocks[i].status.valid)
2359                         continue;
2360                 /* displays are handled in phase1 */
2361                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
2362                         continue;
2363                 /* PSP lost connection when err_event_athub occurs */
2364                 if (amdgpu_ras_intr_triggered() &&
2365                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
2366                         adev->ip_blocks[i].status.hw = false;
2367                         continue;
2368                 }
2369                 /* XXX handle errors */
2370                 r = adev->ip_blocks[i].version->funcs->suspend(adev);
2371                 /* XXX handle errors */
2372                 if (r) {
2373                         DRM_ERROR("suspend of IP block <%s> failed %d\n",
2374                                   adev->ip_blocks[i].version->funcs->name, r);
2375                 }
2376                 adev->ip_blocks[i].status.hw = false;
2377                 /* handle putting the SMC in the appropriate state */
2378                 if(!amdgpu_sriov_vf(adev)){
2379                         if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2380                                 r = amdgpu_dpm_set_mp1_state(adev, adev->mp1_state);
2381                                 if (r) {
2382                                         DRM_ERROR("SMC failed to set mp1 state %d, %d\n",
2383                                                         adev->mp1_state, r);
2384                                         return r;
2385                                 }
2386                         }
2387                 }
2388                 adev->ip_blocks[i].status.hw = false;
2389         }
2390
2391         return 0;
2392 }
2393
2394 /**
2395  * amdgpu_device_ip_suspend - run suspend for hardware IPs
2396  *
2397  * @adev: amdgpu_device pointer
2398  *
2399  * Main suspend function for hardware IPs.  The list of all the hardware
2400  * IPs that make up the asic is walked, clockgating is disabled and the
2401  * suspend callbacks are run.  suspend puts the hardware and software state
2402  * in each IP into a state suitable for suspend.
2403  * Returns 0 on success, negative error code on failure.
2404  */
2405 int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
2406 {
2407         int r;
2408
2409         if (amdgpu_sriov_vf(adev))
2410                 amdgpu_virt_request_full_gpu(adev, false);
2411
2412         r = amdgpu_device_ip_suspend_phase1(adev);
2413         if (r)
2414                 return r;
2415         r = amdgpu_device_ip_suspend_phase2(adev);
2416
2417         if (amdgpu_sriov_vf(adev))
2418                 amdgpu_virt_release_full_gpu(adev, false);
2419
2420         return r;
2421 }
2422
2423 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
2424 {
2425         int i, r;
2426
2427         static enum amd_ip_block_type ip_order[] = {
2428                 AMD_IP_BLOCK_TYPE_GMC,
2429                 AMD_IP_BLOCK_TYPE_COMMON,
2430                 AMD_IP_BLOCK_TYPE_PSP,
2431                 AMD_IP_BLOCK_TYPE_IH,
2432         };
2433
2434         for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2435                 int j;
2436                 struct amdgpu_ip_block *block;
2437
2438                 for (j = 0; j < adev->num_ip_blocks; j++) {
2439                         block = &adev->ip_blocks[j];
2440
2441                         block->status.hw = false;
2442                         if (block->version->type != ip_order[i] ||
2443                                 !block->status.valid)
2444                                 continue;
2445
2446                         r = block->version->funcs->hw_init(adev);
2447                         DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2448                         if (r)
2449                                 return r;
2450                         block->status.hw = true;
2451                 }
2452         }
2453
2454         return 0;
2455 }
2456
2457 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
2458 {
2459         int i, r;
2460
2461         static enum amd_ip_block_type ip_order[] = {
2462                 AMD_IP_BLOCK_TYPE_SMC,
2463                 AMD_IP_BLOCK_TYPE_DCE,
2464                 AMD_IP_BLOCK_TYPE_GFX,
2465                 AMD_IP_BLOCK_TYPE_SDMA,
2466                 AMD_IP_BLOCK_TYPE_UVD,
2467                 AMD_IP_BLOCK_TYPE_VCE,
2468                 AMD_IP_BLOCK_TYPE_VCN
2469         };
2470
2471         for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2472                 int j;
2473                 struct amdgpu_ip_block *block;
2474
2475                 for (j = 0; j < adev->num_ip_blocks; j++) {
2476                         block = &adev->ip_blocks[j];
2477
2478                         if (block->version->type != ip_order[i] ||
2479                                 !block->status.valid ||
2480                                 block->status.hw)
2481                                 continue;
2482
2483                         if (block->version->type == AMD_IP_BLOCK_TYPE_SMC)
2484                                 r = block->version->funcs->resume(adev);
2485                         else
2486                                 r = block->version->funcs->hw_init(adev);
2487
2488                         DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2489                         if (r)
2490                                 return r;
2491                         block->status.hw = true;
2492                 }
2493         }
2494
2495         return 0;
2496 }
2497
2498 /**
2499  * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
2500  *
2501  * @adev: amdgpu_device pointer
2502  *
2503  * First resume function for hardware IPs.  The list of all the hardware
2504  * IPs that make up the asic is walked and the resume callbacks are run for
2505  * COMMON, GMC, and IH.  resume puts the hardware into a functional state
2506  * after a suspend and updates the software state as necessary.  This
2507  * function is also used for restoring the GPU after a GPU reset.
2508  * Returns 0 on success, negative error code on failure.
2509  */
2510 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
2511 {
2512         int i, r;
2513
2514         for (i = 0; i < adev->num_ip_blocks; i++) {
2515                 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
2516                         continue;
2517                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2518                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2519                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2520
2521                         r = adev->ip_blocks[i].version->funcs->resume(adev);
2522                         if (r) {
2523                                 DRM_ERROR("resume of IP block <%s> failed %d\n",
2524                                           adev->ip_blocks[i].version->funcs->name, r);
2525                                 return r;
2526                         }
2527                         adev->ip_blocks[i].status.hw = true;
2528                 }
2529         }
2530
2531         return 0;
2532 }
2533
2534 /**
2535  * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
2536  *
2537  * @adev: amdgpu_device pointer
2538  *
2539  * First resume function for hardware IPs.  The list of all the hardware
2540  * IPs that make up the asic is walked and the resume callbacks are run for
2541  * all blocks except COMMON, GMC, and IH.  resume puts the hardware into a
2542  * functional state after a suspend and updates the software state as
2543  * necessary.  This function is also used for restoring the GPU after a GPU
2544  * reset.
2545  * Returns 0 on success, negative error code on failure.
2546  */
2547 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
2548 {
2549         int i, r;
2550
2551         for (i = 0; i < adev->num_ip_blocks; i++) {
2552                 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
2553                         continue;
2554                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2555                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2556                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
2557                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
2558                         continue;
2559                 r = adev->ip_blocks[i].version->funcs->resume(adev);
2560                 if (r) {
2561                         DRM_ERROR("resume of IP block <%s> failed %d\n",
2562                                   adev->ip_blocks[i].version->funcs->name, r);
2563                         return r;
2564                 }
2565                 adev->ip_blocks[i].status.hw = true;
2566         }
2567
2568         return 0;
2569 }
2570
2571 /**
2572  * amdgpu_device_ip_resume - run resume for hardware IPs
2573  *
2574  * @adev: amdgpu_device pointer
2575  *
2576  * Main resume function for hardware IPs.  The hardware IPs
2577  * are split into two resume functions because they are
2578  * are also used in in recovering from a GPU reset and some additional
2579  * steps need to be take between them.  In this case (S3/S4) they are
2580  * run sequentially.
2581  * Returns 0 on success, negative error code on failure.
2582  */
2583 static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
2584 {
2585         int r;
2586
2587         r = amdgpu_device_ip_resume_phase1(adev);
2588         if (r)
2589                 return r;
2590
2591         r = amdgpu_device_fw_loading(adev);
2592         if (r)
2593                 return r;
2594
2595         r = amdgpu_device_ip_resume_phase2(adev);
2596
2597         return r;
2598 }
2599
2600 /**
2601  * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
2602  *
2603  * @adev: amdgpu_device pointer
2604  *
2605  * Query the VBIOS data tables to determine if the board supports SR-IOV.
2606  */
2607 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
2608 {
2609         if (amdgpu_sriov_vf(adev)) {
2610                 if (adev->is_atom_fw) {
2611                         if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2612                                 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2613                 } else {
2614                         if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2615                                 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2616                 }
2617
2618                 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2619                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
2620         }
2621 }
2622
2623 /**
2624  * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
2625  *
2626  * @asic_type: AMD asic type
2627  *
2628  * Check if there is DC (new modesetting infrastructre) support for an asic.
2629  * returns true if DC has support, false if not.
2630  */
2631 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2632 {
2633         switch (asic_type) {
2634 #if defined(CONFIG_DRM_AMD_DC)
2635         case CHIP_BONAIRE:
2636         case CHIP_KAVERI:
2637         case CHIP_KABINI:
2638         case CHIP_MULLINS:
2639                 /*
2640                  * We have systems in the wild with these ASICs that require
2641                  * LVDS and VGA support which is not supported with DC.
