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 void static inline amdgpu_mm_wreg_mmio(struct amdgpu_device *adev, uint32_t reg, uint32_t v, uint32_t acc_flags)
 310 {
 311         trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
 312
 313         if ((reg * 4) < adev->rmmio_size && !(acc_flags & AMDGPU_REGS_IDX))
 314                 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
 315         else {
 316                 unsigned long flags;
 317
 318                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
 319                 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
 320                 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
 321                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
 322         }
 323
 324         if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
 325                 udelay(500);
 326         }
 327 }
 328
 329 /**
 330  * amdgpu_mm_wreg - write to a memory mapped IO register
 331  *
 332  * @adev: amdgpu_device pointer
 333  * @reg: dword aligned register offset
 334  * @v: 32 bit value to write to the register
 335  * @acc_flags: access flags which require special behavior
 336  *
 337  * Writes the value specified to the offset specified.
 338  */
 339 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
 340                     uint32_t acc_flags)
 341 {
 342         if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
 343                 adev->last_mm_index = v;
 344         }
 345
 346         if ((acc_flags & AMDGPU_REGS_KIQ) || (!(acc_flags & AMDGPU_REGS_NO_KIQ) && amdgpu_sriov_runtime(adev)))
 347                 return amdgpu_kiq_wreg(adev, reg, v);
 348
 349         amdgpu_mm_wreg_mmio(adev, reg, v, acc_flags);
 350 }
 351
 352 /*
 353  * amdgpu_mm_wreg_mmio_rlc -  write register either with mmio or with RLC path if in range
 354  *
 355  * this function is invoked only the debugfs register access
 356  * */
 357 void amdgpu_mm_wreg_mmio_rlc(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
 358                     uint32_t acc_flags)
 359 {
 360         if (amdgpu_sriov_fullaccess(adev) &&
 361                 adev->gfx.rlc.funcs &&
 362                 adev->gfx.rlc.funcs->is_rlcg_access_range) {
 363
 364                 if (adev->gfx.rlc.funcs->is_rlcg_access_range(adev, reg))
 365                         return adev->gfx.rlc.funcs->rlcg_wreg(adev, reg, v);
 366         }
 367
 368         amdgpu_mm_wreg_mmio(adev, reg, v, acc_flags);
 369 }
 370
 371 /**
 372  * amdgpu_io_rreg - read an IO register
 373  *
 374  * @adev: amdgpu_device pointer
 375  * @reg: dword aligned register offset
 376  *
 377  * Returns the 32 bit value from the offset specified.
 378  */
 379 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
 380 {
 381         if ((reg * 4) < adev->rio_mem_size)
 382                 return ioread32(adev->rio_mem + (reg * 4));
 383         else {
 384                 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
 385                 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
 386         }
 387 }
 388
 389 /**
 390  * amdgpu_io_wreg - write to an IO register
 391  *
 392  * @adev: amdgpu_device pointer
 393  * @reg: dword aligned register offset
 394  * @v: 32 bit value to write to the register
 395  *
 396  * Writes the value specified to the offset specified.
 397  */
 398 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
 399 {
 400         if (adev->asic_type >= CHIP_VEGA10 && reg == 0) {
 401                 adev->last_mm_index = v;
 402         }
 403
 404         if ((reg * 4) < adev->rio_mem_size)
 405                 iowrite32(v, adev->rio_mem + (reg * 4));
 406         else {
 407                 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
 408                 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
 409         }
 410
 411         if (adev->asic_type >= CHIP_VEGA10 && reg == 1 && adev->last_mm_index == 0x5702C) {
 412                 udelay(500);
 413         }
 414 }
 415
 416 /**
 417  * amdgpu_mm_rdoorbell - read a doorbell dword
 418  *
 419  * @adev: amdgpu_device pointer
 420  * @index: doorbell index
 421  *
 422  * Returns the value in the doorbell aperture at the
 423  * requested doorbell index (CIK).
 424  */
 425 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
 426 {
 427         if (index < adev->doorbell.num_doorbells) {
 428                 return readl(adev->doorbell.ptr + index);
 429         } else {
 430                 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
 431                 return 0;
 432         }
 433 }
 434
 435 /**
 436  * amdgpu_mm_wdoorbell - write a doorbell dword
 437  *
 438  * @adev: amdgpu_device pointer
 439  * @index: doorbell index
 440  * @v: value to write
 441  *
 442  * Writes @v to the doorbell aperture at the
 443  * requested doorbell index (CIK).
 444  */
 445 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
 446 {
 447         if (index < adev->doorbell.num_doorbells) {
 448                 writel(v, adev->doorbell.ptr + index);
 449         } else {
 450                 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
 451         }
 452 }
 453
 454 /**
 455  * amdgpu_mm_rdoorbell64 - read a doorbell Qword
 456  *
 457  * @adev: amdgpu_device pointer
 458  * @index: doorbell index
 459  *
 460  * Returns the value in the doorbell aperture at the
 461  * requested doorbell index (VEGA10+).
 462  */
 463 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
 464 {
 465         if (index < adev->doorbell.num_doorbells) {
 466                 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
 467         } else {
 468                 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
 469                 return 0;
 470         }
 471 }
 472
 473 /**
 474  * amdgpu_mm_wdoorbell64 - write a doorbell Qword
 475  *
 476  * @adev: amdgpu_device pointer
 477  * @index: doorbell index
 478  * @v: value to write
 479  *
 480  * Writes @v to the doorbell aperture at the
 481  * requested doorbell index (VEGA10+).
 482  */
 483 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
 484 {
 485         if (index < adev->doorbell.num_doorbells) {
 486                 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
 487         } else {
 488                 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
 489         }
 490 }
 491
 492 /**
 493  * amdgpu_invalid_rreg - dummy reg read function
 494  *
 495  * @adev: amdgpu device pointer
 496  * @reg: offset of register
 497  *
 498  * Dummy register read function.  Used for register blocks
 499  * that certain asics don't have (all asics).
 500  * Returns the value in the register.
 501  */
 502 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
 503 {
 504         DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
 505         BUG();
 506         return 0;
 507 }
 508
 509 /**
 510  * amdgpu_invalid_wreg - dummy reg write function
 511  *
 512  * @adev: amdgpu device pointer
 513  * @reg: offset of register
 514  * @v: value to write to the register
 515  *
 516  * Dummy register read function.  Used for register blocks
 517  * that certain asics don't have (all asics).
 518  */
 519 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
 520 {
 521         DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
 522                   reg, v);
 523         BUG();
 524 }
 525
 526 /**
 527  * amdgpu_invalid_rreg64 - dummy 64 bit reg read function
 528  *
 529  * @adev: amdgpu device pointer
 530  * @reg: offset of register
 531  *
 532  * Dummy register read function.  Used for register blocks
 533  * that certain asics don't have (all asics).
 534  * Returns the value in the register.
 535  */
 536 static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg)
 537 {
 538         DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg);
 539         BUG();
 540         return 0;
 541 }
 542
 543 /**
 544  * amdgpu_invalid_wreg64 - dummy reg write function
 545  *
 546  * @adev: amdgpu device pointer
 547  * @reg: offset of register
 548  * @v: value to write to the register
 549  *
 550  * Dummy register read function.  Used for register blocks
 551  * that certain asics don't have (all asics).
 552  */
 553 static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v)
 554 {
 555         DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n",
 556                   reg, v);
 557         BUG();
 558 }
 559
 560 /**
 561  * amdgpu_block_invalid_rreg - dummy reg read function
 562  *
 563  * @adev: amdgpu device pointer
 564  * @block: offset of instance
 565  * @reg: offset of register
 566  *
 567  * Dummy register read function.  Used for register blocks
 568  * that certain asics don't have (all asics).
 569  * Returns the value in the register.
 570  */
 571 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
 572                                           uint32_t block, uint32_t reg)
 573 {
 574         DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
 575                   reg, block);
 576         BUG();
 577         return 0;
 578 }
 579
 580 /**
 581  * amdgpu_block_invalid_wreg - dummy reg write function
 582  *
 583  * @adev: amdgpu device pointer
 584  * @block: offset of instance
 585  * @reg: offset of register
 586  * @v: value to write to the register
 587  *
 588  * Dummy register read function.  Used for register blocks
 589  * that certain asics don't have (all asics).
 590  */
 591 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
 592                                       uint32_t block,
 593                                       uint32_t reg, uint32_t v)
 594 {
 595         DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
 596                   reg, block, v);
 597         BUG();
 598 }
 599
 600 /**
 601  * amdgpu_device_vram_scratch_init - allocate the VRAM scratch page
 602  *
 603  * @adev: amdgpu device pointer
 604  *
 605  * Allocates a scratch page of VRAM for use by various things in the
 606  * driver.
 607  */
 608 static int amdgpu_device_vram_scratch_init(struct amdgpu_device *adev)
 609 {
 610         return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE,
 611                                        PAGE_SIZE, AMDGPU_GEM_DOMAIN_VRAM,
 612                                        &adev->vram_scratch.robj,
 613                                        &adev->vram_scratch.gpu_addr,
 614                                        (void **)&adev->vram_scratch.ptr);
 615 }
 616
 617 /**
 618  * amdgpu_device_vram_scratch_fini - Free the VRAM scratch page
 619  *
 620  * @adev: amdgpu device pointer
 621  *
 622  * Frees the VRAM scratch page.
 623  */
 624 static void amdgpu_device_vram_scratch_fini(struct amdgpu_device *adev)
 625 {
 626         amdgpu_bo_free_kernel(&adev->vram_scratch.robj, NULL, NULL);
 627 }
 628
 629 /**
 630  * amdgpu_device_program_register_sequence - program an array of registers.
 631  *
 632  * @adev: amdgpu_device pointer
 633  * @registers: pointer to the register array
 634  * @array_size: size of the register array
 635  *
 636  * Programs an array or registers with and and or masks.
 637  * This is a helper for setting golden registers.
 638  */
 639 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
 640                                              const u32 *registers,
 641                                              const u32 array_size)
 642 {
 643         u32 tmp, reg, and_mask, or_mask;
 644         int i;
 645
 646         if (array_size % 3)
 647                 return;
 648
 649         for (i = 0; i < array_size; i +=3) {
 650                 reg = registers[i + 0];
 651                 and_mask = registers[i + 1];
 652                 or_mask = registers[i + 2];
 653
 654                 if (and_mask == 0xffffffff) {
 655                         tmp = or_mask;
 656                 } else {
 657                         tmp = RREG32(reg);
 658                         tmp &= ~and_mask;
 659                         if (adev->family >= AMDGPU_FAMILY_AI)
 660                                 tmp |= (or_mask & and_mask);
 661                         else
 662                                 tmp |= or_mask;
 663                 }
 664                 WREG32(reg, tmp);
 665         }
 666 }
 667
 668 /**
 669  * amdgpu_device_pci_config_reset - reset the GPU
 670  *
 671  * @adev: amdgpu_device pointer
 672  *
 673  * Resets the GPU using the pci config reset sequence.
 674  * Only applicable to asics prior to vega10.
 675  */
 676 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
 677 {
 678         pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
 679 }
 680
 681 /*
 682  * GPU doorbell aperture helpers function.
 683  */
 684 /**
 685  * amdgpu_device_doorbell_init - Init doorbell driver information.
 686  *
 687  * @adev: amdgpu_device pointer
 688  *
 689  * Init doorbell driver information (CIK)
 690  * Returns 0 on success, error on failure.
 691  */
 692 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
 693 {
 694
 695         /* No doorbell on SI hardware generation */
 696         if (adev->asic_type < CHIP_BONAIRE) {
 697                 adev->doorbell.base = 0;
 698                 adev->doorbell.size = 0;
 699                 adev->doorbell.num_doorbells = 0;
 700                 adev->doorbell.ptr = NULL;
 701                 return 0;
 702         }
 703
 704         if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
 705                 return -EINVAL;
 706
 707         amdgpu_asic_init_doorbell_index(adev);
 708
 709         /* doorbell bar mapping */
 710         adev->doorbell.base = pci_resource_start(adev->pdev, 2);
 711         adev->doorbell.size = pci_resource_len(adev->pdev, 2);
 712
 713         adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
 714                                              adev->doorbell_index.max_assignment+1);
 715         if (adev->doorbell.num_doorbells == 0)
 716                 return -EINVAL;
 717
 718         /* For Vega, reserve and map two pages on doorbell BAR since SDMA
 719          * paging queue doorbell use the second page. The
 720          * AMDGPU_DOORBELL64_MAX_ASSIGNMENT definition assumes all the
 721          * doorbells are in the first page. So with paging queue enabled,
 722          * the max num_doorbells should + 1 page (0x400 in dword)
 723          */
 724         if (adev->asic_type >= CHIP_VEGA10)
 725                 adev->doorbell.num_doorbells += 0x400;
 726
 727         adev->doorbell.ptr = ioremap(adev->doorbell.base,
 728                                      adev->doorbell.num_doorbells *
 729                                      sizeof(u32));
 730         if (adev->doorbell.ptr == NULL)
 731                 return -ENOMEM;
 732
 733         return 0;
 734 }
 735
 736 /**
 737  * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
 738  *
 739  * @adev: amdgpu_device pointer
 740  *
 741  * Tear down doorbell driver information (CIK)
 742  */
 743 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
 744 {
 745         iounmap(adev->doorbell.ptr);
 746         adev->doorbell.ptr = NULL;
 747 }
 748
 749
 750
 751 /*
 752  * amdgpu_device_wb_*()
 753  * Writeback is the method by which the GPU updates special pages in memory
 754  * with the status of certain GPU events (fences, ring pointers,etc.).
 755  */
 756
 757 /**
 758  * amdgpu_device_wb_fini - Disable Writeback and free memory
 759  *
 760  * @adev: amdgpu_device pointer
 761  *
 762  * Disables Writeback and frees the Writeback memory (all asics).
 763  * Used at driver shutdown.
 764  */
 765 static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
 766 {
 767         if (adev->wb.wb_obj) {
 768                 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
 769                                       &adev->wb.gpu_addr,
 770                                       (void **)&adev->wb.wb);
 771                 adev->wb.wb_obj = NULL;
 772         }
 773 }
 774
 775 /**
 776  * amdgpu_device_wb_init- Init Writeback driver info and allocate memory
 777  *
 778  * @adev: amdgpu_device pointer
 779  *
 780  * Initializes writeback and allocates writeback memory (all asics).
 781  * Used at driver startup.
 782  * Returns 0 on success or an -error on failure.
 783  */
 784 static int amdgpu_device_wb_init(struct amdgpu_device *adev)
 785 {
 786         int r;
 787
 788         if (adev->wb.wb_obj == NULL) {
 789                 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
 790                 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
 791                                             PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
 792                                             &adev->wb.wb_obj, &adev->wb.gpu_addr,
 793                                             (void **)&adev->wb.wb);
 794                 if (r) {
 795                         dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
 796                         return r;
 797                 }
 798
 799                 adev->wb.num_wb = AMDGPU_MAX_WB;
 800                 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
 801
 802                 /* clear wb memory */
 803                 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
 804         }
 805
 806         return 0;
 807 }
 808
 809 /**
 810  * amdgpu_device_wb_get - Allocate a wb entry
 811  *
 812  * @adev: amdgpu_device pointer
 813  * @wb: wb index
 814  *
 815  * Allocate a wb slot for use by the driver (all asics).
 816  * Returns 0 on success or -EINVAL on failure.
 817  */
 818 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
 819 {
 820         unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
 821
 822         if (offset < adev->wb.num_wb) {
 823                 __set_bit(offset, adev->wb.used);
 824                 *wb = offset << 3; /* convert to dw offset */
 825                 return 0;
 826         } else {
 827                 return -EINVAL;
 828         }
 829 }
 830
 831 /**
 832  * amdgpu_device_wb_free - Free a wb entry
 833  *
 834  * @adev: amdgpu_device pointer
 835  * @wb: wb index
 836  *
 837  * Free a wb slot allocated for use by the driver (all asics)
 838  */
 839 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
 840 {
 841         wb >>= 3;
 842         if (wb < adev->wb.num_wb)
 843                 __clear_bit(wb, adev->wb.used);
 844 }
 845
 846 /**
 847  * amdgpu_device_resize_fb_bar - try to resize FB BAR
 848  *
 849  * @adev: amdgpu_device pointer
 850  *
 851  * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
 852  * to fail, but if any of the BARs is not accessible after the size we abort
 853  * driver loading by returning -ENODEV.
