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 #include <linux/iommu.h>
  34 #include <linux/pci.h>
  35 #include <linux/devcoredump.h>
  36 #include <generated/utsrelease.h>
  37 #include <linux/pci-p2pdma.h>
  38 #include <linux/apple-gmux.h>
  39
  40 #include <drm/drm_aperture.h>
  41 #include <drm/drm_atomic_helper.h>
  42 #include <drm/drm_crtc_helper.h>
  43 #include <drm/drm_fb_helper.h>
  44 #include <drm/drm_probe_helper.h>
  45 #include <drm/amdgpu_drm.h>
  46 #include <linux/vgaarb.h>
  47 #include <linux/vga_switcheroo.h>
  48 #include <linux/efi.h>
  49 #include "amdgpu.h"
  50 #include "amdgpu_trace.h"
  51 #include "amdgpu_i2c.h"
  52 #include "atom.h"
  53 #include "amdgpu_atombios.h"
  54 #include "amdgpu_atomfirmware.h"
  55 #include "amd_pcie.h"
  56 #ifdef CONFIG_DRM_AMDGPU_SI
  57 #include "si.h"
  58 #endif
  59 #ifdef CONFIG_DRM_AMDGPU_CIK
  60 #include "cik.h"
  61 #endif
  62 #include "vi.h"
  63 #include "soc15.h"
  64 #include "nv.h"
  65 #include "bif/bif_4_1_d.h"
  66 #include <linux/firmware.h>
  67 #include "amdgpu_vf_error.h"
  68
  69 #include "amdgpu_amdkfd.h"
  70 #include "amdgpu_pm.h"
  71
  72 #include "amdgpu_xgmi.h"
  73 #include "amdgpu_ras.h"
  74 #include "amdgpu_pmu.h"
  75 #include "amdgpu_fru_eeprom.h"
  76 #include "amdgpu_reset.h"
  77
  78 #include <linux/suspend.h>
  79 #include <drm/task_barrier.h>
  80 #include <linux/pm_runtime.h>
  81
  82 #include <drm/drm_drv.h>
  83
  84 #if IS_ENABLED(CONFIG_X86)
  85 #include <asm/intel-family.h>
  86 #endif
  87
  88 MODULE_FIRMWARE("amdgpu/vega10_gpu_info.bin");
  89 MODULE_FIRMWARE("amdgpu/vega12_gpu_info.bin");
  90 MODULE_FIRMWARE("amdgpu/raven_gpu_info.bin");
  91 MODULE_FIRMWARE("amdgpu/picasso_gpu_info.bin");
  92 MODULE_FIRMWARE("amdgpu/raven2_gpu_info.bin");
  93 MODULE_FIRMWARE("amdgpu/arcturus_gpu_info.bin");
  94 MODULE_FIRMWARE("amdgpu/navi12_gpu_info.bin");
  95
  96 #define AMDGPU_RESUME_MS                2000
  97 #define AMDGPU_MAX_RETRY_LIMIT          2
  98 #define AMDGPU_RETRY_SRIOV_RESET(r) ((r) == -EBUSY || (r) == -ETIMEDOUT || (r) == -EINVAL)
  99
 100 static const struct drm_driver amdgpu_kms_driver;
 101
 102 const char *amdgpu_asic_name[] = {
 103         "TAHITI",
 104         "PITCAIRN",
 105         "VERDE",
 106         "OLAND",
 107         "HAINAN",
 108         "BONAIRE",
 109         "KAVERI",
 110         "KABINI",
 111         "HAWAII",
 112         "MULLINS",
 113         "TOPAZ",
 114         "TONGA",
 115         "FIJI",
 116         "CARRIZO",
 117         "STONEY",
 118         "POLARIS10",
 119         "POLARIS11",
 120         "POLARIS12",
 121         "VEGAM",
 122         "VEGA10",
 123         "VEGA12",
 124         "VEGA20",
 125         "RAVEN",
 126         "ARCTURUS",
 127         "RENOIR",
 128         "ALDEBARAN",
 129         "NAVI10",
 130         "CYAN_SKILLFISH",
 131         "NAVI14",
 132         "NAVI12",
 133         "SIENNA_CICHLID",
 134         "NAVY_FLOUNDER",
 135         "VANGOGH",
 136         "DIMGREY_CAVEFISH",
 137         "BEIGE_GOBY",
 138         "YELLOW_CARP",
 139         "IP DISCOVERY",
 140         "LAST",
 141 };
 142
 143 /**
 144  * DOC: pcie_replay_count
 145  *
 146  * The amdgpu driver provides a sysfs API for reporting the total number
 147  * of PCIe replays (NAKs)
 148  * The file pcie_replay_count is used for this and returns the total
 149  * number of replays as a sum of the NAKs generated and NAKs received
 150  */
 151
 152 static ssize_t amdgpu_device_get_pcie_replay_count(struct device *dev,
 153                 struct device_attribute *attr, char *buf)
 154 {
 155         struct drm_device *ddev = dev_get_drvdata(dev);
 156         struct amdgpu_device *adev = drm_to_adev(ddev);
 157         uint64_t cnt = amdgpu_asic_get_pcie_replay_count(adev);
 158
 159         return sysfs_emit(buf, "%llu\n", cnt);
 160 }
 161
 162 static DEVICE_ATTR(pcie_replay_count, S_IRUGO,
 163                 amdgpu_device_get_pcie_replay_count, NULL);
 164
 165 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev);
 166
 167 /**
 168  * DOC: product_name
 169  *
 170  * The amdgpu driver provides a sysfs API for reporting the product name
 171  * for the device
 172  * The file product_name is used for this and returns the product name
 173  * as returned from the FRU.
 174  * NOTE: This is only available for certain server cards
 175  */
 176
 177 static ssize_t amdgpu_device_get_product_name(struct device *dev,
 178                 struct device_attribute *attr, char *buf)
 179 {
 180         struct drm_device *ddev = dev_get_drvdata(dev);
 181         struct amdgpu_device *adev = drm_to_adev(ddev);
 182
 183         return sysfs_emit(buf, "%s\n", adev->product_name);
 184 }
 185
 186 static DEVICE_ATTR(product_name, S_IRUGO,
 187                 amdgpu_device_get_product_name, NULL);
 188
 189 /**
 190  * DOC: product_number
 191  *
 192  * The amdgpu driver provides a sysfs API for reporting the part number
 193  * for the device
 194  * The file product_number is used for this and returns the part number
 195  * as returned from the FRU.
 196  * NOTE: This is only available for certain server cards
 197  */
 198
 199 static ssize_t amdgpu_device_get_product_number(struct device *dev,
 200                 struct device_attribute *attr, char *buf)
 201 {
 202         struct drm_device *ddev = dev_get_drvdata(dev);
 203         struct amdgpu_device *adev = drm_to_adev(ddev);
 204
 205         return sysfs_emit(buf, "%s\n", adev->product_number);
 206 }
 207
 208 static DEVICE_ATTR(product_number, S_IRUGO,
 209                 amdgpu_device_get_product_number, NULL);
 210
 211 /**
 212  * DOC: serial_number
 213  *
 214  * The amdgpu driver provides a sysfs API for reporting the serial number
 215  * for the device
 216  * The file serial_number is used for this and returns the serial number
 217  * as returned from the FRU.
 218  * NOTE: This is only available for certain server cards
 219  */
 220
 221 static ssize_t amdgpu_device_get_serial_number(struct device *dev,
 222                 struct device_attribute *attr, char *buf)
 223 {
 224         struct drm_device *ddev = dev_get_drvdata(dev);
 225         struct amdgpu_device *adev = drm_to_adev(ddev);
 226
 227         return sysfs_emit(buf, "%s\n", adev->serial);
 228 }
 229
 230 static DEVICE_ATTR(serial_number, S_IRUGO,
 231                 amdgpu_device_get_serial_number, NULL);
 232
 233 /**
 234  * amdgpu_device_supports_px - Is the device a dGPU with ATPX power control
 235  *
 236  * @dev: drm_device pointer
 237  *
 238  * Returns true if the device is a dGPU with ATPX power control,
 239  * otherwise return false.
 240  */
 241 bool amdgpu_device_supports_px(struct drm_device *dev)
 242 {
 243         struct amdgpu_device *adev = drm_to_adev(dev);
 244
 245         if ((adev->flags & AMD_IS_PX) && !amdgpu_is_atpx_hybrid())
 246                 return true;
 247         return false;
 248 }
 249
 250 /**
 251  * amdgpu_device_supports_boco - Is the device a dGPU with ACPI power resources
 252  *
 253  * @dev: drm_device pointer
 254  *
 255  * Returns true if the device is a dGPU with ACPI power control,
 256  * otherwise return false.
 257  */
 258 bool amdgpu_device_supports_boco(struct drm_device *dev)
 259 {
 260         struct amdgpu_device *adev = drm_to_adev(dev);
 261
 262         if (adev->has_pr3 ||
 263             ((adev->flags & AMD_IS_PX) && amdgpu_is_atpx_hybrid()))
 264                 return true;
 265         return false;
 266 }
 267
 268 /**
 269  * amdgpu_device_supports_baco - Does the device support BACO
 270  *
 271  * @dev: drm_device pointer
 272  *
 273  * Returns true if the device supporte BACO,
 274  * otherwise return false.
 275  */
 276 bool amdgpu_device_supports_baco(struct drm_device *dev)
 277 {
 278         struct amdgpu_device *adev = drm_to_adev(dev);
 279
 280         return amdgpu_asic_supports_baco(adev);
 281 }
 282
 283 /**
 284  * amdgpu_device_supports_smart_shift - Is the device dGPU with
 285  * smart shift support
 286  *
 287  * @dev: drm_device pointer
 288  *
 289  * Returns true if the device is a dGPU with Smart Shift support,
 290  * otherwise returns false.
 291  */
 292 bool amdgpu_device_supports_smart_shift(struct drm_device *dev)
 293 {
 294         return (amdgpu_device_supports_boco(dev) &&
 295                 amdgpu_acpi_is_power_shift_control_supported());
 296 }
 297
 298 /*
 299  * VRAM access helper functions
 300  */
 301
 302 /**
 303  * amdgpu_device_mm_access - access vram by MM_INDEX/MM_DATA
 304  *
 305  * @adev: amdgpu_device pointer
 306  * @pos: offset of the buffer in vram
 307  * @buf: virtual address of the buffer in system memory
 308  * @size: read/write size, sizeof(@buf) must > @size
 309  * @write: true - write to vram, otherwise - read from vram
 310  */
 311 void amdgpu_device_mm_access(struct amdgpu_device *adev, loff_t pos,
 312                              void *buf, size_t size, bool write)
 313 {
 314         unsigned long flags;
 315         uint32_t hi = ~0, tmp = 0;
 316         uint32_t *data = buf;
 317         uint64_t last;
 318         int idx;
 319
 320         if (!drm_dev_enter(adev_to_drm(adev), &idx))
 321                 return;
 322
 323         BUG_ON(!IS_ALIGNED(pos, 4) || !IS_ALIGNED(size, 4));
 324
 325         spin_lock_irqsave(&adev->mmio_idx_lock, flags);
 326         for (last = pos + size; pos < last; pos += 4) {
 327                 tmp = pos >> 31;
 328
 329                 WREG32_NO_KIQ(mmMM_INDEX, ((uint32_t)pos) | 0x80000000);
 330                 if (tmp != hi) {
 331                         WREG32_NO_KIQ(mmMM_INDEX_HI, tmp);
 332                         hi = tmp;
 333                 }
 334                 if (write)
 335                         WREG32_NO_KIQ(mmMM_DATA, *data++);
 336                 else
 337                         *data++ = RREG32_NO_KIQ(mmMM_DATA);
 338         }
 339
 340         spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
 341         drm_dev_exit(idx);
 342 }
 343
 344 /**
 345  * amdgpu_device_aper_access - access vram by vram aperature
 346  *
 347  * @adev: amdgpu_device pointer
 348  * @pos: offset of the buffer in vram
 349  * @buf: virtual address of the buffer in system memory
 350  * @size: read/write size, sizeof(@buf) must > @size
 351  * @write: true - write to vram, otherwise - read from vram
 352  *
 353  * The return value means how many bytes have been transferred.
 354  */
 355 size_t amdgpu_device_aper_access(struct amdgpu_device *adev, loff_t pos,
 356                                  void *buf, size_t size, bool write)
 357 {
 358 #ifdef CONFIG_64BIT
 359         void __iomem *addr;
 360         size_t count = 0;
 361         uint64_t last;
 362
 363         if (!adev->mman.aper_base_kaddr)
 364                 return 0;
 365
 366         last = min(pos + size, adev->gmc.visible_vram_size);
 367         if (last > pos) {
 368                 addr = adev->mman.aper_base_kaddr + pos;
 369                 count = last - pos;
 370
 371                 if (write) {
 372                         memcpy_toio(addr, buf, count);
 373                         mb();
 374                         amdgpu_device_flush_hdp(adev, NULL);
 375                 } else {
 376                         amdgpu_device_invalidate_hdp(adev, NULL);
 377                         mb();
 378                         memcpy_fromio(buf, addr, count);
 379                 }
 380
 381         }
 382
 383         return count;
 384 #else
 385         return 0;
 386 #endif
 387 }
 388
 389 /**
 390  * amdgpu_device_vram_access - read/write a buffer in vram
 391  *
 392  * @adev: amdgpu_device pointer
 393  * @pos: offset of the buffer in vram
 394  * @buf: virtual address of the buffer in system memory
 395  * @size: read/write size, sizeof(@buf) must > @size
 396  * @write: true - write to vram, otherwise - read from vram
 397  */
 398 void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos,
 399                                void *buf, size_t size, bool write)
 400 {
 401         size_t count;
 402
 403         /* try to using vram apreature to access vram first */
 404         count = amdgpu_device_aper_access(adev, pos, buf, size, write);
 405         size -= count;
 406         if (size) {
 407                 /* using MM to access rest vram */
 408                 pos += count;
 409                 buf += count;
 410                 amdgpu_device_mm_access(adev, pos, buf, size, write);
 411         }
 412 }
 413
 414 /*
 415  * register access helper functions.
 416  */
 417
 418 /* Check if hw access should be skipped because of hotplug or device error */
 419 bool amdgpu_device_skip_hw_access(struct amdgpu_device *adev)
 420 {
 421         if (adev->no_hw_access)
 422                 return true;
 423
 424 #ifdef CONFIG_LOCKDEP
 425         /*
 426          * This is a bit complicated to understand, so worth a comment. What we assert
 427          * here is that the GPU reset is not running on another thread in parallel.
 428          *
 429          * For this we trylock the read side of the reset semaphore, if that succeeds
 430          * we know that the reset is not running in paralell.
 431          *
 432          * If the trylock fails we assert that we are either already holding the read
 433          * side of the lock or are the reset thread itself and hold the write side of
 434          * the lock.
 435          */
 436         if (in_task()) {
 437                 if (down_read_trylock(&adev->reset_domain->sem))
 438                         up_read(&adev->reset_domain->sem);
 439                 else
 440                         lockdep_assert_held(&adev->reset_domain->sem);
 441         }
 442 #endif
 443         return false;
 444 }
 445
 446 /**
 447  * amdgpu_device_rreg - read a memory mapped IO or indirect register
 448  *
 449  * @adev: amdgpu_device pointer
 450  * @reg: dword aligned register offset
 451  * @acc_flags: access flags which require special behavior
 452  *
 453  * Returns the 32 bit value from the offset specified.
 454  */
 455 uint32_t amdgpu_device_rreg(struct amdgpu_device *adev,
 456                             uint32_t reg, uint32_t acc_flags)
 457 {
 458         uint32_t ret;
 459
 460         if (amdgpu_device_skip_hw_access(adev))
 461                 return 0;
 462
 463         if ((reg * 4) < adev->rmmio_size) {
 464                 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) &&
 465                     amdgpu_sriov_runtime(adev) &&
 466                     down_read_trylock(&adev->reset_domain->sem)) {
 467                         ret = amdgpu_kiq_rreg(adev, reg);
 468                         up_read(&adev->reset_domain->sem);
 469                 } else {
 470                         ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
 471                 }
 472         } else {
 473                 ret = adev->pcie_rreg(adev, reg * 4);
 474         }
 475
 476         trace_amdgpu_device_rreg(adev->pdev->device, reg, ret);
 477
 478         return ret;
 479 }
 480
 481 /*
 482  * MMIO register read with bytes helper functions
 483  * @offset:bytes offset from MMIO start
 484  *
 485 */
 486
 487 /**
 488  * amdgpu_mm_rreg8 - read a memory mapped IO register
 489  *
 490  * @adev: amdgpu_device pointer
 491  * @offset: byte aligned register offset
 492  *
 493  * Returns the 8 bit value from the offset specified.
 494  */
 495 uint8_t amdgpu_mm_rreg8(struct amdgpu_device *adev, uint32_t offset)
 496 {
 497         if (amdgpu_device_skip_hw_access(adev))
 498                 return 0;
 499
 500         if (offset < adev->rmmio_size)
 501                 return (readb(adev->rmmio + offset));
 502         BUG();
 503 }
 504
 505 /*
 506  * MMIO register write with bytes helper functions
 507  * @offset:bytes offset from MMIO start
 508  * @value: the value want to be written to the register
 509  *
 510 */
 511 /**
 512  * amdgpu_mm_wreg8 - read a memory mapped IO register
 513  *
 514  * @adev: amdgpu_device pointer
 515  * @offset: byte aligned register offset
 516  * @value: 8 bit value to write
 517  *
 518  * Writes the value specified to the offset specified.
 519  */
 520 void amdgpu_mm_wreg8(struct amdgpu_device *adev, uint32_t offset, uint8_t value)
 521 {
 522         if (amdgpu_device_skip_hw_access(adev))
 523                 return;
 524
 525         if (offset < adev->rmmio_size)
 526                 writeb(value, adev->rmmio + offset);
 527         else
 528                 BUG();
 529 }
 530
 531 /**
 532  * amdgpu_device_wreg - write to a memory mapped IO or indirect register
 533  *
 534  * @adev: amdgpu_device pointer
 535  * @reg: dword aligned register offset
 536  * @v: 32 bit value to write to the register
 537  * @acc_flags: access flags which require special behavior
 538  *
 539  * Writes the value specified to the offset specified.
 540  */
 541 void amdgpu_device_wreg(struct amdgpu_device *adev,
 542                         uint32_t reg, uint32_t v,
 543                         uint32_t acc_flags)
 544 {
 545         if (amdgpu_device_skip_hw_access(adev))
 546                 return;
 547
 548         if ((reg * 4) < adev->rmmio_size) {
 549                 if (!(acc_flags & AMDGPU_REGS_NO_KIQ) &&
 550                     amdgpu_sriov_runtime(adev) &&
 551                     down_read_trylock(&adev->reset_domain->sem)) {
 552                         amdgpu_kiq_wreg(adev, reg, v);
 553                         up_read(&adev->reset_domain->sem);
 554                 } else {
 555                         writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
 556                 }
 557         } else {
 558                 adev->pcie_wreg(adev, reg * 4, v);
 559         }
 560
 561         trace_amdgpu_device_wreg(adev->pdev->device, reg, v);
 562 }
 563
 564 /**
 565  * amdgpu_mm_wreg_mmio_rlc -  write register either with direct/indirect mmio or with RLC path if in range
 566  *
 567  * @adev: amdgpu_device pointer
 568  * @reg: mmio/rlc register
 569  * @v: value to write
 570  *
 571  * this function is invoked only for the debugfs register access
 572  */
 573 void amdgpu_mm_wreg_mmio_rlc(struct amdgpu_device *adev,
 574                              uint32_t reg, uint32_t v)
 575 {
 576         if (amdgpu_device_skip_hw_access(adev))
 577                 return;
 578
 579         if (amdgpu_sriov_fullaccess(adev) &&
 580             adev->gfx.rlc.funcs &&
 581             adev->gfx.rlc.funcs->is_rlcg_access_range) {
 582                 if (adev->gfx.rlc.funcs->is_rlcg_access_range(adev, reg))
 583                         return amdgpu_sriov_wreg(adev, reg, v, 0, 0);
 584         } else if ((reg * 4) >= adev->rmmio_size) {
 585                 adev->pcie_wreg(adev, reg * 4, v);
 586         } else {
 587                 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
 588         }
 589 }
 590
 591 /**
 592  * amdgpu_mm_rdoorbell - read a doorbell dword
 593  *
 594  * @adev: amdgpu_device pointer
 595  * @index: doorbell index
 596  *
 597  * Returns the value in the doorbell aperture at the
 598  * requested doorbell index (CIK).
 599  */
 600 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
 601 {
 602         if (amdgpu_device_skip_hw_access(adev))
 603                 return 0;
 604
 605         if (index < adev->doorbell.num_kernel_doorbells) {
 606                 return readl(adev->doorbell.ptr + index);
 607         } else {
 608                 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
 609                 return 0;
 610         }
 611 }
 612
 613 /**
 614  * amdgpu_mm_wdoorbell - write a doorbell dword
 615  *
 616  * @adev: amdgpu_device pointer
 617  * @index: doorbell index
 618  * @v: value to write
 619  *
 620  * Writes @v to the doorbell aperture at the
 621  * requested doorbell index (CIK).