2642                  *
2643                  * Fallback to the non-DC driver here by default so as not to
2644                  * cause regressions.
2645                  */
2646                 return amdgpu_dc > 0;
2647         case CHIP_HAWAII:
2648         case CHIP_CARRIZO:
2649         case CHIP_STONEY:
2650         case CHIP_POLARIS10:
2651         case CHIP_POLARIS11:
2652         case CHIP_POLARIS12:
2653         case CHIP_VEGAM:
2654         case CHIP_TONGA:
2655         case CHIP_FIJI:
2656         case CHIP_VEGA10:
2657         case CHIP_VEGA12:
2658         case CHIP_VEGA20:
2659 #if defined(CONFIG_DRM_AMD_DC_DCN)
2660         case CHIP_RAVEN:
2661         case CHIP_NAVI10:
2662         case CHIP_NAVI14:
2663         case CHIP_NAVI12:
2664         case CHIP_RENOIR:
2665 #endif
2666                 return amdgpu_dc != 0;
2667 #endif
2668         default:
2669                 if (amdgpu_dc > 0)
2670                         DRM_INFO("Display Core has been requested via kernel parameter "
2671                                          "but isn't supported by ASIC, ignoring\n");
2672                 return false;
2673         }
2674 }
2675
2676 /**
2677  * amdgpu_device_has_dc_support - check if dc is supported
2678  *
2679  * @adev: amdgpu_device_pointer
2680  *
2681  * Returns true for supported, false for not supported
2682  */
2683 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2684 {
2685         if (amdgpu_sriov_vf(adev))
2686                 return false;
2687
2688         return amdgpu_device_asic_has_dc_support(adev->asic_type);
2689 }
2690
2691
2692 static void amdgpu_device_xgmi_reset_func(struct work_struct *__work)
2693 {
2694         struct amdgpu_device *adev =
2695                 container_of(__work, struct amdgpu_device, xgmi_reset_work);
2696         struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev, 0);
2697
2698         /* It's a bug to not have a hive within this function */
2699         if (WARN_ON(!hive))
2700                 return;
2701
2702         /*
2703          * Use task barrier to synchronize all xgmi reset works across the
2704          * hive. task_barrier_enter and task_barrier_exit will block
2705          * until all the threads running the xgmi reset works reach
2706          * those points. task_barrier_full will do both blocks.
2707          */
2708         if (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) {
2709
2710                 task_barrier_enter(&hive->tb);
2711                 adev->asic_reset_res = amdgpu_device_baco_enter(adev->ddev);
2712
2713                 if (adev->asic_reset_res)
2714                         goto fail;
2715
2716                 task_barrier_exit(&hive->tb);
2717                 adev->asic_reset_res = amdgpu_device_baco_exit(adev->ddev);
2718
2719                 if (adev->asic_reset_res)
2720                         goto fail;
2721         } else {
2722
2723                 task_barrier_full(&hive->tb);
2724                 adev->asic_reset_res =  amdgpu_asic_reset(adev);
2725         }
2726
2727 fail:
2728         if (adev->asic_reset_res)
2729                 DRM_WARN("ASIC reset failed with error, %d for drm dev, %s",
2730                          adev->asic_reset_res, adev->ddev->unique);
2731 }
2732
2733 static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev)
2734 {
2735         char *input = amdgpu_lockup_timeout;
2736         char *timeout_setting = NULL;
2737         int index = 0;
2738         long timeout;
2739         int ret = 0;
2740
2741         /*
2742          * By default timeout for non compute jobs is 10000.
2743          * And there is no timeout enforced on compute jobs.
2744          * In SR-IOV or passthrough mode, timeout for compute
2745          * jobs are 10000 by default.
2746          */
2747         adev->gfx_timeout = msecs_to_jiffies(10000);
2748         adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
2749         if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
2750                 adev->compute_timeout = adev->gfx_timeout;
2751         else
2752                 adev->compute_timeout = MAX_SCHEDULE_TIMEOUT;
2753
2754         if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
2755                 while ((timeout_setting = strsep(&input, ",")) &&
2756                                 strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
2757                         ret = kstrtol(timeout_setting, 0, &timeout);
2758                         if (ret)
2759                                 return ret;
2760
2761                         if (timeout == 0) {
2762                                 index++;
2763                                 continue;
2764                         } else if (timeout < 0) {
2765                                 timeout = MAX_SCHEDULE_TIMEOUT;
2766                         } else {
2767                                 timeout = msecs_to_jiffies(timeout);
2768                         }
2769
2770                         switch (index++) {
2771                         case 0:
2772                                 adev->gfx_timeout = timeout;
2773                                 break;
2774                         case 1:
2775                                 adev->compute_timeout = timeout;
2776                                 break;
2777                         case 2:
2778                                 adev->sdma_timeout = timeout;
2779                                 break;
2780                         case 3:
2781                                 adev->video_timeout = timeout;
2782                                 break;
2783                         default:
2784                                 break;
2785                         }
2786                 }
2787                 /*
2788                  * There is only one value specified and
2789                  * it should apply to all non-compute jobs.
2790                  */
2791                 if (index == 1) {
2792                         adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
2793                         if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
2794                                 adev->compute_timeout = adev->gfx_timeout;
2795                 }
2796         }
2797
2798         return ret;
2799 }
2800
2801 /**
2802  * amdgpu_device_init - initialize the driver
2803  *
2804  * @adev: amdgpu_device pointer
2805  * @ddev: drm dev pointer
2806  * @pdev: pci dev pointer
2807  * @flags: driver flags
2808  *
2809  * Initializes the driver info and hw (all asics).
2810  * Returns 0 for success or an error on failure.
2811  * Called at driver startup.
2812  */
2813 int amdgpu_device_init(struct amdgpu_device *adev,
2814                        struct drm_device *ddev,
2815                        struct pci_dev *pdev,
2816                        uint32_t flags)
2817 {
2818         int r, i;
2819         bool boco = false;
2820         u32 max_MBps;
2821
2822         adev->shutdown = false;
2823         adev->dev = &pdev->dev;
2824         adev->ddev = ddev;
2825         adev->pdev = pdev;
2826         adev->flags = flags;
2827
2828         if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST)
2829                 adev->asic_type = amdgpu_force_asic_type;
2830         else
2831                 adev->asic_type = flags & AMD_ASIC_MASK;
2832
2833         adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
2834         if (amdgpu_emu_mode == 1)
2835                 adev->usec_timeout *= 10;
2836         adev->gmc.gart_size = 512 * 1024 * 1024;
2837         adev->accel_working = false;
2838         adev->num_rings = 0;
2839         adev->mman.buffer_funcs = NULL;
2840         adev->mman.buffer_funcs_ring = NULL;
2841         adev->vm_manager.vm_pte_funcs = NULL;
2842         adev->vm_manager.vm_pte_num_scheds = 0;
2843         adev->gmc.gmc_funcs = NULL;
2844         adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2845         bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
2846
2847         adev->smc_rreg = &amdgpu_invalid_rreg;
2848         adev->smc_wreg = &amdgpu_invalid_wreg;
2849         adev->pcie_rreg = &amdgpu_invalid_rreg;
2850         adev->pcie_wreg = &amdgpu_invalid_wreg;
2851         adev->pciep_rreg = &amdgpu_invalid_rreg;
2852         adev->pciep_wreg = &amdgpu_invalid_wreg;
2853         adev->pcie_rreg64 = &amdgpu_invalid_rreg64;
2854         adev->pcie_wreg64 = &amdgpu_invalid_wreg64;
2855         adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
2856         adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
2857         adev->didt_rreg = &amdgpu_invalid_rreg;
2858         adev->didt_wreg = &amdgpu_invalid_wreg;
2859         adev->gc_cac_rreg = &amdgpu_invalid_rreg;
2860         adev->gc_cac_wreg = &amdgpu_invalid_wreg;
2861         adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
2862         adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
2863
2864         DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
2865                  amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
2866                  pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
2867
2868         /* mutex initialization are all done here so we
2869          * can recall function without having locking issues */
2870         atomic_set(&adev->irq.ih.lock, 0);
2871         mutex_init(&adev->firmware.mutex);
2872         mutex_init(&adev->pm.mutex);
2873         mutex_init(&adev->gfx.gpu_clock_mutex);
2874         mutex_init(&adev->srbm_mutex);
2875         mutex_init(&adev->gfx.pipe_reserve_mutex);
2876         mutex_init(&adev->gfx.gfx_off_mutex);
2877         mutex_init(&adev->grbm_idx_mutex);
2878         mutex_init(&adev->mn_lock);
2879         mutex_init(&adev->virt.vf_errors.lock);
2880         hash_init(adev->mn_hash);
2881         mutex_init(&adev->lock_reset);
2882         mutex_init(&adev->psp.mutex);
2883         mutex_init(&adev->notifier_lock);
2884
2885         r = amdgpu_device_check_arguments(adev);
2886         if (r)
2887                 return r;
2888
2889         spin_lock_init(&adev->mmio_idx_lock);
2890         spin_lock_init(&adev->smc_idx_lock);
2891         spin_lock_init(&adev->pcie_idx_lock);
2892         spin_lock_init(&adev->uvd_ctx_idx_lock);
2893         spin_lock_init(&adev->didt_idx_lock);
2894         spin_lock_init(&adev->gc_cac_idx_lock);
2895         spin_lock_init(&adev->se_cac_idx_lock);
2896         spin_lock_init(&adev->audio_endpt_idx_lock);
2897         spin_lock_init(&adev->mm_stats.lock);
2898
2899         INIT_LIST_HEAD(&adev->shadow_list);
2900         mutex_init(&adev->shadow_list_lock);
2901
2902         INIT_LIST_HEAD(&adev->ring_lru_list);