 854  */
 855 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
 856 {
 857         u64 space_needed = roundup_pow_of_two(adev->gmc.real_vram_size);
 858         u32 rbar_size = order_base_2(((space_needed >> 20) | 1)) - 1;
 859         struct pci_bus *root;
 860         struct resource *res;
 861         unsigned i;
 862         u16 cmd;
 863         int r;
 864
 865         /* Bypass for VF */
 866         if (amdgpu_sriov_vf(adev))
 867                 return 0;
 868
 869         /* Check if the root BUS has 64bit memory resources */
 870         root = adev->pdev->bus;
 871         while (root->parent)
 872                 root = root->parent;
 873
 874         pci_bus_for_each_resource(root, res, i) {
 875                 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
 876                     res->start > 0x100000000ull)
 877                         break;
 878         }
 879
 880         /* Trying to resize is pointless without a root hub window above 4GB */
 881         if (!res)
 882                 return 0;
 883
 884         /* Disable memory decoding while we change the BAR addresses and size */
 885         pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
 886         pci_write_config_word(adev->pdev, PCI_COMMAND,
 887                               cmd & ~PCI_COMMAND_MEMORY);
 888
 889         /* Free the VRAM and doorbell BAR, we most likely need to move both. */
 890         amdgpu_device_doorbell_fini(adev);
 891         if (adev->asic_type >= CHIP_BONAIRE)
 892                 pci_release_resource(adev->pdev, 2);
 893
 894         pci_release_resource(adev->pdev, 0);
 895
 896         r = pci_resize_resource(adev->pdev, 0, rbar_size);
 897         if (r == -ENOSPC)
 898                 DRM_INFO("Not enough PCI address space for a large BAR.");
 899         else if (r && r != -ENOTSUPP)
 900                 DRM_ERROR("Problem resizing BAR0 (%d).", r);
 901
 902         pci_assign_unassigned_bus_resources(adev->pdev->bus);
 903
 904         /* When the doorbell or fb BAR isn't available we have no chance of
 905          * using the device.
 906          */
 907         r = amdgpu_device_doorbell_init(adev);
 908         if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
 909                 return -ENODEV;
 910
 911         pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
 912
 913         return 0;
 914 }
 915
 916 /*
 917  * GPU helpers function.
 918  */
 919 /**
 920  * amdgpu_device_need_post - check if the hw need post or not
 921  *
 922  * @adev: amdgpu_device pointer
 923  *
 924  * Check if the asic has been initialized (all asics) at driver startup
 925  * or post is needed if  hw reset is performed.
 926  * Returns true if need or false if not.
 927  */
 928 bool amdgpu_device_need_post(struct amdgpu_device *adev)
 929 {
 930         uint32_t reg;
 931
 932         if (amdgpu_sriov_vf(adev))
 933                 return false;
 934
 935         if (amdgpu_passthrough(adev)) {
 936                 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
 937                  * some old smc fw still need driver do vPost otherwise gpu hang, while
 938                  * those smc fw version above 22.15 doesn't have this flaw, so we force
 939                  * vpost executed for smc version below 22.15
 940                  */
 941                 if (adev->asic_type == CHIP_FIJI) {
 942                         int err;
 943                         uint32_t fw_ver;
 944                         err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
 945                         /* force vPost if error occured */
 946                         if (err)
 947                                 return true;
 948
 949                         fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
 950                         if (fw_ver < 0x00160e00)
 951                                 return true;
 952                 }
 953         }
 954
 955         if (adev->has_hw_reset) {
 956                 adev->has_hw_reset = false;
 957                 return true;
 958         }
 959
 960         /* bios scratch used on CIK+ */
 961         if (adev->asic_type >= CHIP_BONAIRE)
 962                 return amdgpu_atombios_scratch_need_asic_init(adev);
 963
 964         /* check MEM_SIZE for older asics */
 965         reg = amdgpu_asic_get_config_memsize(adev);
 966
 967         if ((reg != 0) && (reg != 0xffffffff))
 968                 return false;
 969
 970         return true;
 971 }
 972
 973 /* if we get transitioned to only one device, take VGA back */
 974 /**
 975  * amdgpu_device_vga_set_decode - enable/disable vga decode
 976  *
 977  * @cookie: amdgpu_device pointer
 978  * @state: enable/disable vga decode
 979  *
 980  * Enable/disable vga decode (all asics).
 981  * Returns VGA resource flags.
 982  */
 983 static unsigned int amdgpu_device_vga_set_decode(void *cookie, bool state)
 984 {
 985         struct amdgpu_device *adev = cookie;
 986         amdgpu_asic_set_vga_state(adev, state);
 987         if (state)
 988                 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
 989                        VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
 990         else
 991                 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
 992 }
 993
 994 /**
 995  * amdgpu_device_check_block_size - validate the vm block size
 996  *
 997  * @adev: amdgpu_device pointer
 998  *
 999  * Validates the vm block size specified via module parameter.
1000  * The vm block size defines number of bits in page table versus page directory,
1001  * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
1002  * page table and the remaining bits are in the page directory.
1003  */
1004 static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
1005 {
1006         /* defines number of bits in page table versus page directory,
1007          * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
1008          * page table and the remaining bits are in the page directory */
1009         if (amdgpu_vm_block_size == -1)
1010                 return;
1011
1012         if (amdgpu_vm_block_size < 9) {
1013                 dev_warn(adev->dev, "VM page table size (%d) too small\n",
1014                          amdgpu_vm_block_size);
1015                 amdgpu_vm_block_size = -1;
1016         }
1017 }
1018
1019 /**
1020  * amdgpu_device_check_vm_size - validate the vm size
1021  *
1022  * @adev: amdgpu_device pointer
1023  *
1024  * Validates the vm size in GB specified via module parameter.
1025  * The VM size is the size of the GPU virtual memory space in GB.
1026  */
1027 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
1028 {
1029         /* no need to check the default value */
1030         if (amdgpu_vm_size == -1)
1031                 return;
1032
1033         if (amdgpu_vm_size < 1) {
1034                 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
1035                          amdgpu_vm_size);
1036                 amdgpu_vm_size = -1;
1037         }
1038 }
1039
1040 static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
1041 {
1042         struct sysinfo si;
1043         bool is_os_64 = (sizeof(void *) == 8);
1044         uint64_t total_memory;
1045         uint64_t dram_size_seven_GB = 0x1B8000000;
1046         uint64_t dram_size_three_GB = 0xB8000000;
1047
1048         if (amdgpu_smu_memory_pool_size == 0)
1049                 return;
1050
1051         if (!is_os_64) {
1052                 DRM_WARN("Not 64-bit OS, feature not supported\n");
1053                 goto def_value;
1054         }
1055         si_meminfo(&si);
1056         total_memory = (uint64_t)si.totalram * si.mem_unit;
1057
1058         if ((amdgpu_smu_memory_pool_size == 1) ||
1059                 (amdgpu_smu_memory_pool_size == 2)) {
1060                 if (total_memory < dram_size_three_GB)
1061                         goto def_value1;
1062         } else if ((amdgpu_smu_memory_pool_size == 4) ||
1063                 (amdgpu_smu_memory_pool_size == 8)) {
1064                 if (total_memory < dram_size_seven_GB)
1065                         goto def_value1;
1066         } else {
1067                 DRM_WARN("Smu memory pool size not supported\n");
1068                 goto def_value;
1069         }
1070         adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
1071
1072         return;
1073
1074 def_value1:
1075         DRM_WARN("No enough system memory\n");
1076 def_value:
1077         adev->pm.smu_prv_buffer_size = 0;
1078 }
1079
1080 /**
1081  * amdgpu_device_check_arguments - validate module params
1082  *
1083  * @adev: amdgpu_device pointer
1084  *
1085  * Validates certain module parameters and updates
1086  * the associated values used by the driver (all asics).
1087  */
1088 static int amdgpu_device_check_arguments(struct amdgpu_device *adev)
1089 {
1090         if (amdgpu_sched_jobs < 4) {
1091                 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
1092                          amdgpu_sched_jobs);
1093                 amdgpu_sched_jobs = 4;
1094         } else if (!is_power_of_2(amdgpu_sched_jobs)){
1095                 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
1096                          amdgpu_sched_jobs);
1097                 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
1098         }
1099
1100         if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
1101                 /* gart size must be greater or equal to 32M */
1102                 dev_warn(adev->dev, "gart size (%d) too small\n",
1103                          amdgpu_gart_size);
1104                 amdgpu_gart_size = -1;
1105         }
1106
1107         if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
1108                 /* gtt size must be greater or equal to 32M */
1109                 dev_warn(adev->dev, "gtt size (%d) too small\n",
1110                                  amdgpu_gtt_size);
1111                 amdgpu_gtt_size = -1;
1112         }
1113
1114         /* valid range is between 4 and 9 inclusive */
1115         if (amdgpu_vm_fragment_size != -1 &&
1116             (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
1117                 dev_warn(adev->dev, "valid range is between 4 and 9\n");
1118                 amdgpu_vm_fragment_size = -1;
1119         }
1120
1121         amdgpu_device_check_smu_prv_buffer_size(adev);
1122
1123         amdgpu_device_check_vm_size(adev);
1124
1125         amdgpu_device_check_block_size(adev);
1126
1127         adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
1128
1129         return 0;
1130 }
1131
1132 /**
1133  * amdgpu_switcheroo_set_state - set switcheroo state
1134  *
1135  * @pdev: pci dev pointer
1136  * @state: vga_switcheroo state
1137  *
1138  * Callback for the switcheroo driver.  Suspends or resumes the
1139  * the asics before or after it is powered up using ACPI methods.
1140  */
1141 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
1142 {
1143         struct drm_device *dev = pci_get_drvdata(pdev);
1144         int r;
1145
1146         if (amdgpu_device_supports_boco(dev) && state == VGA_SWITCHEROO_OFF)
1147                 return;
1148
1149         if (state == VGA_SWITCHEROO_ON) {
1150                 pr_info("amdgpu: switched on\n");
1151                 /* don't suspend or resume card normally */
1152                 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1153
1154                 pci_set_power_state(dev->pdev, PCI_D0);
1155                 pci_restore_state(dev->pdev);
1156                 r = pci_enable_device(dev->pdev);
1157                 if (r)
1158                         DRM_WARN("pci_enable_device failed (%d)\n", r);
1159                 amdgpu_device_resume(dev, true);
1160
1161                 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1162                 drm_kms_helper_poll_enable(dev);
1163         } else {
1164                 pr_info("amdgpu: switched off\n");
1165                 drm_kms_helper_poll_disable(dev);
1166                 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1167                 amdgpu_device_suspend(dev, true);
1168                 pci_save_state(dev->pdev);
1169                 /* Shut down the device */
1170                 pci_disable_device(dev->pdev);
1171                 pci_set_power_state(dev->pdev, PCI_D3cold);
1172                 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1173         }
1174 }
1175
1176 /**
1177  * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1178  *
1179  * @pdev: pci dev pointer
1180  *
1181  * Callback for the switcheroo driver.  Check of the switcheroo
1182  * state can be changed.
1183  * Returns true if the state can be changed, false if not.
1184  */
1185 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1186 {
1187         struct drm_device *dev = pci_get_drvdata(pdev);
1188
1189         /*
1190         * FIXME: open_count is protected by drm_global_mutex but that would lead to
1191         * locking inversion with the driver load path. And the access here is
1192         * completely racy anyway. So don't bother with locking for now.
1193         */
1194         return atomic_read(&dev->open_count) == 0;
1195 }
1196
1197 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1198         .set_gpu_state = amdgpu_switcheroo_set_state,
1199         .reprobe = NULL,
1200         .can_switch = amdgpu_switcheroo_can_switch,
1201 };
1202
1203 /**
1204  * amdgpu_device_ip_set_clockgating_state - set the CG state
1205  *
1206  * @dev: amdgpu_device pointer
1207  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1208  * @state: clockgating state (gate or ungate)
1209  *
1210  * Sets the requested clockgating state for all instances of
1211  * the hardware IP specified.
1212  * Returns the error code from the last instance.
1213  */
1214 int amdgpu_device_ip_set_clockgating_state(void *dev,
1215                                            enum amd_ip_block_type block_type,
1216                                            enum amd_clockgating_state state)
1217 {
1218         struct amdgpu_device *adev = dev;
1219         int i, r = 0;
1220
1221         for (i = 0; i < adev->num_ip_blocks; i++) {
1222                 if (!adev->ip_blocks[i].status.valid)
1223                         continue;
1224                 if (adev->ip_blocks[i].version->type != block_type)
1225                         continue;
1226                 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1227                         continue;
1228                 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1229                         (void *)adev, state);
1230                 if (r)
1231                         DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1232                                   adev->ip_blocks[i].version->funcs->name, r);
1233         }
1234         return r;
1235 }
1236
1237 /**
1238  * amdgpu_device_ip_set_powergating_state - set the PG state
1239  *
1240  * @dev: amdgpu_device pointer
1241  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1242  * @state: powergating state (gate or ungate)
1243  *
1244  * Sets the requested powergating state for all instances of
1245  * the hardware IP specified.
1246  * Returns the error code from the last instance.
1247  */
1248 int amdgpu_device_ip_set_powergating_state(void *dev,
1249                                            enum amd_ip_block_type block_type,
1250                                            enum amd_powergating_state state)
1251 {
1252         struct amdgpu_device *adev = dev;
1253         int i, r = 0;
1254
1255         for (i = 0; i < adev->num_ip_blocks; i++) {
1256                 if (!adev->ip_blocks[i].status.valid)
1257                         continue;
1258                 if (adev->ip_blocks[i].version->type != block_type)
1259                         continue;
1260                 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1261                         continue;
1262                 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1263                         (void *)adev, state);
1264                 if (r)
1265                         DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1266                                   adev->ip_blocks[i].version->funcs->name, r);
1267         }
1268         return r;
1269 }
1270
1271 /**
1272  * amdgpu_device_ip_get_clockgating_state - get the CG state
1273  *
1274  * @adev: amdgpu_device pointer
1275  * @flags: clockgating feature flags
1276  *
1277  * Walks the list of IPs on the device and updates the clockgating
1278  * flags for each IP.
1279  * Updates @flags with the feature flags for each hardware IP where
1280  * clockgating is enabled.
1281  */
1282 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1283                                             u32 *flags)
1284 {
1285         int i;
1286
1287         for (i = 0; i < adev->num_ip_blocks; i++) {
1288                 if (!adev->ip_blocks[i].status.valid)
1289                         continue;
1290                 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1291                         adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1292         }
1293 }
1294
1295 /**
1296  * amdgpu_device_ip_wait_for_idle - wait for idle
1297  *
1298  * @adev: amdgpu_device pointer
1299  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1300  *
1301  * Waits for the request hardware IP to be idle.
1302  * Returns 0 for success or a negative error code on failure.
1303  */
1304 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1305                                    enum amd_ip_block_type block_type)
1306 {
1307         int i, r;
1308
1309         for (i = 0; i < adev->num_ip_blocks; i++) {
1310                 if (!adev->ip_blocks[i].status.valid)
1311                         continue;
1312                 if (adev->ip_blocks[i].version->type == block_type) {
1313                         r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1314                         if (r)
1315                                 return r;
1316                         break;
1317                 }
1318         }
1319         return 0;
1320
1321 }
1322
1323 /**
1324  * amdgpu_device_ip_is_idle - is the hardware IP idle
1325  *
1326  * @adev: amdgpu_device pointer
1327  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1328  *
1329  * Check if the hardware IP is idle or not.
1330  * Returns true if it the IP is idle, false if not.
1331  */
1332 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1333                               enum amd_ip_block_type block_type)
1334 {
1335         int i;
1336
1337         for (i = 0; i < adev->num_ip_blocks; i++) {
1338                 if (!adev->ip_blocks[i].status.valid)
1339                         continue;
1340                 if (adev->ip_blocks[i].version->type == block_type)
1341                         return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1342         }
1343         return true;
1344
1345 }
1346
1347 /**
1348  * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1349  *
1350  * @adev: amdgpu_device pointer
1351  * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
1352  *
1353  * Returns a pointer to the hardware IP block structure
1354  * if it exists for the asic, otherwise NULL.
1355  */
1356 struct amdgpu_ip_block *
1357 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1358                               enum amd_ip_block_type type)
1359 {
1360         int i;
1361
1362         for (i = 0; i < adev->num_ip_blocks; i++)
1363                 if (adev->ip_blocks[i].version->type == type)
1364                         return &adev->ip_blocks[i];
1365
1366         return NULL;
1367 }
1368
1369 /**
1370  * amdgpu_device_ip_block_version_cmp
1371  *
1372  * @adev: amdgpu_device pointer
1373  * @type: enum amd_ip_block_type
1374  * @major: major version
1375  * @minor: minor version
1376  *
1377  * return 0 if equal or greater
1378  * return 1 if smaller or the ip_block doesn't exist
1379  */
1380 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1381                                        enum amd_ip_block_type type,
1382                                        u32 major, u32 minor)
1383 {
1384         struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1385
1386         if (ip_block && ((ip_block->version->major > major) ||
1387                         ((ip_block->version->major == major) &&
1388                         (ip_block->version->minor >= minor))))
1389                 return 0;
1390
1391         return 1;
1392 }
1393
1394 /**
1395  * amdgpu_device_ip_block_add
1396  *
1397  * @adev: amdgpu_device pointer
1398  * @ip_block_version: pointer to the IP to add
1399  *
1400  * Adds the IP block driver information to the collection of IPs
1401  * on the asic.
1402  */
1403 int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1404                                const struct amdgpu_ip_block_version *ip_block_version)
1405 {
1406         if (!ip_block_version)
1407                 return -EINVAL;
1408
1409         DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1410                   ip_block_version->funcs->name);
1411
1412         adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1413
1414         return 0;
1415 }
1416
1417 /**
1418  * amdgpu_device_enable_virtual_display - enable virtual display feature
1419  *
1420  * @adev: amdgpu_device pointer
1421  *
1422  * Enabled the virtual display feature if the user has enabled it via
1423  * the module parameter virtual_display.  This feature provides a virtual
1424  * display hardware on headless boards or in virtualized environments.