 622  */
 623 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
 624 {
 625         if (amdgpu_device_skip_hw_access(adev))
 626                 return;
 627
 628         if (index < adev->doorbell.num_kernel_doorbells) {
 629                 writel(v, adev->doorbell.ptr + index);
 630         } else {
 631                 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
 632         }
 633 }
 634
 635 /**
 636  * amdgpu_mm_rdoorbell64 - read a doorbell Qword
 637  *
 638  * @adev: amdgpu_device pointer
 639  * @index: doorbell index
 640  *
 641  * Returns the value in the doorbell aperture at the
 642  * requested doorbell index (VEGA10+).
 643  */
 644 u64 amdgpu_mm_rdoorbell64(struct amdgpu_device *adev, u32 index)
 645 {
 646         if (amdgpu_device_skip_hw_access(adev))
 647                 return 0;
 648
 649         if (index < adev->doorbell.num_kernel_doorbells) {
 650                 return atomic64_read((atomic64_t *)(adev->doorbell.ptr + index));
 651         } else {
 652                 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
 653                 return 0;
 654         }
 655 }
 656
 657 /**
 658  * amdgpu_mm_wdoorbell64 - write a doorbell Qword
 659  *
 660  * @adev: amdgpu_device pointer
 661  * @index: doorbell index
 662  * @v: value to write
 663  *
 664  * Writes @v to the doorbell aperture at the
 665  * requested doorbell index (VEGA10+).
 666  */
 667 void amdgpu_mm_wdoorbell64(struct amdgpu_device *adev, u32 index, u64 v)
 668 {
 669         if (amdgpu_device_skip_hw_access(adev))
 670                 return;
 671
 672         if (index < adev->doorbell.num_kernel_doorbells) {
 673                 atomic64_set((atomic64_t *)(adev->doorbell.ptr + index), v);
 674         } else {
 675                 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
 676         }
 677 }
 678
 679 /**
 680  * amdgpu_device_indirect_rreg - read an indirect register
 681  *
 682  * @adev: amdgpu_device pointer
 683  * @reg_addr: indirect register address to read from
 684  *
 685  * Returns the value of indirect register @reg_addr
 686  */
 687 u32 amdgpu_device_indirect_rreg(struct amdgpu_device *adev,
 688                                 u32 reg_addr)
 689 {
 690         unsigned long flags, pcie_index, pcie_data;
 691         void __iomem *pcie_index_offset;
 692         void __iomem *pcie_data_offset;
 693         u32 r;
 694
 695         pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
 696         pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
 697
 698         spin_lock_irqsave(&adev->pcie_idx_lock, flags);
 699         pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
 700         pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
 701
 702         writel(reg_addr, pcie_index_offset);
 703         readl(pcie_index_offset);
 704         r = readl(pcie_data_offset);
 705         spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
 706
 707         return r;
 708 }
 709
 710 /**
 711  * amdgpu_device_indirect_rreg64 - read a 64bits indirect register
 712  *
 713  * @adev: amdgpu_device pointer
 714  * @reg_addr: indirect register address to read from
 715  *
 716  * Returns the value of indirect register @reg_addr
 717  */
 718 u64 amdgpu_device_indirect_rreg64(struct amdgpu_device *adev,
 719                                   u32 reg_addr)
 720 {
 721         unsigned long flags, pcie_index, pcie_data;
 722         void __iomem *pcie_index_offset;
 723         void __iomem *pcie_data_offset;
 724         u64 r;
 725
 726         pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
 727         pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
 728
 729         spin_lock_irqsave(&adev->pcie_idx_lock, flags);
 730         pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
 731         pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
 732
 733         /* read low 32 bits */
 734         writel(reg_addr, pcie_index_offset);
 735         readl(pcie_index_offset);
 736         r = readl(pcie_data_offset);
 737         /* read high 32 bits */
 738         writel(reg_addr + 4, pcie_index_offset);
 739         readl(pcie_index_offset);
 740         r |= ((u64)readl(pcie_data_offset) << 32);
 741         spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
 742
 743         return r;
 744 }
 745
 746 /**
 747  * amdgpu_device_indirect_wreg - write an indirect register address
 748  *
 749  * @adev: amdgpu_device pointer
 750  * @pcie_index: mmio register offset
 751  * @pcie_data: mmio register offset
 752  * @reg_addr: indirect register offset
 753  * @reg_data: indirect register data
 754  *
 755  */
 756 void amdgpu_device_indirect_wreg(struct amdgpu_device *adev,
 757                                  u32 reg_addr, u32 reg_data)
 758 {
 759         unsigned long flags, pcie_index, pcie_data;
 760         void __iomem *pcie_index_offset;
 761         void __iomem *pcie_data_offset;
 762
 763         pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
 764         pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
 765
 766         spin_lock_irqsave(&adev->pcie_idx_lock, flags);
 767         pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
 768         pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
 769
 770         writel(reg_addr, pcie_index_offset);
 771         readl(pcie_index_offset);
 772         writel(reg_data, pcie_data_offset);
 773         readl(pcie_data_offset);
 774         spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
 775 }
 776
 777 /**
 778  * amdgpu_device_indirect_wreg64 - write a 64bits indirect register address
 779  *
 780  * @adev: amdgpu_device pointer
 781  * @pcie_index: mmio register offset
 782  * @pcie_data: mmio register offset
 783  * @reg_addr: indirect register offset
 784  * @reg_data: indirect register data
 785  *
 786  */
 787 void amdgpu_device_indirect_wreg64(struct amdgpu_device *adev,
 788                                    u32 reg_addr, u64 reg_data)
 789 {
 790         unsigned long flags, pcie_index, pcie_data;
 791         void __iomem *pcie_index_offset;
 792         void __iomem *pcie_data_offset;
 793
 794         pcie_index = adev->nbio.funcs->get_pcie_index_offset(adev);
 795         pcie_data = adev->nbio.funcs->get_pcie_data_offset(adev);
 796
 797         spin_lock_irqsave(&adev->pcie_idx_lock, flags);
 798         pcie_index_offset = (void __iomem *)adev->rmmio + pcie_index * 4;
 799         pcie_data_offset = (void __iomem *)adev->rmmio + pcie_data * 4;
 800
 801         /* write low 32 bits */
 802         writel(reg_addr, pcie_index_offset);
 803         readl(pcie_index_offset);
 804         writel((u32)(reg_data & 0xffffffffULL), pcie_data_offset);
 805         readl(pcie_data_offset);
 806         /* write high 32 bits */
 807         writel(reg_addr + 4, pcie_index_offset);
 808         readl(pcie_index_offset);
 809         writel((u32)(reg_data >> 32), pcie_data_offset);
 810         readl(pcie_data_offset);
 811         spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
 812 }
 813
 814 /**
 815  * amdgpu_device_get_rev_id - query device rev_id
 816  *
 817  * @adev: amdgpu_device pointer
 818  *
 819  * Return device rev_id
 820  */
 821 u32 amdgpu_device_get_rev_id(struct amdgpu_device *adev)
 822 {
 823         return adev->nbio.funcs->get_rev_id(adev);
 824 }
 825
 826 /**
 827  * amdgpu_invalid_rreg - dummy reg read function
 828  *
 829  * @adev: amdgpu_device pointer
 830  * @reg: offset of register
 831  *
 832  * Dummy register read function.  Used for register blocks
 833  * that certain asics don't have (all asics).
 834  * Returns the value in the register.
 835  */
 836 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
 837 {
 838         DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
 839         BUG();
 840         return 0;
 841 }
 842
 843 /**
 844  * amdgpu_invalid_wreg - dummy reg write function
 845  *
 846  * @adev: amdgpu_device pointer
 847  * @reg: offset of register
 848  * @v: value to write to the register
 849  *
 850  * Dummy register read function.  Used for register blocks
 851  * that certain asics don't have (all asics).
 852  */
 853 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
 854 {
 855         DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
 856                   reg, v);
 857         BUG();
 858 }
 859
 860 /**
 861  * amdgpu_invalid_rreg64 - dummy 64 bit reg read function
 862  *
 863  * @adev: amdgpu_device pointer
 864  * @reg: offset of register
 865  *
 866  * Dummy register read function.  Used for register blocks
 867  * that certain asics don't have (all asics).
 868  * Returns the value in the register.
 869  */
 870 static uint64_t amdgpu_invalid_rreg64(struct amdgpu_device *adev, uint32_t reg)
 871 {
 872         DRM_ERROR("Invalid callback to read 64 bit register 0x%04X\n", reg);
 873         BUG();
 874         return 0;
 875 }
 876
 877 /**
 878  * amdgpu_invalid_wreg64 - dummy reg write function
 879  *
 880  * @adev: amdgpu_device pointer
 881  * @reg: offset of register
 882  * @v: value to write to the register
 883  *
 884  * Dummy register read function.  Used for register blocks
 885  * that certain asics don't have (all asics).
 886  */
 887 static void amdgpu_invalid_wreg64(struct amdgpu_device *adev, uint32_t reg, uint64_t v)
 888 {
 889         DRM_ERROR("Invalid callback to write 64 bit register 0x%04X with 0x%08llX\n",
 890                   reg, v);
 891         BUG();
 892 }
 893
 894 /**
 895  * amdgpu_block_invalid_rreg - dummy reg read function
 896  *
 897  * @adev: amdgpu_device pointer
 898  * @block: offset of instance
 899  * @reg: offset of register
 900  *
 901  * Dummy register read function.  Used for register blocks
 902  * that certain asics don't have (all asics).
 903  * Returns the value in the register.
 904  */
 905 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
 906                                           uint32_t block, uint32_t reg)
 907 {
 908         DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
 909                   reg, block);
 910         BUG();
 911         return 0;
 912 }
 913
 914 /**
 915  * amdgpu_block_invalid_wreg - dummy reg write function
 916  *
 917  * @adev: amdgpu_device pointer
 918  * @block: offset of instance
 919  * @reg: offset of register
 920  * @v: value to write to the register
 921  *
 922  * Dummy register read function.  Used for register blocks
 923  * that certain asics don't have (all asics).
 924  */
 925 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
 926                                       uint32_t block,
 927                                       uint32_t reg, uint32_t v)
 928 {
 929         DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
 930                   reg, block, v);
 931         BUG();
 932 }
 933
 934 /**
 935  * amdgpu_device_asic_init - Wrapper for atom asic_init
 936  *
 937  * @adev: amdgpu_device pointer
 938  *
 939  * Does any asic specific work and then calls atom asic init.
 940  */
 941 static int amdgpu_device_asic_init(struct amdgpu_device *adev)
 942 {
 943         amdgpu_asic_pre_asic_init(adev);
 944
 945         if (adev->ip_versions[GC_HWIP][0] >= IP_VERSION(11, 0, 0))
 946                 return amdgpu_atomfirmware_asic_init(adev, true);
 947         else
 948                 return amdgpu_atom_asic_init(adev->mode_info.atom_context);
 949 }
 950
 951 /**
 952  * amdgpu_device_mem_scratch_init - allocate the VRAM scratch page
 953  *
 954  * @adev: amdgpu_device pointer
 955  *
 956  * Allocates a scratch page of VRAM for use by various things in the
 957  * driver.
 958  */
 959 static int amdgpu_device_mem_scratch_init(struct amdgpu_device *adev)
 960 {
 961         return amdgpu_bo_create_kernel(adev, AMDGPU_GPU_PAGE_SIZE, PAGE_SIZE,
 962                                        AMDGPU_GEM_DOMAIN_VRAM |
 963                                        AMDGPU_GEM_DOMAIN_GTT,
 964                                        &adev->mem_scratch.robj,
 965                                        &adev->mem_scratch.gpu_addr,
 966                                        (void **)&adev->mem_scratch.ptr);
 967 }
 968
 969 /**
 970  * amdgpu_device_mem_scratch_fini - Free the VRAM scratch page
 971  *
 972  * @adev: amdgpu_device pointer
 973  *
 974  * Frees the VRAM scratch page.
 975  */
 976 static void amdgpu_device_mem_scratch_fini(struct amdgpu_device *adev)
 977 {
 978         amdgpu_bo_free_kernel(&adev->mem_scratch.robj, NULL, NULL);
 979 }
 980
 981 /**
 982  * amdgpu_device_program_register_sequence - program an array of registers.
 983  *
 984  * @adev: amdgpu_device pointer
 985  * @registers: pointer to the register array
 986  * @array_size: size of the register array
 987  *
 988  * Programs an array or registers with and and or masks.
 989  * This is a helper for setting golden registers.
 990  */
 991 void amdgpu_device_program_register_sequence(struct amdgpu_device *adev,
 992                                              const u32 *registers,
 993                                              const u32 array_size)
 994 {
 995         u32 tmp, reg, and_mask, or_mask;
 996         int i;
 997
 998         if (array_size % 3)
 999                 return;
1000
1001         for (i = 0; i < array_size; i += 3) {
1002                 reg = registers[i + 0];
1003                 and_mask = registers[i + 1];
1004                 or_mask = registers[i + 2];
1005
1006                 if (and_mask == 0xffffffff) {
1007                         tmp = or_mask;
1008                 } else {
1009                         tmp = RREG32(reg);
1010                         tmp &= ~and_mask;
1011                         if (adev->family >= AMDGPU_FAMILY_AI)
1012                                 tmp |= (or_mask & and_mask);
1013                         else
1014                                 tmp |= or_mask;
1015                 }
1016                 WREG32(reg, tmp);
1017         }
1018 }
1019
1020 /**
1021  * amdgpu_device_pci_config_reset - reset the GPU
1022  *
1023  * @adev: amdgpu_device pointer
1024  *
1025  * Resets the GPU using the pci config reset sequence.
1026  * Only applicable to asics prior to vega10.
1027  */
1028 void amdgpu_device_pci_config_reset(struct amdgpu_device *adev)
1029 {
1030         pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
1031 }
1032
1033 /**
1034  * amdgpu_device_pci_reset - reset the GPU using generic PCI means
1035  *
1036  * @adev: amdgpu_device pointer
1037  *
1038  * Resets the GPU using generic pci reset interfaces (FLR, SBR, etc.).
1039  */
1040 int amdgpu_device_pci_reset(struct amdgpu_device *adev)
1041 {
1042         return pci_reset_function(adev->pdev);
1043 }
1044
1045 /*
1046  * GPU doorbell aperture helpers function.
1047  */
1048 /**
1049  * amdgpu_device_doorbell_init - Init doorbell driver information.
1050  *
1051  * @adev: amdgpu_device pointer
1052  *
1053  * Init doorbell driver information (CIK)
1054  * Returns 0 on success, error on failure.
1055  */
1056 static int amdgpu_device_doorbell_init(struct amdgpu_device *adev)
1057 {
1058
1059         /* No doorbell on SI hardware generation */
1060         if (adev->asic_type < CHIP_BONAIRE) {
1061                 adev->doorbell.base = 0;
1062                 adev->doorbell.size = 0;
1063                 adev->doorbell.num_kernel_doorbells = 0;
1064                 adev->doorbell.ptr = NULL;
1065                 return 0;
1066         }
1067
1068         if (pci_resource_flags(adev->pdev, 2) & IORESOURCE_UNSET)
1069                 return -EINVAL;
1070
1071         amdgpu_asic_init_doorbell_index(adev);
1072
1073         /* doorbell bar mapping */
1074         adev->doorbell.base = pci_resource_start(adev->pdev, 2);
1075         adev->doorbell.size = pci_resource_len(adev->pdev, 2);
1076
1077         if (adev->enable_mes) {
1078                 adev->doorbell.num_kernel_doorbells =
1079                         adev->doorbell.size / sizeof(u32);
1080         } else {
1081                 adev->doorbell.num_kernel_doorbells =
1082                         min_t(u32, adev->doorbell.size / sizeof(u32),
1083                               adev->doorbell_index.max_assignment+1);
1084                 if (adev->doorbell.num_kernel_doorbells == 0)
1085                         return -EINVAL;
1086
1087                 /* For Vega, reserve and map two pages on doorbell BAR since SDMA
1088                  * paging queue doorbell use the second page. The
1089                  * AMDGPU_DOORBELL64_MAX_ASSIGNMENT definition assumes all the
1090                  * doorbells are in the first page. So with paging queue enabled,
1091                  * the max num_kernel_doorbells should + 1 page (0x400 in dword)
1092                  */
1093                 if (adev->asic_type >= CHIP_VEGA10)
1094                         adev->doorbell.num_kernel_doorbells += 0x400;
1095         }
1096
1097         adev->doorbell.ptr = ioremap(adev->doorbell.base,
1098                                      adev->doorbell.num_kernel_doorbells *
1099                                      sizeof(u32));
1100         if (adev->doorbell.ptr == NULL)
1101                 return -ENOMEM;
1102
1103         return 0;
1104 }
1105
1106 /**
1107  * amdgpu_device_doorbell_fini - Tear down doorbell driver information.
1108  *
1109  * @adev: amdgpu_device pointer
1110  *
1111  * Tear down doorbell driver information (CIK)
1112  */
1113 static void amdgpu_device_doorbell_fini(struct amdgpu_device *adev)
1114 {
1115         iounmap(adev->doorbell.ptr);
1116         adev->doorbell.ptr = NULL;
1117 }
1118
1119
1120
1121 /*
1122  * amdgpu_device_wb_*()
1123  * Writeback is the method by which the GPU updates special pages in memory
1124  * with the status of certain GPU events (fences, ring pointers,etc.).
1125  */
1126
1127 /**
1128  * amdgpu_device_wb_fini - Disable Writeback and free memory
1129  *
1130  * @adev: amdgpu_device pointer
1131  *
1132  * Disables Writeback and frees the Writeback memory (all asics).
1133  * Used at driver shutdown.
1134  */
1135 static void amdgpu_device_wb_fini(struct amdgpu_device *adev)
1136 {
1137         if (adev->wb.wb_obj) {
1138                 amdgpu_bo_free_kernel(&adev->wb.wb_obj,
1139                                       &adev->wb.gpu_addr,
1140                                       (void **)&adev->wb.wb);
1141                 adev->wb.wb_obj = NULL;
1142         }
1143 }
1144
1145 /**
1146  * amdgpu_device_wb_init - Init Writeback driver info and allocate memory
1147  *
1148  * @adev: amdgpu_device pointer
1149  *
1150  * Initializes writeback and allocates writeback memory (all asics).
1151  * Used at driver startup.
1152  * Returns 0 on success or an -error on failure.
1153  */
1154 static int amdgpu_device_wb_init(struct amdgpu_device *adev)
1155 {
1156         int r;
1157
1158         if (adev->wb.wb_obj == NULL) {
1159                 /* AMDGPU_MAX_WB * sizeof(uint32_t) * 8 = AMDGPU_MAX_WB 256bit slots */
1160                 r = amdgpu_bo_create_kernel(adev, AMDGPU_MAX_WB * sizeof(uint32_t) * 8,
1161                                             PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT,
1162                                             &adev->wb.wb_obj, &adev->wb.gpu_addr,
1163                                             (void **)&adev->wb.wb);
1164                 if (r) {
1165                         dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
1166                         return r;
1167                 }
1168
1169                 adev->wb.num_wb = AMDGPU_MAX_WB;
1170                 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
1171
1172                 /* clear wb memory */
1173                 memset((char *)adev->wb.wb, 0, AMDGPU_MAX_WB * sizeof(uint32_t) * 8);
1174         }
1175
1176         return 0;
1177 }
1178
1179 /**
1180  * amdgpu_device_wb_get - Allocate a wb entry
1181  *
1182  * @adev: amdgpu_device pointer
1183  * @wb: wb index
1184  *
1185  * Allocate a wb slot for use by the driver (all asics).
1186  * Returns 0 on success or -EINVAL on failure.
1187  */
1188 int amdgpu_device_wb_get(struct amdgpu_device *adev, u32 *wb)
1189 {
1190         unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
1191
1192         if (offset < adev->wb.num_wb) {
1193                 __set_bit(offset, adev->wb.used);
1194                 *wb = offset << 3; /* convert to dw offset */
1195                 return 0;
1196         } else {
1197                 return -EINVAL;
1198         }
1199 }
1200
1201 /**
1202  * amdgpu_device_wb_free - Free a wb entry
1203  *
1204  * @adev: amdgpu_device pointer
1205  * @wb: wb index
1206  *
1207  * Free a wb slot allocated for use by the driver (all asics)
1208  */
1209 void amdgpu_device_wb_free(struct amdgpu_device *adev, u32 wb)
1210 {
1211         wb >>= 3;
1212         if (wb < adev->wb.num_wb)
1213                 __clear_bit(wb, adev->wb.used);
1214 }
1215
1216 /**
1217  * amdgpu_device_resize_fb_bar - try to resize FB BAR
1218  *
1219  * @adev: amdgpu_device pointer
1220  *
1221  * Try to resize FB BAR to make all VRAM CPU accessible. We try very hard not
1222  * to fail, but if any of the BARs is not accessible after the size we abort
1223  * driver loading by returning -ENODEV.