2903         spin_lock_init(&adev->ring_lru_list_lock);
2904
2905         INIT_DELAYED_WORK(&adev->delayed_init_work,
2906                           amdgpu_device_delayed_init_work_handler);
2907         INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
2908                           amdgpu_device_delay_enable_gfx_off);
2909
2910         INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func);
2911
2912         adev->gfx.gfx_off_req_count = 1;
2913         adev->pm.ac_power = power_supply_is_system_supplied() > 0 ? true : false;
2914
2915         /* Registers mapping */
2916         /* TODO: block userspace mapping of io register */
2917         if (adev->asic_type >= CHIP_BONAIRE) {
2918                 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
2919                 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
2920         } else {
2921                 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
2922                 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
2923         }
2924
2925         adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
2926         if (adev->rmmio == NULL) {
2927                 return -ENOMEM;
2928         }
2929         DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
2930         DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
2931
2932         /* io port mapping */
2933         for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
2934                 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
2935                         adev->rio_mem_size = pci_resource_len(adev->pdev, i);
2936                         adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
2937                         break;
2938                 }
2939         }
2940         if (adev->rio_mem == NULL)
2941                 DRM_INFO("PCI I/O BAR is not found.\n");
2942
2943         /* enable PCIE atomic ops */
2944         r = pci_enable_atomic_ops_to_root(adev->pdev,
2945                                           PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
2946                                           PCI_EXP_DEVCAP2_ATOMIC_COMP64);
2947         if (r) {
2948                 adev->have_atomics_support = false;
2949                 DRM_INFO("PCIE atomic ops is not supported\n");
2950         } else {
2951                 adev->have_atomics_support = true;
2952         }
2953
2954         amdgpu_device_get_pcie_info(adev);
2955
2956         if (amdgpu_mcbp)
2957                 DRM_INFO("MCBP is enabled\n");
2958
2959         if (amdgpu_mes && adev->asic_type >= CHIP_NAVI10)
2960                 adev->enable_mes = true;
2961
2962         if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10) {
2963                 r = amdgpu_discovery_init(adev);
2964                 if (r) {
2965                         dev_err(adev->dev, "amdgpu_discovery_init failed\n");
2966                         return r;
2967                 }
2968         }
2969
2970         /* early init functions */
2971         r = amdgpu_device_ip_early_init(adev);
2972         if (r)
2973                 return r;
2974
2975         r = amdgpu_device_get_job_timeout_settings(adev);
2976         if (r) {
2977                 dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n");
2978                 return r;
2979         }
2980
2981         /* doorbell bar mapping and doorbell index init*/
2982         amdgpu_device_doorbell_init(adev);
2983
2984         /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
2985         /* this will fail for cards that aren't VGA class devices, just
2986          * ignore it */
2987         vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
2988
2989         if (amdgpu_device_supports_boco(ddev))
2990                 boco = true;
2991         if (amdgpu_has_atpx() &&
2992             (amdgpu_is_atpx_hybrid() ||
2993              amdgpu_has_atpx_dgpu_power_cntl()) &&
2994             !pci_is_thunderbolt_attached(adev->pdev))
2995                 vga_switcheroo_register_client(adev->pdev,
2996                                                &amdgpu_switcheroo_ops, boco);
2997         if (boco)
2998                 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
2999
3000         if (amdgpu_emu_mode == 1) {
3001                 /* post the asic on emulation mode */
3002                 emu_soc_asic_init(adev);
3003                 goto fence_driver_init;
3004         }
3005
3006         /* detect if we are with an SRIOV vbios */
3007         amdgpu_device_detect_sriov_bios(adev);
3008
3009         /* check if we need to reset the asic
3010          *  E.g., driver was not cleanly unloaded previously, etc.
3011          */
3012         if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) {
3013                 r = amdgpu_asic_reset(adev);
3014                 if (r) {
3015                         dev_err(adev->dev, "asic reset on init failed\n");
3016                         goto failed;
3017                 }
3018         }
3019
3020         /* Post card if necessary */
3021         if (amdgpu_device_need_post(adev)) {
3022                 if (!adev->bios) {
3023                         dev_err(adev->dev, "no vBIOS found\n");
3024                         r = -EINVAL;
3025                         goto failed;
3026                 }
3027                 DRM_INFO("GPU posting now...\n");
3028                 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
3029                 if (r) {
3030                         dev_err(adev->dev, "gpu post error!\n");
3031                         goto failed;
3032                 }
3033         }
3034
3035         if (adev->is_atom_fw) {
3036                 /* Initialize clocks */
3037                 r = amdgpu_atomfirmware_get_clock_info(adev);
3038                 if (r) {
3039                         dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
3040                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
3041                         goto failed;
3042                 }
3043         } else {
3044                 /* Initialize clocks */
3045                 r = amdgpu_atombios_get_clock_info(adev);
3046                 if (r) {
3047                         dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
3048                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
3049                         goto failed;
3050                 }
3051                 /* init i2c buses */
3052                 if (!amdgpu_device_has_dc_support(adev))
3053                         amdgpu_atombios_i2c_init(adev);
3054         }
3055
3056 fence_driver_init:
3057         /* Fence driver */
3058         r = amdgpu_fence_driver_init(adev);
3059         if (r) {
3060                 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
3061                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
3062                 goto failed;
3063         }
3064
3065         /* init the mode config */
3066         drm_mode_config_init(adev->ddev);
3067
3068         r = amdgpu_device_ip_init(adev);
3069         if (r) {
3070                 /* failed in exclusive mode due to timeout */
3071                 if (amdgpu_sriov_vf(adev) &&
3072                     !amdgpu_sriov_runtime(adev) &&
3073                     amdgpu_virt_mmio_blocked(adev) &&
3074                     !amdgpu_virt_wait_reset(adev)) {
3075                         dev_err(adev->dev, "VF exclusive mode timeout\n");
3076                         /* Don't send request since VF is inactive. */
3077                         adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
3078                         adev->virt.ops = NULL;
3079                         r = -EAGAIN;
3080                         goto failed;
3081                 }
3082                 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
3083                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
3084                 goto failed;
3085         }
3086
3087         DRM_DEBUG("SE %d, SH per SE %d, CU per SH %d, active_cu_number %d\n",
3088                         adev->gfx.config.max_shader_engines,
3089                         adev->gfx.config.max_sh_per_se,
3090                         adev->gfx.config.max_cu_per_sh,
3091                         adev->gfx.cu_info.number);
3092
3093         amdgpu_ctx_init_sched(adev);
3094
3095         adev->accel_working = true;
3096
3097         amdgpu_vm_check_compute_bug(adev);
3098
3099         /* Initialize the buffer migration limit. */
3100         if (amdgpu_moverate >= 0)
3101                 max_MBps = amdgpu_moverate;
3102         else
3103                 max_MBps = 8; /* Allow 8 MB/s. */
3104         /* Get a log2 for easy divisions. */
3105         adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
3106
3107         amdgpu_fbdev_init(adev);
3108
3109         r = amdgpu_pm_sysfs_init(adev);
3110         if (r) {
3111                 adev->pm_sysfs_en = false;
3112                 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
3113         } else
3114                 adev->pm_sysfs_en = true;
3115
3116         r = amdgpu_ucode_sysfs_init(adev);
3117         if (r) {
3118                 adev->ucode_sysfs_en = false;
3119                 DRM_ERROR("Creating firmware sysfs failed (%d).\n", r);
3120         } else
3121                 adev->ucode_sysfs_en = true;
3122
3123         r = amdgpu_debugfs_gem_init(adev);
3124         if (r)
3125                 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
3126
3127         r = amdgpu_debugfs_regs_init(adev);
3128         if (r)
3129                 DRM_ERROR("registering register debugfs failed (%d).\n", r);
3130
3131         r = amdgpu_debugfs_firmware_init(adev);
3132         if (r)
3133                 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
3134
3135         r = amdgpu_debugfs_init(adev);
3136         if (r)
3137                 DRM_ERROR("Creating debugfs files failed (%d).\n", r);
3138
3139         if ((amdgpu_testing & 1)) {
3140                 if (adev->accel_working)
3141                         amdgpu_test_moves(adev);
3142                 else
3143                         DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
3144         }
3145         if (amdgpu_benchmarking) {
3146                 if (adev->accel_working)
3147                         amdgpu_benchmark(adev, amdgpu_benchmarking);
3148                 else
3149                         DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
3150         }
3151
3152         /*
3153          * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost.