1425  * This function parses and validates the configuration string specified by
1426  * the user and configues the virtual display configuration (number of
1427  * virtual connectors, crtcs, etc.) specified.
1428  */
1429 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1430 {
1431         adev->enable_virtual_display = false;
1432
1433         if (amdgpu_virtual_display) {
1434                 struct drm_device *ddev = adev->ddev;
1435                 const char *pci_address_name = pci_name(ddev->pdev);
1436                 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1437
1438                 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1439                 pciaddstr_tmp = pciaddstr;
1440                 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1441                         pciaddname = strsep(&pciaddname_tmp, ",");
1442                         if (!strcmp("all", pciaddname)
1443                             || !strcmp(pci_address_name, pciaddname)) {
1444                                 long num_crtc;
1445                                 int res = -1;
1446
1447                                 adev->enable_virtual_display = true;
1448
1449                                 if (pciaddname_tmp)
1450                                         res = kstrtol(pciaddname_tmp, 10,
1451                                                       &num_crtc);
1452
1453                                 if (!res) {
1454                                         if (num_crtc < 1)
1455                                                 num_crtc = 1;
1456                                         if (num_crtc > 6)
1457                                                 num_crtc = 6;
1458                                         adev->mode_info.num_crtc = num_crtc;
1459                                 } else {
1460                                         adev->mode_info.num_crtc = 1;
1461                                 }
1462                                 break;
1463                         }
1464                 }
1465
1466                 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1467                          amdgpu_virtual_display, pci_address_name,
1468                          adev->enable_virtual_display, adev->mode_info.num_crtc);
1469
1470                 kfree(pciaddstr);
1471         }
1472 }
1473
1474 /**
1475  * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1476  *
1477  * @adev: amdgpu_device pointer
1478  *
1479  * Parses the asic configuration parameters specified in the gpu info
1480  * firmware and makes them availale to the driver for use in configuring
1481  * the asic.
1482  * Returns 0 on success, -EINVAL on failure.
1483  */
1484 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1485 {
1486         const char *chip_name;
1487         char fw_name[30];
1488         int err;
1489         const struct gpu_info_firmware_header_v1_0 *hdr;
1490
1491         adev->firmware.gpu_info_fw = NULL;
1492
1493         switch (adev->asic_type) {
1494         case CHIP_TOPAZ:
1495         case CHIP_TONGA:
1496         case CHIP_FIJI:
1497         case CHIP_POLARIS10:
1498         case CHIP_POLARIS11:
1499         case CHIP_POLARIS12:
1500         case CHIP_VEGAM:
1501         case CHIP_CARRIZO:
1502         case CHIP_STONEY:
1503 #ifdef CONFIG_DRM_AMDGPU_SI
1504         case CHIP_VERDE:
1505         case CHIP_TAHITI:
1506         case CHIP_PITCAIRN:
1507         case CHIP_OLAND:
1508         case CHIP_HAINAN:
1509 #endif
1510 #ifdef CONFIG_DRM_AMDGPU_CIK
1511         case CHIP_BONAIRE:
1512         case CHIP_HAWAII:
1513         case CHIP_KAVERI:
1514         case CHIP_KABINI:
1515         case CHIP_MULLINS:
1516 #endif
1517         case CHIP_VEGA20:
1518         default:
1519                 return 0;
1520         case CHIP_VEGA10:
1521                 chip_name = "vega10";
1522                 break;
1523         case CHIP_VEGA12:
1524                 chip_name = "vega12";
1525                 break;
1526         case CHIP_RAVEN:
1527                 if (adev->rev_id >= 8)
1528                         chip_name = "raven2";
1529                 else if (adev->pdev->device == 0x15d8)
1530                         chip_name = "picasso";
1531                 else
1532                         chip_name = "raven";
1533                 break;
1534         case CHIP_ARCTURUS:
1535                 chip_name = "arcturus";
1536                 break;
1537         case CHIP_RENOIR:
1538                 chip_name = "renoir";
1539                 break;
1540         case CHIP_NAVI10:
1541                 chip_name = "navi10";
1542                 break;
1543         case CHIP_NAVI14:
1544                 chip_name = "navi14";
1545                 break;
1546         case CHIP_NAVI12:
1547                 chip_name = "navi12";
1548                 break;
1549         }
1550
1551         snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
1552         err = request_firmware(&adev->firmware.gpu_info_fw, fw_name, adev->dev);
1553         if (err) {
1554                 dev_err(adev->dev,
1555                         "Failed to load gpu_info firmware \"%s\"\n",
1556                         fw_name);
1557                 goto out;
1558         }
1559         err = amdgpu_ucode_validate(adev->firmware.gpu_info_fw);
1560         if (err) {
1561                 dev_err(adev->dev,
1562                         "Failed to validate gpu_info firmware \"%s\"\n",
1563                         fw_name);
1564                 goto out;
1565         }
1566
1567         hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
1568         amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
1569
1570         switch (hdr->version_major) {
1571         case 1:
1572         {
1573                 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
1574                         (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
1575                                                                 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1576
1577                 if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10)
1578                         goto parse_soc_bounding_box;
1579
1580                 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
1581                 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
1582                 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
1583                 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
1584                 adev->gfx.config.max_texture_channel_caches =
1585                         le32_to_cpu(gpu_info_fw->gc_num_tccs);
1586                 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
1587                 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
1588                 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
1589                 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
1590                 adev->gfx.config.double_offchip_lds_buf =
1591                         le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
1592                 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
1593                 adev->gfx.cu_info.max_waves_per_simd =
1594                         le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
1595                 adev->gfx.cu_info.max_scratch_slots_per_cu =
1596                         le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
1597                 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
1598                 if (hdr->version_minor >= 1) {
1599                         const struct gpu_info_firmware_v1_1 *gpu_info_fw =
1600                                 (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data +
1601                                                                         le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1602                         adev->gfx.config.num_sc_per_sh =
1603                                 le32_to_cpu(gpu_info_fw->num_sc_per_sh);
1604                         adev->gfx.config.num_packer_per_sc =
1605                                 le32_to_cpu(gpu_info_fw->num_packer_per_sc);
1606                 }
1607
1608 parse_soc_bounding_box:
1609                 /*
1610                  * soc bounding box info is not integrated in disocovery table,
1611                  * we always need to parse it from gpu info firmware.
1612                  */
1613                 if (hdr->version_minor == 2) {
1614                         const struct gpu_info_firmware_v1_2 *gpu_info_fw =
1615                                 (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data +
1616                                                                         le32_to_cpu(hdr->header.ucode_array_offset_bytes));
1617                         adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box;
1618                 }
1619                 break;
1620         }
1621         default:
1622                 dev_err(adev->dev,
1623                         "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
1624                 err = -EINVAL;
1625                 goto out;
1626         }
1627 out:
1628         return err;
1629 }
1630
1631 /**
1632  * amdgpu_device_ip_early_init - run early init for hardware IPs
1633  *
1634  * @adev: amdgpu_device pointer
1635  *
1636  * Early initialization pass for hardware IPs.  The hardware IPs that make
1637  * up each asic are discovered each IP's early_init callback is run.  This
1638  * is the first stage in initializing the asic.
1639  * Returns 0 on success, negative error code on failure.
1640  */
1641 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
1642 {
1643         int i, r;
1644
1645         amdgpu_device_enable_virtual_display(adev);
1646
1647         switch (adev->asic_type) {
1648         case CHIP_TOPAZ:
1649         case CHIP_TONGA:
1650         case CHIP_FIJI:
1651         case CHIP_POLARIS10:
1652         case CHIP_POLARIS11:
1653         case CHIP_POLARIS12:
1654         case CHIP_VEGAM:
1655         case CHIP_CARRIZO:
1656         case CHIP_STONEY:
1657                 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1658                         adev->family = AMDGPU_FAMILY_CZ;
1659                 else
1660                         adev->family = AMDGPU_FAMILY_VI;
1661
1662                 r = vi_set_ip_blocks(adev);
1663                 if (r)
1664                         return r;
1665                 break;
1666 #ifdef CONFIG_DRM_AMDGPU_SI
1667         case CHIP_VERDE:
1668         case CHIP_TAHITI:
1669         case CHIP_PITCAIRN:
1670         case CHIP_OLAND:
1671         case CHIP_HAINAN:
1672                 adev->family = AMDGPU_FAMILY_SI;
1673                 r = si_set_ip_blocks(adev);
1674                 if (r)
1675                         return r;
1676                 break;
1677 #endif
1678 #ifdef CONFIG_DRM_AMDGPU_CIK
1679         case CHIP_BONAIRE:
1680         case CHIP_HAWAII:
1681         case CHIP_KAVERI:
1682         case CHIP_KABINI:
1683         case CHIP_MULLINS:
1684                 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1685                         adev->family = AMDGPU_FAMILY_CI;
1686                 else
1687                         adev->family = AMDGPU_FAMILY_KV;
1688
1689                 r = cik_set_ip_blocks(adev);
1690                 if (r)
1691                         return r;
1692                 break;
1693 #endif
1694         case CHIP_VEGA10:
1695         case CHIP_VEGA12:
1696         case CHIP_VEGA20:
1697         case CHIP_RAVEN:
1698         case CHIP_ARCTURUS:
1699         case CHIP_RENOIR:
1700                 if (adev->asic_type == CHIP_RAVEN ||
1701                     adev->asic_type == CHIP_RENOIR)
1702                         adev->family = AMDGPU_FAMILY_RV;
1703                 else
1704                         adev->family = AMDGPU_FAMILY_AI;
1705
1706                 r = soc15_set_ip_blocks(adev);
1707                 if (r)
1708                         return r;
1709                 break;
1710         case  CHIP_NAVI10:
1711         case  CHIP_NAVI14:
1712         case  CHIP_NAVI12:
1713                 adev->family = AMDGPU_FAMILY_NV;
1714
1715                 r = nv_set_ip_blocks(adev);
1716                 if (r)
1717                         return r;
1718                 break;
1719         default:
1720                 /* FIXME: not supported yet */
1721                 return -EINVAL;
1722         }
1723
1724         r = amdgpu_device_parse_gpu_info_fw(adev);
1725         if (r)
1726                 return r;
1727
1728         if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10)
1729                 amdgpu_discovery_get_gfx_info(adev);
1730
1731         amdgpu_amdkfd_device_probe(adev);
1732
1733         if (amdgpu_sriov_vf(adev)) {
1734                 r = amdgpu_virt_request_full_gpu(adev, true);
1735                 if (r)
1736                         return -EAGAIN;
1737         }
1738
1739         adev->pm.pp_feature = amdgpu_pp_feature_mask;
1740         if (amdgpu_sriov_vf(adev) || sched_policy == KFD_SCHED_POLICY_NO_HWS)
1741                 adev->pm.pp_feature &= ~PP_GFXOFF_MASK;
1742
1743         for (i = 0; i < adev->num_ip_blocks; i++) {
1744                 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1745                         DRM_ERROR("disabled ip block: %d <%s>\n",
1746                                   i, adev->ip_blocks[i].version->funcs->name);
1747                         adev->ip_blocks[i].status.valid = false;
1748                 } else {
1749                         if (adev->ip_blocks[i].version->funcs->early_init) {
1750                                 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
1751                                 if (r == -ENOENT) {
1752                                         adev->ip_blocks[i].status.valid = false;
1753                                 } else if (r) {
1754                                         DRM_ERROR("early_init of IP block <%s> failed %d\n",
1755                                                   adev->ip_blocks[i].version->funcs->name, r);
1756                                         return r;
1757                                 } else {
1758                                         adev->ip_blocks[i].status.valid = true;
1759                                 }
1760                         } else {
1761                                 adev->ip_blocks[i].status.valid = true;
1762                         }
1763                 }
1764                 /* get the vbios after the asic_funcs are set up */
1765                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) {
1766                         /* Read BIOS */
1767                         if (!amdgpu_get_bios(adev))
1768                                 return -EINVAL;
1769
1770                         r = amdgpu_atombios_init(adev);
1771                         if (r) {
1772                                 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
1773                                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
1774                                 return r;
1775                         }
1776                 }
1777         }
1778
1779         adev->cg_flags &= amdgpu_cg_mask;
1780         adev->pg_flags &= amdgpu_pg_mask;
1781
1782         return 0;
1783 }
1784
1785 static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev)
1786 {
1787         int i, r;
1788
1789         for (i = 0; i < adev->num_ip_blocks; i++) {
1790                 if (!adev->ip_blocks[i].status.sw)
1791                         continue;
1792                 if (adev->ip_blocks[i].status.hw)
1793                         continue;
1794                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
1795                     (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) ||
1796                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
1797                         r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1798                         if (r) {
1799                                 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1800                                           adev->ip_blocks[i].version->funcs->name, r);
1801                                 return r;
1802                         }
1803                         adev->ip_blocks[i].status.hw = true;
1804                 }
1805         }
1806
1807         return 0;
1808 }
1809
1810 static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev)
1811 {
1812         int i, r;
1813
1814         for (i = 0; i < adev->num_ip_blocks; i++) {
1815                 if (!adev->ip_blocks[i].status.sw)
1816                         continue;
1817                 if (adev->ip_blocks[i].status.hw)
1818                         continue;
1819                 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1820                 if (r) {
1821                         DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1822                                   adev->ip_blocks[i].version->funcs->name, r);
1823                         return r;
1824                 }
1825                 adev->ip_blocks[i].status.hw = true;
1826         }
1827
1828         return 0;
1829 }
1830
1831 static int amdgpu_device_fw_loading(struct amdgpu_device *adev)
1832 {
1833         int r = 0;
1834         int i;
1835         uint32_t smu_version;
1836
1837         if (adev->asic_type >= CHIP_VEGA10) {
1838                 for (i = 0; i < adev->num_ip_blocks; i++) {
1839                         if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP)
1840                                 continue;
1841
1842                         /* no need to do the fw loading again if already done*/
1843                         if (adev->ip_blocks[i].status.hw == true)
1844                                 break;
1845
1846                         if (adev->in_gpu_reset || adev->in_suspend) {
1847                                 r = adev->ip_blocks[i].version->funcs->resume(adev);
1848                                 if (r) {
1849                                         DRM_ERROR("resume of IP block <%s> failed %d\n",
1850                                                           adev->ip_blocks[i].version->funcs->name, r);
1851                                         return r;
1852                                 }
1853                         } else {
1854                                 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
1855                                 if (r) {
1856                                         DRM_ERROR("hw_init of IP block <%s> failed %d\n",
1857                                                           adev->ip_blocks[i].version->funcs->name, r);
1858                                         return r;
1859                                 }
1860                         }
1861
1862                         adev->ip_blocks[i].status.hw = true;
1863                         break;
1864                 }
1865         }
1866
1867         if (!amdgpu_sriov_vf(adev) || adev->asic_type == CHIP_TONGA)
1868                 r = amdgpu_pm_load_smu_firmware(adev, &smu_version);
1869
1870         return r;
1871 }
1872
1873 /**
1874  * amdgpu_device_ip_init - run init for hardware IPs
1875  *
1876  * @adev: amdgpu_device pointer
1877  *
1878  * Main initialization pass for hardware IPs.  The list of all the hardware
1879  * IPs that make up the asic is walked and the sw_init and hw_init callbacks
1880  * are run.  sw_init initializes the software state associated with each IP
1881  * and hw_init initializes the hardware associated with each IP.
1882  * Returns 0 on success, negative error code on failure.
1883  */
1884 static int amdgpu_device_ip_init(struct amdgpu_device *adev)
1885 {
1886         int i, r;
1887
1888         r = amdgpu_ras_init(adev);
1889         if (r)
1890                 return r;
1891
1892         for (i = 0; i < adev->num_ip_blocks; i++) {
1893                 if (!adev->ip_blocks[i].status.valid)
1894                         continue;
1895                 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
1896                 if (r) {
1897                         DRM_ERROR("sw_init of IP block <%s> failed %d\n",
1898                                   adev->ip_blocks[i].version->funcs->name, r);
1899                         goto init_failed;
1900                 }
1901                 adev->ip_blocks[i].status.sw = true;
1902
1903                 /* need to do gmc hw init early so we can allocate gpu mem */
1904                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
1905                         r = amdgpu_device_vram_scratch_init(adev);
1906                         if (r) {
1907                                 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1908                                 goto init_failed;
1909                         }
1910                         r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
1911                         if (r) {
1912                                 DRM_ERROR("hw_init %d failed %d\n", i, r);
1913                                 goto init_failed;
1914                         }
1915                         r = amdgpu_device_wb_init(adev);
1916                         if (r) {
1917                                 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
1918                                 goto init_failed;
1919                         }
1920                         adev->ip_blocks[i].status.hw = true;
1921
1922                         /* right after GMC hw init, we create CSA */
1923                         if (amdgpu_mcbp || amdgpu_sriov_vf(adev)) {
1924                                 r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj,
1925                                                                 AMDGPU_GEM_DOMAIN_VRAM,
1926                                                                 AMDGPU_CSA_SIZE);
1927                                 if (r) {
1928                                         DRM_ERROR("allocate CSA failed %d\n", r);
1929                                         goto init_failed;
1930                                 }
1931                         }
1932                 }
1933         }
1934
1935         if (amdgpu_sriov_vf(adev))
1936                 amdgpu_virt_init_data_exchange(adev);
1937
1938         r = amdgpu_ib_pool_init(adev);
1939         if (r) {
1940                 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
1941                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
1942                 goto init_failed;
1943         }
1944
1945         r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/
1946         if (r)
1947                 goto init_failed;
1948
1949         r = amdgpu_device_ip_hw_init_phase1(adev);
1950         if (r)
1951                 goto init_failed;
1952
1953         r = amdgpu_device_fw_loading(adev);
1954         if (r)
1955                 goto init_failed;
1956
1957         r = amdgpu_device_ip_hw_init_phase2(adev);
1958         if (r)
1959                 goto init_failed;
1960
1961         /*
1962          * retired pages will be loaded from eeprom and reserved here,
1963          * it should be called after amdgpu_device_ip_hw_init_phase2  since
1964          * for some ASICs the RAS EEPROM code relies on SMU fully functioning
1965          * for I2C communication which only true at this point.