1224  */
1225 int amdgpu_device_resize_fb_bar(struct amdgpu_device *adev)
1226 {
1227         int rbar_size = pci_rebar_bytes_to_size(adev->gmc.real_vram_size);
1228         struct pci_bus *root;
1229         struct resource *res;
1230         unsigned i;
1231         u16 cmd;
1232         int r;
1233
1234         /* Bypass for VF */
1235         if (amdgpu_sriov_vf(adev))
1236                 return 0;
1237
1238         /* skip if the bios has already enabled large BAR */
1239         if (adev->gmc.real_vram_size &&
1240             (pci_resource_len(adev->pdev, 0) >= adev->gmc.real_vram_size))
1241                 return 0;
1242
1243         /* Check if the root BUS has 64bit memory resources */
1244         root = adev->pdev->bus;
1245         while (root->parent)
1246                 root = root->parent;
1247
1248         pci_bus_for_each_resource(root, res, i) {
1249                 if (res && res->flags & (IORESOURCE_MEM | IORESOURCE_MEM_64) &&
1250                     res->start > 0x100000000ull)
1251                         break;
1252         }
1253
1254         /* Trying to resize is pointless without a root hub window above 4GB */
1255         if (!res)
1256                 return 0;
1257
1258         /* Limit the BAR size to what is available */
1259         rbar_size = min(fls(pci_rebar_get_possible_sizes(adev->pdev, 0)) - 1,
1260                         rbar_size);
1261
1262         /* Disable memory decoding while we change the BAR addresses and size */
1263         pci_read_config_word(adev->pdev, PCI_COMMAND, &cmd);
1264         pci_write_config_word(adev->pdev, PCI_COMMAND,
1265                               cmd & ~PCI_COMMAND_MEMORY);
1266
1267         /* Free the VRAM and doorbell BAR, we most likely need to move both. */
1268         amdgpu_device_doorbell_fini(adev);
1269         if (adev->asic_type >= CHIP_BONAIRE)
1270                 pci_release_resource(adev->pdev, 2);
1271
1272         pci_release_resource(adev->pdev, 0);
1273
1274         r = pci_resize_resource(adev->pdev, 0, rbar_size);
1275         if (r == -ENOSPC)
1276                 DRM_INFO("Not enough PCI address space for a large BAR.");
1277         else if (r && r != -ENOTSUPP)
1278                 DRM_ERROR("Problem resizing BAR0 (%d).", r);
1279
1280         pci_assign_unassigned_bus_resources(adev->pdev->bus);
1281
1282         /* When the doorbell or fb BAR isn't available we have no chance of
1283          * using the device.
1284          */
1285         r = amdgpu_device_doorbell_init(adev);
1286         if (r || (pci_resource_flags(adev->pdev, 0) & IORESOURCE_UNSET))
1287                 return -ENODEV;
1288
1289         pci_write_config_word(adev->pdev, PCI_COMMAND, cmd);
1290
1291         return 0;
1292 }
1293
1294 /*
1295  * GPU helpers function.
1296  */
1297 /**
1298  * amdgpu_device_need_post - check if the hw need post or not
1299  *
1300  * @adev: amdgpu_device pointer
1301  *
1302  * Check if the asic has been initialized (all asics) at driver startup
1303  * or post is needed if  hw reset is performed.
1304  * Returns true if need or false if not.
1305  */
1306 bool amdgpu_device_need_post(struct amdgpu_device *adev)
1307 {
1308         uint32_t reg;
1309
1310         if (amdgpu_sriov_vf(adev))
1311                 return false;
1312
1313         if (amdgpu_passthrough(adev)) {
1314                 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
1315                  * some old smc fw still need driver do vPost otherwise gpu hang, while
1316                  * those smc fw version above 22.15 doesn't have this flaw, so we force
1317                  * vpost executed for smc version below 22.15
1318                  */
1319                 if (adev->asic_type == CHIP_FIJI) {
1320                         int err;
1321                         uint32_t fw_ver;
1322                         err = request_firmware(&adev->pm.fw, "amdgpu/fiji_smc.bin", adev->dev);
1323                         /* force vPost if error occured */
1324                         if (err)
1325                                 return true;
1326
1327                         fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
1328                         if (fw_ver < 0x00160e00)
1329                                 return true;
1330                 }
1331         }
1332
1333         /* Don't post if we need to reset whole hive on init */
1334         if (adev->gmc.xgmi.pending_reset)
1335                 return false;
1336
1337         if (adev->has_hw_reset) {
1338                 adev->has_hw_reset = false;
1339                 return true;
1340         }
1341
1342         /* bios scratch used on CIK+ */
1343         if (adev->asic_type >= CHIP_BONAIRE)
1344                 return amdgpu_atombios_scratch_need_asic_init(adev);
1345
1346         /* check MEM_SIZE for older asics */
1347         reg = amdgpu_asic_get_config_memsize(adev);
1348
1349         if ((reg != 0) && (reg != 0xffffffff))
1350                 return false;
1351
1352         return true;
1353 }
1354
1355 /**
1356  * amdgpu_device_should_use_aspm - check if the device should program ASPM
1357  *
1358  * @adev: amdgpu_device pointer
1359  *
1360  * Confirm whether the module parameter and pcie bridge agree that ASPM should
1361  * be set for this device.
1362  *
1363  * Returns true if it should be used or false if not.
1364  */
1365 bool amdgpu_device_should_use_aspm(struct amdgpu_device *adev)
1366 {
1367         switch (amdgpu_aspm) {
1368         case -1:
1369                 break;
1370         case 0:
1371                 return false;
1372         case 1:
1373                 return true;
1374         default:
1375                 return false;
1376         }
1377         return pcie_aspm_enabled(adev->pdev);
1378 }
1379
1380 bool amdgpu_device_aspm_support_quirk(void)
1381 {
1382 #if IS_ENABLED(CONFIG_X86)
1383         struct cpuinfo_x86 *c = &cpu_data(0);
1384
1385         return !(c->x86 == 6 && c->x86_model == INTEL_FAM6_ALDERLAKE);
1386 #else
1387         return true;
1388 #endif
1389 }
1390
1391 /* if we get transitioned to only one device, take VGA back */
1392 /**
1393  * amdgpu_device_vga_set_decode - enable/disable vga decode
1394  *
1395  * @pdev: PCI device pointer
1396  * @state: enable/disable vga decode
1397  *
1398  * Enable/disable vga decode (all asics).
1399  * Returns VGA resource flags.
1400  */
1401 static unsigned int amdgpu_device_vga_set_decode(struct pci_dev *pdev,
1402                 bool state)
1403 {
1404         struct amdgpu_device *adev = drm_to_adev(pci_get_drvdata(pdev));
1405         amdgpu_asic_set_vga_state(adev, state);
1406         if (state)
1407                 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
1408                        VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
1409         else
1410                 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
1411 }
1412
1413 /**
1414  * amdgpu_device_check_block_size - validate the vm block size
1415  *
1416  * @adev: amdgpu_device pointer
1417  *
1418  * Validates the vm block size specified via module parameter.
1419  * The vm block size defines number of bits in page table versus page directory,
1420  * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
1421  * page table and the remaining bits are in the page directory.
1422  */
1423 static void amdgpu_device_check_block_size(struct amdgpu_device *adev)
1424 {
1425         /* defines number of bits in page table versus page directory,
1426          * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
1427          * page table and the remaining bits are in the page directory */
1428         if (amdgpu_vm_block_size == -1)
1429                 return;
1430
1431         if (amdgpu_vm_block_size < 9) {
1432                 dev_warn(adev->dev, "VM page table size (%d) too small\n",
1433                          amdgpu_vm_block_size);
1434                 amdgpu_vm_block_size = -1;
1435         }
1436 }
1437
1438 /**
1439  * amdgpu_device_check_vm_size - validate the vm size
1440  *
1441  * @adev: amdgpu_device pointer
1442  *
1443  * Validates the vm size in GB specified via module parameter.
1444  * The VM size is the size of the GPU virtual memory space in GB.
1445  */
1446 static void amdgpu_device_check_vm_size(struct amdgpu_device *adev)
1447 {
1448         /* no need to check the default value */
1449         if (amdgpu_vm_size == -1)
1450                 return;
1451
1452         if (amdgpu_vm_size < 1) {
1453                 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
1454                          amdgpu_vm_size);
1455                 amdgpu_vm_size = -1;
1456         }
1457 }
1458
1459 static void amdgpu_device_check_smu_prv_buffer_size(struct amdgpu_device *adev)
1460 {
1461         struct sysinfo si;
1462         bool is_os_64 = (sizeof(void *) == 8);
1463         uint64_t total_memory;
1464         uint64_t dram_size_seven_GB = 0x1B8000000;
1465         uint64_t dram_size_three_GB = 0xB8000000;
1466
1467         if (amdgpu_smu_memory_pool_size == 0)
1468                 return;
1469
1470         if (!is_os_64) {
1471                 DRM_WARN("Not 64-bit OS, feature not supported\n");
1472                 goto def_value;
1473         }
1474         si_meminfo(&si);
1475         total_memory = (uint64_t)si.totalram * si.mem_unit;
1476
1477         if ((amdgpu_smu_memory_pool_size == 1) ||
1478                 (amdgpu_smu_memory_pool_size == 2)) {
1479                 if (total_memory < dram_size_three_GB)
1480                         goto def_value1;
1481         } else if ((amdgpu_smu_memory_pool_size == 4) ||
1482                 (amdgpu_smu_memory_pool_size == 8)) {
1483                 if (total_memory < dram_size_seven_GB)
1484                         goto def_value1;
1485         } else {
1486                 DRM_WARN("Smu memory pool size not supported\n");
1487                 goto def_value;
1488         }
1489         adev->pm.smu_prv_buffer_size = amdgpu_smu_memory_pool_size << 28;
1490
1491         return;
1492
1493 def_value1:
1494         DRM_WARN("No enough system memory\n");
1495 def_value:
1496         adev->pm.smu_prv_buffer_size = 0;
1497 }
1498
1499 static int amdgpu_device_init_apu_flags(struct amdgpu_device *adev)
1500 {
1501         if (!(adev->flags & AMD_IS_APU) ||
1502             adev->asic_type < CHIP_RAVEN)
1503                 return 0;
1504
1505         switch (adev->asic_type) {
1506         case CHIP_RAVEN:
1507                 if (adev->pdev->device == 0x15dd)
1508                         adev->apu_flags |= AMD_APU_IS_RAVEN;
1509                 if (adev->pdev->device == 0x15d8)
1510                         adev->apu_flags |= AMD_APU_IS_PICASSO;
1511                 break;
1512         case CHIP_RENOIR:
1513                 if ((adev->pdev->device == 0x1636) ||
1514                     (adev->pdev->device == 0x164c))
1515                         adev->apu_flags |= AMD_APU_IS_RENOIR;
1516                 else
1517                         adev->apu_flags |= AMD_APU_IS_GREEN_SARDINE;
1518                 break;
1519         case CHIP_VANGOGH:
1520                 adev->apu_flags |= AMD_APU_IS_VANGOGH;
1521                 break;
1522         case CHIP_YELLOW_CARP:
1523                 break;
1524         case CHIP_CYAN_SKILLFISH:
1525                 if ((adev->pdev->device == 0x13FE) ||
1526                     (adev->pdev->device == 0x143F))
1527                         adev->apu_flags |= AMD_APU_IS_CYAN_SKILLFISH2;
1528                 break;
1529         default:
1530                 break;
1531         }
1532
1533         return 0;
1534 }
1535
1536 /**
1537  * amdgpu_device_check_arguments - validate module params
1538  *
1539  * @adev: amdgpu_device pointer
1540  *
1541  * Validates certain module parameters and updates
1542  * the associated values used by the driver (all asics).
1543  */
1544 static int amdgpu_device_check_arguments(struct amdgpu_device *adev)
1545 {
1546         if (amdgpu_sched_jobs < 4) {
1547                 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
1548                          amdgpu_sched_jobs);
1549                 amdgpu_sched_jobs = 4;
1550         } else if (!is_power_of_2(amdgpu_sched_jobs)) {
1551                 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
1552                          amdgpu_sched_jobs);
1553                 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
1554         }
1555
1556         if (amdgpu_gart_size != -1 && amdgpu_gart_size < 32) {
1557                 /* gart size must be greater or equal to 32M */
1558                 dev_warn(adev->dev, "gart size (%d) too small\n",
1559                          amdgpu_gart_size);
1560                 amdgpu_gart_size = -1;
1561         }
1562
1563         if (amdgpu_gtt_size != -1 && amdgpu_gtt_size < 32) {
1564                 /* gtt size must be greater or equal to 32M */
1565                 dev_warn(adev->dev, "gtt size (%d) too small\n",
1566                                  amdgpu_gtt_size);
1567                 amdgpu_gtt_size = -1;
1568         }
1569
1570         /* valid range is between 4 and 9 inclusive */
1571         if (amdgpu_vm_fragment_size != -1 &&
1572             (amdgpu_vm_fragment_size > 9 || amdgpu_vm_fragment_size < 4)) {
1573                 dev_warn(adev->dev, "valid range is between 4 and 9\n");
1574                 amdgpu_vm_fragment_size = -1;
1575         }
1576
1577         if (amdgpu_sched_hw_submission < 2) {
1578                 dev_warn(adev->dev, "sched hw submission jobs (%d) must be at least 2\n",
1579                          amdgpu_sched_hw_submission);
1580                 amdgpu_sched_hw_submission = 2;
1581         } else if (!is_power_of_2(amdgpu_sched_hw_submission)) {
1582                 dev_warn(adev->dev, "sched hw submission jobs (%d) must be a power of 2\n",
1583                          amdgpu_sched_hw_submission);
1584                 amdgpu_sched_hw_submission = roundup_pow_of_two(amdgpu_sched_hw_submission);
1585         }
1586
1587         if (amdgpu_reset_method < -1 || amdgpu_reset_method > 4) {
1588                 dev_warn(adev->dev, "invalid option for reset method, reverting to default\n");
1589                 amdgpu_reset_method = -1;
1590         }
1591
1592         amdgpu_device_check_smu_prv_buffer_size(adev);
1593
1594         amdgpu_device_check_vm_size(adev);
1595
1596         amdgpu_device_check_block_size(adev);
1597
1598         adev->firmware.load_type = amdgpu_ucode_get_load_type(adev, amdgpu_fw_load_type);
1599
1600         return 0;
1601 }
1602
1603 /**
1604  * amdgpu_switcheroo_set_state - set switcheroo state
1605  *
1606  * @pdev: pci dev pointer
1607  * @state: vga_switcheroo state
1608  *
1609  * Callback for the switcheroo driver.  Suspends or resumes
1610  * the asics before or after it is powered up using ACPI methods.
1611  */
1612 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev,
1613                                         enum vga_switcheroo_state state)
1614 {
1615         struct drm_device *dev = pci_get_drvdata(pdev);
1616         int r;
1617
1618         if (amdgpu_device_supports_px(dev) && state == VGA_SWITCHEROO_OFF)
1619                 return;
1620
1621         if (state == VGA_SWITCHEROO_ON) {
1622                 pr_info("switched on\n");
1623                 /* don't suspend or resume card normally */
1624                 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1625
1626                 pci_set_power_state(pdev, PCI_D0);
1627                 amdgpu_device_load_pci_state(pdev);
1628                 r = pci_enable_device(pdev);
1629                 if (r)
1630                         DRM_WARN("pci_enable_device failed (%d)\n", r);
1631                 amdgpu_device_resume(dev, true);
1632
1633                 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1634         } else {
1635                 pr_info("switched off\n");
1636                 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1637                 amdgpu_device_suspend(dev, true);
1638                 amdgpu_device_cache_pci_state(pdev);
1639                 /* Shut down the device */
1640                 pci_disable_device(pdev);
1641                 pci_set_power_state(pdev, PCI_D3cold);
1642                 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1643         }
1644 }
1645
1646 /**
1647  * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1648  *
1649  * @pdev: pci dev pointer
1650  *
1651  * Callback for the switcheroo driver.  Check of the switcheroo
1652  * state can be changed.
1653  * Returns true if the state can be changed, false if not.
1654  */
1655 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1656 {
1657         struct drm_device *dev = pci_get_drvdata(pdev);
1658
1659         /*
1660         * FIXME: open_count is protected by drm_global_mutex but that would lead to
1661         * locking inversion with the driver load path. And the access here is
1662         * completely racy anyway. So don't bother with locking for now.
1663         */
1664         return atomic_read(&dev->open_count) == 0;
1665 }
1666
1667 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1668         .set_gpu_state = amdgpu_switcheroo_set_state,
1669         .reprobe = NULL,
1670         .can_switch = amdgpu_switcheroo_can_switch,
1671 };
1672
1673 /**
1674  * amdgpu_device_ip_set_clockgating_state - set the CG state
1675  *
1676  * @dev: amdgpu_device pointer
1677  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1678  * @state: clockgating state (gate or ungate)
1679  *
1680  * Sets the requested clockgating state for all instances of
1681  * the hardware IP specified.
1682  * Returns the error code from the last instance.
1683  */
1684 int amdgpu_device_ip_set_clockgating_state(void *dev,
1685                                            enum amd_ip_block_type block_type,
1686                                            enum amd_clockgating_state state)
1687 {
1688         struct amdgpu_device *adev = dev;
1689         int i, r = 0;
1690
1691         for (i = 0; i < adev->num_ip_blocks; i++) {
1692                 if (!adev->ip_blocks[i].status.valid)
1693                         continue;
1694                 if (adev->ip_blocks[i].version->type != block_type)
1695                         continue;
1696                 if (!adev->ip_blocks[i].version->funcs->set_clockgating_state)
1697                         continue;
1698                 r = adev->ip_blocks[i].version->funcs->set_clockgating_state(
1699                         (void *)adev, state);
1700                 if (r)
1701                         DRM_ERROR("set_clockgating_state of IP block <%s> failed %d\n",
1702                                   adev->ip_blocks[i].version->funcs->name, r);
1703         }
1704         return r;
1705 }
1706
1707 /**
1708  * amdgpu_device_ip_set_powergating_state - set the PG state
1709  *
1710  * @dev: amdgpu_device pointer
1711  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1712  * @state: powergating state (gate or ungate)
1713  *
1714  * Sets the requested powergating state for all instances of
1715  * the hardware IP specified.
1716  * Returns the error code from the last instance.
1717  */
1718 int amdgpu_device_ip_set_powergating_state(void *dev,
1719                                            enum amd_ip_block_type block_type,
1720                                            enum amd_powergating_state state)
1721 {
1722         struct amdgpu_device *adev = dev;
1723         int i, r = 0;
1724
1725         for (i = 0; i < adev->num_ip_blocks; i++) {
1726                 if (!adev->ip_blocks[i].status.valid)
1727                         continue;
1728                 if (adev->ip_blocks[i].version->type != block_type)
1729                         continue;
1730                 if (!adev->ip_blocks[i].version->funcs->set_powergating_state)
1731                         continue;
1732                 r = adev->ip_blocks[i].version->funcs->set_powergating_state(
1733                         (void *)adev, state);
1734                 if (r)
1735                         DRM_ERROR("set_powergating_state of IP block <%s> failed %d\n",
1736                                   adev->ip_blocks[i].version->funcs->name, r);
1737         }
1738         return r;
1739 }
1740
1741 /**
1742  * amdgpu_device_ip_get_clockgating_state - get the CG state
1743  *
1744  * @adev: amdgpu_device pointer
1745  * @flags: clockgating feature flags
1746  *
1747  * Walks the list of IPs on the device and updates the clockgating
1748  * flags for each IP.
1749  * Updates @flags with the feature flags for each hardware IP where
1750  * clockgating is enabled.
1751  */
1752 void amdgpu_device_ip_get_clockgating_state(struct amdgpu_device *adev,
1753                                             u64 *flags)
1754 {
1755         int i;
1756
1757         for (i = 0; i < adev->num_ip_blocks; i++) {
1758                 if (!adev->ip_blocks[i].status.valid)
1759                         continue;
1760                 if (adev->ip_blocks[i].version->funcs->get_clockgating_state)
1761                         adev->ip_blocks[i].version->funcs->get_clockgating_state((void *)adev, flags);
1762         }
1763 }
1764
1765 /**
1766  * amdgpu_device_ip_wait_for_idle - wait for idle
1767  *
1768  * @adev: amdgpu_device pointer
1769  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1770  *
1771  * Waits for the request hardware IP to be idle.
1772  * Returns 0 for success or a negative error code on failure.
1773  */
1774 int amdgpu_device_ip_wait_for_idle(struct amdgpu_device *adev,
1775                                    enum amd_ip_block_type block_type)
1776 {
1777         int i, r;
1778
1779         for (i = 0; i < adev->num_ip_blocks; i++) {
1780                 if (!adev->ip_blocks[i].status.valid)
1781                         continue;
1782                 if (adev->ip_blocks[i].version->type == block_type) {
1783                         r = adev->ip_blocks[i].version->funcs->wait_for_idle((void *)adev);
1784                         if (r)
1785                                 return r;
1786                         break;
1787                 }
1788         }
1789         return 0;
1790
1791 }
1792
1793 /**
1794  * amdgpu_device_ip_is_idle - is the hardware IP idle
1795  *
1796  * @adev: amdgpu_device pointer
1797  * @block_type: Type of hardware IP (SMU, GFX, UVD, etc.)