3154          * Otherwise the mgpu fan boost feature will be skipped due to the
3155          * gpu instance is counted less.
3156          */
3157         amdgpu_register_gpu_instance(adev);
3158
3159         /* enable clockgating, etc. after ib tests, etc. since some blocks require
3160          * explicit gating rather than handling it automatically.
3161          */
3162         r = amdgpu_device_ip_late_init(adev);
3163         if (r) {
3164                 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
3165                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
3166                 goto failed;
3167         }
3168
3169         /* must succeed. */
3170         amdgpu_ras_resume(adev);
3171
3172         queue_delayed_work(system_wq, &adev->delayed_init_work,
3173                            msecs_to_jiffies(AMDGPU_RESUME_MS));
3174
3175         r = device_create_file(adev->dev, &dev_attr_pcie_replay_count);
3176         if (r) {
3177                 dev_err(adev->dev, "Could not create pcie_replay_count");
3178                 return r;
3179         }
3180
3181         if (IS_ENABLED(CONFIG_PERF_EVENTS))
3182                 r = amdgpu_pmu_init(adev);
3183         if (r)
3184                 dev_err(adev->dev, "amdgpu_pmu_init failed\n");
3185
3186         return 0;
3187
3188 failed:
3189         amdgpu_vf_error_trans_all(adev);
3190         if (boco)
3191                 vga_switcheroo_fini_domain_pm_ops(adev->dev);
3192
3193         return r;
3194 }
3195
3196 /**
3197  * amdgpu_device_fini - tear down the driver
3198  *
3199  * @adev: amdgpu_device pointer
3200  *
3201  * Tear down the driver info (all asics).
3202  * Called at driver shutdown.
3203  */
3204 void amdgpu_device_fini(struct amdgpu_device *adev)
3205 {
3206         int r;
3207
3208         DRM_INFO("amdgpu: finishing device.\n");
3209         flush_delayed_work(&adev->delayed_init_work);
3210         adev->shutdown = true;
3211
3212         /* disable all interrupts */
3213         amdgpu_irq_disable_all(adev);
3214         if (adev->mode_info.mode_config_initialized){
3215                 if (!amdgpu_device_has_dc_support(adev))
3216                         drm_helper_force_disable_all(adev->ddev);
3217                 else
3218                         drm_atomic_helper_shutdown(adev->ddev);
3219         }
3220         amdgpu_fence_driver_fini(adev);
3221         if (adev->pm_sysfs_en)
3222                 amdgpu_pm_sysfs_fini(adev);
3223         amdgpu_fbdev_fini(adev);
3224         r = amdgpu_device_ip_fini(adev);
3225         if (adev->firmware.gpu_info_fw) {
3226                 release_firmware(adev->firmware.gpu_info_fw);
3227                 adev->firmware.gpu_info_fw = NULL;
3228         }
3229         adev->accel_working = false;
3230         /* free i2c buses */
3231         if (!amdgpu_device_has_dc_support(adev))
3232                 amdgpu_i2c_fini(adev);
3233
3234         if (amdgpu_emu_mode != 1)
3235                 amdgpu_atombios_fini(adev);
3236
3237         kfree(adev->bios);
3238         adev->bios = NULL;
3239         if (amdgpu_has_atpx() &&
3240             (amdgpu_is_atpx_hybrid() ||
3241              amdgpu_has_atpx_dgpu_power_cntl()) &&
3242             !pci_is_thunderbolt_attached(adev->pdev))
3243                 vga_switcheroo_unregister_client(adev->pdev);
3244         if (amdgpu_device_supports_boco(adev->ddev))
3245                 vga_switcheroo_fini_domain_pm_ops(adev->dev);
3246         vga_client_register(adev->pdev, NULL, NULL, NULL);
3247         if (adev->rio_mem)
3248                 pci_iounmap(adev->pdev, adev->rio_mem);
3249         adev->rio_mem = NULL;
3250         iounmap(adev->rmmio);
3251         adev->rmmio = NULL;
3252         amdgpu_device_doorbell_fini(adev);
3253
3254         amdgpu_debugfs_regs_cleanup(adev);
3255         device_remove_file(adev->dev, &dev_attr_pcie_replay_count);
3256         if (adev->ucode_sysfs_en)
3257                 amdgpu_ucode_sysfs_fini(adev);
3258         if (IS_ENABLED(CONFIG_PERF_EVENTS))
3259                 amdgpu_pmu_fini(adev);
3260         amdgpu_debugfs_preempt_cleanup(adev);
3261         if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10)
3262                 amdgpu_discovery_fini(adev);
3263 }
3264
3265
3266 /*
3267  * Suspend & resume.
3268  */
3269 /**
3270  * amdgpu_device_suspend - initiate device suspend
3271  *
3272  * @dev: drm dev pointer
3273  * @suspend: suspend state
3274  * @fbcon : notify the fbdev of suspend
3275  *
3276  * Puts the hw in the suspend state (all asics).
3277  * Returns 0 for success or an error on failure.
3278  * Called at driver suspend.
3279  */
3280 int amdgpu_device_suspend(struct drm_device *dev, bool fbcon)
3281 {
3282         struct amdgpu_device *adev;
3283         struct drm_crtc *crtc;
3284         struct drm_connector *connector;
3285         struct drm_connector_list_iter iter;
3286         int r;
3287
3288         if (dev == NULL || dev->dev_private == NULL) {
3289                 return -ENODEV;
3290         }
3291
3292         adev = dev->dev_private;
3293
3294         if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3295                 return 0;
3296
3297         adev->in_suspend = true;
3298         drm_kms_helper_poll_disable(dev);
3299
3300         if (fbcon)
3301                 amdgpu_fbdev_set_suspend(adev, 1);
3302
3303         cancel_delayed_work_sync(&adev->delayed_init_work);
3304
3305         if (!amdgpu_device_has_dc_support(adev)) {
3306                 /* turn off display hw */
3307                 drm_modeset_lock_all(dev);
3308                 drm_connector_list_iter_begin(dev, &iter);
3309                 drm_for_each_connector_iter(connector, &iter)
3310                         drm_helper_connector_dpms(connector,
3311                                                   DRM_MODE_DPMS_OFF);
3312                 drm_connector_list_iter_end(&iter);
3313                 drm_modeset_unlock_all(dev);
3314                         /* unpin the front buffers and cursors */
3315                 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
3316                         struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
3317                         struct drm_framebuffer *fb = crtc->primary->fb;
3318                         struct amdgpu_bo *robj;
3319
3320                         if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) {
3321                                 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
3322                                 r = amdgpu_bo_reserve(aobj, true);
3323                                 if (r == 0) {
3324                                         amdgpu_bo_unpin(aobj);
3325                                         amdgpu_bo_unreserve(aobj);
3326                                 }
3327                         }
3328
3329                         if (fb == NULL || fb->obj[0] == NULL) {
3330                                 continue;
3331                         }
3332                         robj = gem_to_amdgpu_bo(fb->obj[0]);
3333                         /* don't unpin kernel fb objects */
3334                         if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
3335                                 r = amdgpu_bo_reserve(robj, true);
3336                                 if (r == 0) {
3337                                         amdgpu_bo_unpin(robj);
3338                                         amdgpu_bo_unreserve(robj);
3339                                 }
3340                         }
3341                 }
3342         }
3343
3344         amdgpu_amdkfd_suspend(adev, !fbcon);
3345
3346         amdgpu_ras_suspend(adev);
3347
3348         r = amdgpu_device_ip_suspend_phase1(adev);
3349
3350         /* evict vram memory */
3351         amdgpu_bo_evict_vram(adev);
3352
3353         amdgpu_fence_driver_suspend(adev);
3354
3355         r = amdgpu_device_ip_suspend_phase2(adev);
3356
3357         /* evict remaining vram memory
3358          * This second call to evict vram is to evict the gart page table
3359          * using the CPU.