1966          * recovery_init may fail, but it can free all resources allocated by
1967          * itself and its failure should not stop amdgpu init process.
1968          *
1969          * Note: theoretically, this should be called before all vram allocations
1970          * to protect retired page from abusing
1971          */
1972         amdgpu_ras_recovery_init(adev);
1973
1974         if (adev->gmc.xgmi.num_physical_nodes > 1)
1975                 amdgpu_xgmi_add_device(adev);
1976         amdgpu_amdkfd_device_init(adev);
1977
1978 init_failed:
1979         if (amdgpu_sriov_vf(adev))
1980                 amdgpu_virt_release_full_gpu(adev, true);
1981
1982         return r;
1983 }
1984
1985 /**
1986  * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
1987  *
1988  * @adev: amdgpu_device pointer
1989  *
1990  * Writes a reset magic value to the gart pointer in VRAM.  The driver calls
1991  * this function before a GPU reset.  If the value is retained after a
1992  * GPU reset, VRAM has not been lost.  Some GPU resets may destry VRAM contents.
1993  */
1994 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
1995 {
1996         memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
1997 }
1998
1999 /**
2000  * amdgpu_device_check_vram_lost - check if vram is valid
2001  *
2002  * @adev: amdgpu_device pointer
2003  *
2004  * Checks the reset magic value written to the gart pointer in VRAM.
2005  * The driver calls this after a GPU reset to see if the contents of
2006  * VRAM is lost or now.
2007  * returns true if vram is lost, false if not.
2008  */
2009 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
2010 {
2011         if (memcmp(adev->gart.ptr, adev->reset_magic,
2012                         AMDGPU_RESET_MAGIC_NUM))
2013                 return true;
2014
2015         if (!adev->in_gpu_reset)
2016                 return false;
2017
2018         /*
2019          * For all ASICs with baco/mode1 reset, the VRAM is
2020          * always assumed to be lost.
2021          */
2022         switch (amdgpu_asic_reset_method(adev)) {
2023         case AMD_RESET_METHOD_BACO:
2024         case AMD_RESET_METHOD_MODE1:
2025                 return true;
2026         default:
2027                 return false;
2028         }
2029 }
2030
2031 /**
2032  * amdgpu_device_set_cg_state - set clockgating for amdgpu device
2033  *
2034  * @adev: amdgpu_device pointer
2035  * @state: clockgating state (gate or ungate)
2036  *
2037  * The list of all the hardware IPs that make up the asic is walked and the
2038  * set_clockgating_state callbacks are run.
2039  * Late initialization pass enabling clockgating for hardware IPs.
2040  * Fini or suspend, pass disabling clockgating for hardware IPs.
2041  * Returns 0 on success, negative error code on failure.
2042  */
2043
2044 static int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
2045                                                 enum amd_clockgating_state state)
2046 {
2047         int i, j, r;
2048
2049         if (amdgpu_emu_mode == 1)
2050                 return 0;
2051
2052         for (j = 0; j < adev->num_ip_blocks; j++) {
2053                 i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2054                 if (!adev->ip_blocks[i].status.late_initialized)
2055                         continue;
2056                 /* skip CG for VCE/UVD, it's handled specially */
2057                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2058                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2059                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2060                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2061                     adev->ip_blocks[i].version->funcs->set_clockgating_state) {
2062                         /* enable clockgating to save power */
2063                         r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
2064                                                                                      state);
2065                         if (r) {
2066                                 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
2067                                           adev->ip_blocks[i].version->funcs->name, r);
2068                                 return r;
2069                         }
2070                 }
2071         }
2072
2073         return 0;
2074 }
2075
2076 static int amdgpu_device_set_pg_state(struct amdgpu_device *adev, enum amd_powergating_state state)
2077 {
2078         int i, j, r;
2079
2080         if (amdgpu_emu_mode == 1)
2081                 return 0;
2082
2083         for (j = 0; j < adev->num_ip_blocks; j++) {
2084                 i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2085                 if (!adev->ip_blocks[i].status.late_initialized)
2086                         continue;
2087                 /* skip CG for VCE/UVD, it's handled specially */
2088                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2089                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2090                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2091                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2092                     adev->ip_blocks[i].version->funcs->set_powergating_state) {
2093                         /* enable powergating to save power */
2094                         r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
2095                                                                                         state);
2096                         if (r) {
2097                                 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
2098                                           adev->ip_blocks[i].version->funcs->name, r);
2099                                 return r;
2100                         }
2101                 }
2102         }
2103         return 0;
2104 }
2105
2106 static int amdgpu_device_enable_mgpu_fan_boost(void)
2107 {
2108         struct amdgpu_gpu_instance *gpu_ins;
2109         struct amdgpu_device *adev;
2110         int i, ret = 0;
2111
2112         mutex_lock(&mgpu_info.mutex);
2113
2114         /*
2115          * MGPU fan boost feature should be enabled
2116          * only when there are two or more dGPUs in
2117          * the system
2118          */
2119         if (mgpu_info.num_dgpu < 2)
2120                 goto out;
2121
2122         for (i = 0; i < mgpu_info.num_dgpu; i++) {
2123                 gpu_ins = &(mgpu_info.gpu_ins[i]);
2124                 adev = gpu_ins->adev;
2125                 if (!(adev->flags & AMD_IS_APU) &&
2126                     !gpu_ins->mgpu_fan_enabled &&
2127                     adev->powerplay.pp_funcs &&
2128                     adev->powerplay.pp_funcs->enable_mgpu_fan_boost) {
2129                         ret = amdgpu_dpm_enable_mgpu_fan_boost(adev);
2130                         if (ret)
2131                                 break;
2132
2133                         gpu_ins->mgpu_fan_enabled = 1;
2134                 }
2135         }
2136
2137 out:
2138         mutex_unlock(&mgpu_info.mutex);
2139
2140         return ret;
2141 }
2142
2143 /**
2144  * amdgpu_device_ip_late_init - run late init for hardware IPs
2145  *
2146  * @adev: amdgpu_device pointer
2147  *
2148  * Late initialization pass for hardware IPs.  The list of all the hardware
2149  * IPs that make up the asic is walked and the late_init callbacks are run.
2150  * late_init covers any special initialization that an IP requires
2151  * after all of the have been initialized or something that needs to happen
2152  * late in the init process.
2153  * Returns 0 on success, negative error code on failure.
2154  */
2155 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
2156 {
2157         struct amdgpu_gpu_instance *gpu_instance;
2158         int i = 0, r;
2159
2160         for (i = 0; i < adev->num_ip_blocks; i++) {
2161                 if (!adev->ip_blocks[i].status.hw)
2162                         continue;
2163                 if (adev->ip_blocks[i].version->funcs->late_init) {
2164                         r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
2165                         if (r) {
2166                                 DRM_ERROR("late_init of IP block <%s> failed %d\n",
2167                                           adev->ip_blocks[i].version->funcs->name, r);
2168                                 return r;
2169                         }
2170                 }
2171                 adev->ip_blocks[i].status.late_initialized = true;
2172         }
2173
2174         amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
2175         amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
2176
2177         amdgpu_device_fill_reset_magic(adev);
2178
2179         r = amdgpu_device_enable_mgpu_fan_boost();
2180         if (r)
2181                 DRM_ERROR("enable mgpu fan boost failed (%d).\n", r);
2182
2183
2184         if (adev->gmc.xgmi.num_physical_nodes > 1) {
2185                 mutex_lock(&mgpu_info.mutex);
2186
2187                 /*
2188                  * Reset device p-state to low as this was booted with high.
2189                  *
2190                  * This should be performed only after all devices from the same
2191                  * hive get initialized.
2192                  *
2193                  * However, it's unknown how many device in the hive in advance.
2194                  * As this is counted one by one during devices initializations.
2195                  *
2196                  * So, we wait for all XGMI interlinked devices initialized.
2197                  * This may bring some delays as those devices may come from
2198                  * different hives. But that should be OK.
2199                  */
2200                 if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) {
2201                         for (i = 0; i < mgpu_info.num_gpu; i++) {
2202                                 gpu_instance = &(mgpu_info.gpu_ins[i]);
2203                                 if (gpu_instance->adev->flags & AMD_IS_APU)
2204                                         continue;
2205
2206                                 r = amdgpu_xgmi_set_pstate(gpu_instance->adev, 0);
2207                                 if (r) {
2208                                         DRM_ERROR("pstate setting failed (%d).\n", r);
2209                                         break;
2210                                 }
2211                         }
2212                 }
2213
2214                 mutex_unlock(&mgpu_info.mutex);
2215         }
2216
2217         return 0;
2218 }
2219
2220 /**
2221  * amdgpu_device_ip_fini - run fini for hardware IPs
2222  *
2223  * @adev: amdgpu_device pointer
2224  *
2225  * Main teardown pass for hardware IPs.  The list of all the hardware
2226  * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
2227  * are run.  hw_fini tears down the hardware associated with each IP
2228  * and sw_fini tears down any software state associated with each IP.
2229  * Returns 0 on success, negative error code on failure.
2230  */
2231 static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
2232 {
2233         int i, r;
2234
2235         amdgpu_ras_pre_fini(adev);
2236
2237         if (adev->gmc.xgmi.num_physical_nodes > 1)
2238                 amdgpu_xgmi_remove_device(adev);
2239
2240         amdgpu_amdkfd_device_fini(adev);
2241
2242         amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2243         amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2244
2245         /* need to disable SMC first */
2246         for (i = 0; i < adev->num_ip_blocks; i++) {
2247                 if (!adev->ip_blocks[i].status.hw)
2248                         continue;
2249                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2250                         r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2251                         /* XXX handle errors */
2252                         if (r) {
2253                                 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2254                                           adev->ip_blocks[i].version->funcs->name, r);
2255                         }
2256                         adev->ip_blocks[i].status.hw = false;
2257                         break;
2258                 }
2259         }
2260
2261         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2262                 if (!adev->ip_blocks[i].status.hw)
2263                         continue;
2264
2265                 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2266                 /* XXX handle errors */
2267                 if (r) {
2268                         DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2269                                   adev->ip_blocks[i].version->funcs->name, r);
2270                 }
2271
2272                 adev->ip_blocks[i].status.hw = false;
2273         }
2274
2275
2276         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2277                 if (!adev->ip_blocks[i].status.sw)
2278                         continue;
2279
2280                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
2281                         amdgpu_ucode_free_bo(adev);
2282                         amdgpu_free_static_csa(&adev->virt.csa_obj);
2283                         amdgpu_device_wb_fini(adev);
2284                         amdgpu_device_vram_scratch_fini(adev);
2285                         amdgpu_ib_pool_fini(adev);
2286                 }
2287
2288                 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
2289                 /* XXX handle errors */
2290                 if (r) {
2291                         DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
2292                                   adev->ip_blocks[i].version->funcs->name, r);
2293                 }
2294                 adev->ip_blocks[i].status.sw = false;
2295                 adev->ip_blocks[i].status.valid = false;
2296         }
2297
2298         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2299                 if (!adev->ip_blocks[i].status.late_initialized)
2300                         continue;
2301                 if (adev->ip_blocks[i].version->funcs->late_fini)
2302                         adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
2303                 adev->ip_blocks[i].status.late_initialized = false;
2304         }
2305
2306         amdgpu_ras_fini(adev);
2307
2308         if (amdgpu_sriov_vf(adev))
2309                 if (amdgpu_virt_release_full_gpu(adev, false))
2310                         DRM_ERROR("failed to release exclusive mode on fini\n");
2311
2312         return 0;
2313 }
2314
2315 /**
2316  * amdgpu_device_delayed_init_work_handler - work handler for IB tests
2317  *
2318  * @work: work_struct.
2319  */
2320 static void amdgpu_device_delayed_init_work_handler(struct work_struct *work)
2321 {
2322         struct amdgpu_device *adev =
2323                 container_of(work, struct amdgpu_device, delayed_init_work.work);
2324         int r;
2325
2326         r = amdgpu_ib_ring_tests(adev);
2327         if (r)
2328                 DRM_ERROR("ib ring test failed (%d).\n", r);
2329 }
2330
2331 static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
2332 {
2333         struct amdgpu_device *adev =
2334                 container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
2335
2336         mutex_lock(&adev->gfx.gfx_off_mutex);
2337         if (!adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
2338                 if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
2339                         adev->gfx.gfx_off_state = true;
2340         }
2341         mutex_unlock(&adev->gfx.gfx_off_mutex);
2342 }
2343
2344 /**
2345  * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
2346  *
2347  * @adev: amdgpu_device pointer
2348  *
2349  * Main suspend function for hardware IPs.  The list of all the hardware
2350  * IPs that make up the asic is walked, clockgating is disabled and the
2351  * suspend callbacks are run.  suspend puts the hardware and software state
2352  * in each IP into a state suitable for suspend.
2353  * Returns 0 on success, negative error code on failure.
2354  */
2355 static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
2356 {
2357         int i, r;
2358
2359         amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2360         amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2361
2362         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2363                 if (!adev->ip_blocks[i].status.valid)
2364                         continue;
2365                 /* displays are handled separately */
2366                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) {
2367                         /* XXX handle errors */
2368                         r = adev->ip_blocks[i].version->funcs->suspend(adev);
2369                         /* XXX handle errors */
2370                         if (r) {
2371                                 DRM_ERROR("suspend of IP block <%s> failed %d\n",
2372                                           adev->ip_blocks[i].version->funcs->name, r);
2373                                 return r;
2374                         }
2375                         adev->ip_blocks[i].status.hw = false;
2376                 }
2377         }
2378
2379         return 0;
2380 }
2381
2382 /**
2383  * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
2384  *
2385  * @adev: amdgpu_device pointer
2386  *
2387  * Main suspend function for hardware IPs.  The list of all the hardware
2388  * IPs that make up the asic is walked, clockgating is disabled and the
2389  * suspend callbacks are run.  suspend puts the hardware and software state
2390  * in each IP into a state suitable for suspend.
2391  * Returns 0 on success, negative error code on failure.
2392  */
2393 static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
2394 {
2395         int i, r;
2396
2397         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2398                 if (!adev->ip_blocks[i].status.valid)
2399                         continue;
2400                 /* displays are handled in phase1 */
2401                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
2402                         continue;
2403                 /* PSP lost connection when err_event_athub occurs */
2404                 if (amdgpu_ras_intr_triggered() &&
2405                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
2406                         adev->ip_blocks[i].status.hw = false;
2407                         continue;
2408                 }
2409                 /* XXX handle errors */
2410                 r = adev->ip_blocks[i].version->funcs->suspend(adev);
2411                 /* XXX handle errors */
2412                 if (r) {
2413                         DRM_ERROR("suspend of IP block <%s> failed %d\n",
2414                                   adev->ip_blocks[i].version->funcs->name, r);
2415                 }
2416                 adev->ip_blocks[i].status.hw = false;
2417                 /* handle putting the SMC in the appropriate state */
2418                 if(!amdgpu_sriov_vf(adev)){
2419                         if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2420                                 r = amdgpu_dpm_set_mp1_state(adev, adev->mp1_state);
2421                                 if (r) {
2422                                         DRM_ERROR("SMC failed to set mp1 state %d, %d\n",
2423                                                         adev->mp1_state, r);
2424                                         return r;
2425                                 }
2426                         }
2427                 }
2428                 adev->ip_blocks[i].status.hw = false;
2429         }
2430
2431         return 0;
2432 }
2433
2434 /**
2435  * amdgpu_device_ip_suspend - run suspend for hardware IPs
2436  *
2437  * @adev: amdgpu_device pointer
2438  *
2439  * Main suspend function for hardware IPs.  The list of all the hardware
2440  * IPs that make up the asic is walked, clockgating is disabled and the
2441  * suspend callbacks are run.  suspend puts the hardware and software state
2442  * in each IP into a state suitable for suspend.
2443  * Returns 0 on success, negative error code on failure.