1798  *
1799  * Check if the hardware IP is idle or not.
1800  * Returns true if it the IP is idle, false if not.
1801  */
1802 bool amdgpu_device_ip_is_idle(struct amdgpu_device *adev,
1803                               enum amd_ip_block_type block_type)
1804 {
1805         int i;
1806
1807         for (i = 0; i < adev->num_ip_blocks; i++) {
1808                 if (!adev->ip_blocks[i].status.valid)
1809                         continue;
1810                 if (adev->ip_blocks[i].version->type == block_type)
1811                         return adev->ip_blocks[i].version->funcs->is_idle((void *)adev);
1812         }
1813         return true;
1814
1815 }
1816
1817 /**
1818  * amdgpu_device_ip_get_ip_block - get a hw IP pointer
1819  *
1820  * @adev: amdgpu_device pointer
1821  * @type: Type of hardware IP (SMU, GFX, UVD, etc.)
1822  *
1823  * Returns a pointer to the hardware IP block structure
1824  * if it exists for the asic, otherwise NULL.
1825  */
1826 struct amdgpu_ip_block *
1827 amdgpu_device_ip_get_ip_block(struct amdgpu_device *adev,
1828                               enum amd_ip_block_type type)
1829 {
1830         int i;
1831
1832         for (i = 0; i < adev->num_ip_blocks; i++)
1833                 if (adev->ip_blocks[i].version->type == type)
1834                         return &adev->ip_blocks[i];
1835
1836         return NULL;
1837 }
1838
1839 /**
1840  * amdgpu_device_ip_block_version_cmp
1841  *
1842  * @adev: amdgpu_device pointer
1843  * @type: enum amd_ip_block_type
1844  * @major: major version
1845  * @minor: minor version
1846  *
1847  * return 0 if equal or greater
1848  * return 1 if smaller or the ip_block doesn't exist
1849  */
1850 int amdgpu_device_ip_block_version_cmp(struct amdgpu_device *adev,
1851                                        enum amd_ip_block_type type,
1852                                        u32 major, u32 minor)
1853 {
1854         struct amdgpu_ip_block *ip_block = amdgpu_device_ip_get_ip_block(adev, type);
1855
1856         if (ip_block && ((ip_block->version->major > major) ||
1857                         ((ip_block->version->major == major) &&
1858                         (ip_block->version->minor >= minor))))
1859                 return 0;
1860
1861         return 1;
1862 }
1863
1864 /**
1865  * amdgpu_device_ip_block_add
1866  *
1867  * @adev: amdgpu_device pointer
1868  * @ip_block_version: pointer to the IP to add
1869  *
1870  * Adds the IP block driver information to the collection of IPs
1871  * on the asic.
1872  */
1873 int amdgpu_device_ip_block_add(struct amdgpu_device *adev,
1874                                const struct amdgpu_ip_block_version *ip_block_version)
1875 {
1876         if (!ip_block_version)
1877                 return -EINVAL;
1878
1879         switch (ip_block_version->type) {
1880         case AMD_IP_BLOCK_TYPE_VCN:
1881                 if (adev->harvest_ip_mask & AMD_HARVEST_IP_VCN_MASK)
1882                         return 0;
1883                 break;
1884         case AMD_IP_BLOCK_TYPE_JPEG:
1885                 if (adev->harvest_ip_mask & AMD_HARVEST_IP_JPEG_MASK)
1886                         return 0;
1887                 break;
1888         default:
1889                 break;
1890         }
1891
1892         DRM_INFO("add ip block number %d <%s>\n", adev->num_ip_blocks,
1893                   ip_block_version->funcs->name);
1894
1895         adev->ip_blocks[adev->num_ip_blocks++].version = ip_block_version;
1896
1897         return 0;
1898 }
1899
1900 /**
1901  * amdgpu_device_enable_virtual_display - enable virtual display feature
1902  *
1903  * @adev: amdgpu_device pointer
1904  *
1905  * Enabled the virtual display feature if the user has enabled it via
1906  * the module parameter virtual_display.  This feature provides a virtual
1907  * display hardware on headless boards or in virtualized environments.
1908  * This function parses and validates the configuration string specified by
1909  * the user and configues the virtual display configuration (number of
1910  * virtual connectors, crtcs, etc.) specified.
1911  */
1912 static void amdgpu_device_enable_virtual_display(struct amdgpu_device *adev)
1913 {
1914         adev->enable_virtual_display = false;
1915
1916         if (amdgpu_virtual_display) {
1917                 const char *pci_address_name = pci_name(adev->pdev);
1918                 char *pciaddstr, *pciaddstr_tmp, *pciaddname_tmp, *pciaddname;
1919
1920                 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1921                 pciaddstr_tmp = pciaddstr;
1922                 while ((pciaddname_tmp = strsep(&pciaddstr_tmp, ";"))) {
1923                         pciaddname = strsep(&pciaddname_tmp, ",");
1924                         if (!strcmp("all", pciaddname)
1925                             || !strcmp(pci_address_name, pciaddname)) {
1926                                 long num_crtc;
1927                                 int res = -1;
1928
1929                                 adev->enable_virtual_display = true;
1930
1931                                 if (pciaddname_tmp)
1932                                         res = kstrtol(pciaddname_tmp, 10,
1933                                                       &num_crtc);
1934
1935                                 if (!res) {
1936                                         if (num_crtc < 1)
1937                                                 num_crtc = 1;
1938                                         if (num_crtc > 6)
1939                                                 num_crtc = 6;
1940                                         adev->mode_info.num_crtc = num_crtc;
1941                                 } else {
1942                                         adev->mode_info.num_crtc = 1;
1943                                 }
1944                                 break;
1945                         }
1946                 }
1947
1948                 DRM_INFO("virtual display string:%s, %s:virtual_display:%d, num_crtc:%d\n",
1949                          amdgpu_virtual_display, pci_address_name,
1950                          adev->enable_virtual_display, adev->mode_info.num_crtc);
1951
1952                 kfree(pciaddstr);
1953         }
1954 }
1955
1956 void amdgpu_device_set_sriov_virtual_display(struct amdgpu_device *adev)
1957 {
1958         if (amdgpu_sriov_vf(adev) && !adev->enable_virtual_display) {
1959                 adev->mode_info.num_crtc = 1;
1960                 adev->enable_virtual_display = true;
1961                 DRM_INFO("virtual_display:%d, num_crtc:%d\n",
1962                          adev->enable_virtual_display, adev->mode_info.num_crtc);
1963         }
1964 }
1965
1966 /**
1967  * amdgpu_device_parse_gpu_info_fw - parse gpu info firmware
1968  *
1969  * @adev: amdgpu_device pointer
1970  *
1971  * Parses the asic configuration parameters specified in the gpu info
1972  * firmware and makes them availale to the driver for use in configuring
1973  * the asic.
1974  * Returns 0 on success, -EINVAL on failure.
1975  */
1976 static int amdgpu_device_parse_gpu_info_fw(struct amdgpu_device *adev)
1977 {
1978         const char *chip_name;
1979         char fw_name[40];
1980         int err;
1981         const struct gpu_info_firmware_header_v1_0 *hdr;
1982
1983         adev->firmware.gpu_info_fw = NULL;
1984
1985         if (adev->mman.discovery_bin) {
1986                 /*
1987                  * FIXME: The bounding box is still needed by Navi12, so
1988                  * temporarily read it from gpu_info firmware. Should be dropped
1989                  * when DAL no longer needs it.
1990                  */
1991                 if (adev->asic_type != CHIP_NAVI12)
1992                         return 0;
1993         }
1994
1995         switch (adev->asic_type) {
1996         default:
1997                 return 0;
1998         case CHIP_VEGA10:
1999                 chip_name = "vega10";
2000                 break;
2001         case CHIP_VEGA12:
2002                 chip_name = "vega12";
2003                 break;
2004         case CHIP_RAVEN:
2005                 if (adev->apu_flags & AMD_APU_IS_RAVEN2)
2006                         chip_name = "raven2";
2007                 else if (adev->apu_flags & AMD_APU_IS_PICASSO)
2008                         chip_name = "picasso";
2009                 else
2010                         chip_name = "raven";
2011                 break;
2012         case CHIP_ARCTURUS:
2013                 chip_name = "arcturus";
2014                 break;
2015         case CHIP_NAVI12:
2016                 chip_name = "navi12";
2017                 break;
2018         }
2019
2020         snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_gpu_info.bin", chip_name);
2021         err = amdgpu_ucode_request(adev, &adev->firmware.gpu_info_fw, fw_name);
2022         if (err) {
2023                 dev_err(adev->dev,
2024                         "Failed to get gpu_info firmware \"%s\"\n",
2025                         fw_name);
2026                 goto out;
2027         }
2028
2029         hdr = (const struct gpu_info_firmware_header_v1_0 *)adev->firmware.gpu_info_fw->data;
2030         amdgpu_ucode_print_gpu_info_hdr(&hdr->header);
2031
2032         switch (hdr->version_major) {
2033         case 1:
2034         {
2035                 const struct gpu_info_firmware_v1_0 *gpu_info_fw =
2036                         (const struct gpu_info_firmware_v1_0 *)(adev->firmware.gpu_info_fw->data +
2037                                                                 le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2038
2039                 /*
2040                  * Should be droped when DAL no longer needs it.
2041                  */
2042                 if (adev->asic_type == CHIP_NAVI12)
2043                         goto parse_soc_bounding_box;
2044
2045                 adev->gfx.config.max_shader_engines = le32_to_cpu(gpu_info_fw->gc_num_se);
2046                 adev->gfx.config.max_cu_per_sh = le32_to_cpu(gpu_info_fw->gc_num_cu_per_sh);
2047                 adev->gfx.config.max_sh_per_se = le32_to_cpu(gpu_info_fw->gc_num_sh_per_se);
2048                 adev->gfx.config.max_backends_per_se = le32_to_cpu(gpu_info_fw->gc_num_rb_per_se);
2049                 adev->gfx.config.max_texture_channel_caches =
2050                         le32_to_cpu(gpu_info_fw->gc_num_tccs);
2051                 adev->gfx.config.max_gprs = le32_to_cpu(gpu_info_fw->gc_num_gprs);
2052                 adev->gfx.config.max_gs_threads = le32_to_cpu(gpu_info_fw->gc_num_max_gs_thds);
2053                 adev->gfx.config.gs_vgt_table_depth = le32_to_cpu(gpu_info_fw->gc_gs_table_depth);
2054                 adev->gfx.config.gs_prim_buffer_depth = le32_to_cpu(gpu_info_fw->gc_gsprim_buff_depth);
2055                 adev->gfx.config.double_offchip_lds_buf =
2056                         le32_to_cpu(gpu_info_fw->gc_double_offchip_lds_buffer);
2057                 adev->gfx.cu_info.wave_front_size = le32_to_cpu(gpu_info_fw->gc_wave_size);
2058                 adev->gfx.cu_info.max_waves_per_simd =
2059                         le32_to_cpu(gpu_info_fw->gc_max_waves_per_simd);
2060                 adev->gfx.cu_info.max_scratch_slots_per_cu =
2061                         le32_to_cpu(gpu_info_fw->gc_max_scratch_slots_per_cu);
2062                 adev->gfx.cu_info.lds_size = le32_to_cpu(gpu_info_fw->gc_lds_size);
2063                 if (hdr->version_minor >= 1) {
2064                         const struct gpu_info_firmware_v1_1 *gpu_info_fw =
2065                                 (const struct gpu_info_firmware_v1_1 *)(adev->firmware.gpu_info_fw->data +
2066                                                                         le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2067                         adev->gfx.config.num_sc_per_sh =
2068                                 le32_to_cpu(gpu_info_fw->num_sc_per_sh);
2069                         adev->gfx.config.num_packer_per_sc =
2070                                 le32_to_cpu(gpu_info_fw->num_packer_per_sc);
2071                 }
2072
2073 parse_soc_bounding_box:
2074                 /*
2075                  * soc bounding box info is not integrated in disocovery table,
2076                  * we always need to parse it from gpu info firmware if needed.
2077                  */
2078                 if (hdr->version_minor == 2) {
2079                         const struct gpu_info_firmware_v1_2 *gpu_info_fw =
2080                                 (const struct gpu_info_firmware_v1_2 *)(adev->firmware.gpu_info_fw->data +
2081                                                                         le32_to_cpu(hdr->header.ucode_array_offset_bytes));
2082                         adev->dm.soc_bounding_box = &gpu_info_fw->soc_bounding_box;
2083                 }
2084                 break;
2085         }
2086         default:
2087                 dev_err(adev->dev,
2088                         "Unsupported gpu_info table %d\n", hdr->header.ucode_version);
2089                 err = -EINVAL;
2090                 goto out;
2091         }
2092 out:
2093         return err;
2094 }
2095
2096 /**
2097  * amdgpu_device_ip_early_init - run early init for hardware IPs
2098  *
2099  * @adev: amdgpu_device pointer
2100  *
2101  * Early initialization pass for hardware IPs.  The hardware IPs that make
2102  * up each asic are discovered each IP's early_init callback is run.  This
2103  * is the first stage in initializing the asic.
2104  * Returns 0 on success, negative error code on failure.
2105  */
2106 static int amdgpu_device_ip_early_init(struct amdgpu_device *adev)
2107 {
2108         struct drm_device *dev = adev_to_drm(adev);
2109         struct pci_dev *parent;
2110         int i, r;
2111         bool total;
2112
2113         amdgpu_device_enable_virtual_display(adev);
2114
2115         if (amdgpu_sriov_vf(adev)) {
2116                 r = amdgpu_virt_request_full_gpu(adev, true);
2117                 if (r)
2118                         return r;
2119         }
2120
2121         switch (adev->asic_type) {
2122 #ifdef CONFIG_DRM_AMDGPU_SI
2123         case CHIP_VERDE:
2124         case CHIP_TAHITI:
2125         case CHIP_PITCAIRN:
2126         case CHIP_OLAND:
2127         case CHIP_HAINAN:
2128                 adev->family = AMDGPU_FAMILY_SI;
2129                 r = si_set_ip_blocks(adev);
2130                 if (r)
2131                         return r;
2132                 break;
2133 #endif
2134 #ifdef CONFIG_DRM_AMDGPU_CIK
2135         case CHIP_BONAIRE:
2136         case CHIP_HAWAII:
2137         case CHIP_KAVERI:
2138         case CHIP_KABINI:
2139         case CHIP_MULLINS:
2140                 if (adev->flags & AMD_IS_APU)
2141                         adev->family = AMDGPU_FAMILY_KV;
2142                 else
2143                         adev->family = AMDGPU_FAMILY_CI;
2144
2145                 r = cik_set_ip_blocks(adev);
2146                 if (r)
2147                         return r;
2148                 break;
2149 #endif
2150         case CHIP_TOPAZ:
2151         case CHIP_TONGA:
2152         case CHIP_FIJI:
2153         case CHIP_POLARIS10:
2154         case CHIP_POLARIS11:
2155         case CHIP_POLARIS12:
2156         case CHIP_VEGAM:
2157         case CHIP_CARRIZO:
2158         case CHIP_STONEY:
2159                 if (adev->flags & AMD_IS_APU)
2160                         adev->family = AMDGPU_FAMILY_CZ;
2161                 else
2162                         adev->family = AMDGPU_FAMILY_VI;
2163
2164                 r = vi_set_ip_blocks(adev);
2165                 if (r)
2166                         return r;
2167                 break;
2168         default:
2169                 r = amdgpu_discovery_set_ip_blocks(adev);
2170                 if (r)
2171                         return r;
2172                 break;
2173         }
2174
2175         if (amdgpu_has_atpx() &&
2176             (amdgpu_is_atpx_hybrid() ||
2177              amdgpu_has_atpx_dgpu_power_cntl()) &&
2178             ((adev->flags & AMD_IS_APU) == 0) &&
2179             !pci_is_thunderbolt_attached(to_pci_dev(dev->dev)))
2180                 adev->flags |= AMD_IS_PX;
2181
2182         if (!(adev->flags & AMD_IS_APU)) {
2183                 parent = pci_upstream_bridge(adev->pdev);
2184                 adev->has_pr3 = parent ? pci_pr3_present(parent) : false;
2185         }
2186
2187
2188         adev->pm.pp_feature = amdgpu_pp_feature_mask;
2189         if (amdgpu_sriov_vf(adev) || sched_policy == KFD_SCHED_POLICY_NO_HWS)
2190                 adev->pm.pp_feature &= ~PP_GFXOFF_MASK;
2191         if (amdgpu_sriov_vf(adev) && adev->asic_type == CHIP_SIENNA_CICHLID)
2192                 adev->pm.pp_feature &= ~PP_OVERDRIVE_MASK;
2193
2194         total = true;
2195         for (i = 0; i < adev->num_ip_blocks; i++) {
2196                 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
2197                         DRM_ERROR("disabled ip block: %d <%s>\n",
2198                                   i, adev->ip_blocks[i].version->funcs->name);
2199                         adev->ip_blocks[i].status.valid = false;
2200                 } else {
2201                         if (adev->ip_blocks[i].version->funcs->early_init) {
2202                                 r = adev->ip_blocks[i].version->funcs->early_init((void *)adev);
2203                                 if (r == -ENOENT) {
2204                                         adev->ip_blocks[i].status.valid = false;
2205                                 } else if (r) {
2206                                         DRM_ERROR("early_init of IP block <%s> failed %d\n",
2207                                                   adev->ip_blocks[i].version->funcs->name, r);
2208                                         total = false;
2209                                 } else {
2210                                         adev->ip_blocks[i].status.valid = true;
2211                                 }
2212                         } else {
2213                                 adev->ip_blocks[i].status.valid = true;
2214                         }
2215                 }
2216                 /* get the vbios after the asic_funcs are set up */
2217                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) {
2218                         r = amdgpu_device_parse_gpu_info_fw(adev);
2219                         if (r)
2220                                 return r;
2221
2222                         /* Read BIOS */
2223                         if (!amdgpu_get_bios(adev))
2224                                 return -EINVAL;
2225
2226                         r = amdgpu_atombios_init(adev);
2227                         if (r) {
2228                                 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
2229                                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_INIT_FAIL, 0, 0);
2230                                 return r;
2231                         }
2232
2233                         /*get pf2vf msg info at it's earliest time*/
2234                         if (amdgpu_sriov_vf(adev))
2235                                 amdgpu_virt_init_data_exchange(adev);
2236
2237                 }
2238         }
2239         if (!total)
2240                 return -ENODEV;
2241
2242         amdgpu_amdkfd_device_probe(adev);
2243         adev->cg_flags &= amdgpu_cg_mask;
2244         adev->pg_flags &= amdgpu_pg_mask;
2245
2246         return 0;
2247 }
2248
2249 static int amdgpu_device_ip_hw_init_phase1(struct amdgpu_device *adev)
2250 {
2251         int i, r;
2252
2253         for (i = 0; i < adev->num_ip_blocks; i++) {
2254                 if (!adev->ip_blocks[i].status.sw)
2255                         continue;
2256                 if (adev->ip_blocks[i].status.hw)
2257                         continue;
2258                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
2259                     (amdgpu_sriov_vf(adev) && (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)) ||
2260                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH) {
2261                         r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2262                         if (r) {
2263                                 DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2264                                           adev->ip_blocks[i].version->funcs->name, r);
2265                                 return r;
2266                         }
2267                         adev->ip_blocks[i].status.hw = true;
2268                 }
2269         }
2270
2271         return 0;
2272 }
2273
2274 static int amdgpu_device_ip_hw_init_phase2(struct amdgpu_device *adev)
2275 {
2276         int i, r;
2277
2278         for (i = 0; i < adev->num_ip_blocks; i++) {
2279                 if (!adev->ip_blocks[i].status.sw)
2280                         continue;
2281                 if (adev->ip_blocks[i].status.hw)
2282                         continue;
2283                 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2284                 if (r) {
2285                         DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2286                                   adev->ip_blocks[i].version->funcs->name, r);
2287                         return r;
2288                 }
2289                 adev->ip_blocks[i].status.hw = true;
2290         }
2291
2292         return 0;
2293 }
2294
2295 static int amdgpu_device_fw_loading(struct amdgpu_device *adev)
2296 {
2297         int r = 0;
2298         int i;
2299         uint32_t smu_version;
2300
2301         if (adev->asic_type >= CHIP_VEGA10) {
2302                 for (i = 0; i < adev->num_ip_blocks; i++) {
2303                         if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_PSP)
2304                                 continue;
2305
2306                         if (!adev->ip_blocks[i].status.sw)
2307                                 continue;
2308
2309                         /* no need to do the fw loading again if already done*/
2310                         if (adev->ip_blocks[i].status.hw == true)
2311                                 break;
2312
2313                         if (amdgpu_in_reset(adev) || adev->in_suspend) {
2314                                 r = adev->ip_blocks[i].version->funcs->resume(adev);
2315                                 if (r) {
2316                                         DRM_ERROR("resume of IP block <%s> failed %d\n",
2317                                                           adev->ip_blocks[i].version->funcs->name, r);
2318                                         return r;
2319                                 }
2320                         } else {
2321                                 r = adev->ip_blocks[i].version->funcs->hw_init(adev);
2322                                 if (r) {
2323                                         DRM_ERROR("hw_init of IP block <%s> failed %d\n",
2324                                                           adev->ip_blocks[i].version->funcs->name, r);
2325                                         return r;
2326                                 }
2327                         }
2328
2329                         adev->ip_blocks[i].status.hw = true;
2330                         break;
2331                 }
2332         }
2333
2334         if (!amdgpu_sriov_vf(adev) || adev->asic_type == CHIP_TONGA)
2335                 r = amdgpu_pm_load_smu_firmware(adev, &smu_version);
2336
2337         return r;
2338 }
2339
2340 static int amdgpu_device_init_schedulers(struct amdgpu_device *adev)
2341 {
2342         long timeout;
2343         int r, i;
2344
2345         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2346                 struct amdgpu_ring *ring = adev->rings[i];
2347
2348                 /* No need to setup the GPU scheduler for rings that don't need it */
2349                 if (!ring || ring->no_scheduler)
2350                         continue;
2351
2352                 switch (ring->funcs->type) {
2353                 case AMDGPU_RING_TYPE_GFX:
2354                         timeout = adev->gfx_timeout;
2355                         break;
2356                 case AMDGPU_RING_TYPE_COMPUTE:
2357                         timeout = adev->compute_timeout;
2358                         break;
2359                 case AMDGPU_RING_TYPE_SDMA:
2360                         timeout = adev->sdma_timeout;
2361                         break;
2362                 default:
2363                         timeout = adev->video_timeout;
2364                         break;
2365                 }
2366
2367                 r = drm_sched_init(&ring->sched, &amdgpu_sched_ops,
2368                                    ring->num_hw_submission, 0,
2369                                    timeout, adev->reset_domain->wq,
2370                                    ring->sched_score, ring->name,
2371                                    adev->dev);
2372                 if (r) {
2373                         DRM_ERROR("Failed to create scheduler on ring %s.\n",
2374                                   ring->name);
2375                         return r;
2376                 }
2377         }
2378
2379         return 0;
2380 }
2381
2382
2383 /**
2384  * amdgpu_device_ip_init - run init for hardware IPs
2385  *
2386  * @adev: amdgpu_device pointer
2387  *
2388  * Main initialization pass for hardware IPs.  The list of all the hardware
2389  * IPs that make up the asic is walked and the sw_init and hw_init callbacks
2390  * are run.  sw_init initializes the software state associated with each IP
2391  * and hw_init initializes the hardware associated with each IP.