3360          */
3361         amdgpu_bo_evict_vram(adev);
3362
3363         return 0;
3364 }
3365
3366 /**
3367  * amdgpu_device_resume - initiate device resume
3368  *
3369  * @dev: drm dev pointer
3370  * @resume: resume state
3371  * @fbcon : notify the fbdev of resume
3372  *
3373  * Bring the hw back to operating state (all asics).
3374  * Returns 0 for success or an error on failure.
3375  * Called at driver resume.
3376  */
3377 int amdgpu_device_resume(struct drm_device *dev, bool fbcon)
3378 {
3379         struct drm_connector *connector;
3380         struct drm_connector_list_iter iter;
3381         struct amdgpu_device *adev = dev->dev_private;
3382         struct drm_crtc *crtc;
3383         int r = 0;
3384
3385         if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3386                 return 0;
3387
3388         /* post card */
3389         if (amdgpu_device_need_post(adev)) {
3390                 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
3391                 if (r)
3392                         DRM_ERROR("amdgpu asic init failed\n");
3393         }
3394
3395         r = amdgpu_device_ip_resume(adev);
3396         if (r) {
3397                 DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
3398                 return r;
3399         }
3400         amdgpu_fence_driver_resume(adev);
3401
3402
3403         r = amdgpu_device_ip_late_init(adev);
3404         if (r)
3405                 return r;
3406
3407         queue_delayed_work(system_wq, &adev->delayed_init_work,
3408                            msecs_to_jiffies(AMDGPU_RESUME_MS));
3409
3410         if (!amdgpu_device_has_dc_support(adev)) {
3411                 /* pin cursors */
3412                 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
3413                         struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
3414
3415                         if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) {
3416                                 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
3417                                 r = amdgpu_bo_reserve(aobj, true);
3418                                 if (r == 0) {
3419                                         r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM);
3420                                         if (r != 0)
3421                                                 DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
3422                                         amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj);
3423                                         amdgpu_bo_unreserve(aobj);
3424                                 }
3425                         }
3426                 }
3427         }
3428         r = amdgpu_amdkfd_resume(adev, !fbcon);
3429         if (r)
3430                 return r;
3431
3432         /* Make sure IB tests flushed */
3433         flush_delayed_work(&adev->delayed_init_work);
3434
3435         /* blat the mode back in */
3436         if (fbcon) {
3437                 if (!amdgpu_device_has_dc_support(adev)) {
3438                         /* pre DCE11 */
3439                         drm_helper_resume_force_mode(dev);
3440
3441                         /* turn on display hw */
3442                         drm_modeset_lock_all(dev);
3443
3444                         drm_connector_list_iter_begin(dev, &iter);
3445                         drm_for_each_connector_iter(connector, &iter)
3446                                 drm_helper_connector_dpms(connector,
3447                                                           DRM_MODE_DPMS_ON);
3448                         drm_connector_list_iter_end(&iter);
3449
3450                         drm_modeset_unlock_all(dev);
3451                 }
3452                 amdgpu_fbdev_set_suspend(adev, 0);
3453         }
3454
3455         drm_kms_helper_poll_enable(dev);
3456
3457         amdgpu_ras_resume(adev);
3458
3459         /*
3460          * Most of the connector probing functions try to acquire runtime pm
3461          * refs to ensure that the GPU is powered on when connector polling is
3462          * performed. Since we're calling this from a runtime PM callback,
3463          * trying to acquire rpm refs will cause us to deadlock.
3464          *
3465          * Since we're guaranteed to be holding the rpm lock, it's safe to
3466          * temporarily disable the rpm helpers so this doesn't deadlock us.
3467          */
3468 #ifdef CONFIG_PM
3469         dev->dev->power.disable_depth++;
3470 #endif
3471         if (!amdgpu_device_has_dc_support(adev))
3472                 drm_helper_hpd_irq_event(dev);
3473         else
3474                 drm_kms_helper_hotplug_event(dev);
3475 #ifdef CONFIG_PM
3476         dev->dev->power.disable_depth--;
3477 #endif
3478         adev->in_suspend = false;
3479
3480         return 0;
3481 }
3482
3483 /**
3484  * amdgpu_device_ip_check_soft_reset - did soft reset succeed
3485  *
3486  * @adev: amdgpu_device pointer
3487  *
3488  * The list of all the hardware IPs that make up the asic is walked and
3489  * the check_soft_reset callbacks are run.  check_soft_reset determines
3490  * if the asic is still hung or not.
3491  * Returns true if any of the IPs are still in a hung state, false if not.
3492  */
3493 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
3494 {
3495         int i;
3496         bool asic_hang = false;
3497
3498         if (amdgpu_sriov_vf(adev))
3499                 return true;
3500
3501         if (amdgpu_asic_need_full_reset(adev))
3502                 return true;
3503
3504         for (i = 0; i < adev->num_ip_blocks; i++) {
3505                 if (!adev->ip_blocks[i].status.valid)
3506                         continue;
3507                 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
3508                         adev->ip_blocks[i].status.hang =
3509                                 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
3510                 if (adev->ip_blocks[i].status.hang) {
3511                         DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
3512                         asic_hang = true;
3513                 }
3514         }
3515         return asic_hang;
3516 }
3517
3518 /**
3519  * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
3520  *
3521  * @adev: amdgpu_device pointer
3522  *
3523  * The list of all the hardware IPs that make up the asic is walked and the
3524  * pre_soft_reset callbacks are run if the block is hung.  pre_soft_reset
3525  * handles any IP specific hardware or software state changes that are
3526  * necessary for a soft reset to succeed.
3527  * Returns 0 on success, negative error code on failure.
3528  */
3529 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
3530 {
3531         int i, r = 0;
3532
3533         for (i = 0; i < adev->num_ip_blocks; i++) {
3534                 if (!adev->ip_blocks[i].status.valid)
3535                         continue;
3536                 if (adev->ip_blocks[i].status.hang &&
3537                     adev->ip_blocks[i].version->funcs->pre_soft_reset) {
3538                         r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
3539                         if (r)
3540                                 return r;
3541                 }
3542         }
3543
3544         return 0;
3545 }
3546
3547 /**
3548  * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
3549  *
3550  * @adev: amdgpu_device pointer
3551  *
3552  * Some hardware IPs cannot be soft reset.  If they are hung, a full gpu
3553  * reset is necessary to recover.
3554  * Returns true if a full asic reset is required, false if not.
3555  */
3556 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
3557 {
3558         int i;
3559
3560         if (amdgpu_asic_need_full_reset(adev))
3561                 return true;
3562
3563         for (i = 0; i < adev->num_ip_blocks; i++) {
3564                 if (!adev->ip_blocks[i].status.valid)
3565                         continue;
3566                 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
3567                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
3568                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
3569                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
3570                      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
3571                         if (adev->ip_blocks[i].status.hang) {
3572                                 DRM_INFO("Some block need full reset!\n");
3573                                 return true;
3574                         }
3575                 }
3576         }
3577         return false;
3578 }
3579
3580 /**
3581  * amdgpu_device_ip_soft_reset - do a soft reset
3582  *
3583  * @adev: amdgpu_device pointer
3584  *
3585  * The list of all the hardware IPs that make up the asic is walked and the
3586  * soft_reset callbacks are run if the block is hung.  soft_reset handles any
3587  * IP specific hardware or software state changes that are necessary to soft
3588  * reset the IP.
3589  * Returns 0 on success, negative error code on failure.
3590  */
3591 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
3592 {
3593         int i, r = 0;
3594
3595         for (i = 0; i < adev->num_ip_blocks; i++) {
3596                 if (!adev->ip_blocks[i].status.valid)
3597                         continue;
3598                 if (adev->ip_blocks[i].status.hang &&
3599                     adev->ip_blocks[i].version->funcs->soft_reset) {
3600                         r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
3601                         if (r)
3602                                 return r;
3603                 }
3604         }
3605
3606         return 0;
3607 }
3608
3609 /**
3610  * amdgpu_device_ip_post_soft_reset - clean up from soft reset
3611  *
3612  * @adev: amdgpu_device pointer
3613  *
3614  * The list of all the hardware IPs that make up the asic is walked and the
3615  * post_soft_reset callbacks are run if the asic was hung.  post_soft_reset
3616  * handles any IP specific hardware or software state changes that are
3617  * necessary after the IP has been soft reset.
3618  * Returns 0 on success, negative error code on failure.
3619  */
3620 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
3621 {
3622         int i, r = 0;
3623
3624         for (i = 0; i < adev->num_ip_blocks; i++) {
3625                 if (!adev->ip_blocks[i].status.valid)
3626                         continue;
3627                 if (adev->ip_blocks[i].status.hang &&
3628                     adev->ip_blocks[i].version->funcs->post_soft_reset)
3629                         r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
3630                 if (r)
3631                         return r;
3632         }
3633
3634         return 0;
3635 }
3636
3637 /**
3638  * amdgpu_device_recover_vram - Recover some VRAM contents
3639  *
3640  * @adev: amdgpu_device pointer
3641  *
3642  * Restores the contents of VRAM buffers from the shadows in GTT.  Used to
3643  * restore things like GPUVM page tables after a GPU reset where
3644  * the contents of VRAM might be lost.