2444  */
2445 int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
2446 {
2447         int r;
2448
2449         if (amdgpu_sriov_vf(adev))
2450                 amdgpu_virt_request_full_gpu(adev, false);
2451
2452         r = amdgpu_device_ip_suspend_phase1(adev);
2453         if (r)
2454                 return r;
2455         r = amdgpu_device_ip_suspend_phase2(adev);
2456
2457         if (amdgpu_sriov_vf(adev))
2458                 amdgpu_virt_release_full_gpu(adev, false);
2459
2460         return r;
2461 }
2462
2463 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
2464 {
2465         int i, r;
2466
2467         static enum amd_ip_block_type ip_order[] = {
2468                 AMD_IP_BLOCK_TYPE_GMC,
2469                 AMD_IP_BLOCK_TYPE_COMMON,
2470                 AMD_IP_BLOCK_TYPE_PSP,
2471                 AMD_IP_BLOCK_TYPE_IH,
2472         };
2473
2474         for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2475                 int j;
2476                 struct amdgpu_ip_block *block;
2477
2478                 for (j = 0; j < adev->num_ip_blocks; j++) {
2479                         block = &adev->ip_blocks[j];
2480
2481                         block->status.hw = false;
2482                         if (block->version->type != ip_order[i] ||
2483                                 !block->status.valid)
2484                                 continue;
2485
2486                         r = block->version->funcs->hw_init(adev);
2487                         DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2488                         if (r)
2489                                 return r;
2490                         block->status.hw = true;
2491                 }
2492         }
2493
2494         return 0;
2495 }
2496
2497 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
2498 {
2499         int i, r;
2500
2501         static enum amd_ip_block_type ip_order[] = {
2502                 AMD_IP_BLOCK_TYPE_SMC,
2503                 AMD_IP_BLOCK_TYPE_DCE,
2504                 AMD_IP_BLOCK_TYPE_GFX,
2505                 AMD_IP_BLOCK_TYPE_SDMA,
2506                 AMD_IP_BLOCK_TYPE_UVD,
2507                 AMD_IP_BLOCK_TYPE_VCE,
2508                 AMD_IP_BLOCK_TYPE_VCN
2509         };
2510
2511         for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
2512                 int j;
2513                 struct amdgpu_ip_block *block;
2514
2515                 for (j = 0; j < adev->num_ip_blocks; j++) {
2516                         block = &adev->ip_blocks[j];
2517
2518                         if (block->version->type != ip_order[i] ||
2519                                 !block->status.valid ||
2520                                 block->status.hw)
2521                                 continue;
2522
2523                         if (block->version->type == AMD_IP_BLOCK_TYPE_SMC)
2524                                 r = block->version->funcs->resume(adev);
2525                         else
2526                                 r = block->version->funcs->hw_init(adev);
2527
2528                         DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
2529                         if (r)
2530                                 return r;
2531                         block->status.hw = true;
2532                 }
2533         }
2534
2535         return 0;
2536 }
2537
2538 /**
2539  * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
2540  *
2541  * @adev: amdgpu_device pointer
2542  *
2543  * First resume function for hardware IPs.  The list of all the hardware
2544  * IPs that make up the asic is walked and the resume callbacks are run for
2545  * COMMON, GMC, and IH.  resume puts the hardware into a functional state
2546  * after a suspend and updates the software state as necessary.  This
2547  * function is also used for restoring the GPU after a GPU reset.
2548  * Returns 0 on success, negative error code on failure.
2549  */
2550 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
2551 {
2552         int i, r;
2553
2554         for (i = 0; i < adev->num_ip_blocks; i++) {
2555                 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
2556                         continue;
2557                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2558                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2559                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2560
2561                         r = adev->ip_blocks[i].version->funcs->resume(adev);
2562                         if (r) {
2563                                 DRM_ERROR("resume of IP block <%s> failed %d\n",
2564                                           adev->ip_blocks[i].version->funcs->name, r);
2565                                 return r;
2566                         }
2567                         adev->ip_blocks[i].status.hw = true;
2568                 }
2569         }
2570
2571         return 0;
2572 }
2573
2574 /**
2575  * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
2576  *
2577  * @adev: amdgpu_device pointer
2578  *
2579  * First resume function for hardware IPs.  The list of all the hardware
2580  * IPs that make up the asic is walked and the resume callbacks are run for
2581  * all blocks except COMMON, GMC, and IH.  resume puts the hardware into a
2582  * functional state after a suspend and updates the software state as
2583  * necessary.  This function is also used for restoring the GPU after a GPU
2584  * reset.
2585  * Returns 0 on success, negative error code on failure.
2586  */
2587 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
2588 {
2589         int i, r;
2590
2591         for (i = 0; i < adev->num_ip_blocks; i++) {
2592                 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
2593                         continue;
2594                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2595                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
2596                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
2597                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
2598                         continue;
2599                 r = adev->ip_blocks[i].version->funcs->resume(adev);
2600                 if (r) {
2601                         DRM_ERROR("resume of IP block <%s> failed %d\n",
2602                                   adev->ip_blocks[i].version->funcs->name, r);
2603                         return r;
2604                 }
2605                 adev->ip_blocks[i].status.hw = true;
2606         }
2607
2608         return 0;
2609 }
2610
2611 /**
2612  * amdgpu_device_ip_resume - run resume for hardware IPs
2613  *
2614  * @adev: amdgpu_device pointer
2615  *
2616  * Main resume function for hardware IPs.  The hardware IPs
2617  * are split into two resume functions because they are
2618  * are also used in in recovering from a GPU reset and some additional
2619  * steps need to be take between them.  In this case (S3/S4) they are
2620  * run sequentially.
2621  * Returns 0 on success, negative error code on failure.
2622  */
2623 static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
2624 {
2625         int r;
2626
2627         r = amdgpu_device_ip_resume_phase1(adev);
2628         if (r)
2629                 return r;
2630
2631         r = amdgpu_device_fw_loading(adev);
2632         if (r)
2633                 return r;
2634
2635         r = amdgpu_device_ip_resume_phase2(adev);
2636
2637         return r;
2638 }
2639
2640 /**
2641  * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
2642  *
2643  * @adev: amdgpu_device pointer
2644  *
2645  * Query the VBIOS data tables to determine if the board supports SR-IOV.
2646  */
2647 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
2648 {
2649         if (amdgpu_sriov_vf(adev)) {
2650                 if (adev->is_atom_fw) {
2651                         if (amdgpu_atomfirmware_gpu_supports_virtualization(adev))
2652                                 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2653                 } else {
2654                         if (amdgpu_atombios_has_gpu_virtualization_table(adev))
2655                                 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
2656                 }
2657
2658                 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
2659                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
2660         }
2661 }
2662
2663 /**
2664  * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
2665  *
2666  * @asic_type: AMD asic type
2667  *
2668  * Check if there is DC (new modesetting infrastructre) support for an asic.
2669  * returns true if DC has support, false if not.
2670  */
2671 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
2672 {
2673         switch (asic_type) {
2674 #if defined(CONFIG_DRM_AMD_DC)
2675         case CHIP_BONAIRE:
2676         case CHIP_KAVERI:
2677         case CHIP_KABINI:
2678         case CHIP_MULLINS:
2679                 /*
2680                  * We have systems in the wild with these ASICs that require
2681                  * LVDS and VGA support which is not supported with DC.
2682                  *
2683                  * Fallback to the non-DC driver here by default so as not to
2684                  * cause regressions.
2685                  */
2686                 return amdgpu_dc > 0;
2687         case CHIP_HAWAII:
2688         case CHIP_CARRIZO:
2689         case CHIP_STONEY:
2690         case CHIP_POLARIS10:
2691         case CHIP_POLARIS11:
2692         case CHIP_POLARIS12:
2693         case CHIP_VEGAM:
2694         case CHIP_TONGA:
2695         case CHIP_FIJI:
2696         case CHIP_VEGA10:
2697         case CHIP_VEGA12:
2698         case CHIP_VEGA20:
2699 #if defined(CONFIG_DRM_AMD_DC_DCN)
2700         case CHIP_RAVEN:
2701         case CHIP_NAVI10:
2702         case CHIP_NAVI14:
2703         case CHIP_NAVI12:
2704         case CHIP_RENOIR:
2705 #endif
2706                 return amdgpu_dc != 0;
2707 #endif
2708         default:
2709                 if (amdgpu_dc > 0)
2710                         DRM_INFO("Display Core has been requested via kernel parameter "
2711                                          "but isn't supported by ASIC, ignoring\n");
2712                 return false;
2713         }
2714 }
2715
2716 /**
2717  * amdgpu_device_has_dc_support - check if dc is supported
2718  *
2719  * @adev: amdgpu_device_pointer
2720  *
2721  * Returns true for supported, false for not supported
2722  */
2723 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
2724 {
2725         if (amdgpu_sriov_vf(adev))
2726                 return false;
2727
2728         return amdgpu_device_asic_has_dc_support(adev->asic_type);
2729 }
2730
2731
2732 static void amdgpu_device_xgmi_reset_func(struct work_struct *__work)
2733 {
2734         struct amdgpu_device *adev =
2735                 container_of(__work, struct amdgpu_device, xgmi_reset_work);
2736         struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev, 0);
2737
2738         /* It's a bug to not have a hive within this function */
2739         if (WARN_ON(!hive))
2740                 return;
2741
2742         /*
2743          * Use task barrier to synchronize all xgmi reset works across the
2744          * hive. task_barrier_enter and task_barrier_exit will block
2745          * until all the threads running the xgmi reset works reach
2746          * those points. task_barrier_full will do both blocks.
2747          */
2748         if (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) {
2749
2750                 task_barrier_enter(&hive->tb);
2751                 adev->asic_reset_res = amdgpu_device_baco_enter(adev->ddev);
2752
2753                 if (adev->asic_reset_res)
2754                         goto fail;
2755
2756                 task_barrier_exit(&hive->tb);
2757                 adev->asic_reset_res = amdgpu_device_baco_exit(adev->ddev);
2758
2759                 if (adev->asic_reset_res)
2760                         goto fail;
2761
2762                 if (adev->mmhub.funcs && adev->mmhub.funcs->reset_ras_error_count)
2763                         adev->mmhub.funcs->reset_ras_error_count(adev);
2764         } else {
2765
2766                 task_barrier_full(&hive->tb);
2767                 adev->asic_reset_res =  amdgpu_asic_reset(adev);
2768         }
2769
2770 fail:
2771         if (adev->asic_reset_res)
2772                 DRM_WARN("ASIC reset failed with error, %d for drm dev, %s",
2773                          adev->asic_reset_res, adev->ddev->unique);
2774 }
2775
2776 static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev)
2777 {
2778         char *input = amdgpu_lockup_timeout;
2779         char *timeout_setting = NULL;
2780         int index = 0;
2781         long timeout;
2782         int ret = 0;
2783
2784         /*
2785          * By default timeout for non compute jobs is 10000.
2786          * And there is no timeout enforced on compute jobs.
2787          * In SR-IOV or passthrough mode, timeout for compute
2788          * jobs are 10000 by default.
2789          */
2790         adev->gfx_timeout = msecs_to_jiffies(10000);
2791         adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
2792         if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
2793                 adev->compute_timeout = adev->gfx_timeout;
2794         else
2795                 adev->compute_timeout = MAX_SCHEDULE_TIMEOUT;
2796
2797         if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
2798                 while ((timeout_setting = strsep(&input, ",")) &&
2799                                 strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
2800                         ret = kstrtol(timeout_setting, 0, &timeout);
2801                         if (ret)
2802                                 return ret;
2803
2804                         if (timeout == 0) {
2805                                 index++;
2806                                 continue;
2807                         } else if (timeout < 0) {
2808                                 timeout = MAX_SCHEDULE_TIMEOUT;
2809                         } else {
2810                                 timeout = msecs_to_jiffies(timeout);
2811                         }
2812
2813                         switch (index++) {
2814                         case 0:
2815                                 adev->gfx_timeout = timeout;
2816                                 break;
2817                         case 1:
2818                                 adev->compute_timeout = timeout;
2819                                 break;
2820                         case 2:
2821                                 adev->sdma_timeout = timeout;
2822                                 break;
2823                         case 3:
2824                                 adev->video_timeout = timeout;
2825                                 break;
2826                         default:
2827                                 break;
2828                         }
2829                 }
2830                 /*
2831                  * There is only one value specified and
2832                  * it should apply to all non-compute jobs.
2833                  */
2834                 if (index == 1) {
2835                         adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
2836                         if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
2837                                 adev->compute_timeout = adev->gfx_timeout;
2838                 }
2839         }
2840
2841         return ret;
2842 }
2843
2844 /**
2845  * amdgpu_device_init - initialize the driver
2846  *
2847  * @adev: amdgpu_device pointer
2848  * @ddev: drm dev pointer
2849  * @pdev: pci dev pointer
2850  * @flags: driver flags
2851  *
2852  * Initializes the driver info and hw (all asics).
2853  * Returns 0 for success or an error on failure.
2854  * Called at driver startup.
2855  */
2856 int amdgpu_device_init(struct amdgpu_device *adev,
2857                        struct drm_device *ddev,
2858                        struct pci_dev *pdev,
2859                        uint32_t flags)
2860 {
2861         int r, i;
2862         bool boco = false;
2863         u32 max_MBps;
2864
2865         adev->shutdown = false;
2866         adev->dev = &pdev->dev;
2867         adev->ddev = ddev;
2868         adev->pdev = pdev;
2869         adev->flags = flags;
2870
2871         if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST)
2872                 adev->asic_type = amdgpu_force_asic_type;
2873         else
2874                 adev->asic_type = flags & AMD_ASIC_MASK;
2875
2876         adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
2877         if (amdgpu_emu_mode == 1)
2878                 adev->usec_timeout *= 10;
2879         adev->gmc.gart_size = 512 * 1024 * 1024;
2880         adev->accel_working = false;
2881         adev->num_rings = 0;
2882         adev->mman.buffer_funcs = NULL;
2883         adev->mman.buffer_funcs_ring = NULL;
2884         adev->vm_manager.vm_pte_funcs = NULL;
2885         adev->vm_manager.vm_pte_num_scheds = 0;
2886         adev->gmc.gmc_funcs = NULL;
2887         adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
2888         bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
2889
2890         adev->smc_rreg = &amdgpu_invalid_rreg;
2891         adev->smc_wreg = &amdgpu_invalid_wreg;
2892         adev->pcie_rreg = &amdgpu_invalid_rreg;
2893         adev->pcie_wreg = &amdgpu_invalid_wreg;
2894         adev->pciep_rreg = &amdgpu_invalid_rreg;
2895         adev->pciep_wreg = &amdgpu_invalid_wreg;
2896         adev->pcie_rreg64 = &amdgpu_invalid_rreg64;
2897         adev->pcie_wreg64 = &amdgpu_invalid_wreg64;
2898         adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
2899         adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
2900         adev->didt_rreg = &amdgpu_invalid_rreg;
2901         adev->didt_wreg = &amdgpu_invalid_wreg;
2902         adev->gc_cac_rreg = &amdgpu_invalid_rreg;
2903         adev->gc_cac_wreg = &amdgpu_invalid_wreg;
2904         adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
2905         adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
2906
2907         DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
2908                  amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
2909                  pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
2910
2911         /* mutex initialization are all done here so we
2912          * can recall function without having locking issues */
2913         atomic_set(&adev->irq.ih.lock, 0);
2914         mutex_init(&adev->firmware.mutex);
2915         mutex_init(&adev->pm.mutex);
2916         mutex_init(&adev->gfx.gpu_clock_mutex);
2917         mutex_init(&adev->srbm_mutex);
2918         mutex_init(&adev->gfx.pipe_reserve_mutex);
2919         mutex_init(&adev->gfx.gfx_off_mutex);
2920         mutex_init(&adev->grbm_idx_mutex);
2921         mutex_init(&adev->mn_lock);
2922         mutex_init(&adev->virt.vf_errors.lock);
2923         hash_init(adev->mn_hash);
2924         mutex_init(&adev->lock_reset);
2925         mutex_init(&adev->psp.mutex);
2926         mutex_init(&adev->notifier_lock);
2927
2928         r = amdgpu_device_check_arguments(adev);
2929         if (r)
2930                 return r;
2931
2932         spin_lock_init(&adev->mmio_idx_lock);
2933         spin_lock_init(&adev->smc_idx_lock);
2934         spin_lock_init(&adev->pcie_idx_lock);
2935         spin_lock_init(&adev->uvd_ctx_idx_lock);
2936         spin_lock_init(&adev->didt_idx_lock);
2937         spin_lock_init(&adev->gc_cac_idx_lock);
2938         spin_lock_init(&adev->se_cac_idx_lock);
2939         spin_lock_init(&adev->audio_endpt_idx_lock);
2940         spin_lock_init(&adev->mm_stats.lock);
2941
2942         INIT_LIST_HEAD(&adev->shadow_list);
2943         mutex_init(&adev->shadow_list_lock);
2944
2945         INIT_LIST_HEAD(&adev->ring_lru_list);
2946         spin_lock_init(&adev->ring_lru_list_lock);
2947
2948         INIT_DELAYED_WORK(&adev->delayed_init_work,
2949                           amdgpu_device_delayed_init_work_handler);
2950         INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
2951                           amdgpu_device_delay_enable_gfx_off);
2952
2953         INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func);
2954
2955         adev->gfx.gfx_off_req_count = 1;
2956         adev->pm.ac_power = power_supply_is_system_supplied() > 0 ? true : false;
2957
2958         /* Registers mapping */
2959         /* TODO: block userspace mapping of io register */
2960         if (adev->asic_type >= CHIP_BONAIRE) {
2961                 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
2962                 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
2963         } else {
2964                 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
2965                 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
2966         }
2967
2968         adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
2969         if (adev->rmmio == NULL) {
2970                 return -ENOMEM;
2971         }
2972         DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
2973         DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
2974
2975         /* io port mapping */
2976         for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
2977                 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
2978                         adev->rio_mem_size = pci_resource_len(adev->pdev, i);
2979                         adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
2980                         break;
2981                 }
2982         }
2983         if (adev->rio_mem == NULL)
2984                 DRM_INFO("PCI I/O BAR is not found.\n");
2985
2986         /* enable PCIE atomic ops */
2987         r = pci_enable_atomic_ops_to_root(adev->pdev,
2988                                           PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
2989                                           PCI_EXP_DEVCAP2_ATOMIC_COMP64);
2990         if (r) {
2991                 adev->have_atomics_support = false;
2992                 DRM_INFO("PCIE atomic ops is not supported\n");
2993         } else {
2994                 adev->have_atomics_support = true;
2995         }
2996
2997         amdgpu_device_get_pcie_info(adev);
2998
2999         if (amdgpu_mcbp)
3000                 DRM_INFO("MCBP is enabled\n");
3001
3002         if (amdgpu_mes && adev->asic_type >= CHIP_NAVI10)
3003                 adev->enable_mes = true;
3004
3005         if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10) {
3006                 r = amdgpu_discovery_init(adev);
3007                 if (r) {
3008                         dev_err(adev->dev, "amdgpu_discovery_init failed\n");
3009                         return r;
3010                 }
3011         }
3012
3013         /* early init functions */
3014         r = amdgpu_device_ip_early_init(adev);
3015         if (r)
3016                 return r;
3017
3018         r = amdgpu_device_get_job_timeout_settings(adev);
3019         if (r) {
3020                 dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n");
3021                 return r;
3022         }
3023
3024         /* doorbell bar mapping and doorbell index init*/
3025         amdgpu_device_doorbell_init(adev);
3026
3027         /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
3028         /* this will fail for cards that aren't VGA class devices, just
3029          * ignore it */
3030         vga_client_register(adev->pdev, adev, NULL, amdgpu_device_vga_set_decode);
3031
3032         if (amdgpu_device_supports_boco(ddev))
3033                 boco = true;
3034         if (amdgpu_has_atpx() &&
3035             (amdgpu_is_atpx_hybrid() ||
3036              amdgpu_has_atpx_dgpu_power_cntl()) &&
3037             !pci_is_thunderbolt_attached(adev->pdev))
3038                 vga_switcheroo_register_client(adev->pdev,
3039                                                &amdgpu_switcheroo_ops, boco);
3040         if (boco)
3041                 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
3042
3043         if (amdgpu_emu_mode == 1) {
3044                 /* post the asic on emulation mode */
3045                 emu_soc_asic_init(adev);
3046                 goto fence_driver_init;
3047         }
3048
3049         /* detect if we are with an SRIOV vbios */
3050         amdgpu_device_detect_sriov_bios(adev);
3051
3052         /* check if we need to reset the asic
3053          *  E.g., driver was not cleanly unloaded previously, etc.