2392  * Returns 0 on success, negative error code on failure.
2393  */
2394 static int amdgpu_device_ip_init(struct amdgpu_device *adev)
2395 {
2396         int i, r;
2397
2398         r = amdgpu_ras_init(adev);
2399         if (r)
2400                 return r;
2401
2402         for (i = 0; i < adev->num_ip_blocks; i++) {
2403                 if (!adev->ip_blocks[i].status.valid)
2404                         continue;
2405                 r = adev->ip_blocks[i].version->funcs->sw_init((void *)adev);
2406                 if (r) {
2407                         DRM_ERROR("sw_init of IP block <%s> failed %d\n",
2408                                   adev->ip_blocks[i].version->funcs->name, r);
2409                         goto init_failed;
2410                 }
2411                 adev->ip_blocks[i].status.sw = true;
2412
2413                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON) {
2414                         /* need to do common hw init early so everything is set up for gmc */
2415                         r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
2416                         if (r) {
2417                                 DRM_ERROR("hw_init %d failed %d\n", i, r);
2418                                 goto init_failed;
2419                         }
2420                         adev->ip_blocks[i].status.hw = true;
2421                 } else if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
2422                         /* need to do gmc hw init early so we can allocate gpu mem */
2423                         /* Try to reserve bad pages early */
2424                         if (amdgpu_sriov_vf(adev))
2425                                 amdgpu_virt_exchange_data(adev);
2426
2427                         r = amdgpu_device_mem_scratch_init(adev);
2428                         if (r) {
2429                                 DRM_ERROR("amdgpu_mem_scratch_init failed %d\n", r);
2430                                 goto init_failed;
2431                         }
2432                         r = adev->ip_blocks[i].version->funcs->hw_init((void *)adev);
2433                         if (r) {
2434                                 DRM_ERROR("hw_init %d failed %d\n", i, r);
2435                                 goto init_failed;
2436                         }
2437                         r = amdgpu_device_wb_init(adev);
2438                         if (r) {
2439                                 DRM_ERROR("amdgpu_device_wb_init failed %d\n", r);
2440                                 goto init_failed;
2441                         }
2442                         adev->ip_blocks[i].status.hw = true;
2443
2444                         /* right after GMC hw init, we create CSA */
2445                         if (amdgpu_mcbp) {
2446                                 r = amdgpu_allocate_static_csa(adev, &adev->virt.csa_obj,
2447                                                                AMDGPU_GEM_DOMAIN_VRAM |
2448                                                                AMDGPU_GEM_DOMAIN_GTT,
2449                                                                AMDGPU_CSA_SIZE);
2450                                 if (r) {
2451                                         DRM_ERROR("allocate CSA failed %d\n", r);
2452                                         goto init_failed;
2453                                 }
2454                         }
2455                 }
2456         }
2457
2458         if (amdgpu_sriov_vf(adev))
2459                 amdgpu_virt_init_data_exchange(adev);
2460
2461         r = amdgpu_ib_pool_init(adev);
2462         if (r) {
2463                 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
2464                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_IB_INIT_FAIL, 0, r);
2465                 goto init_failed;
2466         }
2467
2468         r = amdgpu_ucode_create_bo(adev); /* create ucode bo when sw_init complete*/
2469         if (r)
2470                 goto init_failed;
2471
2472         r = amdgpu_device_ip_hw_init_phase1(adev);
2473         if (r)
2474                 goto init_failed;
2475
2476         r = amdgpu_device_fw_loading(adev);
2477         if (r)
2478                 goto init_failed;
2479
2480         r = amdgpu_device_ip_hw_init_phase2(adev);
2481         if (r)
2482                 goto init_failed;
2483
2484         /*
2485          * retired pages will be loaded from eeprom and reserved here,
2486          * it should be called after amdgpu_device_ip_hw_init_phase2  since
2487          * for some ASICs the RAS EEPROM code relies on SMU fully functioning
2488          * for I2C communication which only true at this point.
2489          *
2490          * amdgpu_ras_recovery_init may fail, but the upper only cares the
2491          * failure from bad gpu situation and stop amdgpu init process
2492          * accordingly. For other failed cases, it will still release all
2493          * the resource and print error message, rather than returning one
2494          * negative value to upper level.
2495          *
2496          * Note: theoretically, this should be called before all vram allocations
2497          * to protect retired page from abusing
2498          */
2499         r = amdgpu_ras_recovery_init(adev);
2500         if (r)
2501                 goto init_failed;
2502
2503         /**
2504          * In case of XGMI grab extra reference for reset domain for this device
2505          */
2506         if (adev->gmc.xgmi.num_physical_nodes > 1) {
2507                 if (amdgpu_xgmi_add_device(adev) == 0) {
2508                         if (!amdgpu_sriov_vf(adev)) {
2509                                 struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
2510
2511                                 if (WARN_ON(!hive)) {
2512                                         r = -ENOENT;
2513                                         goto init_failed;
2514                                 }
2515
2516                                 if (!hive->reset_domain ||
2517                                     !amdgpu_reset_get_reset_domain(hive->reset_domain)) {
2518                                         r = -ENOENT;
2519                                         amdgpu_put_xgmi_hive(hive);
2520                                         goto init_failed;
2521                                 }
2522
2523                                 /* Drop the early temporary reset domain we created for device */
2524                                 amdgpu_reset_put_reset_domain(adev->reset_domain);
2525                                 adev->reset_domain = hive->reset_domain;
2526                                 amdgpu_put_xgmi_hive(hive);
2527                         }
2528                 }
2529         }
2530
2531         r = amdgpu_device_init_schedulers(adev);
2532         if (r)
2533                 goto init_failed;
2534
2535         /* Don't init kfd if whole hive need to be reset during init */
2536         if (!adev->gmc.xgmi.pending_reset)
2537                 amdgpu_amdkfd_device_init(adev);
2538
2539         amdgpu_fru_get_product_info(adev);
2540
2541 init_failed:
2542
2543         return r;
2544 }
2545
2546 /**
2547  * amdgpu_device_fill_reset_magic - writes reset magic to gart pointer
2548  *
2549  * @adev: amdgpu_device pointer
2550  *
2551  * Writes a reset magic value to the gart pointer in VRAM.  The driver calls
2552  * this function before a GPU reset.  If the value is retained after a
2553  * GPU reset, VRAM has not been lost.  Some GPU resets may destry VRAM contents.
2554  */
2555 static void amdgpu_device_fill_reset_magic(struct amdgpu_device *adev)
2556 {
2557         memcpy(adev->reset_magic, adev->gart.ptr, AMDGPU_RESET_MAGIC_NUM);
2558 }
2559
2560 /**
2561  * amdgpu_device_check_vram_lost - check if vram is valid
2562  *
2563  * @adev: amdgpu_device pointer
2564  *
2565  * Checks the reset magic value written to the gart pointer in VRAM.
2566  * The driver calls this after a GPU reset to see if the contents of
2567  * VRAM is lost or now.
2568  * returns true if vram is lost, false if not.
2569  */
2570 static bool amdgpu_device_check_vram_lost(struct amdgpu_device *adev)
2571 {
2572         if (memcmp(adev->gart.ptr, adev->reset_magic,
2573                         AMDGPU_RESET_MAGIC_NUM))
2574                 return true;
2575
2576         if (!amdgpu_in_reset(adev))
2577                 return false;
2578
2579         /*
2580          * For all ASICs with baco/mode1 reset, the VRAM is
2581          * always assumed to be lost.
2582          */
2583         switch (amdgpu_asic_reset_method(adev)) {
2584         case AMD_RESET_METHOD_BACO:
2585         case AMD_RESET_METHOD_MODE1:
2586                 return true;
2587         default:
2588                 return false;
2589         }
2590 }
2591
2592 /**
2593  * amdgpu_device_set_cg_state - set clockgating for amdgpu device
2594  *
2595  * @adev: amdgpu_device pointer
2596  * @state: clockgating state (gate or ungate)
2597  *
2598  * The list of all the hardware IPs that make up the asic is walked and the
2599  * set_clockgating_state callbacks are run.
2600  * Late initialization pass enabling clockgating for hardware IPs.
2601  * Fini or suspend, pass disabling clockgating for hardware IPs.
2602  * Returns 0 on success, negative error code on failure.
2603  */
2604
2605 int amdgpu_device_set_cg_state(struct amdgpu_device *adev,
2606                                enum amd_clockgating_state state)
2607 {
2608         int i, j, r;
2609
2610         if (amdgpu_emu_mode == 1)
2611                 return 0;
2612
2613         for (j = 0; j < adev->num_ip_blocks; j++) {
2614                 i = state == AMD_CG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2615                 if (!adev->ip_blocks[i].status.late_initialized)
2616                         continue;
2617                 /* skip CG for GFX, SDMA on S0ix */
2618                 if (adev->in_s0ix &&
2619                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX ||
2620                      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA))
2621                         continue;
2622                 /* skip CG for VCE/UVD, it's handled specially */
2623                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2624                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2625                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2626                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2627                     adev->ip_blocks[i].version->funcs->set_clockgating_state) {
2628                         /* enable clockgating to save power */
2629                         r = adev->ip_blocks[i].version->funcs->set_clockgating_state((void *)adev,
2630                                                                                      state);
2631                         if (r) {
2632                                 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
2633                                           adev->ip_blocks[i].version->funcs->name, r);
2634                                 return r;
2635                         }
2636                 }
2637         }
2638
2639         return 0;
2640 }
2641
2642 int amdgpu_device_set_pg_state(struct amdgpu_device *adev,
2643                                enum amd_powergating_state state)
2644 {
2645         int i, j, r;
2646
2647         if (amdgpu_emu_mode == 1)
2648                 return 0;
2649
2650         for (j = 0; j < adev->num_ip_blocks; j++) {
2651                 i = state == AMD_PG_STATE_GATE ? j : adev->num_ip_blocks - j - 1;
2652                 if (!adev->ip_blocks[i].status.late_initialized)
2653                         continue;
2654                 /* skip PG for GFX, SDMA on S0ix */
2655                 if (adev->in_s0ix &&
2656                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX ||
2657                      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA))
2658                         continue;
2659                 /* skip CG for VCE/UVD, it's handled specially */
2660                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_UVD &&
2661                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCE &&
2662                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_VCN &&
2663                     adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_JPEG &&
2664                     adev->ip_blocks[i].version->funcs->set_powergating_state) {
2665                         /* enable powergating to save power */
2666                         r = adev->ip_blocks[i].version->funcs->set_powergating_state((void *)adev,
2667                                                                                         state);
2668                         if (r) {
2669                                 DRM_ERROR("set_powergating_state(gate) of IP block <%s> failed %d\n",
2670                                           adev->ip_blocks[i].version->funcs->name, r);
2671                                 return r;
2672                         }
2673                 }
2674         }
2675         return 0;
2676 }
2677
2678 static int amdgpu_device_enable_mgpu_fan_boost(void)
2679 {
2680         struct amdgpu_gpu_instance *gpu_ins;
2681         struct amdgpu_device *adev;
2682         int i, ret = 0;
2683
2684         mutex_lock(&mgpu_info.mutex);
2685
2686         /*
2687          * MGPU fan boost feature should be enabled
2688          * only when there are two or more dGPUs in
2689          * the system
2690          */
2691         if (mgpu_info.num_dgpu < 2)
2692                 goto out;
2693
2694         for (i = 0; i < mgpu_info.num_dgpu; i++) {
2695                 gpu_ins = &(mgpu_info.gpu_ins[i]);
2696                 adev = gpu_ins->adev;
2697                 if (!(adev->flags & AMD_IS_APU) &&
2698                     !gpu_ins->mgpu_fan_enabled) {
2699                         ret = amdgpu_dpm_enable_mgpu_fan_boost(adev);
2700                         if (ret)
2701                                 break;
2702
2703                         gpu_ins->mgpu_fan_enabled = 1;
2704                 }
2705         }
2706
2707 out:
2708         mutex_unlock(&mgpu_info.mutex);
2709
2710         return ret;
2711 }
2712
2713 /**
2714  * amdgpu_device_ip_late_init - run late init for hardware IPs
2715  *
2716  * @adev: amdgpu_device pointer
2717  *
2718  * Late initialization pass for hardware IPs.  The list of all the hardware
2719  * IPs that make up the asic is walked and the late_init callbacks are run.
2720  * late_init covers any special initialization that an IP requires
2721  * after all of the have been initialized or something that needs to happen
2722  * late in the init process.
2723  * Returns 0 on success, negative error code on failure.
2724  */
2725 static int amdgpu_device_ip_late_init(struct amdgpu_device *adev)
2726 {
2727         struct amdgpu_gpu_instance *gpu_instance;
2728         int i = 0, r;
2729
2730         for (i = 0; i < adev->num_ip_blocks; i++) {
2731                 if (!adev->ip_blocks[i].status.hw)
2732                         continue;
2733                 if (adev->ip_blocks[i].version->funcs->late_init) {
2734                         r = adev->ip_blocks[i].version->funcs->late_init((void *)adev);
2735                         if (r) {
2736                                 DRM_ERROR("late_init of IP block <%s> failed %d\n",
2737                                           adev->ip_blocks[i].version->funcs->name, r);
2738                                 return r;
2739                         }
2740                 }
2741                 adev->ip_blocks[i].status.late_initialized = true;
2742         }
2743
2744         r = amdgpu_ras_late_init(adev);
2745         if (r) {
2746                 DRM_ERROR("amdgpu_ras_late_init failed %d", r);
2747                 return r;
2748         }
2749
2750         amdgpu_ras_set_error_query_ready(adev, true);
2751
2752         amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE);
2753         amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE);
2754
2755         amdgpu_device_fill_reset_magic(adev);
2756
2757         r = amdgpu_device_enable_mgpu_fan_boost();
2758         if (r)
2759                 DRM_ERROR("enable mgpu fan boost failed (%d).\n", r);
2760
2761         /* For passthrough configuration on arcturus and aldebaran, enable special handling SBR */
2762         if (amdgpu_passthrough(adev) &&
2763             ((adev->asic_type == CHIP_ARCTURUS && adev->gmc.xgmi.num_physical_nodes > 1) ||
2764              adev->asic_type == CHIP_ALDEBARAN))
2765                 amdgpu_dpm_handle_passthrough_sbr(adev, true);
2766
2767         if (adev->gmc.xgmi.num_physical_nodes > 1) {
2768                 mutex_lock(&mgpu_info.mutex);
2769
2770                 /*
2771                  * Reset device p-state to low as this was booted with high.
2772                  *
2773                  * This should be performed only after all devices from the same
2774                  * hive get initialized.
2775                  *
2776                  * However, it's unknown how many device in the hive in advance.
2777                  * As this is counted one by one during devices initializations.
2778                  *
2779                  * So, we wait for all XGMI interlinked devices initialized.
2780                  * This may bring some delays as those devices may come from
2781                  * different hives. But that should be OK.
2782                  */
2783                 if (mgpu_info.num_dgpu == adev->gmc.xgmi.num_physical_nodes) {
2784                         for (i = 0; i < mgpu_info.num_gpu; i++) {
2785                                 gpu_instance = &(mgpu_info.gpu_ins[i]);
2786                                 if (gpu_instance->adev->flags & AMD_IS_APU)
2787                                         continue;
2788
2789                                 r = amdgpu_xgmi_set_pstate(gpu_instance->adev,
2790                                                 AMDGPU_XGMI_PSTATE_MIN);
2791                                 if (r) {
2792                                         DRM_ERROR("pstate setting failed (%d).\n", r);
2793                                         break;
2794                                 }
2795                         }
2796                 }
2797
2798                 mutex_unlock(&mgpu_info.mutex);
2799         }
2800
2801         return 0;
2802 }
2803
2804 /**
2805  * amdgpu_device_smu_fini_early - smu hw_fini wrapper
2806  *
2807  * @adev: amdgpu_device pointer
2808  *
2809  * For ASICs need to disable SMC first
2810  */
2811 static void amdgpu_device_smu_fini_early(struct amdgpu_device *adev)
2812 {
2813         int i, r;
2814
2815         if (adev->ip_versions[GC_HWIP][0] > IP_VERSION(9, 0, 0))
2816                 return;
2817
2818         for (i = 0; i < adev->num_ip_blocks; i++) {
2819                 if (!adev->ip_blocks[i].status.hw)
2820                         continue;
2821                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
2822                         r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2823                         /* XXX handle errors */
2824                         if (r) {
2825                                 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2826                                           adev->ip_blocks[i].version->funcs->name, r);
2827                         }
2828                         adev->ip_blocks[i].status.hw = false;
2829                         break;
2830                 }
2831         }
2832 }
2833
2834 static int amdgpu_device_ip_fini_early(struct amdgpu_device *adev)
2835 {
2836         int i, r;
2837
2838         for (i = 0; i < adev->num_ip_blocks; i++) {
2839                 if (!adev->ip_blocks[i].version->funcs->early_fini)
2840                         continue;
2841
2842                 r = adev->ip_blocks[i].version->funcs->early_fini((void *)adev);
2843                 if (r) {
2844                         DRM_DEBUG("early_fini of IP block <%s> failed %d\n",
2845                                   adev->ip_blocks[i].version->funcs->name, r);
2846                 }
2847         }
2848
2849         amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2850         amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2851
2852         amdgpu_amdkfd_suspend(adev, false);
2853
2854         /* Workaroud for ASICs need to disable SMC first */
2855         amdgpu_device_smu_fini_early(adev);
2856
2857         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2858                 if (!adev->ip_blocks[i].status.hw)
2859                         continue;
2860
2861                 r = adev->ip_blocks[i].version->funcs->hw_fini((void *)adev);
2862                 /* XXX handle errors */
2863                 if (r) {
2864                         DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
2865                                   adev->ip_blocks[i].version->funcs->name, r);
2866                 }
2867
2868                 adev->ip_blocks[i].status.hw = false;
2869         }
2870
2871         if (amdgpu_sriov_vf(adev)) {
2872                 if (amdgpu_virt_release_full_gpu(adev, false))
2873                         DRM_ERROR("failed to release exclusive mode on fini\n");
2874         }
2875
2876         return 0;
2877 }
2878
2879 /**
2880  * amdgpu_device_ip_fini - run fini for hardware IPs
2881  *
2882  * @adev: amdgpu_device pointer
2883  *
2884  * Main teardown pass for hardware IPs.  The list of all the hardware
2885  * IPs that make up the asic is walked and the hw_fini and sw_fini callbacks
2886  * are run.  hw_fini tears down the hardware associated with each IP
2887  * and sw_fini tears down any software state associated with each IP.