3645  *
3646  * Returns:
3647  * 0 on success, negative error code on failure.
3648  */
3649 static int amdgpu_device_recover_vram(struct amdgpu_device *adev)
3650 {
3651         struct dma_fence *fence = NULL, *next = NULL;
3652         struct amdgpu_bo *shadow;
3653         long r = 1, tmo;
3654
3655         if (amdgpu_sriov_runtime(adev))
3656                 tmo = msecs_to_jiffies(8000);
3657         else
3658                 tmo = msecs_to_jiffies(100);
3659
3660         DRM_INFO("recover vram bo from shadow start\n");
3661         mutex_lock(&adev->shadow_list_lock);
3662         list_for_each_entry(shadow, &adev->shadow_list, shadow_list) {
3663
3664                 /* No need to recover an evicted BO */
3665                 if (shadow->tbo.mem.mem_type != TTM_PL_TT ||
3666                     shadow->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET ||
3667                     shadow->parent->tbo.mem.mem_type != TTM_PL_VRAM)
3668                         continue;
3669
3670                 r = amdgpu_bo_restore_shadow(shadow, &next);
3671                 if (r)
3672                         break;
3673
3674                 if (fence) {
3675                         tmo = dma_fence_wait_timeout(fence, false, tmo);
3676                         dma_fence_put(fence);
3677                         fence = next;
3678                         if (tmo == 0) {
3679                                 r = -ETIMEDOUT;
3680                                 break;
3681                         } else if (tmo < 0) {
3682                                 r = tmo;
3683                                 break;
3684                         }
3685                 } else {
3686                         fence = next;
3687                 }
3688         }
3689         mutex_unlock(&adev->shadow_list_lock);
3690
3691         if (fence)
3692                 tmo = dma_fence_wait_timeout(fence, false, tmo);
3693         dma_fence_put(fence);
3694
3695         if (r < 0 || tmo <= 0) {
3696                 DRM_ERROR("recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo);
3697                 return -EIO;
3698         }
3699
3700         DRM_INFO("recover vram bo from shadow done\n");
3701         return 0;
3702 }
3703
3704
3705 /**
3706  * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
3707  *
3708  * @adev: amdgpu device pointer
3709  * @from_hypervisor: request from hypervisor
3710  *
3711  * do VF FLR and reinitialize Asic
3712  * return 0 means succeeded otherwise failed
3713  */
3714 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
3715                                      bool from_hypervisor)
3716 {
3717         int r;
3718
3719         if (from_hypervisor)
3720                 r = amdgpu_virt_request_full_gpu(adev, true);
3721         else
3722                 r = amdgpu_virt_reset_gpu(adev);
3723         if (r)
3724                 return r;
3725
3726         /* Resume IP prior to SMC */
3727         r = amdgpu_device_ip_reinit_early_sriov(adev);
3728         if (r)
3729                 goto error;
3730
3731         amdgpu_virt_init_data_exchange(adev);
3732         /* we need recover gart prior to run SMC/CP/SDMA resume */
3733         amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
3734
3735         r = amdgpu_device_fw_loading(adev);
3736         if (r)
3737                 return r;
3738
3739         /* now we are okay to resume SMC/CP/SDMA */
3740         r = amdgpu_device_ip_reinit_late_sriov(adev);
3741         if (r)
3742                 goto error;
3743
3744         amdgpu_irq_gpu_reset_resume_helper(adev);
3745         r = amdgpu_ib_ring_tests(adev);
3746         amdgpu_amdkfd_post_reset(adev);
3747
3748 error:
3749         amdgpu_virt_release_full_gpu(adev, true);
3750         if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
3751                 amdgpu_inc_vram_lost(adev);
3752                 r = amdgpu_device_recover_vram(adev);
3753         }
3754
3755         return r;
3756 }
3757
3758 /**
3759  * amdgpu_device_should_recover_gpu - check if we should try GPU recovery
3760  *
3761  * @adev: amdgpu device pointer
3762  *
3763  * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover
3764  * a hung GPU.
3765  */
3766 bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev)
3767 {
3768         if (!amdgpu_device_ip_check_soft_reset(adev)) {
3769                 DRM_INFO("Timeout, but no hardware hang detected.\n");
3770                 return false;
3771         }
3772
3773         if (amdgpu_gpu_recovery == 0)
3774                 goto disabled;
3775
3776         if (amdgpu_sriov_vf(adev))
3777                 return true;
3778
3779         if (amdgpu_gpu_recovery == -1) {
3780                 switch (adev->asic_type) {
3781                 case CHIP_BONAIRE:
3782                 case CHIP_HAWAII:
3783                 case CHIP_TOPAZ:
3784                 case CHIP_TONGA:
3785                 case CHIP_FIJI:
3786                 case CHIP_POLARIS10:
3787                 case CHIP_POLARIS11:
3788                 case CHIP_POLARIS12:
3789                 case CHIP_VEGAM:
3790                 case CHIP_VEGA20:
3791                 case CHIP_VEGA10:
3792                 case CHIP_VEGA12:
3793                 case CHIP_RAVEN:
3794                 case CHIP_ARCTURUS:
3795                 case CHIP_RENOIR:
3796                 case CHIP_NAVI10:
3797                 case CHIP_NAVI14:
3798                 case CHIP_NAVI12:
3799                         break;
3800                 default:
3801                         goto disabled;
3802                 }
3803         }
3804
3805         return true;
3806
3807 disabled:
3808                 DRM_INFO("GPU recovery disabled.\n");
3809                 return false;
3810 }
3811
3812
3813 static int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
3814                                         struct amdgpu_job *job,
3815                                         bool *need_full_reset_arg)
3816 {
3817         int i, r = 0;
3818         bool need_full_reset  = *need_full_reset_arg;
3819
3820         /* block all schedulers and reset given job's ring */
3821         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3822                 struct amdgpu_ring *ring = adev->rings[i];
3823
3824                 if (!ring || !ring->sched.thread)
3825                         continue;
3826
3827                 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
3828                 amdgpu_fence_driver_force_completion(ring);
3829         }
3830
3831         if(job)
3832                 drm_sched_increase_karma(&job->base);
3833
3834         /* Don't suspend on bare metal if we are not going to HW reset the ASIC */
3835         if (!amdgpu_sriov_vf(adev)) {
3836
3837                 if (!need_full_reset)
3838                         need_full_reset = amdgpu_device_ip_need_full_reset(adev);
3839
3840                 if (!need_full_reset) {
3841                         amdgpu_device_ip_pre_soft_reset(adev);
3842                         r = amdgpu_device_ip_soft_reset(adev);
3843                         amdgpu_device_ip_post_soft_reset(adev);
3844                         if (r || amdgpu_device_ip_check_soft_reset(adev)) {
3845                                 DRM_INFO("soft reset failed, will fallback to full reset!\n");
3846                                 need_full_reset = true;
3847                         }
3848                 }
3849
3850                 if (need_full_reset)
3851                         r = amdgpu_device_ip_suspend(adev);
3852
3853                 *need_full_reset_arg = need_full_reset;
3854         }
3855
3856         return r;
3857 }
3858
3859 static int amdgpu_do_asic_reset(struct amdgpu_hive_info *hive,
3860                                struct list_head *device_list_handle,
3861                                bool *need_full_reset_arg)
3862 {
3863         struct amdgpu_device *tmp_adev = NULL;
3864         bool need_full_reset = *need_full_reset_arg, vram_lost = false;
3865         int r = 0;
3866
3867         /*
3868          * ASIC reset has to be done on all HGMI hive nodes ASAP
3869          * to allow proper links negotiation in FW (within 1 sec)
3870          */
3871         if (need_full_reset) {
3872                 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3873                         /* For XGMI run all resets in parallel to speed up the process */
3874                         if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
3875                                 if (!queue_work(system_unbound_wq, &tmp_adev->xgmi_reset_work))
3876                                         r = -EALREADY;
3877                         } else
3878                                 r = amdgpu_asic_reset(tmp_adev);
3879
3880                         if (r) {
3881                                 DRM_ERROR("ASIC reset failed with error, %d for drm dev, %s",
3882                                          r, tmp_adev->ddev->unique);
3883                                 break;
3884                         }
3885                 }
3886
3887                 /* For XGMI wait for all resets to complete before proceed */
3888                 if (!r) {
3889                         list_for_each_entry(tmp_adev, device_list_handle,
3890                                             gmc.xgmi.head) {
3891                                 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
3892                                         flush_work(&tmp_adev->xgmi_reset_work);
3893                                         r = tmp_adev->asic_reset_res;
3894                                         if (r)
3895                                                 break;
3896                                 }
3897                         }
3898                 }
3899         }
3900
3901         if (!r && amdgpu_ras_intr_triggered())
3902                 amdgpu_ras_intr_cleared();
3903
3904         list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3905                 if (need_full_reset) {
3906                         /* post card */
3907                         if (amdgpu_atom_asic_init(tmp_adev->mode_info.atom_context))
3908                                 DRM_WARN("asic atom init failed!");
3909
3910                         if (!r) {
3911                                 dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n");
3912                                 r = amdgpu_device_ip_resume_phase1(tmp_adev);
3913                                 if (r)
3914                                         goto out;
3915
3916                                 vram_lost = amdgpu_device_check_vram_lost(tmp_adev);
3917                                 if (vram_lost) {
3918                                         DRM_INFO("VRAM is lost due to GPU reset!\n");
3919                                         amdgpu_inc_vram_lost(tmp_adev);
3920                                 }
3921
3922                                 r = amdgpu_gtt_mgr_recover(
3923                                         &tmp_adev->mman.bdev.man[TTM_PL_TT]);
3924                                 if (r)
3925                                         goto out;
3926
3927                                 r = amdgpu_device_fw_loading(tmp_adev);
3928                                 if (r)
3929                                         return r;
3930
3931                                 r = amdgpu_device_ip_resume_phase2(tmp_adev);
3932                                 if (r)
3933                                         goto out;
3934
3935                                 if (vram_lost)
3936                                         amdgpu_device_fill_reset_magic(tmp_adev);
3937
3938                                 /*
3939                                  * Add this ASIC as tracked as reset was already
3940                                  * complete successfully.