3054          */
3055         if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) {
3056                 r = amdgpu_asic_reset(adev);
3057                 if (r) {
3058                         dev_err(adev->dev, "asic reset on init failed\n");
3059                         goto failed;
3060                 }
3061         }
3062
3063         /* Post card if necessary */
3064         if (amdgpu_device_need_post(adev)) {
3065                 if (!adev->bios) {
3066                         dev_err(adev->dev, "no vBIOS found\n");
3067                         r = -EINVAL;
3068                         goto failed;
3069                 }
3070                 DRM_INFO("GPU posting now...\n");
3071                 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
3072                 if (r) {
3073                         dev_err(adev->dev, "gpu post error!\n");
3074                         goto failed;
3075                 }
3076         }
3077
3078         if (adev->is_atom_fw) {
3079                 /* Initialize clocks */
3080                 r = amdgpu_atomfirmware_get_clock_info(adev);
3081                 if (r) {
3082                         dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
3083                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
3084                         goto failed;
3085                 }
3086         } else {
3087                 /* Initialize clocks */
3088                 r = amdgpu_atombios_get_clock_info(adev);
3089                 if (r) {
3090                         dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
3091                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
3092                         goto failed;
3093                 }
3094                 /* init i2c buses */
3095                 if (!amdgpu_device_has_dc_support(adev))
3096                         amdgpu_atombios_i2c_init(adev);
3097         }
3098
3099 fence_driver_init:
3100         /* Fence driver */
3101         r = amdgpu_fence_driver_init(adev);
3102         if (r) {
3103                 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
3104                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
3105                 goto failed;
3106         }
3107
3108         /* init the mode config */
3109         drm_mode_config_init(adev->ddev);
3110
3111         r = amdgpu_device_ip_init(adev);
3112         if (r) {
3113                 /* failed in exclusive mode due to timeout */
3114                 if (amdgpu_sriov_vf(adev) &&
3115                     !amdgpu_sriov_runtime(adev) &&
3116                     amdgpu_virt_mmio_blocked(adev) &&
3117                     !amdgpu_virt_wait_reset(adev)) {
3118                         dev_err(adev->dev, "VF exclusive mode timeout\n");
3119                         /* Don't send request since VF is inactive. */
3120                         adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
3121                         adev->virt.ops = NULL;
3122                         r = -EAGAIN;
3123                         goto failed;
3124                 }
3125                 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
3126                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
3127                 goto failed;
3128         }
3129
3130         DRM_DEBUG("SE %d, SH per SE %d, CU per SH %d, active_cu_number %d\n",
3131                         adev->gfx.config.max_shader_engines,
3132                         adev->gfx.config.max_sh_per_se,
3133                         adev->gfx.config.max_cu_per_sh,
3134                         adev->gfx.cu_info.number);
3135
3136         amdgpu_ctx_init_sched(adev);
3137
3138         adev->accel_working = true;
3139
3140         amdgpu_vm_check_compute_bug(adev);
3141
3142         /* Initialize the buffer migration limit. */
3143         if (amdgpu_moverate >= 0)
3144                 max_MBps = amdgpu_moverate;
3145         else
3146                 max_MBps = 8; /* Allow 8 MB/s. */
3147         /* Get a log2 for easy divisions. */
3148         adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
3149
3150         amdgpu_fbdev_init(adev);
3151
3152         r = amdgpu_pm_sysfs_init(adev);
3153         if (r) {
3154                 adev->pm_sysfs_en = false;
3155                 DRM_ERROR("registering pm debugfs failed (%d).\n", r);
3156         } else
3157                 adev->pm_sysfs_en = true;
3158
3159         r = amdgpu_ucode_sysfs_init(adev);
3160         if (r) {
3161                 adev->ucode_sysfs_en = false;
3162                 DRM_ERROR("Creating firmware sysfs failed (%d).\n", r);
3163         } else
3164                 adev->ucode_sysfs_en = true;
3165
3166         if ((amdgpu_testing & 1)) {
3167                 if (adev->accel_working)
3168                         amdgpu_test_moves(adev);
3169                 else
3170                         DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
3171         }
3172         if (amdgpu_benchmarking) {
3173                 if (adev->accel_working)
3174                         amdgpu_benchmark(adev, amdgpu_benchmarking);
3175                 else
3176                         DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
3177         }
3178
3179         /*
3180          * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost.
3181          * Otherwise the mgpu fan boost feature will be skipped due to the
3182          * gpu instance is counted less.
3183          */
3184         amdgpu_register_gpu_instance(adev);
3185
3186         /* enable clockgating, etc. after ib tests, etc. since some blocks require
3187          * explicit gating rather than handling it automatically.
3188          */
3189         r = amdgpu_device_ip_late_init(adev);
3190         if (r) {
3191                 dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
3192                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
3193                 goto failed;
3194         }
3195
3196         /* must succeed. */
3197         amdgpu_ras_resume(adev);
3198
3199         queue_delayed_work(system_wq, &adev->delayed_init_work,
3200                            msecs_to_jiffies(AMDGPU_RESUME_MS));
3201
3202         r = device_create_file(adev->dev, &dev_attr_pcie_replay_count);
3203         if (r) {
3204                 dev_err(adev->dev, "Could not create pcie_replay_count");
3205                 return r;
3206         }
3207
3208         if (IS_ENABLED(CONFIG_PERF_EVENTS))
3209                 r = amdgpu_pmu_init(adev);
3210         if (r)
3211                 dev_err(adev->dev, "amdgpu_pmu_init failed\n");
3212
3213         return 0;
3214
3215 failed:
3216         amdgpu_vf_error_trans_all(adev);
3217         if (boco)
3218                 vga_switcheroo_fini_domain_pm_ops(adev->dev);
3219
3220         return r;
3221 }
3222
3223 /**
3224  * amdgpu_device_fini - tear down the driver
3225  *
3226  * @adev: amdgpu_device pointer
3227  *
3228  * Tear down the driver info (all asics).
3229  * Called at driver shutdown.
3230  */
3231 void amdgpu_device_fini(struct amdgpu_device *adev)
3232 {
3233         int r;
3234
3235         DRM_INFO("amdgpu: finishing device.\n");
3236         flush_delayed_work(&adev->delayed_init_work);
3237         adev->shutdown = true;
3238
3239         /* make sure IB test finished before entering exclusive mode
3240          * to avoid preemption on IB test
3241          * */
3242         if (amdgpu_sriov_vf(adev))
3243                 amdgpu_virt_request_full_gpu(adev, false);
3244
3245         /* disable all interrupts */
3246         amdgpu_irq_disable_all(adev);
3247         if (adev->mode_info.mode_config_initialized){
3248                 if (!amdgpu_device_has_dc_support(adev))
3249                         drm_helper_force_disable_all(adev->ddev);
3250                 else
3251                         drm_atomic_helper_shutdown(adev->ddev);
3252         }
3253         amdgpu_fence_driver_fini(adev);
3254         if (adev->pm_sysfs_en)
3255                 amdgpu_pm_sysfs_fini(adev);
3256         amdgpu_fbdev_fini(adev);
3257         r = amdgpu_device_ip_fini(adev);
3258         if (adev->firmware.gpu_info_fw) {
3259                 release_firmware(adev->firmware.gpu_info_fw);
3260                 adev->firmware.gpu_info_fw = NULL;
3261         }
3262         adev->accel_working = false;
3263         /* free i2c buses */
3264         if (!amdgpu_device_has_dc_support(adev))
3265                 amdgpu_i2c_fini(adev);
3266
3267         if (amdgpu_emu_mode != 1)
3268                 amdgpu_atombios_fini(adev);
3269
3270         kfree(adev->bios);
3271         adev->bios = NULL;
3272         if (amdgpu_has_atpx() &&
3273             (amdgpu_is_atpx_hybrid() ||
3274              amdgpu_has_atpx_dgpu_power_cntl()) &&
3275             !pci_is_thunderbolt_attached(adev->pdev))
3276                 vga_switcheroo_unregister_client(adev->pdev);
3277         if (amdgpu_device_supports_boco(adev->ddev))
3278                 vga_switcheroo_fini_domain_pm_ops(adev->dev);
3279         vga_client_register(adev->pdev, NULL, NULL, NULL);
3280         if (adev->rio_mem)
3281                 pci_iounmap(adev->pdev, adev->rio_mem);
3282         adev->rio_mem = NULL;
3283         iounmap(adev->rmmio);
3284         adev->rmmio = NULL;
3285         amdgpu_device_doorbell_fini(adev);
3286
3287         device_remove_file(adev->dev, &dev_attr_pcie_replay_count);
3288         if (adev->ucode_sysfs_en)
3289                 amdgpu_ucode_sysfs_fini(adev);
3290         if (IS_ENABLED(CONFIG_PERF_EVENTS))
3291                 amdgpu_pmu_fini(adev);
3292         if (amdgpu_discovery && adev->asic_type >= CHIP_NAVI10)
3293                 amdgpu_discovery_fini(adev);
3294 }
3295
3296
3297 /*
3298  * Suspend & resume.
3299  */
3300 /**
3301  * amdgpu_device_suspend - initiate device suspend
3302  *
3303  * @dev: drm dev pointer
3304  * @suspend: suspend state
3305  * @fbcon : notify the fbdev of suspend
3306  *
3307  * Puts the hw in the suspend state (all asics).
3308  * Returns 0 for success or an error on failure.
3309  * Called at driver suspend.
3310  */
3311 int amdgpu_device_suspend(struct drm_device *dev, bool fbcon)
3312 {
3313         struct amdgpu_device *adev;
3314         struct drm_crtc *crtc;
3315         struct drm_connector *connector;
3316         struct drm_connector_list_iter iter;
3317         int r;
3318
3319         if (dev == NULL || dev->dev_private == NULL) {
3320                 return -ENODEV;
3321         }
3322
3323         adev = dev->dev_private;
3324
3325         if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3326                 return 0;
3327
3328         adev->in_suspend = true;
3329         drm_kms_helper_poll_disable(dev);
3330
3331         if (fbcon)
3332                 amdgpu_fbdev_set_suspend(adev, 1);
3333
3334         cancel_delayed_work_sync(&adev->delayed_init_work);
3335
3336         if (!amdgpu_device_has_dc_support(adev)) {
3337                 /* turn off display hw */
3338                 drm_modeset_lock_all(dev);
3339                 drm_connector_list_iter_begin(dev, &iter);
3340                 drm_for_each_connector_iter(connector, &iter)
3341                         drm_helper_connector_dpms(connector,
3342                                                   DRM_MODE_DPMS_OFF);
3343                 drm_connector_list_iter_end(&iter);
3344                 drm_modeset_unlock_all(dev);
3345                         /* unpin the front buffers and cursors */
3346                 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
3347                         struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
3348                         struct drm_framebuffer *fb = crtc->primary->fb;
3349                         struct amdgpu_bo *robj;
3350
3351                         if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) {
3352                                 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
3353                                 r = amdgpu_bo_reserve(aobj, true);
3354                                 if (r == 0) {
3355                                         amdgpu_bo_unpin(aobj);
3356                                         amdgpu_bo_unreserve(aobj);
3357                                 }
3358                         }
3359
3360                         if (fb == NULL || fb->obj[0] == NULL) {
3361                                 continue;
3362                         }
3363                         robj = gem_to_amdgpu_bo(fb->obj[0]);
3364                         /* don't unpin kernel fb objects */
3365                         if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
3366                                 r = amdgpu_bo_reserve(robj, true);
3367                                 if (r == 0) {
3368                                         amdgpu_bo_unpin(robj);
3369                                         amdgpu_bo_unreserve(robj);
3370                                 }
3371                         }
3372                 }
3373         }
3374
3375         amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
3376         amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
3377
3378         amdgpu_amdkfd_suspend(adev, !fbcon);
3379
3380         amdgpu_ras_suspend(adev);
3381
3382         r = amdgpu_device_ip_suspend_phase1(adev);
3383
3384         /* evict vram memory */
3385         amdgpu_bo_evict_vram(adev);
3386
3387         amdgpu_fence_driver_suspend(adev);
3388
3389         r = amdgpu_device_ip_suspend_phase2(adev);
3390
3391         /* evict remaining vram memory
3392          * This second call to evict vram is to evict the gart page table
3393          * using the CPU.
3394          */
3395         amdgpu_bo_evict_vram(adev);
3396
3397         return 0;
3398 }
3399
3400 /**
3401  * amdgpu_device_resume - initiate device resume
3402  *
3403  * @dev: drm dev pointer
3404  * @resume: resume state
3405  * @fbcon : notify the fbdev of resume
3406  *
3407  * Bring the hw back to operating state (all asics).
3408  * Returns 0 for success or an error on failure.
3409  * Called at driver resume.