2888  * Returns 0 on success, negative error code on failure.
2889  */
2890 static int amdgpu_device_ip_fini(struct amdgpu_device *adev)
2891 {
2892         int i, r;
2893
2894         if (amdgpu_sriov_vf(adev) && adev->virt.ras_init_done)
2895                 amdgpu_virt_release_ras_err_handler_data(adev);
2896
2897         if (adev->gmc.xgmi.num_physical_nodes > 1)
2898                 amdgpu_xgmi_remove_device(adev);
2899
2900         amdgpu_amdkfd_device_fini_sw(adev);
2901
2902         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2903                 if (!adev->ip_blocks[i].status.sw)
2904                         continue;
2905
2906                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) {
2907                         amdgpu_ucode_free_bo(adev);
2908                         amdgpu_free_static_csa(&adev->virt.csa_obj);
2909                         amdgpu_device_wb_fini(adev);
2910                         amdgpu_device_mem_scratch_fini(adev);
2911                         amdgpu_ib_pool_fini(adev);
2912                 }
2913
2914                 r = adev->ip_blocks[i].version->funcs->sw_fini((void *)adev);
2915                 /* XXX handle errors */
2916                 if (r) {
2917                         DRM_DEBUG("sw_fini of IP block <%s> failed %d\n",
2918                                   adev->ip_blocks[i].version->funcs->name, r);
2919                 }
2920                 adev->ip_blocks[i].status.sw = false;
2921                 adev->ip_blocks[i].status.valid = false;
2922         }
2923
2924         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2925                 if (!adev->ip_blocks[i].status.late_initialized)
2926                         continue;
2927                 if (adev->ip_blocks[i].version->funcs->late_fini)
2928                         adev->ip_blocks[i].version->funcs->late_fini((void *)adev);
2929                 adev->ip_blocks[i].status.late_initialized = false;
2930         }
2931
2932         amdgpu_ras_fini(adev);
2933
2934         return 0;
2935 }
2936
2937 /**
2938  * amdgpu_device_delayed_init_work_handler - work handler for IB tests
2939  *
2940  * @work: work_struct.
2941  */
2942 static void amdgpu_device_delayed_init_work_handler(struct work_struct *work)
2943 {
2944         struct amdgpu_device *adev =
2945                 container_of(work, struct amdgpu_device, delayed_init_work.work);
2946         int r;
2947
2948         r = amdgpu_ib_ring_tests(adev);
2949         if (r)
2950                 DRM_ERROR("ib ring test failed (%d).\n", r);
2951 }
2952
2953 static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
2954 {
2955         struct amdgpu_device *adev =
2956                 container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
2957
2958         WARN_ON_ONCE(adev->gfx.gfx_off_state);
2959         WARN_ON_ONCE(adev->gfx.gfx_off_req_count);
2960
2961         if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
2962                 adev->gfx.gfx_off_state = true;
2963 }
2964
2965 /**
2966  * amdgpu_device_ip_suspend_phase1 - run suspend for hardware IPs (phase 1)
2967  *
2968  * @adev: amdgpu_device pointer
2969  *
2970  * Main suspend function for hardware IPs.  The list of all the hardware
2971  * IPs that make up the asic is walked, clockgating is disabled and the
2972  * suspend callbacks are run.  suspend puts the hardware and software state
2973  * in each IP into a state suitable for suspend.
2974  * Returns 0 on success, negative error code on failure.
2975  */
2976 static int amdgpu_device_ip_suspend_phase1(struct amdgpu_device *adev)
2977 {
2978         int i, r;
2979
2980         amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE);
2981         amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE);
2982
2983         /*
2984          * Per PMFW team's suggestion, driver needs to handle gfxoff
2985          * and df cstate features disablement for gpu reset(e.g. Mode1Reset)
2986          * scenario. Add the missing df cstate disablement here.
2987          */
2988         if (amdgpu_dpm_set_df_cstate(adev, DF_CSTATE_DISALLOW))
2989                 dev_warn(adev->dev, "Failed to disallow df cstate");
2990
2991         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
2992                 if (!adev->ip_blocks[i].status.valid)
2993                         continue;
2994
2995                 /* displays are handled separately */
2996                 if (adev->ip_blocks[i].version->type != AMD_IP_BLOCK_TYPE_DCE)
2997                         continue;
2998
2999                 /* XXX handle errors */
3000                 r = adev->ip_blocks[i].version->funcs->suspend(adev);
3001                 /* XXX handle errors */
3002                 if (r) {
3003                         DRM_ERROR("suspend of IP block <%s> failed %d\n",
3004                                   adev->ip_blocks[i].version->funcs->name, r);
3005                         return r;
3006                 }
3007
3008                 adev->ip_blocks[i].status.hw = false;
3009         }
3010
3011         return 0;
3012 }
3013
3014 /**
3015  * amdgpu_device_ip_suspend_phase2 - run suspend for hardware IPs (phase 2)
3016  *
3017  * @adev: amdgpu_device pointer
3018  *
3019  * Main suspend function for hardware IPs.  The list of all the hardware
3020  * IPs that make up the asic is walked, clockgating is disabled and the
3021  * suspend callbacks are run.  suspend puts the hardware and software state
3022  * in each IP into a state suitable for suspend.
3023  * Returns 0 on success, negative error code on failure.
3024  */
3025 static int amdgpu_device_ip_suspend_phase2(struct amdgpu_device *adev)
3026 {
3027         int i, r;
3028
3029         if (adev->in_s0ix)
3030                 amdgpu_dpm_gfx_state_change(adev, sGpuChangeState_D3Entry);
3031
3032         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
3033                 if (!adev->ip_blocks[i].status.valid)
3034                         continue;
3035                 /* displays are handled in phase1 */
3036                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE)
3037                         continue;
3038                 /* PSP lost connection when err_event_athub occurs */
3039                 if (amdgpu_ras_intr_triggered() &&
3040                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
3041                         adev->ip_blocks[i].status.hw = false;
3042                         continue;
3043                 }
3044
3045                 /* skip unnecessary suspend if we do not initialize them yet */
3046                 if (adev->gmc.xgmi.pending_reset &&
3047                     !(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3048                       adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC ||
3049                       adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3050                       adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH)) {
3051                         adev->ip_blocks[i].status.hw = false;
3052                         continue;
3053                 }
3054
3055                 /* skip suspend of gfx/mes and psp for S0ix
3056                  * gfx is in gfxoff state, so on resume it will exit gfxoff just
3057                  * like at runtime. PSP is also part of the always on hardware
3058                  * so no need to suspend it.
3059                  */
3060                 if (adev->in_s0ix &&
3061                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP ||
3062                      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GFX ||
3063                      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_MES))
3064                         continue;
3065
3066                 /* SDMA 5.x+ is part of GFX power domain so it's covered by GFXOFF */
3067                 if (adev->in_s0ix &&
3068                     (adev->ip_versions[SDMA0_HWIP][0] >= IP_VERSION(5, 0, 0)) &&
3069                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SDMA))
3070                         continue;
3071
3072                 /* Once swPSP provides the IMU, RLC FW binaries to TOS during cold-boot.
3073                  * These are in TMR, hence are expected to be reused by PSP-TOS to reload
3074                  * from this location and RLC Autoload automatically also gets loaded
3075                  * from here based on PMFW -> PSP message during re-init sequence.
3076                  * Therefore, the psp suspend & resume should be skipped to avoid destroy
3077                  * the TMR and reload FWs again for IMU enabled APU ASICs.
3078                  */
3079                 if (amdgpu_in_reset(adev) &&
3080                     (adev->flags & AMD_IS_APU) && adev->gfx.imu.funcs &&
3081                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
3082                         continue;
3083
3084                 /* XXX handle errors */
3085                 r = adev->ip_blocks[i].version->funcs->suspend(adev);
3086                 /* XXX handle errors */
3087                 if (r) {
3088                         DRM_ERROR("suspend of IP block <%s> failed %d\n",
3089                                   adev->ip_blocks[i].version->funcs->name, r);
3090                 }
3091                 adev->ip_blocks[i].status.hw = false;
3092                 /* handle putting the SMC in the appropriate state */
3093                 if (!amdgpu_sriov_vf(adev)) {
3094                         if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) {
3095                                 r = amdgpu_dpm_set_mp1_state(adev, adev->mp1_state);
3096                                 if (r) {
3097                                         DRM_ERROR("SMC failed to set mp1 state %d, %d\n",
3098                                                         adev->mp1_state, r);
3099                                         return r;
3100                                 }
3101                         }
3102                 }
3103         }
3104
3105         return 0;
3106 }
3107
3108 /**
3109  * amdgpu_device_ip_suspend - run suspend for hardware IPs
3110  *
3111  * @adev: amdgpu_device pointer
3112  *
3113  * Main suspend function for hardware IPs.  The list of all the hardware
3114  * IPs that make up the asic is walked, clockgating is disabled and the
3115  * suspend callbacks are run.  suspend puts the hardware and software state
3116  * in each IP into a state suitable for suspend.
3117  * Returns 0 on success, negative error code on failure.
3118  */
3119 int amdgpu_device_ip_suspend(struct amdgpu_device *adev)
3120 {
3121         int r;
3122
3123         if (amdgpu_sriov_vf(adev)) {
3124                 amdgpu_virt_fini_data_exchange(adev);
3125                 amdgpu_virt_request_full_gpu(adev, false);
3126         }
3127
3128         r = amdgpu_device_ip_suspend_phase1(adev);
3129         if (r)
3130                 return r;
3131         r = amdgpu_device_ip_suspend_phase2(adev);
3132
3133         if (amdgpu_sriov_vf(adev))
3134                 amdgpu_virt_release_full_gpu(adev, false);
3135
3136         return r;
3137 }
3138
3139 static int amdgpu_device_ip_reinit_early_sriov(struct amdgpu_device *adev)
3140 {
3141         int i, r;
3142
3143         static enum amd_ip_block_type ip_order[] = {
3144                 AMD_IP_BLOCK_TYPE_COMMON,
3145                 AMD_IP_BLOCK_TYPE_GMC,
3146                 AMD_IP_BLOCK_TYPE_PSP,
3147                 AMD_IP_BLOCK_TYPE_IH,
3148         };
3149
3150         for (i = 0; i < adev->num_ip_blocks; i++) {
3151                 int j;
3152                 struct amdgpu_ip_block *block;
3153
3154                 block = &adev->ip_blocks[i];
3155                 block->status.hw = false;
3156
3157                 for (j = 0; j < ARRAY_SIZE(ip_order); j++) {
3158
3159                         if (block->version->type != ip_order[j] ||
3160                                 !block->status.valid)
3161                                 continue;
3162
3163                         r = block->version->funcs->hw_init(adev);
3164                         DRM_INFO("RE-INIT-early: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
3165                         if (r)
3166                                 return r;
3167                         block->status.hw = true;
3168                 }
3169         }
3170
3171         return 0;
3172 }
3173
3174 static int amdgpu_device_ip_reinit_late_sriov(struct amdgpu_device *adev)
3175 {
3176         int i, r;
3177
3178         static enum amd_ip_block_type ip_order[] = {
3179                 AMD_IP_BLOCK_TYPE_SMC,
3180                 AMD_IP_BLOCK_TYPE_DCE,
3181                 AMD_IP_BLOCK_TYPE_GFX,
3182                 AMD_IP_BLOCK_TYPE_SDMA,
3183                 AMD_IP_BLOCK_TYPE_MES,
3184                 AMD_IP_BLOCK_TYPE_UVD,
3185                 AMD_IP_BLOCK_TYPE_VCE,
3186                 AMD_IP_BLOCK_TYPE_VCN,
3187                 AMD_IP_BLOCK_TYPE_JPEG
3188         };
3189
3190         for (i = 0; i < ARRAY_SIZE(ip_order); i++) {
3191                 int j;
3192                 struct amdgpu_ip_block *block;
3193
3194                 for (j = 0; j < adev->num_ip_blocks; j++) {
3195                         block = &adev->ip_blocks[j];
3196
3197                         if (block->version->type != ip_order[i] ||
3198                                 !block->status.valid ||
3199                                 block->status.hw)
3200                                 continue;
3201
3202                         if (block->version->type == AMD_IP_BLOCK_TYPE_SMC)
3203                                 r = block->version->funcs->resume(adev);
3204                         else
3205                                 r = block->version->funcs->hw_init(adev);
3206
3207                         DRM_INFO("RE-INIT-late: %s %s\n", block->version->funcs->name, r?"failed":"succeeded");
3208                         if (r)
3209                                 return r;
3210                         block->status.hw = true;
3211                 }
3212         }
3213
3214         return 0;
3215 }
3216
3217 /**
3218  * amdgpu_device_ip_resume_phase1 - run resume for hardware IPs
3219  *
3220  * @adev: amdgpu_device pointer
3221  *
3222  * First resume function for hardware IPs.  The list of all the hardware
3223  * IPs that make up the asic is walked and the resume callbacks are run for
3224  * COMMON, GMC, and IH.  resume puts the hardware into a functional state
3225  * after a suspend and updates the software state as necessary.  This
3226  * function is also used for restoring the GPU after a GPU reset.
3227  * Returns 0 on success, negative error code on failure.
3228  */
3229 static int amdgpu_device_ip_resume_phase1(struct amdgpu_device *adev)
3230 {
3231         int i, r;
3232
3233         for (i = 0; i < adev->num_ip_blocks; i++) {
3234                 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
3235                         continue;
3236                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3237                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3238                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
3239                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP && amdgpu_sriov_vf(adev))) {
3240
3241                         r = adev->ip_blocks[i].version->funcs->resume(adev);
3242                         if (r) {
3243                                 DRM_ERROR("resume of IP block <%s> failed %d\n",
3244                                           adev->ip_blocks[i].version->funcs->name, r);
3245                                 return r;
3246                         }
3247                         adev->ip_blocks[i].status.hw = true;
3248                 }
3249         }
3250
3251         return 0;
3252 }
3253
3254 /**
3255  * amdgpu_device_ip_resume_phase2 - run resume for hardware IPs
3256  *
3257  * @adev: amdgpu_device pointer
3258  *
3259  * First resume function for hardware IPs.  The list of all the hardware
3260  * IPs that make up the asic is walked and the resume callbacks are run for
3261  * all blocks except COMMON, GMC, and IH.  resume puts the hardware into a
3262  * functional state after a suspend and updates the software state as
3263  * necessary.  This function is also used for restoring the GPU after a GPU
3264  * reset.
3265  * Returns 0 on success, negative error code on failure.
3266  */
3267 static int amdgpu_device_ip_resume_phase2(struct amdgpu_device *adev)
3268 {
3269         int i, r;
3270
3271         for (i = 0; i < adev->num_ip_blocks; i++) {
3272                 if (!adev->ip_blocks[i].status.valid || adev->ip_blocks[i].status.hw)
3273                         continue;
3274                 if (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3275                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3276                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
3277                     adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP)
3278                         continue;
3279                 r = adev->ip_blocks[i].version->funcs->resume(adev);
3280                 if (r) {
3281                         DRM_ERROR("resume of IP block <%s> failed %d\n",
3282                                   adev->ip_blocks[i].version->funcs->name, r);
3283                         return r;
3284                 }
3285                 adev->ip_blocks[i].status.hw = true;
3286         }
3287
3288         return 0;
3289 }
3290
3291 /**
3292  * amdgpu_device_ip_resume - run resume for hardware IPs
3293  *
3294  * @adev: amdgpu_device pointer
3295  *
3296  * Main resume function for hardware IPs.  The hardware IPs
3297  * are split into two resume functions because they are
3298  * are also used in in recovering from a GPU reset and some additional
3299  * steps need to be take between them.  In this case (S3/S4) they are
3300  * run sequentially.
3301  * Returns 0 on success, negative error code on failure.
3302  */
3303 static int amdgpu_device_ip_resume(struct amdgpu_device *adev)
3304 {
3305         int r;
3306
3307         if (!adev->in_s0ix) {
3308                 r = amdgpu_amdkfd_resume_iommu(adev);
3309                 if (r)
3310                         return r;
3311         }
3312
3313         r = amdgpu_device_ip_resume_phase1(adev);
3314         if (r)
3315                 return r;
3316
3317         r = amdgpu_device_fw_loading(adev);
3318         if (r)
3319                 return r;
3320
3321         r = amdgpu_device_ip_resume_phase2(adev);
3322
3323         return r;
3324 }
3325
3326 /**
3327  * amdgpu_device_detect_sriov_bios - determine if the board supports SR-IOV
3328  *
3329  * @adev: amdgpu_device pointer
3330  *
3331  * Query the VBIOS data tables to determine if the board supports SR-IOV.
3332  */
3333 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
3334 {
3335         if (amdgpu_sriov_vf(adev)) {
3336                 if (adev->is_atom_fw) {
3337                         if (amdgpu_atomfirmware_gpu_virtualization_supported(adev))
3338                                 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
3339                 } else {
3340                         if (amdgpu_atombios_has_gpu_virtualization_table(adev))
3341                                 adev->virt.caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
3342                 }
3343
3344                 if (!(adev->virt.caps & AMDGPU_SRIOV_CAPS_SRIOV_VBIOS))
3345                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_NO_VBIOS, 0, 0);
3346         }
3347 }
3348
3349 /**
3350  * amdgpu_device_asic_has_dc_support - determine if DC supports the asic
3351  *
3352  * @asic_type: AMD asic type
3353  *
3354  * Check if there is DC (new modesetting infrastructre) support for an asic.
3355  * returns true if DC has support, false if not.
3356  */
3357 bool amdgpu_device_asic_has_dc_support(enum amd_asic_type asic_type)
3358 {
3359         switch (asic_type) {
3360 #ifdef CONFIG_DRM_AMDGPU_SI
3361         case CHIP_HAINAN:
3362 #endif
3363         case CHIP_TOPAZ:
3364                 /* chips with no display hardware */
3365                 return false;
3366 #if defined(CONFIG_DRM_AMD_DC)
3367         case CHIP_TAHITI:
3368         case CHIP_PITCAIRN:
3369         case CHIP_VERDE:
3370         case CHIP_OLAND:
3371                 /*
3372                  * We have systems in the wild with these ASICs that require
3373                  * LVDS and VGA support which is not supported with DC.
3374                  *
3375                  * Fallback to the non-DC driver here by default so as not to
3376                  * cause regressions.
3377                  */
3378 #if defined(CONFIG_DRM_AMD_DC_SI)
3379                 return amdgpu_dc > 0;
3380 #else
3381                 return false;
3382 #endif
3383         case CHIP_BONAIRE:
3384         case CHIP_KAVERI:
3385         case CHIP_KABINI:
3386         case CHIP_MULLINS:
3387                 /*
3388                  * We have systems in the wild with these ASICs that require
3389                  * VGA support which is not supported with DC.
3390                  *
3391                  * Fallback to the non-DC driver here by default so as not to
3392                  * cause regressions.
3393                  */
3394                 return amdgpu_dc > 0;
3395         default:
3396                 return amdgpu_dc != 0;
3397 #else
3398         default:
3399                 if (amdgpu_dc > 0)
3400                         DRM_INFO_ONCE("Display Core has been requested via kernel parameter "
3401                                          "but isn't supported by ASIC, ignoring\n");
3402                 return false;
3403 #endif
3404         }
3405 }
3406
3407 /**
3408  * amdgpu_device_has_dc_support - check if dc is supported
3409  *
3410  * @adev: amdgpu_device pointer
3411  *
3412  * Returns true for supported, false for not supported
3413  */
3414 bool amdgpu_device_has_dc_support(struct amdgpu_device *adev)
3415 {
3416         if (adev->enable_virtual_display ||
3417             (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK))
3418                 return false;
3419
3420         return amdgpu_device_asic_has_dc_support(adev->asic_type);
3421 }
3422
3423 static void amdgpu_device_xgmi_reset_func(struct work_struct *__work)
3424 {
3425         struct amdgpu_device *adev =
3426                 container_of(__work, struct amdgpu_device, xgmi_reset_work);
3427         struct amdgpu_hive_info *hive = amdgpu_get_xgmi_hive(adev);
3428
3429         /* It's a bug to not have a hive within this function */
3430         if (WARN_ON(!hive))
3431                 return;
3432
3433         /*
3434          * Use task barrier to synchronize all xgmi reset works across the
3435          * hive. task_barrier_enter and task_barrier_exit will block
3436          * until all the threads running the xgmi reset works reach
3437          * those points. task_barrier_full will do both blocks.