3941                                  */
3942                                 amdgpu_register_gpu_instance(tmp_adev);
3943
3944                                 r = amdgpu_device_ip_late_init(tmp_adev);
3945                                 if (r)
3946                                         goto out;
3947
3948                                 /* must succeed. */
3949                                 amdgpu_ras_resume(tmp_adev);
3950
3951                                 /* Update PSP FW topology after reset */
3952                                 if (hive && tmp_adev->gmc.xgmi.num_physical_nodes > 1)
3953                                         r = amdgpu_xgmi_update_topology(hive, tmp_adev);
3954                         }
3955                 }
3956
3957
3958 out:
3959                 if (!r) {
3960                         amdgpu_irq_gpu_reset_resume_helper(tmp_adev);
3961                         r = amdgpu_ib_ring_tests(tmp_adev);
3962                         if (r) {
3963                                 dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r);
3964                                 r = amdgpu_device_ip_suspend(tmp_adev);
3965                                 need_full_reset = true;
3966                                 r = -EAGAIN;
3967                                 goto end;
3968                         }
3969                 }
3970
3971                 if (!r)
3972                         r = amdgpu_device_recover_vram(tmp_adev);
3973                 else
3974                         tmp_adev->asic_reset_res = r;
3975         }
3976
3977 end:
3978         *need_full_reset_arg = need_full_reset;
3979         return r;
3980 }
3981
3982 static bool amdgpu_device_lock_adev(struct amdgpu_device *adev, bool trylock)
3983 {
3984         if (trylock) {
3985                 if (!mutex_trylock(&adev->lock_reset))
3986                         return false;
3987         } else
3988                 mutex_lock(&adev->lock_reset);
3989
3990         atomic_inc(&adev->gpu_reset_counter);
3991         adev->in_gpu_reset = true;
3992         switch (amdgpu_asic_reset_method(adev)) {
3993         case AMD_RESET_METHOD_MODE1:
3994                 adev->mp1_state = PP_MP1_STATE_SHUTDOWN;
3995                 break;
3996         case AMD_RESET_METHOD_MODE2:
3997                 adev->mp1_state = PP_MP1_STATE_RESET;
3998                 break;
3999         default:
4000                 adev->mp1_state = PP_MP1_STATE_NONE;
4001                 break;
4002         }
4003
4004         return true;
4005 }
4006
4007 static void amdgpu_device_unlock_adev(struct amdgpu_device *adev)
4008 {
4009         amdgpu_vf_error_trans_all(adev);
4010         adev->mp1_state = PP_MP1_STATE_NONE;
4011         adev->in_gpu_reset = false;
4012         mutex_unlock(&adev->lock_reset);
4013 }
4014
4015 /**
4016  * amdgpu_device_gpu_recover - reset the asic and recover scheduler
4017  *
4018  * @adev: amdgpu device pointer
4019  * @job: which job trigger hang
4020  *
4021  * Attempt to reset the GPU if it has hung (all asics).
4022  * Attempt to do soft-reset or full-reset and reinitialize Asic
4023  * Returns 0 for success or an error on failure.
4024  */
4025
4026 int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
4027                               struct amdgpu_job *job)
4028 {
4029         struct list_head device_list, *device_list_handle =  NULL;
4030         bool need_full_reset, job_signaled;
4031         struct amdgpu_hive_info *hive = NULL;
4032         struct amdgpu_device *tmp_adev = NULL;
4033         int i, r = 0;
4034         bool in_ras_intr = amdgpu_ras_intr_triggered();
4035         bool use_baco =
4036                 (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) ?
4037                 true : false;
4038
4039         /*
4040          * Flush RAM to disk so that after reboot
4041          * the user can read log and see why the system rebooted.
4042          */
4043         if (in_ras_intr && !use_baco && amdgpu_ras_get_context(adev)->reboot) {
4044
4045                 DRM_WARN("Emergency reboot.");
4046
4047                 ksys_sync_helper();
4048                 emergency_restart();
4049         }
4050
4051         need_full_reset = job_signaled = false;
4052         INIT_LIST_HEAD(&device_list);
4053
4054         dev_info(adev->dev, "GPU %s begin!\n",
4055                 (in_ras_intr && !use_baco) ? "jobs stop":"reset");
4056
4057         cancel_delayed_work_sync(&adev->delayed_init_work);
4058
4059         hive = amdgpu_get_xgmi_hive(adev, false);
4060
4061         /*
4062          * Here we trylock to avoid chain of resets executing from
4063          * either trigger by jobs on different adevs in XGMI hive or jobs on
4064          * different schedulers for same device while this TO handler is running.
4065          * We always reset all schedulers for device and all devices for XGMI
4066          * hive so that should take care of them too.
4067          */
4068
4069         if (hive && !mutex_trylock(&hive->reset_lock)) {
4070                 DRM_INFO("Bailing on TDR for s_job:%llx, hive: %llx as another already in progress",
4071                           job ? job->base.id : -1, hive->hive_id);
4072                 return 0;
4073         }
4074
4075         /* Start with adev pre asic reset first for soft reset check.*/
4076         if (!amdgpu_device_lock_adev(adev, !hive)) {
4077                 DRM_INFO("Bailing on TDR for s_job:%llx, as another already in progress",
4078                           job ? job->base.id : -1);
4079                 return 0;
4080         }
4081
4082         /* Block kfd: SRIOV would do it separately */
4083         if (!amdgpu_sriov_vf(adev))
4084                 amdgpu_amdkfd_pre_reset(adev);
4085
4086         /* Build list of devices to reset */
4087         if  (adev->gmc.xgmi.num_physical_nodes > 1) {
4088                 if (!hive) {
4089                         /*unlock kfd: SRIOV would do it separately */
4090                         if (!amdgpu_sriov_vf(adev))
4091                                 amdgpu_amdkfd_post_reset(adev);
4092                         amdgpu_device_unlock_adev(adev);
4093                         return -ENODEV;
4094                 }
4095
4096                 /*
4097                  * In case we are in XGMI hive mode device reset is done for all the
4098                  * nodes in the hive to retrain all XGMI links and hence the reset
4099                  * sequence is executed in loop on all nodes.
4100                  */
4101                 device_list_handle = &hive->device_list;
4102         } else {
4103                 list_add_tail(&adev->gmc.xgmi.head, &device_list);
4104                 device_list_handle = &device_list;
4105         }
4106
4107         /* block all schedulers and reset given job's ring */
4108         list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4109                 if (tmp_adev != adev) {
4110                         amdgpu_device_lock_adev(tmp_adev, false);
4111                         if (!amdgpu_sriov_vf(tmp_adev))
4112                                         amdgpu_amdkfd_pre_reset(tmp_adev);
4113                 }
4114
4115                 /*
4116                  * Mark these ASICs to be reseted as untracked first
4117                  * And add them back after reset completed
4118                  */
4119                 amdgpu_unregister_gpu_instance(tmp_adev);
4120
4121                 /* disable ras on ALL IPs */
4122                 if (!(in_ras_intr && !use_baco) &&
4123                       amdgpu_device_ip_need_full_reset(tmp_adev))
4124                         amdgpu_ras_suspend(tmp_adev);
4125
4126                 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4127                         struct amdgpu_ring *ring = tmp_adev->rings[i];
4128
4129                         if (!ring || !ring->sched.thread)
4130                                 continue;
4131
4132                         drm_sched_stop(&ring->sched, job ? &job->base : NULL);
4133
4134                         if (in_ras_intr && !use_baco)
4135                                 amdgpu_job_stop_all_jobs_on_sched(&ring->sched);
4136                 }
4137         }
4138
4139
4140         if (in_ras_intr && !use_baco)
4141                 goto skip_sched_resume;
4142
4143         /*
4144          * Must check guilty signal here since after this point all old
4145          * HW fences are force signaled.