3410  */
3411 int amdgpu_device_resume(struct drm_device *dev, bool fbcon)
3412 {
3413         struct drm_connector *connector;
3414         struct drm_connector_list_iter iter;
3415         struct amdgpu_device *adev = dev->dev_private;
3416         struct drm_crtc *crtc;
3417         int r = 0;
3418
3419         if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
3420                 return 0;
3421
3422         /* post card */
3423         if (amdgpu_device_need_post(adev)) {
3424                 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
3425                 if (r)
3426                         DRM_ERROR("amdgpu asic init failed\n");
3427         }
3428
3429         r = amdgpu_device_ip_resume(adev);
3430         if (r) {
3431                 DRM_ERROR("amdgpu_device_ip_resume failed (%d).\n", r);
3432                 return r;
3433         }
3434         amdgpu_fence_driver_resume(adev);
3435
3436
3437         r = amdgpu_device_ip_late_init(adev);
3438         if (r)
3439                 return r;
3440
3441         queue_delayed_work(system_wq, &adev->delayed_init_work,
3442                            msecs_to_jiffies(AMDGPU_RESUME_MS));
3443
3444         if (!amdgpu_device_has_dc_support(adev)) {
3445                 /* pin cursors */
3446                 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
3447                         struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
3448
3449                         if (amdgpu_crtc->cursor_bo && !adev->enable_virtual_display) {
3450                                 struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
3451                                 r = amdgpu_bo_reserve(aobj, true);
3452                                 if (r == 0) {
3453                                         r = amdgpu_bo_pin(aobj, AMDGPU_GEM_DOMAIN_VRAM);
3454                                         if (r != 0)
3455                                                 DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
3456                                         amdgpu_crtc->cursor_addr = amdgpu_bo_gpu_offset(aobj);
3457                                         amdgpu_bo_unreserve(aobj);
3458                                 }
3459                         }
3460                 }
3461         }
3462         r = amdgpu_amdkfd_resume(adev, !fbcon);
3463         if (r)
3464                 return r;
3465
3466         /* Make sure IB tests flushed */
3467         flush_delayed_work(&adev->delayed_init_work);
3468
3469         /* blat the mode back in */
3470         if (fbcon) {
3471                 if (!amdgpu_device_has_dc_support(adev)) {
3472                         /* pre DCE11 */
3473                         drm_helper_resume_force_mode(dev);
3474
3475                         /* turn on display hw */
3476                         drm_modeset_lock_all(dev);
3477
3478                         drm_connector_list_iter_begin(dev, &iter);
3479                         drm_for_each_connector_iter(connector, &iter)
3480                                 drm_helper_connector_dpms(connector,
3481                                                           DRM_MODE_DPMS_ON);
3482                         drm_connector_list_iter_end(&iter);
3483
3484                         drm_modeset_unlock_all(dev);
3485                 }
3486                 amdgpu_fbdev_set_suspend(adev, 0);
3487         }
3488
3489         drm_kms_helper_poll_enable(dev);
3490
3491         amdgpu_ras_resume(adev);
3492
3493         /*
3494          * Most of the connector probing functions try to acquire runtime pm
3495          * refs to ensure that the GPU is powered on when connector polling is
3496          * performed. Since we're calling this from a runtime PM callback,
3497          * trying to acquire rpm refs will cause us to deadlock.
3498          *
3499          * Since we're guaranteed to be holding the rpm lock, it's safe to
3500          * temporarily disable the rpm helpers so this doesn't deadlock us.
3501          */
3502 #ifdef CONFIG_PM
3503         dev->dev->power.disable_depth++;
3504 #endif
3505         if (!amdgpu_device_has_dc_support(adev))
3506                 drm_helper_hpd_irq_event(dev);
3507         else
3508                 drm_kms_helper_hotplug_event(dev);
3509 #ifdef CONFIG_PM
3510         dev->dev->power.disable_depth--;
3511 #endif
3512         adev->in_suspend = false;
3513
3514         return 0;
3515 }
3516
3517 /**
3518  * amdgpu_device_ip_check_soft_reset - did soft reset succeed
3519  *
3520  * @adev: amdgpu_device pointer
3521  *
3522  * The list of all the hardware IPs that make up the asic is walked and
3523  * the check_soft_reset callbacks are run.  check_soft_reset determines
3524  * if the asic is still hung or not.
3525  * Returns true if any of the IPs are still in a hung state, false if not.
3526  */
3527 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
3528 {
3529         int i;
3530         bool asic_hang = false;
3531
3532         if (amdgpu_sriov_vf(adev))
3533                 return true;
3534
3535         if (amdgpu_asic_need_full_reset(adev))
3536                 return true;
3537
3538         for (i = 0; i < adev->num_ip_blocks; i++) {
3539                 if (!adev->ip_blocks[i].status.valid)
3540                         continue;
3541                 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
3542                         adev->ip_blocks[i].status.hang =
3543                                 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
3544                 if (adev->ip_blocks[i].status.hang) {
3545                         DRM_INFO("IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
3546                         asic_hang = true;
3547                 }
3548         }
3549         return asic_hang;
3550 }
3551
3552 /**
3553  * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
3554  *
3555  * @adev: amdgpu_device pointer
3556  *
3557  * The list of all the hardware IPs that make up the asic is walked and the
3558  * pre_soft_reset callbacks are run if the block is hung.  pre_soft_reset
3559  * handles any IP specific hardware or software state changes that are
3560  * necessary for a soft reset to succeed.
3561  * Returns 0 on success, negative error code on failure.
3562  */
3563 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
3564 {
3565         int i, r = 0;
3566
3567         for (i = 0; i < adev->num_ip_blocks; i++) {
3568                 if (!adev->ip_blocks[i].status.valid)
3569                         continue;
3570                 if (adev->ip_blocks[i].status.hang &&
3571                     adev->ip_blocks[i].version->funcs->pre_soft_reset) {
3572                         r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
3573                         if (r)
3574                                 return r;
3575                 }
3576         }
3577
3578         return 0;
3579 }
3580
3581 /**
3582  * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
3583  *
3584  * @adev: amdgpu_device pointer
3585  *
3586  * Some hardware IPs cannot be soft reset.  If they are hung, a full gpu
3587  * reset is necessary to recover.
3588  * Returns true if a full asic reset is required, false if not.
3589  */
3590 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
3591 {
3592         int i;
3593
3594         if (amdgpu_asic_need_full_reset(adev))
3595                 return true;
3596
3597         for (i = 0; i < adev->num_ip_blocks; i++) {
3598                 if (!adev->ip_blocks[i].status.valid)
3599                         continue;
3600                 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
3601                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
3602                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
3603                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
3604                      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
3605                         if (adev->ip_blocks[i].status.hang) {
3606                                 DRM_INFO("Some block need full reset!\n");
3607                                 return true;
3608                         }
3609                 }
3610         }
3611         return false;
3612 }
3613
3614 /**
3615  * amdgpu_device_ip_soft_reset - do a soft reset
3616  *
3617  * @adev: amdgpu_device pointer
3618  *
3619  * The list of all the hardware IPs that make up the asic is walked and the
3620  * soft_reset callbacks are run if the block is hung.  soft_reset handles any
3621  * IP specific hardware or software state changes that are necessary to soft
3622  * reset the IP.
3623  * Returns 0 on success, negative error code on failure.
3624  */
3625 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
3626 {
3627         int i, r = 0;
3628
3629         for (i = 0; i < adev->num_ip_blocks; i++) {
3630                 if (!adev->ip_blocks[i].status.valid)
3631                         continue;
3632                 if (adev->ip_blocks[i].status.hang &&
3633                     adev->ip_blocks[i].version->funcs->soft_reset) {
3634                         r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
3635                         if (r)
3636                                 return r;
3637                 }
3638         }
3639
3640         return 0;
3641 }
3642
3643 /**
3644  * amdgpu_device_ip_post_soft_reset - clean up from soft reset
3645  *
3646  * @adev: amdgpu_device pointer
3647  *
3648  * The list of all the hardware IPs that make up the asic is walked and the
3649  * post_soft_reset callbacks are run if the asic was hung.  post_soft_reset
3650  * handles any IP specific hardware or software state changes that are
3651  * necessary after the IP has been soft reset.
3652  * Returns 0 on success, negative error code on failure.
3653  */
3654 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
3655 {
3656         int i, r = 0;
3657
3658         for (i = 0; i < adev->num_ip_blocks; i++) {
3659                 if (!adev->ip_blocks[i].status.valid)
3660                         continue;
3661                 if (adev->ip_blocks[i].status.hang &&
3662                     adev->ip_blocks[i].version->funcs->post_soft_reset)
3663                         r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
3664                 if (r)
3665                         return r;
3666         }
3667
3668         return 0;
3669 }
3670
3671 /**
3672  * amdgpu_device_recover_vram - Recover some VRAM contents
3673  *
3674  * @adev: amdgpu_device pointer
3675  *
3676  * Restores the contents of VRAM buffers from the shadows in GTT.  Used to
3677  * restore things like GPUVM page tables after a GPU reset where
3678  * the contents of VRAM might be lost.
3679  *
3680  * Returns:
3681  * 0 on success, negative error code on failure.
3682  */
3683 static int amdgpu_device_recover_vram(struct amdgpu_device *adev)
3684 {
3685         struct dma_fence *fence = NULL, *next = NULL;
3686         struct amdgpu_bo *shadow;
3687         long r = 1, tmo;
3688
3689         if (amdgpu_sriov_runtime(adev))
3690                 tmo = msecs_to_jiffies(8000);
3691         else
3692                 tmo = msecs_to_jiffies(100);
3693
3694         DRM_INFO("recover vram bo from shadow start\n");
3695         mutex_lock(&adev->shadow_list_lock);
3696         list_for_each_entry(shadow, &adev->shadow_list, shadow_list) {
3697
3698                 /* No need to recover an evicted BO */
3699                 if (shadow->tbo.mem.mem_type != TTM_PL_TT ||
3700                     shadow->tbo.mem.start == AMDGPU_BO_INVALID_OFFSET ||
3701                     shadow->parent->tbo.mem.mem_type != TTM_PL_VRAM)
3702                         continue;
3703
3704                 r = amdgpu_bo_restore_shadow(shadow, &next);
3705                 if (r)
3706                         break;
3707
3708                 if (fence) {
3709                         tmo = dma_fence_wait_timeout(fence, false, tmo);
3710                         dma_fence_put(fence);
3711                         fence = next;
3712                         if (tmo == 0) {
3713                                 r = -ETIMEDOUT;
3714                                 break;
3715                         } else if (tmo < 0) {
3716                                 r = tmo;
3717                                 break;
3718                         }
3719                 } else {
3720                         fence = next;
3721                 }
3722         }
3723         mutex_unlock(&adev->shadow_list_lock);
3724
3725         if (fence)
3726                 tmo = dma_fence_wait_timeout(fence, false, tmo);
3727         dma_fence_put(fence);
3728
3729         if (r < 0 || tmo <= 0) {
3730                 DRM_ERROR("recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo);
3731                 return -EIO;
3732         }
3733
3734         DRM_INFO("recover vram bo from shadow done\n");
3735         return 0;
3736 }
3737
3738
3739 /**
3740  * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
3741  *
3742  * @adev: amdgpu device pointer
3743  * @from_hypervisor: request from hypervisor
3744  *
3745  * do VF FLR and reinitialize Asic
3746  * return 0 means succeeded otherwise failed
3747  */
3748 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
3749                                      bool from_hypervisor)
3750 {
3751         int r;
3752
3753         if (from_hypervisor)
3754                 r = amdgpu_virt_request_full_gpu(adev, true);
3755         else
3756                 r = amdgpu_virt_reset_gpu(adev);
3757         if (r)
3758                 return r;
3759
3760         /* Resume IP prior to SMC */
3761         r = amdgpu_device_ip_reinit_early_sriov(adev);
3762         if (r)
3763                 goto error;
3764
3765         amdgpu_virt_init_data_exchange(adev);
3766         /* we need recover gart prior to run SMC/CP/SDMA resume */
3767         amdgpu_gtt_mgr_recover(&adev->mman.bdev.man[TTM_PL_TT]);
3768
3769         r = amdgpu_device_fw_loading(adev);
3770         if (r)
3771                 return r;
3772
3773         /* now we are okay to resume SMC/CP/SDMA */
3774         r = amdgpu_device_ip_reinit_late_sriov(adev);
3775         if (r)
3776                 goto error;
3777
3778         amdgpu_irq_gpu_reset_resume_helper(adev);
3779         r = amdgpu_ib_ring_tests(adev);
3780         amdgpu_amdkfd_post_reset(adev);
3781
3782 error:
3783         amdgpu_virt_release_full_gpu(adev, true);
3784         if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
3785                 amdgpu_inc_vram_lost(adev);
3786                 r = amdgpu_device_recover_vram(adev);
3787         }
3788
3789         return r;
3790 }
3791
3792 /**
3793  * amdgpu_device_should_recover_gpu - check if we should try GPU recovery
3794  *
3795  * @adev: amdgpu device pointer
3796  *
3797  * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover
3798  * a hung GPU.
3799  */
3800 bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev)
3801 {
3802         if (!amdgpu_device_ip_check_soft_reset(adev)) {
3803                 DRM_INFO("Timeout, but no hardware hang detected.\n");
3804                 return false;
3805         }
3806
3807         if (amdgpu_gpu_recovery == 0)
3808                 goto disabled;
3809
3810         if (amdgpu_sriov_vf(adev))
3811                 return true;
3812
3813         if (amdgpu_gpu_recovery == -1) {
3814                 switch (adev->asic_type) {
3815                 case CHIP_BONAIRE:
3816                 case CHIP_HAWAII:
3817                 case CHIP_TOPAZ:
3818                 case CHIP_TONGA:
3819                 case CHIP_FIJI:
3820                 case CHIP_POLARIS10:
3821                 case CHIP_POLARIS11:
3822                 case CHIP_POLARIS12:
3823                 case CHIP_VEGAM:
3824                 case CHIP_VEGA20:
3825                 case CHIP_VEGA10:
3826                 case CHIP_VEGA12:
3827                 case CHIP_RAVEN:
3828                 case CHIP_ARCTURUS:
3829                 case CHIP_RENOIR:
3830                 case CHIP_NAVI10:
3831                 case CHIP_NAVI14:
3832                 case CHIP_NAVI12:
3833                         break;
3834                 default:
3835                         goto disabled;
3836                 }
3837         }
3838
3839         return true;
3840
3841 disabled:
3842                 DRM_INFO("GPU recovery disabled.\n");
3843                 return false;
3844 }
3845
3846
3847 static int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
3848                                         struct amdgpu_job *job,
3849                                         bool *need_full_reset_arg)
3850 {
3851         int i, r = 0;
3852         bool need_full_reset  = *need_full_reset_arg;
3853
3854         /* block all schedulers and reset given job's ring */
3855         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
3856                 struct amdgpu_ring *ring = adev->rings[i];
3857
3858                 if (!ring || !ring->sched.thread)
3859                         continue;
3860
3861                 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
3862                 amdgpu_fence_driver_force_completion(ring);
3863         }
3864
3865         if(job)
3866                 drm_sched_increase_karma(&job->base);
3867
3868         /* Don't suspend on bare metal if we are not going to HW reset the ASIC */
3869         if (!amdgpu_sriov_vf(adev)) {
3870
3871                 if (!need_full_reset)
3872                         need_full_reset = amdgpu_device_ip_need_full_reset(adev);
3873
3874                 if (!need_full_reset) {
3875                         amdgpu_device_ip_pre_soft_reset(adev);
3876                         r = amdgpu_device_ip_soft_reset(adev);
3877                         amdgpu_device_ip_post_soft_reset(adev);
3878                         if (r || amdgpu_device_ip_check_soft_reset(adev)) {
3879                                 DRM_INFO("soft reset failed, will fallback to full reset!\n");
3880                                 need_full_reset = true;
3881                         }
3882                 }
3883
3884                 if (need_full_reset)
3885                         r = amdgpu_device_ip_suspend(adev);
3886
3887                 *need_full_reset_arg = need_full_reset;
3888         }
3889
3890         return r;
3891 }
3892
3893 static int amdgpu_do_asic_reset(struct amdgpu_hive_info *hive,
3894                                struct list_head *device_list_handle,
3895                                bool *need_full_reset_arg)
3896 {
3897         struct amdgpu_device *tmp_adev = NULL;
3898         bool need_full_reset = *need_full_reset_arg, vram_lost = false;
3899         int r = 0;
3900
3901         /*
3902          * ASIC reset has to be done on all HGMI hive nodes ASAP
3903          * to allow proper links negotiation in FW (within 1 sec)
3904          */
3905         if (need_full_reset) {
3906                 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3907                         /* For XGMI run all resets in parallel to speed up the process */
3908                         if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
3909                                 if (!queue_work(system_unbound_wq, &tmp_adev->xgmi_reset_work))
3910                                         r = -EALREADY;
3911                         } else
3912                                 r = amdgpu_asic_reset(tmp_adev);
3913
3914                         if (r) {
3915                                 DRM_ERROR("ASIC reset failed with error, %d for drm dev, %s",
3916                                          r, tmp_adev->ddev->unique);
3917                                 break;
3918                         }
3919                 }
3920
3921                 /* For XGMI wait for all resets to complete before proceed */
3922                 if (!r) {
3923                         list_for_each_entry(tmp_adev, device_list_handle,
3924                                             gmc.xgmi.head) {
3925                                 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
3926                                         flush_work(&tmp_adev->xgmi_reset_work);
3927                                         r = tmp_adev->asic_reset_res;
3928                                         if (r)
3929                                                 break;
3930                                 }
3931                         }
3932                 }
3933         }
3934
3935         if (!r && amdgpu_ras_intr_triggered()) {
3936                 list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3937                         if (tmp_adev->mmhub.funcs &&
3938                             tmp_adev->mmhub.funcs->reset_ras_error_count)
3939                                 tmp_adev->mmhub.funcs->reset_ras_error_count(tmp_adev);
3940                 }
3941
3942                 amdgpu_ras_intr_cleared();
3943         }
3944
3945         list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
3946                 if (need_full_reset) {
3947                         /* post card */
3948                         if (amdgpu_atom_asic_init(tmp_adev->mode_info.atom_context))
3949                                 DRM_WARN("asic atom init failed!");
3950
3951                         if (!r) {
3952                                 dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n");
3953                                 r = amdgpu_device_ip_resume_phase1(tmp_adev);
3954                                 if (r)
3955                                         goto out;
3956
3957                                 vram_lost = amdgpu_device_check_vram_lost(tmp_adev);
3958                                 if (vram_lost) {
3959                                         DRM_INFO("VRAM is lost due to GPU reset!\n");
3960                                         amdgpu_inc_vram_lost(tmp_adev);
3961                                 }
3962
3963                                 r = amdgpu_gtt_mgr_recover(
3964                                         &tmp_adev->mman.bdev.man[TTM_PL_TT]);
3965                                 if (r)
3966                                         goto out;
3967
3968                                 r = amdgpu_device_fw_loading(tmp_adev);
3969                                 if (r)
3970                                         return r;
3971
3972                                 r = amdgpu_device_ip_resume_phase2(tmp_adev);
3973                                 if (r)
3974                                         goto out;
3975
3976                                 if (vram_lost)
3977                                         amdgpu_device_fill_reset_magic(tmp_adev);
3978
3979                                 /*
3980                                  * Add this ASIC as tracked as reset was already
3981                                  * complete successfully.