3438          */
3439         if (amdgpu_asic_reset_method(adev) == AMD_RESET_METHOD_BACO) {
3440
3441                 task_barrier_enter(&hive->tb);
3442                 adev->asic_reset_res = amdgpu_device_baco_enter(adev_to_drm(adev));
3443
3444                 if (adev->asic_reset_res)
3445                         goto fail;
3446
3447                 task_barrier_exit(&hive->tb);
3448                 adev->asic_reset_res = amdgpu_device_baco_exit(adev_to_drm(adev));
3449
3450                 if (adev->asic_reset_res)
3451                         goto fail;
3452
3453                 if (adev->mmhub.ras && adev->mmhub.ras->ras_block.hw_ops &&
3454                     adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count)
3455                         adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count(adev);
3456         } else {
3457
3458                 task_barrier_full(&hive->tb);
3459                 adev->asic_reset_res =  amdgpu_asic_reset(adev);
3460         }
3461
3462 fail:
3463         if (adev->asic_reset_res)
3464                 DRM_WARN("ASIC reset failed with error, %d for drm dev, %s",
3465                          adev->asic_reset_res, adev_to_drm(adev)->unique);
3466         amdgpu_put_xgmi_hive(hive);
3467 }
3468
3469 static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev)
3470 {
3471         char *input = amdgpu_lockup_timeout;
3472         char *timeout_setting = NULL;
3473         int index = 0;
3474         long timeout;
3475         int ret = 0;
3476
3477         /*
3478          * By default timeout for non compute jobs is 10000
3479          * and 60000 for compute jobs.
3480          * In SR-IOV or passthrough mode, timeout for compute
3481          * jobs are 60000 by default.
3482          */
3483         adev->gfx_timeout = msecs_to_jiffies(10000);
3484         adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
3485         if (amdgpu_sriov_vf(adev))
3486                 adev->compute_timeout = amdgpu_sriov_is_pp_one_vf(adev) ?
3487                                         msecs_to_jiffies(60000) : msecs_to_jiffies(10000);
3488         else
3489                 adev->compute_timeout =  msecs_to_jiffies(60000);
3490
3491         if (strnlen(input, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
3492                 while ((timeout_setting = strsep(&input, ",")) &&
3493                                 strnlen(timeout_setting, AMDGPU_MAX_TIMEOUT_PARAM_LENGTH)) {
3494                         ret = kstrtol(timeout_setting, 0, &timeout);
3495                         if (ret)
3496                                 return ret;
3497
3498                         if (timeout == 0) {
3499                                 index++;
3500                                 continue;
3501                         } else if (timeout < 0) {
3502                                 timeout = MAX_SCHEDULE_TIMEOUT;
3503                                 dev_warn(adev->dev, "lockup timeout disabled");
3504                                 add_taint(TAINT_SOFTLOCKUP, LOCKDEP_STILL_OK);
3505                         } else {
3506                                 timeout = msecs_to_jiffies(timeout);
3507                         }
3508
3509                         switch (index++) {
3510                         case 0:
3511                                 adev->gfx_timeout = timeout;
3512                                 break;
3513                         case 1:
3514                                 adev->compute_timeout = timeout;
3515                                 break;
3516                         case 2:
3517                                 adev->sdma_timeout = timeout;
3518                                 break;
3519                         case 3:
3520                                 adev->video_timeout = timeout;
3521                                 break;
3522                         default:
3523                                 break;
3524                         }
3525                 }
3526                 /*
3527                  * There is only one value specified and
3528                  * it should apply to all non-compute jobs.
3529                  */
3530                 if (index == 1) {
3531                         adev->sdma_timeout = adev->video_timeout = adev->gfx_timeout;
3532                         if (amdgpu_sriov_vf(adev) || amdgpu_passthrough(adev))
3533                                 adev->compute_timeout = adev->gfx_timeout;
3534                 }
3535         }
3536
3537         return ret;
3538 }
3539
3540 /**
3541  * amdgpu_device_check_iommu_direct_map - check if RAM direct mapped to GPU
3542  *
3543  * @adev: amdgpu_device pointer
3544  *
3545  * RAM direct mapped to GPU if IOMMU is not enabled or is pass through mode
3546  */
3547 static void amdgpu_device_check_iommu_direct_map(struct amdgpu_device *adev)
3548 {
3549         struct iommu_domain *domain;
3550
3551         domain = iommu_get_domain_for_dev(adev->dev);
3552         if (!domain || domain->type == IOMMU_DOMAIN_IDENTITY)
3553                 adev->ram_is_direct_mapped = true;
3554 }
3555
3556 static const struct attribute *amdgpu_dev_attributes[] = {
3557         &dev_attr_product_name.attr,
3558         &dev_attr_product_number.attr,
3559         &dev_attr_serial_number.attr,
3560         &dev_attr_pcie_replay_count.attr,
3561         NULL
3562 };
3563
3564 /**
3565  * amdgpu_device_init - initialize the driver
3566  *
3567  * @adev: amdgpu_device pointer
3568  * @flags: driver flags
3569  *
3570  * Initializes the driver info and hw (all asics).
3571  * Returns 0 for success or an error on failure.
3572  * Called at driver startup.
3573  */
3574 int amdgpu_device_init(struct amdgpu_device *adev,
3575                        uint32_t flags)
3576 {
3577         struct drm_device *ddev = adev_to_drm(adev);
3578         struct pci_dev *pdev = adev->pdev;
3579         int r, i;
3580         bool px = false;
3581         u32 max_MBps;
3582         int tmp;
3583
3584         adev->shutdown = false;
3585         adev->flags = flags;
3586
3587         if (amdgpu_force_asic_type >= 0 && amdgpu_force_asic_type < CHIP_LAST)
3588                 adev->asic_type = amdgpu_force_asic_type;
3589         else
3590                 adev->asic_type = flags & AMD_ASIC_MASK;
3591
3592         adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
3593         if (amdgpu_emu_mode == 1)
3594                 adev->usec_timeout *= 10;
3595         adev->gmc.gart_size = 512 * 1024 * 1024;
3596         adev->accel_working = false;
3597         adev->num_rings = 0;
3598         RCU_INIT_POINTER(adev->gang_submit, dma_fence_get_stub());
3599         adev->mman.buffer_funcs = NULL;
3600         adev->mman.buffer_funcs_ring = NULL;
3601         adev->vm_manager.vm_pte_funcs = NULL;
3602         adev->vm_manager.vm_pte_num_scheds = 0;
3603         adev->gmc.gmc_funcs = NULL;
3604         adev->harvest_ip_mask = 0x0;
3605         adev->fence_context = dma_fence_context_alloc(AMDGPU_MAX_RINGS);
3606         bitmap_zero(adev->gfx.pipe_reserve_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
3607
3608         adev->smc_rreg = &amdgpu_invalid_rreg;
3609         adev->smc_wreg = &amdgpu_invalid_wreg;
3610         adev->pcie_rreg = &amdgpu_invalid_rreg;
3611         adev->pcie_wreg = &amdgpu_invalid_wreg;
3612         adev->pciep_rreg = &amdgpu_invalid_rreg;
3613         adev->pciep_wreg = &amdgpu_invalid_wreg;
3614         adev->pcie_rreg64 = &amdgpu_invalid_rreg64;
3615         adev->pcie_wreg64 = &amdgpu_invalid_wreg64;
3616         adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
3617         adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
3618         adev->didt_rreg = &amdgpu_invalid_rreg;
3619         adev->didt_wreg = &amdgpu_invalid_wreg;
3620         adev->gc_cac_rreg = &amdgpu_invalid_rreg;
3621         adev->gc_cac_wreg = &amdgpu_invalid_wreg;
3622         adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
3623         adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
3624
3625         DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
3626                  amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
3627                  pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
3628
3629         /* mutex initialization are all done here so we
3630          * can recall function without having locking issues */
3631         mutex_init(&adev->firmware.mutex);
3632         mutex_init(&adev->pm.mutex);
3633         mutex_init(&adev->gfx.gpu_clock_mutex);
3634         mutex_init(&adev->srbm_mutex);
3635         mutex_init(&adev->gfx.pipe_reserve_mutex);
3636         mutex_init(&adev->gfx.gfx_off_mutex);
3637         mutex_init(&adev->grbm_idx_mutex);
3638         mutex_init(&adev->mn_lock);
3639         mutex_init(&adev->virt.vf_errors.lock);
3640         hash_init(adev->mn_hash);
3641         mutex_init(&adev->psp.mutex);
3642         mutex_init(&adev->notifier_lock);
3643         mutex_init(&adev->pm.stable_pstate_ctx_lock);
3644         mutex_init(&adev->benchmark_mutex);
3645
3646         amdgpu_device_init_apu_flags(adev);
3647
3648         r = amdgpu_device_check_arguments(adev);
3649         if (r)
3650                 return r;
3651
3652         spin_lock_init(&adev->mmio_idx_lock);
3653         spin_lock_init(&adev->smc_idx_lock);
3654         spin_lock_init(&adev->pcie_idx_lock);
3655         spin_lock_init(&adev->uvd_ctx_idx_lock);
3656         spin_lock_init(&adev->didt_idx_lock);
3657         spin_lock_init(&adev->gc_cac_idx_lock);
3658         spin_lock_init(&adev->se_cac_idx_lock);
3659         spin_lock_init(&adev->audio_endpt_idx_lock);
3660         spin_lock_init(&adev->mm_stats.lock);
3661
3662         INIT_LIST_HEAD(&adev->shadow_list);
3663         mutex_init(&adev->shadow_list_lock);
3664
3665         INIT_LIST_HEAD(&adev->reset_list);
3666
3667         INIT_LIST_HEAD(&adev->ras_list);
3668
3669         INIT_DELAYED_WORK(&adev->delayed_init_work,
3670                           amdgpu_device_delayed_init_work_handler);
3671         INIT_DELAYED_WORK(&adev->gfx.gfx_off_delay_work,
3672                           amdgpu_device_delay_enable_gfx_off);
3673
3674         INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func);
3675
3676         adev->gfx.gfx_off_req_count = 1;
3677         adev->gfx.gfx_off_residency = 0;
3678         adev->gfx.gfx_off_entrycount = 0;
3679         adev->pm.ac_power = power_supply_is_system_supplied() > 0;
3680
3681         atomic_set(&adev->throttling_logging_enabled, 1);
3682         /*
3683          * If throttling continues, logging will be performed every minute
3684          * to avoid log flooding. "-1" is subtracted since the thermal
3685          * throttling interrupt comes every second. Thus, the total logging
3686          * interval is 59 seconds(retelimited printk interval) + 1(waiting
3687          * for throttling interrupt) = 60 seconds.
3688          */
3689         ratelimit_state_init(&adev->throttling_logging_rs, (60 - 1) * HZ, 1);
3690         ratelimit_set_flags(&adev->throttling_logging_rs, RATELIMIT_MSG_ON_RELEASE);
3691
3692         /* Registers mapping */
3693         /* TODO: block userspace mapping of io register */
3694         if (adev->asic_type >= CHIP_BONAIRE) {
3695                 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
3696                 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
3697         } else {
3698                 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
3699                 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
3700         }
3701
3702         for (i = 0; i < AMD_IP_BLOCK_TYPE_NUM; i++)
3703                 atomic_set(&adev->pm.pwr_state[i], POWER_STATE_UNKNOWN);
3704
3705         adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
3706         if (adev->rmmio == NULL) {
3707                 return -ENOMEM;
3708         }
3709         DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
3710         DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
3711
3712         amdgpu_device_get_pcie_info(adev);
3713
3714         if (amdgpu_mcbp)
3715                 DRM_INFO("MCBP is enabled\n");
3716
3717         /*
3718          * Reset domain needs to be present early, before XGMI hive discovered
3719          * (if any) and intitialized to use reset sem and in_gpu reset flag
3720          * early on during init and before calling to RREG32.
3721          */
3722         adev->reset_domain = amdgpu_reset_create_reset_domain(SINGLE_DEVICE, "amdgpu-reset-dev");
3723         if (!adev->reset_domain)
3724                 return -ENOMEM;
3725
3726         /* detect hw virtualization here */
3727         amdgpu_detect_virtualization(adev);
3728
3729         r = amdgpu_device_get_job_timeout_settings(adev);
3730         if (r) {
3731                 dev_err(adev->dev, "invalid lockup_timeout parameter syntax\n");
3732                 return r;
3733         }
3734
3735         /* early init functions */
3736         r = amdgpu_device_ip_early_init(adev);
3737         if (r)
3738                 return r;
3739
3740         /* Get rid of things like offb */
3741         r = drm_aperture_remove_conflicting_pci_framebuffers(adev->pdev, &amdgpu_kms_driver);
3742         if (r)
3743                 return r;
3744
3745         /* Enable TMZ based on IP_VERSION */
3746         amdgpu_gmc_tmz_set(adev);
3747
3748         amdgpu_gmc_noretry_set(adev);
3749         /* Need to get xgmi info early to decide the reset behavior*/
3750         if (adev->gmc.xgmi.supported) {
3751                 r = adev->gfxhub.funcs->get_xgmi_info(adev);
3752                 if (r)
3753                         return r;
3754         }
3755
3756         /* enable PCIE atomic ops */
3757         if (amdgpu_sriov_vf(adev))
3758                 adev->have_atomics_support = ((struct amd_sriov_msg_pf2vf_info *)
3759                         adev->virt.fw_reserve.p_pf2vf)->pcie_atomic_ops_support_flags ==
3760                         (PCI_EXP_DEVCAP2_ATOMIC_COMP32 | PCI_EXP_DEVCAP2_ATOMIC_COMP64);
3761         /* APUs w/ gfx9 onwards doesn't reply on PCIe atomics, rather it is a
3762          * internal path natively support atomics, set have_atomics_support to true.
3763          */
3764         else if ((adev->flags & AMD_IS_APU) &&
3765                 (adev->ip_versions[GC_HWIP][0] > IP_VERSION(9, 0, 0)))
3766                 adev->have_atomics_support = true;
3767         else
3768                 adev->have_atomics_support =
3769                         !pci_enable_atomic_ops_to_root(adev->pdev,
3770                                           PCI_EXP_DEVCAP2_ATOMIC_COMP32 |
3771                                           PCI_EXP_DEVCAP2_ATOMIC_COMP64);
3772         if (!adev->have_atomics_support)
3773                 dev_info(adev->dev, "PCIE atomic ops is not supported\n");
3774
3775         /* doorbell bar mapping and doorbell index init*/
3776         amdgpu_device_doorbell_init(adev);
3777
3778         if (amdgpu_emu_mode == 1) {
3779                 /* post the asic on emulation mode */
3780                 emu_soc_asic_init(adev);
3781                 goto fence_driver_init;
3782         }
3783
3784         amdgpu_reset_init(adev);
3785
3786         /* detect if we are with an SRIOV vbios */
3787         amdgpu_device_detect_sriov_bios(adev);
3788
3789         /* check if we need to reset the asic
3790          *  E.g., driver was not cleanly unloaded previously, etc.
3791          */
3792         if (!amdgpu_sriov_vf(adev) && amdgpu_asic_need_reset_on_init(adev)) {
3793                 if (adev->gmc.xgmi.num_physical_nodes) {
3794                         dev_info(adev->dev, "Pending hive reset.\n");
3795                         adev->gmc.xgmi.pending_reset = true;
3796                         /* Only need to init necessary block for SMU to handle the reset */
3797                         for (i = 0; i < adev->num_ip_blocks; i++) {
3798                                 if (!adev->ip_blocks[i].status.valid)
3799                                         continue;
3800                                 if (!(adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC ||
3801                                       adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_COMMON ||
3802                                       adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_IH ||
3803                                       adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC)) {
3804                                         DRM_DEBUG("IP %s disabled for hw_init.\n",
3805                                                 adev->ip_blocks[i].version->funcs->name);
3806                                         adev->ip_blocks[i].status.hw = true;
3807                                 }
3808                         }
3809                 } else {
3810                         tmp = amdgpu_reset_method;
3811                         /* It should do a default reset when loading or reloading the driver,
3812                          * regardless of the module parameter reset_method.
3813                          */
3814                         amdgpu_reset_method = AMD_RESET_METHOD_NONE;
3815                         r = amdgpu_asic_reset(adev);
3816                         amdgpu_reset_method = tmp;
3817                         if (r) {
3818                                 dev_err(adev->dev, "asic reset on init failed\n");
3819                                 goto failed;
3820                         }
3821                 }
3822         }
3823
3824         /* Post card if necessary */
3825         if (amdgpu_device_need_post(adev)) {
3826                 if (!adev->bios) {
3827                         dev_err(adev->dev, "no vBIOS found\n");
3828                         r = -EINVAL;
3829                         goto failed;
3830                 }
3831                 DRM_INFO("GPU posting now...\n");
3832                 r = amdgpu_device_asic_init(adev);
3833                 if (r) {
3834                         dev_err(adev->dev, "gpu post error!\n");
3835                         goto failed;
3836                 }
3837         }
3838
3839         if (adev->is_atom_fw) {
3840                 /* Initialize clocks */
3841                 r = amdgpu_atomfirmware_get_clock_info(adev);
3842                 if (r) {
3843                         dev_err(adev->dev, "amdgpu_atomfirmware_get_clock_info failed\n");
3844                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
3845                         goto failed;
3846                 }
3847         } else {
3848                 /* Initialize clocks */
3849                 r = amdgpu_atombios_get_clock_info(adev);
3850                 if (r) {
3851                         dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
3852                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_ATOMBIOS_GET_CLOCK_FAIL, 0, 0);
3853                         goto failed;
3854                 }
3855                 /* init i2c buses */
3856                 if (!amdgpu_device_has_dc_support(adev))
3857                         amdgpu_atombios_i2c_init(adev);
3858         }
3859
3860 fence_driver_init:
3861         /* Fence driver */
3862         r = amdgpu_fence_driver_sw_init(adev);
3863         if (r) {
3864                 dev_err(adev->dev, "amdgpu_fence_driver_sw_init failed\n");
3865                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_FENCE_INIT_FAIL, 0, 0);
3866                 goto failed;
3867         }
3868
3869         /* init the mode config */
3870         drm_mode_config_init(adev_to_drm(adev));
3871
3872         r = amdgpu_device_ip_init(adev);
3873         if (r) {
3874                 dev_err(adev->dev, "amdgpu_device_ip_init failed\n");
3875                 amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_INIT_FAIL, 0, 0);
3876                 goto release_ras_con;
3877         }
3878
3879         amdgpu_fence_driver_hw_init(adev);
3880
3881         dev_info(adev->dev,
3882                 "SE %d, SH per SE %d, CU per SH %d, active_cu_number %d\n",
3883                         adev->gfx.config.max_shader_engines,
3884                         adev->gfx.config.max_sh_per_se,
3885                         adev->gfx.config.max_cu_per_sh,
3886                         adev->gfx.cu_info.number);
3887
3888         adev->accel_working = true;
3889
3890         amdgpu_vm_check_compute_bug(adev);
3891
3892         /* Initialize the buffer migration limit. */
3893         if (amdgpu_moverate >= 0)
3894                 max_MBps = amdgpu_moverate;
3895         else
3896                 max_MBps = 8; /* Allow 8 MB/s. */
3897         /* Get a log2 for easy divisions. */
3898         adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
3899
3900         r = amdgpu_pm_sysfs_init(adev);
3901         if (r)
3902                 DRM_ERROR("registering pm sysfs failed (%d).\n", r);
3903
3904         r = amdgpu_ucode_sysfs_init(adev);
3905         if (r) {
3906                 adev->ucode_sysfs_en = false;
3907                 DRM_ERROR("Creating firmware sysfs failed (%d).\n", r);
3908         } else
3909                 adev->ucode_sysfs_en = true;
3910
3911         r = amdgpu_psp_sysfs_init(adev);
3912         if (r) {
3913                 adev->psp_sysfs_en = false;
3914                 if (!amdgpu_sriov_vf(adev))
3915                         DRM_ERROR("Creating psp sysfs failed\n");
3916         } else
3917                 adev->psp_sysfs_en = true;
3918
3919         /*
3920          * Register gpu instance before amdgpu_device_enable_mgpu_fan_boost.
3921          * Otherwise the mgpu fan boost feature will be skipped due to the
3922          * gpu instance is counted less.
3923          */
3924         amdgpu_register_gpu_instance(adev);
3925
3926         /* enable clockgating, etc. after ib tests, etc. since some blocks require
3927          * explicit gating rather than handling it automatically.