4146          *
4147          * job->base holds a reference to parent fence
4148          */
4149         if (job && job->base.s_fence->parent &&
4150             dma_fence_is_signaled(job->base.s_fence->parent))
4151                 job_signaled = true;
4152
4153         if (job_signaled) {
4154                 dev_info(adev->dev, "Guilty job already signaled, skipping HW reset");
4155                 goto skip_hw_reset;
4156         }
4157
4158
4159         /* Guilty job will be freed after this*/
4160         r = amdgpu_device_pre_asic_reset(adev, job, &need_full_reset);
4161         if (r) {
4162                 /*TODO Should we stop ?*/
4163                 DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ",
4164                           r, adev->ddev->unique);
4165                 adev->asic_reset_res = r;
4166         }
4167
4168 retry:  /* Rest of adevs pre asic reset from XGMI hive. */
4169         list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4170
4171                 if (tmp_adev == adev)
4172                         continue;
4173
4174                 r = amdgpu_device_pre_asic_reset(tmp_adev,
4175                                                  NULL,
4176                                                  &need_full_reset);
4177                 /*TODO Should we stop ?*/
4178                 if (r) {
4179                         DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ",
4180                                   r, tmp_adev->ddev->unique);
4181                         tmp_adev->asic_reset_res = r;
4182                 }
4183         }
4184
4185         /* Actual ASIC resets if needed.*/
4186         /* TODO Implement XGMI hive reset logic for SRIOV */
4187         if (amdgpu_sriov_vf(adev)) {
4188                 r = amdgpu_device_reset_sriov(adev, job ? false : true);
4189                 if (r)
4190                         adev->asic_reset_res = r;
4191         } else {
4192                 r  = amdgpu_do_asic_reset(hive, device_list_handle, &need_full_reset);
4193                 if (r && r == -EAGAIN)
4194                         goto retry;
4195         }
4196
4197 skip_hw_reset:
4198
4199         /* Post ASIC reset for all devs .*/
4200         list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4201
4202                 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4203                         struct amdgpu_ring *ring = tmp_adev->rings[i];
4204
4205                         if (!ring || !ring->sched.thread)
4206                                 continue;
4207
4208                         /* No point to resubmit jobs if we didn't HW reset*/
4209                         if (!tmp_adev->asic_reset_res && !job_signaled)
4210                                 drm_sched_resubmit_jobs(&ring->sched);
4211
4212                         drm_sched_start(&ring->sched, !tmp_adev->asic_reset_res);
4213                 }
4214
4215                 if (!amdgpu_device_has_dc_support(tmp_adev) && !job_signaled) {
4216                         drm_helper_resume_force_mode(tmp_adev->ddev);
4217                 }
4218
4219                 tmp_adev->asic_reset_res = 0;
4220
4221                 if (r) {
4222                         /* bad news, how to tell it to userspace ? */
4223                         dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&tmp_adev->gpu_reset_counter));
4224                         amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
4225                 } else {
4226                         dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&tmp_adev->gpu_reset_counter));
4227                 }
4228         }
4229
4230 skip_sched_resume:
4231         list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4232                 /*unlock kfd: SRIOV would do it separately */
4233                 if (!(in_ras_intr && !use_baco) && !amdgpu_sriov_vf(tmp_adev))
4234                         amdgpu_amdkfd_post_reset(tmp_adev);
4235                 amdgpu_device_unlock_adev(tmp_adev);
4236         }
4237
4238         if (hive)
4239                 mutex_unlock(&hive->reset_lock);
4240
4241         if (r)
4242                 dev_info(adev->dev, "GPU reset end with ret = %d\n", r);
4243         return r;
4244 }
4245
4246 /**
4247  * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
4248  *
4249  * @adev: amdgpu_device pointer
4250  *
4251  * Fetchs and stores in the driver the PCIE capabilities (gen speed
4252  * and lanes) of the slot the device is in. Handles APUs and
4253  * virtualized environments where PCIE config space may not be available.
4254  */
4255 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
4256 {
4257         struct pci_dev *pdev;
4258         enum pci_bus_speed speed_cap, platform_speed_cap;
4259         enum pcie_link_width platform_link_width;
4260
4261         if (amdgpu_pcie_gen_cap)
4262                 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
4263
4264         if (amdgpu_pcie_lane_cap)
4265                 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
4266
4267         /* covers APUs as well */
4268         if (pci_is_root_bus(adev->pdev->bus)) {
4269                 if (adev->pm.pcie_gen_mask == 0)
4270                         adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
4271                 if (adev->pm.pcie_mlw_mask == 0)
4272                         adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
4273                 return;
4274         }
4275
4276         if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask)
4277                 return;
4278
4279         pcie_bandwidth_available(adev->pdev, NULL,
4280                                  &platform_speed_cap, &platform_link_width);
4281
4282         if (adev->pm.pcie_gen_mask == 0) {
4283                 /* asic caps */
4284                 pdev = adev->pdev;
4285                 speed_cap = pcie_get_speed_cap(pdev);
4286                 if (speed_cap == PCI_SPEED_UNKNOWN) {
4287                         adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4288                                                   CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4289                                                   CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
4290                 } else {
4291                         if (speed_cap == PCIE_SPEED_16_0GT)
4292                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4293                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4294                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
4295                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
4296                         else if (speed_cap == PCIE_SPEED_8_0GT)
4297                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4298                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4299                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
4300                         else if (speed_cap == PCIE_SPEED_5_0GT)
4301                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4302                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
4303                         else
4304                                 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
4305                 }
4306                 /* platform caps */
4307                 if (platform_speed_cap == PCI_SPEED_UNKNOWN) {
4308                         adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4309                                                    CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
4310                 } else {
4311                         if (platform_speed_cap == PCIE_SPEED_16_0GT)
4312                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4313                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4314                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
4315                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
4316                         else if (platform_speed_cap == PCIE_SPEED_8_0GT)
4317                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4318                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4319                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
4320                         else if (platform_speed_cap == PCIE_SPEED_5_0GT)
4321                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4322                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
4323                         else
4324                                 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
4325
4326                 }
4327         }
4328         if (adev->pm.pcie_mlw_mask == 0) {
4329                 if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) {
4330                         adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
4331                 } else {
4332                         switch (platform_link_width) {
4333                         case PCIE_LNK_X32:
4334                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
4335                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
4336                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4337                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4338                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4339                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4340                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4341                                 break;
4342                         case PCIE_LNK_X16:
4343                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
4344                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4345                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4346                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4347                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4348                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4349                                 break;
4350                         case PCIE_LNK_X12:
4351                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4352                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4353                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4354                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4355                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4356                                 break;
4357                         case PCIE_LNK_X8:
4358                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4359                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4360                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4361                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4362                                 break;
4363                         case PCIE_LNK_X4:
4364                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4365                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4366                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4367                                 break;
4368                         case PCIE_LNK_X2:
4369                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4370                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4371                                 break;
4372                         case PCIE_LNK_X1:
4373                                 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
4374                                 break;
4375                         default:
4376                                 break;
4377                         }
4378                 }
4379         }
4380 }
4381
4382 int amdgpu_device_baco_enter(struct drm_device *dev)
4383 {
4384         struct amdgpu_device *adev = dev->dev_private;
4385         struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
4386
4387         if (!amdgpu_device_supports_baco(adev->ddev))
4388                 return -ENOTSUPP;
4389
4390         if (ras && ras->supported)
4391                 adev->nbio.funcs->enable_doorbell_interrupt(adev, false);
4392
4393         return amdgpu_dpm_baco_enter(adev);
4394 }
4395
4396 int amdgpu_device_baco_exit(struct drm_device *dev)
4397 {
4398         struct amdgpu_device *adev = dev->dev_private;
4399         struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
4400         int ret = 0;
4401
4402         if (!amdgpu_device_supports_baco(adev->ddev))
4403                 return -ENOTSUPP;
4404
4405         ret = amdgpu_dpm_baco_exit(adev);
4406         if (ret)
4407                 return ret;
4408
4409         if (ras && ras->supported)
4410                 adev->nbio.funcs->enable_doorbell_interrupt(adev, true);
4411
4412         return 0;
4413 }