3982                                  */
3983                                 amdgpu_register_gpu_instance(tmp_adev);
3984
3985                                 r = amdgpu_device_ip_late_init(tmp_adev);
3986                                 if (r)
3987                                         goto out;
3988
3989                                 amdgpu_fbdev_set_suspend(tmp_adev, 0);
3990
3991                                 /* must succeed. */
3992                                 amdgpu_ras_resume(tmp_adev);
3993
3994                                 /* Update PSP FW topology after reset */
3995                                 if (hive && tmp_adev->gmc.xgmi.num_physical_nodes > 1)
3996                                         r = amdgpu_xgmi_update_topology(hive, tmp_adev);
3997                         }
3998                 }
3999
4000
4001 out:
4002                 if (!r) {
4003                         amdgpu_irq_gpu_reset_resume_helper(tmp_adev);
4004                         r = amdgpu_ib_ring_tests(tmp_adev);
4005                         if (r) {
4006                                 dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r);
4007                                 r = amdgpu_device_ip_suspend(tmp_adev);
4008                                 need_full_reset = true;
4009                                 r = -EAGAIN;
4010                                 goto end;
4011                         }
4012                 }
4013
4014                 if (!r)
4015                         r = amdgpu_device_recover_vram(tmp_adev);
4016                 else
4017                         tmp_adev->asic_reset_res = r;
4018         }
4019
4020 end:
4021         *need_full_reset_arg = need_full_reset;
4022         return r;
4023 }
4024
4025 static bool amdgpu_device_lock_adev(struct amdgpu_device *adev, bool trylock)
4026 {
4027         if (trylock) {
4028                 if (!mutex_trylock(&adev->lock_reset))
4029                         return false;
4030         } else
4031                 mutex_lock(&adev->lock_reset);
4032
4033         atomic_inc(&adev->gpu_reset_counter);
4034         adev->in_gpu_reset = true;
4035         switch (amdgpu_asic_reset_method(adev)) {
4036         case AMD_RESET_METHOD_MODE1:
4037                 adev->mp1_state = PP_MP1_STATE_SHUTDOWN;
4038                 break;
4039         case AMD_RESET_METHOD_MODE2:
4040                 adev->mp1_state = PP_MP1_STATE_RESET;
4041                 break;
4042         default:
4043                 adev->mp1_state = PP_MP1_STATE_NONE;
4044                 break;
4045         }
4046
4047         return true;
4048 }
4049
4050 static void amdgpu_device_unlock_adev(struct amdgpu_device *adev)
4051 {
4052         amdgpu_vf_error_trans_all(adev);
4053         adev->mp1_state = PP_MP1_STATE_NONE;
4054         adev->in_gpu_reset = false;
4055         mutex_unlock(&adev->lock_reset);
4056 }
4057
4058 /**
4059  * amdgpu_device_gpu_recover - reset the asic and recover scheduler
4060  *
4061  * @adev: amdgpu device pointer
4062  * @job: which job trigger hang
4063  *
4064  * Attempt to reset the GPU if it has hung (all asics).
4065  * Attempt to do soft-reset or full-reset and reinitialize Asic
4066  * Returns 0 for success or an error on failure.
4067  */
4068
4069 int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
4070                               struct amdgpu_job *job)
4071 {
4072         struct list_head device_list, *device_list_handle =  NULL;
4073         bool need_full_reset, job_signaled;
4074         struct amdgpu_hive_info *hive = NULL;
4075         struct amdgpu_device *tmp_adev = NULL;
4076         int i, r = 0;
4077         bool in_ras_intr = amdgpu_ras_intr_triggered();
4078         bool use_baco =
4079                 (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) ?
4080                 true : false;
4081
4082         /*
4083          * Flush RAM to disk so that after reboot
4084          * the user can read log and see why the system rebooted.
4085          */
4086         if (in_ras_intr && !use_baco && amdgpu_ras_get_context(adev)->reboot) {
4087
4088                 DRM_WARN("Emergency reboot.");
4089
4090                 ksys_sync_helper();
4091                 emergency_restart();
4092         }
4093
4094         need_full_reset = job_signaled = false;
4095         INIT_LIST_HEAD(&device_list);
4096
4097         dev_info(adev->dev, "GPU %s begin!\n",
4098                 (in_ras_intr && !use_baco) ? "jobs stop":"reset");
4099
4100         cancel_delayed_work_sync(&adev->delayed_init_work);
4101
4102         hive = amdgpu_get_xgmi_hive(adev, false);
4103
4104         /*
4105          * Here we trylock to avoid chain of resets executing from
4106          * either trigger by jobs on different adevs in XGMI hive or jobs on
4107          * different schedulers for same device while this TO handler is running.
4108          * We always reset all schedulers for device and all devices for XGMI
4109          * hive so that should take care of them too.
4110          */
4111
4112         if (hive && !mutex_trylock(&hive->reset_lock)) {
4113                 DRM_INFO("Bailing on TDR for s_job:%llx, hive: %llx as another already in progress",
4114                           job ? job->base.id : -1, hive->hive_id);
4115                 return 0;
4116         }
4117
4118         /* Start with adev pre asic reset first for soft reset check.*/
4119         if (!amdgpu_device_lock_adev(adev, !hive)) {
4120                 DRM_INFO("Bailing on TDR for s_job:%llx, as another already in progress",
4121                           job ? job->base.id : -1);
4122                 return 0;
4123         }
4124
4125         /* Block kfd: SRIOV would do it separately */
4126         if (!amdgpu_sriov_vf(adev))
4127                 amdgpu_amdkfd_pre_reset(adev);
4128
4129         /* Build list of devices to reset */
4130         if  (adev->gmc.xgmi.num_physical_nodes > 1) {
4131                 if (!hive) {
4132                         /*unlock kfd: SRIOV would do it separately */
4133                         if (!amdgpu_sriov_vf(adev))
4134                                 amdgpu_amdkfd_post_reset(adev);
4135                         amdgpu_device_unlock_adev(adev);
4136                         return -ENODEV;
4137                 }
4138
4139                 /*
4140                  * In case we are in XGMI hive mode device reset is done for all the
4141                  * nodes in the hive to retrain all XGMI links and hence the reset
4142                  * sequence is executed in loop on all nodes.
4143                  */
4144                 device_list_handle = &hive->device_list;
4145         } else {
4146                 list_add_tail(&adev->gmc.xgmi.head, &device_list);
4147                 device_list_handle = &device_list;
4148         }
4149
4150         /* block all schedulers and reset given job's ring */
4151         list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4152                 if (tmp_adev != adev) {
4153                         amdgpu_device_lock_adev(tmp_adev, false);
4154                         if (!amdgpu_sriov_vf(tmp_adev))
4155                                         amdgpu_amdkfd_pre_reset(tmp_adev);
4156                 }
4157
4158                 /*
4159                  * Mark these ASICs to be reseted as untracked first
4160                  * And add them back after reset completed
4161                  */
4162                 amdgpu_unregister_gpu_instance(tmp_adev);
4163
4164                 amdgpu_fbdev_set_suspend(adev, 1);
4165
4166                 /* disable ras on ALL IPs */
4167                 if (!(in_ras_intr && !use_baco) &&
4168                       amdgpu_device_ip_need_full_reset(tmp_adev))
4169                         amdgpu_ras_suspend(tmp_adev);
4170
4171                 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4172                         struct amdgpu_ring *ring = tmp_adev->rings[i];
4173
4174                         if (!ring || !ring->sched.thread)
4175                                 continue;
4176
4177                         drm_sched_stop(&ring->sched, job ? &job->base : NULL);
4178
4179                         if (in_ras_intr && !use_baco)
4180                                 amdgpu_job_stop_all_jobs_on_sched(&ring->sched);
4181                 }
4182         }
4183
4184
4185         if (in_ras_intr && !use_baco)
4186                 goto skip_sched_resume;
4187
4188         /*
4189          * Must check guilty signal here since after this point all old
4190          * HW fences are force signaled.
4191          *
4192          * job->base holds a reference to parent fence
4193          */
4194         if (job && job->base.s_fence->parent &&
4195             dma_fence_is_signaled(job->base.s_fence->parent))
4196                 job_signaled = true;
4197
4198         if (job_signaled) {
4199                 dev_info(adev->dev, "Guilty job already signaled, skipping HW reset");
4200                 goto skip_hw_reset;
4201         }
4202
4203
4204         /* Guilty job will be freed after this*/
4205         r = amdgpu_device_pre_asic_reset(adev, job, &need_full_reset);
4206         if (r) {
4207                 /*TODO Should we stop ?*/
4208                 DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ",
4209                           r, adev->ddev->unique);
4210                 adev->asic_reset_res = r;
4211         }
4212
4213 retry:  /* Rest of adevs pre asic reset from XGMI hive. */
4214         list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4215
4216                 if (tmp_adev == adev)
4217                         continue;
4218
4219                 r = amdgpu_device_pre_asic_reset(tmp_adev,
4220                                                  NULL,
4221                                                  &need_full_reset);
4222                 /*TODO Should we stop ?*/
4223                 if (r) {
4224                         DRM_ERROR("GPU pre asic reset failed with err, %d for drm dev, %s ",
4225                                   r, tmp_adev->ddev->unique);
4226                         tmp_adev->asic_reset_res = r;
4227                 }
4228         }
4229
4230         /* Actual ASIC resets if needed.*/
4231         /* TODO Implement XGMI hive reset logic for SRIOV */
4232         if (amdgpu_sriov_vf(adev)) {
4233                 r = amdgpu_device_reset_sriov(adev, job ? false : true);
4234                 if (r)
4235                         adev->asic_reset_res = r;
4236         } else {
4237                 r  = amdgpu_do_asic_reset(hive, device_list_handle, &need_full_reset);
4238                 if (r && r == -EAGAIN)
4239                         goto retry;
4240         }
4241
4242 skip_hw_reset:
4243
4244         /* Post ASIC reset for all devs .*/
4245         list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4246
4247                 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4248                         struct amdgpu_ring *ring = tmp_adev->rings[i];
4249
4250                         if (!ring || !ring->sched.thread)
4251                                 continue;
4252
4253                         /* No point to resubmit jobs if we didn't HW reset*/
4254                         if (!tmp_adev->asic_reset_res && !job_signaled)
4255                                 drm_sched_resubmit_jobs(&ring->sched);
4256
4257                         drm_sched_start(&ring->sched, !tmp_adev->asic_reset_res);
4258                 }
4259
4260                 if (!amdgpu_device_has_dc_support(tmp_adev) && !job_signaled) {
4261                         drm_helper_resume_force_mode(tmp_adev->ddev);
4262                 }
4263
4264                 tmp_adev->asic_reset_res = 0;
4265
4266                 if (r) {
4267                         /* bad news, how to tell it to userspace ? */
4268                         dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&tmp_adev->gpu_reset_counter));
4269                         amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
4270                 } else {
4271                         dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&tmp_adev->gpu_reset_counter));
4272                 }
4273         }
4274
4275 skip_sched_resume:
4276         list_for_each_entry(tmp_adev, device_list_handle, gmc.xgmi.head) {
4277                 /*unlock kfd: SRIOV would do it separately */
4278                 if (!(in_ras_intr && !use_baco) && !amdgpu_sriov_vf(tmp_adev))
4279                         amdgpu_amdkfd_post_reset(tmp_adev);
4280                 amdgpu_device_unlock_adev(tmp_adev);
4281         }
4282
4283         if (hive)
4284                 mutex_unlock(&hive->reset_lock);
4285
4286         if (r)
4287                 dev_info(adev->dev, "GPU reset end with ret = %d\n", r);
4288         return r;
4289 }
4290
4291 /**
4292  * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
4293  *
4294  * @adev: amdgpu_device pointer
4295  *
4296  * Fetchs and stores in the driver the PCIE capabilities (gen speed
4297  * and lanes) of the slot the device is in. Handles APUs and
4298  * virtualized environments where PCIE config space may not be available.
4299  */
4300 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
4301 {
4302         struct pci_dev *pdev;
4303         enum pci_bus_speed speed_cap, platform_speed_cap;
4304         enum pcie_link_width platform_link_width;
4305
4306         if (amdgpu_pcie_gen_cap)
4307                 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
4308
4309         if (amdgpu_pcie_lane_cap)
4310                 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
4311
4312         /* covers APUs as well */
4313         if (pci_is_root_bus(adev->pdev->bus)) {
4314                 if (adev->pm.pcie_gen_mask == 0)
4315                         adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
4316                 if (adev->pm.pcie_mlw_mask == 0)
4317                         adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
4318                 return;
4319         }
4320
4321         if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask)
4322                 return;
4323
4324         pcie_bandwidth_available(adev->pdev, NULL,
4325                                  &platform_speed_cap, &platform_link_width);
4326
4327         if (adev->pm.pcie_gen_mask == 0) {
4328                 /* asic caps */
4329                 pdev = adev->pdev;
4330                 speed_cap = pcie_get_speed_cap(pdev);
4331                 if (speed_cap == PCI_SPEED_UNKNOWN) {
4332                         adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4333                                                   CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4334                                                   CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
4335                 } else {
4336                         if (speed_cap == PCIE_SPEED_16_0GT)
4337                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4338                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4339                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
4340                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
4341                         else if (speed_cap == PCIE_SPEED_8_0GT)
4342                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4343                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4344                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
4345                         else if (speed_cap == PCIE_SPEED_5_0GT)
4346                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4347                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
4348                         else
4349                                 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
4350                 }
4351                 /* platform caps */
4352                 if (platform_speed_cap == PCI_SPEED_UNKNOWN) {
4353                         adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4354                                                    CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
4355                 } else {
4356                         if (platform_speed_cap == PCIE_SPEED_16_0GT)
4357                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4358                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4359                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
4360                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
4361                         else if (platform_speed_cap == PCIE_SPEED_8_0GT)
4362                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4363                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
4364                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
4365                         else if (platform_speed_cap == PCIE_SPEED_5_0GT)
4366                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
4367                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
4368                         else
4369                                 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
4370
4371                 }
4372         }
4373         if (adev->pm.pcie_mlw_mask == 0) {
4374                 if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) {
4375                         adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
4376                 } else {
4377                         switch (platform_link_width) {
4378                         case PCIE_LNK_X32:
4379                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
4380                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
4381                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4382                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4383                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4384                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4385                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4386                                 break;
4387                         case PCIE_LNK_X16:
4388                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
4389                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4390                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4391                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4392                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4393                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4394                                 break;
4395                         case PCIE_LNK_X12:
4396                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
4397                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4398                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4399                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4400                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4401                                 break;
4402                         case PCIE_LNK_X8:
4403                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
4404                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4405                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4406                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4407                                 break;
4408                         case PCIE_LNK_X4:
4409                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
4410                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4411                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4412                                 break;
4413                         case PCIE_LNK_X2:
4414                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
4415                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
4416                                 break;
4417                         case PCIE_LNK_X1:
4418                                 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
4419                                 break;
4420                         default:
4421                                 break;
4422                         }
4423                 }
4424         }
4425 }
4426
4427 int amdgpu_device_baco_enter(struct drm_device *dev)
4428 {
4429         struct amdgpu_device *adev = dev->dev_private;
4430         struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
4431
4432         if (!amdgpu_device_supports_baco(adev->ddev))
4433                 return -ENOTSUPP;
4434
4435         if (ras && ras->supported)
4436                 adev->nbio.funcs->enable_doorbell_interrupt(adev, false);
4437
4438         return amdgpu_dpm_baco_enter(adev);
4439 }
4440
4441 int amdgpu_device_baco_exit(struct drm_device *dev)
4442 {
4443         struct amdgpu_device *adev = dev->dev_private;
4444         struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
4445         int ret = 0;
4446
4447         if (!amdgpu_device_supports_baco(adev->ddev))
4448                 return -ENOTSUPP;
4449
4450         ret = amdgpu_dpm_baco_exit(adev);
4451         if (ret)
4452                 return ret;
4453
4454         if (ras && ras->supported)
4455                 adev->nbio.funcs->enable_doorbell_interrupt(adev, true);
4456
4457         return 0;
4458 }