3928          */
3929         if (!adev->gmc.xgmi.pending_reset) {
3930                 r = amdgpu_device_ip_late_init(adev);
3931                 if (r) {
3932                         dev_err(adev->dev, "amdgpu_device_ip_late_init failed\n");
3933                         amdgpu_vf_error_put(adev, AMDGIM_ERROR_VF_AMDGPU_LATE_INIT_FAIL, 0, r);
3934                         goto release_ras_con;
3935                 }
3936                 /* must succeed. */
3937                 amdgpu_ras_resume(adev);
3938                 queue_delayed_work(system_wq, &adev->delayed_init_work,
3939                                    msecs_to_jiffies(AMDGPU_RESUME_MS));
3940         }
3941
3942         if (amdgpu_sriov_vf(adev)) {
3943                 amdgpu_virt_release_full_gpu(adev, true);
3944                 flush_delayed_work(&adev->delayed_init_work);
3945         }
3946
3947         r = sysfs_create_files(&adev->dev->kobj, amdgpu_dev_attributes);
3948         if (r)
3949                 dev_err(adev->dev, "Could not create amdgpu device attr\n");
3950
3951         if (IS_ENABLED(CONFIG_PERF_EVENTS))
3952                 r = amdgpu_pmu_init(adev);
3953         if (r)
3954                 dev_err(adev->dev, "amdgpu_pmu_init failed\n");
3955
3956         /* Have stored pci confspace at hand for restore in sudden PCI error */
3957         if (amdgpu_device_cache_pci_state(adev->pdev))
3958                 pci_restore_state(pdev);
3959
3960         /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
3961         /* this will fail for cards that aren't VGA class devices, just
3962          * ignore it */
3963         if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA)
3964                 vga_client_register(adev->pdev, amdgpu_device_vga_set_decode);
3965
3966         px = amdgpu_device_supports_px(ddev);
3967
3968         if (px || (!pci_is_thunderbolt_attached(adev->pdev) &&
3969                                 apple_gmux_detect(NULL, NULL)))
3970                 vga_switcheroo_register_client(adev->pdev,
3971                                                &amdgpu_switcheroo_ops, px);
3972
3973         if (px)
3974                 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
3975
3976         if (adev->gmc.xgmi.pending_reset)
3977                 queue_delayed_work(system_wq, &mgpu_info.delayed_reset_work,
3978                                    msecs_to_jiffies(AMDGPU_RESUME_MS));
3979
3980         amdgpu_device_check_iommu_direct_map(adev);
3981
3982         return 0;
3983
3984 release_ras_con:
3985         if (amdgpu_sriov_vf(adev))
3986                 amdgpu_virt_release_full_gpu(adev, true);
3987
3988         /* failed in exclusive mode due to timeout */
3989         if (amdgpu_sriov_vf(adev) &&
3990                 !amdgpu_sriov_runtime(adev) &&
3991                 amdgpu_virt_mmio_blocked(adev) &&
3992                 !amdgpu_virt_wait_reset(adev)) {
3993                 dev_err(adev->dev, "VF exclusive mode timeout\n");
3994                 /* Don't send request since VF is inactive. */
3995                 adev->virt.caps &= ~AMDGPU_SRIOV_CAPS_RUNTIME;
3996                 adev->virt.ops = NULL;
3997                 r = -EAGAIN;
3998         }
3999         amdgpu_release_ras_context(adev);
4000
4001 failed:
4002         amdgpu_vf_error_trans_all(adev);
4003
4004         return r;
4005 }
4006
4007 static void amdgpu_device_unmap_mmio(struct amdgpu_device *adev)
4008 {
4009
4010         /* Clear all CPU mappings pointing to this device */
4011         unmap_mapping_range(adev->ddev.anon_inode->i_mapping, 0, 0, 1);
4012
4013         /* Unmap all mapped bars - Doorbell, registers and VRAM */
4014         amdgpu_device_doorbell_fini(adev);
4015
4016         iounmap(adev->rmmio);
4017         adev->rmmio = NULL;
4018         if (adev->mman.aper_base_kaddr)
4019                 iounmap(adev->mman.aper_base_kaddr);
4020         adev->mman.aper_base_kaddr = NULL;
4021
4022         /* Memory manager related */
4023         if (!adev->gmc.xgmi.connected_to_cpu) {
4024                 arch_phys_wc_del(adev->gmc.vram_mtrr);
4025                 arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
4026         }
4027 }
4028
4029 /**
4030  * amdgpu_device_fini_hw - tear down the driver
4031  *
4032  * @adev: amdgpu_device pointer
4033  *
4034  * Tear down the driver info (all asics).
4035  * Called at driver shutdown.
4036  */
4037 void amdgpu_device_fini_hw(struct amdgpu_device *adev)
4038 {
4039         dev_info(adev->dev, "amdgpu: finishing device.\n");
4040         flush_delayed_work(&adev->delayed_init_work);
4041         adev->shutdown = true;
4042
4043         /* make sure IB test finished before entering exclusive mode
4044          * to avoid preemption on IB test
4045          * */
4046         if (amdgpu_sriov_vf(adev)) {
4047                 amdgpu_virt_request_full_gpu(adev, false);
4048                 amdgpu_virt_fini_data_exchange(adev);
4049         }
4050
4051         /* disable all interrupts */
4052         amdgpu_irq_disable_all(adev);
4053         if (adev->mode_info.mode_config_initialized) {
4054                 if (!drm_drv_uses_atomic_modeset(adev_to_drm(adev)))
4055                         drm_helper_force_disable_all(adev_to_drm(adev));
4056                 else
4057                         drm_atomic_helper_shutdown(adev_to_drm(adev));
4058         }
4059         amdgpu_fence_driver_hw_fini(adev);
4060
4061         if (adev->mman.initialized)
4062                 drain_workqueue(adev->mman.bdev.wq);
4063
4064         if (adev->pm.sysfs_initialized)
4065                 amdgpu_pm_sysfs_fini(adev);
4066         if (adev->ucode_sysfs_en)
4067                 amdgpu_ucode_sysfs_fini(adev);
4068         if (adev->psp_sysfs_en)
4069                 amdgpu_psp_sysfs_fini(adev);
4070         sysfs_remove_files(&adev->dev->kobj, amdgpu_dev_attributes);
4071
4072         /* disable ras feature must before hw fini */
4073         amdgpu_ras_pre_fini(adev);
4074
4075         amdgpu_device_ip_fini_early(adev);
4076
4077         amdgpu_irq_fini_hw(adev);
4078
4079         if (adev->mman.initialized)
4080                 ttm_device_clear_dma_mappings(&adev->mman.bdev);
4081
4082         amdgpu_gart_dummy_page_fini(adev);
4083
4084         if (drm_dev_is_unplugged(adev_to_drm(adev)))
4085                 amdgpu_device_unmap_mmio(adev);
4086
4087 }
4088
4089 void amdgpu_device_fini_sw(struct amdgpu_device *adev)
4090 {
4091         int idx;
4092         bool px;
4093
4094         amdgpu_fence_driver_sw_fini(adev);
4095         amdgpu_device_ip_fini(adev);
4096         amdgpu_ucode_release(&adev->firmware.gpu_info_fw);
4097         adev->accel_working = false;
4098         dma_fence_put(rcu_dereference_protected(adev->gang_submit, true));
4099
4100         amdgpu_reset_fini(adev);
4101
4102         /* free i2c buses */
4103         if (!amdgpu_device_has_dc_support(adev))
4104                 amdgpu_i2c_fini(adev);
4105
4106         if (amdgpu_emu_mode != 1)
4107                 amdgpu_atombios_fini(adev);
4108
4109         kfree(adev->bios);
4110         adev->bios = NULL;
4111
4112         px = amdgpu_device_supports_px(adev_to_drm(adev));
4113
4114         if (px || (!pci_is_thunderbolt_attached(adev->pdev) &&
4115                                 apple_gmux_detect(NULL, NULL)))
4116                 vga_switcheroo_unregister_client(adev->pdev);
4117
4118         if (px)
4119                 vga_switcheroo_fini_domain_pm_ops(adev->dev);
4120
4121         if ((adev->pdev->class >> 8) == PCI_CLASS_DISPLAY_VGA)
4122                 vga_client_unregister(adev->pdev);
4123
4124         if (drm_dev_enter(adev_to_drm(adev), &idx)) {
4125
4126                 iounmap(adev->rmmio);
4127                 adev->rmmio = NULL;
4128                 amdgpu_device_doorbell_fini(adev);
4129                 drm_dev_exit(idx);
4130         }
4131
4132         if (IS_ENABLED(CONFIG_PERF_EVENTS))
4133                 amdgpu_pmu_fini(adev);
4134         if (adev->mman.discovery_bin)
4135                 amdgpu_discovery_fini(adev);
4136
4137         amdgpu_reset_put_reset_domain(adev->reset_domain);
4138         adev->reset_domain = NULL;
4139
4140         kfree(adev->pci_state);
4141
4142 }
4143
4144 /**
4145  * amdgpu_device_evict_resources - evict device resources
4146  * @adev: amdgpu device object
4147  *
4148  * Evicts all ttm device resources(vram BOs, gart table) from the lru list
4149  * of the vram memory type. Mainly used for evicting device resources
4150  * at suspend time.
4151  *
4152  */
4153 static int amdgpu_device_evict_resources(struct amdgpu_device *adev)
4154 {
4155         int ret;
4156
4157         /* No need to evict vram on APUs for suspend to ram or s2idle */
4158         if ((adev->in_s3 || adev->in_s0ix) && (adev->flags & AMD_IS_APU))
4159                 return 0;
4160
4161         ret = amdgpu_ttm_evict_resources(adev, TTM_PL_VRAM);
4162         if (ret)
4163                 DRM_WARN("evicting device resources failed\n");
4164         return ret;
4165 }
4166
4167 /*
4168  * Suspend & resume.
4169  */
4170 /**
4171  * amdgpu_device_suspend - initiate device suspend
4172  *
4173  * @dev: drm dev pointer
4174  * @fbcon : notify the fbdev of suspend
4175  *
4176  * Puts the hw in the suspend state (all asics).
4177  * Returns 0 for success or an error on failure.
4178  * Called at driver suspend.
4179  */
4180 int amdgpu_device_suspend(struct drm_device *dev, bool fbcon)
4181 {
4182         struct amdgpu_device *adev = drm_to_adev(dev);
4183         int r = 0;
4184
4185         if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
4186                 return 0;
4187
4188         adev->in_suspend = true;
4189
4190         /* Evict the majority of BOs before grabbing the full access */
4191         r = amdgpu_device_evict_resources(adev);
4192         if (r)
4193                 return r;
4194
4195         if (amdgpu_sriov_vf(adev)) {
4196                 amdgpu_virt_fini_data_exchange(adev);
4197                 r = amdgpu_virt_request_full_gpu(adev, false);
4198                 if (r)
4199                         return r;
4200         }
4201
4202         if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D3))
4203                 DRM_WARN("smart shift update failed\n");
4204
4205         if (fbcon)
4206                 drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, true);
4207
4208         cancel_delayed_work_sync(&adev->delayed_init_work);
4209
4210         amdgpu_ras_suspend(adev);
4211
4212         amdgpu_device_ip_suspend_phase1(adev);
4213
4214         if (!adev->in_s0ix)
4215                 amdgpu_amdkfd_suspend(adev, adev->in_runpm);
4216
4217         r = amdgpu_device_evict_resources(adev);
4218         if (r)
4219                 return r;
4220
4221         amdgpu_fence_driver_hw_fini(adev);
4222
4223         amdgpu_device_ip_suspend_phase2(adev);
4224
4225         if (amdgpu_sriov_vf(adev))
4226                 amdgpu_virt_release_full_gpu(adev, false);
4227
4228         return 0;
4229 }
4230
4231 /**
4232  * amdgpu_device_resume - initiate device resume
4233  *
4234  * @dev: drm dev pointer
4235  * @fbcon : notify the fbdev of resume
4236  *
4237  * Bring the hw back to operating state (all asics).
4238  * Returns 0 for success or an error on failure.
4239  * Called at driver resume.
4240  */
4241 int amdgpu_device_resume(struct drm_device *dev, bool fbcon)
4242 {
4243         struct amdgpu_device *adev = drm_to_adev(dev);
4244         int r = 0;
4245
4246         if (amdgpu_sriov_vf(adev)) {
4247                 r = amdgpu_virt_request_full_gpu(adev, true);
4248                 if (r)
4249                         return r;
4250         }
4251
4252         if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
4253                 return 0;
4254
4255         if (adev->in_s0ix)
4256                 amdgpu_dpm_gfx_state_change(adev, sGpuChangeState_D0Entry);
4257
4258         /* post card */
4259         if (amdgpu_device_need_post(adev)) {
4260                 r = amdgpu_device_asic_init(adev);
4261                 if (r)
4262                         dev_err(adev->dev, "amdgpu asic init failed\n");
4263         }
4264
4265         r = amdgpu_device_ip_resume(adev);
4266
4267         if (r) {
4268                 dev_err(adev->dev, "amdgpu_device_ip_resume failed (%d).\n", r);
4269                 goto exit;
4270         }
4271         amdgpu_fence_driver_hw_init(adev);
4272
4273         r = amdgpu_device_ip_late_init(adev);
4274         if (r)
4275                 goto exit;
4276
4277         queue_delayed_work(system_wq, &adev->delayed_init_work,
4278                            msecs_to_jiffies(AMDGPU_RESUME_MS));
4279
4280         if (!adev->in_s0ix) {
4281                 r = amdgpu_amdkfd_resume(adev, adev->in_runpm);
4282                 if (r)
4283                         goto exit;
4284         }
4285
4286 exit:
4287         if (amdgpu_sriov_vf(adev)) {
4288                 amdgpu_virt_init_data_exchange(adev);
4289                 amdgpu_virt_release_full_gpu(adev, true);
4290         }
4291
4292         if (r)
4293                 return r;
4294
4295         /* Make sure IB tests flushed */
4296         flush_delayed_work(&adev->delayed_init_work);
4297
4298         if (fbcon)
4299                 drm_fb_helper_set_suspend_unlocked(adev_to_drm(adev)->fb_helper, false);
4300
4301         amdgpu_ras_resume(adev);
4302
4303         if (adev->mode_info.num_crtc) {
4304                 /*
4305                  * Most of the connector probing functions try to acquire runtime pm
4306                  * refs to ensure that the GPU is powered on when connector polling is
4307                  * performed. Since we're calling this from a runtime PM callback,
4308                  * trying to acquire rpm refs will cause us to deadlock.
4309                  *
4310                  * Since we're guaranteed to be holding the rpm lock, it's safe to
4311                  * temporarily disable the rpm helpers so this doesn't deadlock us.
4312                  */
4313 #ifdef CONFIG_PM
4314                 dev->dev->power.disable_depth++;
4315 #endif
4316                 if (!adev->dc_enabled)
4317                         drm_helper_hpd_irq_event(dev);
4318                 else
4319                         drm_kms_helper_hotplug_event(dev);
4320 #ifdef CONFIG_PM
4321                 dev->dev->power.disable_depth--;
4322 #endif
4323         }
4324         adev->in_suspend = false;
4325
4326         if (adev->enable_mes)
4327                 amdgpu_mes_self_test(adev);
4328
4329         if (amdgpu_acpi_smart_shift_update(dev, AMDGPU_SS_DEV_D0))
4330                 DRM_WARN("smart shift update failed\n");
4331
4332         return 0;
4333 }
4334
4335 /**
4336  * amdgpu_device_ip_check_soft_reset - did soft reset succeed
4337  *
4338  * @adev: amdgpu_device pointer
4339  *
4340  * The list of all the hardware IPs that make up the asic is walked and
4341  * the check_soft_reset callbacks are run.  check_soft_reset determines
4342  * if the asic is still hung or not.
4343  * Returns true if any of the IPs are still in a hung state, false if not.
4344  */
4345 static bool amdgpu_device_ip_check_soft_reset(struct amdgpu_device *adev)
4346 {
4347         int i;
4348         bool asic_hang = false;
4349
4350         if (amdgpu_sriov_vf(adev))
4351                 return true;
4352
4353         if (amdgpu_asic_need_full_reset(adev))
4354                 return true;
4355
4356         for (i = 0; i < adev->num_ip_blocks; i++) {
4357                 if (!adev->ip_blocks[i].status.valid)
4358                         continue;
4359                 if (adev->ip_blocks[i].version->funcs->check_soft_reset)
4360                         adev->ip_blocks[i].status.hang =
4361                                 adev->ip_blocks[i].version->funcs->check_soft_reset(adev);
4362                 if (adev->ip_blocks[i].status.hang) {
4363                         dev_info(adev->dev, "IP block:%s is hung!\n", adev->ip_blocks[i].version->funcs->name);
4364                         asic_hang = true;
4365                 }
4366         }
4367         return asic_hang;
4368 }
4369
4370 /**
4371  * amdgpu_device_ip_pre_soft_reset - prepare for soft reset
4372  *
4373  * @adev: amdgpu_device pointer
4374  *
4375  * The list of all the hardware IPs that make up the asic is walked and the
4376  * pre_soft_reset callbacks are run if the block is hung.  pre_soft_reset
4377  * handles any IP specific hardware or software state changes that are
4378  * necessary for a soft reset to succeed.
4379  * Returns 0 on success, negative error code on failure.
4380  */
4381 static int amdgpu_device_ip_pre_soft_reset(struct amdgpu_device *adev)
4382 {
4383         int i, r = 0;
4384
4385         for (i = 0; i < adev->num_ip_blocks; i++) {
4386                 if (!adev->ip_blocks[i].status.valid)
4387                         continue;
4388                 if (adev->ip_blocks[i].status.hang &&
4389                     adev->ip_blocks[i].version->funcs->pre_soft_reset) {
4390                         r = adev->ip_blocks[i].version->funcs->pre_soft_reset(adev);
4391                         if (r)
4392                                 return r;
4393                 }
4394         }
4395
4396         return 0;
4397 }
4398
4399 /**
4400  * amdgpu_device_ip_need_full_reset - check if a full asic reset is needed
4401  *
4402  * @adev: amdgpu_device pointer
4403  *
4404  * Some hardware IPs cannot be soft reset.  If they are hung, a full gpu
4405  * reset is necessary to recover.
4406  * Returns true if a full asic reset is required, false if not.
4407  */
4408 static bool amdgpu_device_ip_need_full_reset(struct amdgpu_device *adev)
4409 {
4410         int i;
4411
4412         if (amdgpu_asic_need_full_reset(adev))
4413                 return true;
4414
4415         for (i = 0; i < adev->num_ip_blocks; i++) {
4416                 if (!adev->ip_blocks[i].status.valid)
4417                         continue;
4418                 if ((adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_GMC) ||
4419                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_SMC) ||
4420                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_ACP) ||
4421                     (adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_DCE) ||
4422                      adev->ip_blocks[i].version->type == AMD_IP_BLOCK_TYPE_PSP) {
4423                         if (adev->ip_blocks[i].status.hang) {
4424                                 dev_info(adev->dev, "Some block need full reset!\n");
4425                                 return true;
4426                         }
4427                 }
4428         }
4429         return false;
4430 }
4431
4432 /**
4433  * amdgpu_device_ip_soft_reset - do a soft reset
4434  *
4435  * @adev: amdgpu_device pointer
4436  *
4437  * The list of all the hardware IPs that make up the asic is walked and the
4438  * soft_reset callbacks are run if the block is hung.  soft_reset handles any
4439  * IP specific hardware or software state changes that are necessary to soft
4440  * reset the IP.
4441  * Returns 0 on success, negative error code on failure.
4442  */
4443 static int amdgpu_device_ip_soft_reset(struct amdgpu_device *adev)
4444 {
4445         int i, r = 0;
4446
4447         for (i = 0; i < adev->num_ip_blocks; i++) {
4448                 if (!adev->ip_blocks[i].status.valid)
4449                         continue;
4450                 if (adev->ip_blocks[i].status.hang &&
4451                     adev->ip_blocks[i].version->funcs->soft_reset) {
4452                         r = adev->ip_blocks[i].version->funcs->soft_reset(adev);
4453                         if (r)
4454                                 return r;
4455                 }
4456         }
4457
4458         return 0;
4459 }
4460
4461 /**
4462  * amdgpu_device_ip_post_soft_reset - clean up from soft reset
4463  *
4464  * @adev: amdgpu_device pointer
4465  *
4466  * The list of all the hardware IPs that make up the asic is walked and the
4467  * post_soft_reset callbacks are run if the asic was hung.  post_soft_reset
4468  * handles any IP specific hardware or software state changes that are
4469  * necessary after the IP has been soft reset.
4470  * Returns 0 on success, negative error code on failure.
4471  */
4472 static int amdgpu_device_ip_post_soft_reset(struct amdgpu_device *adev)
4473 {
4474         int i, r = 0;
4475
4476         for (i = 0; i < adev->num_ip_blocks; i++) {
4477                 if (!adev->ip_blocks[i].status.valid)
4478                         continue;
4479                 if (adev->ip_blocks[i].status.hang &&
4480                     adev->ip_blocks[i].version->funcs->post_soft_reset)
4481                         r = adev->ip_blocks[i].version->funcs->post_soft_reset(adev);
4482                 if (r)
4483                         return r;
4484         }
4485
4486         return 0;
4487 }
4488
4489 /**
4490  * amdgpu_device_recover_vram - Recover some VRAM contents
4491  *
4492  * @adev: amdgpu_device pointer
4493  *
4494  * Restores the contents of VRAM buffers from the shadows in GTT.  Used to
4495  * restore things like GPUVM page tables after a GPU reset where
4496  * the contents of VRAM might be lost.
4497  *
4498  * Returns:
4499  * 0 on success, negative error code on failure.
4500  */
4501 static int amdgpu_device_recover_vram(struct amdgpu_device *adev)
4502 {
4503         struct dma_fence *fence = NULL, *next = NULL;
4504         struct amdgpu_bo *shadow;
4505         struct amdgpu_bo_vm *vmbo;
4506         long r = 1, tmo;
4507
4508         if (amdgpu_sriov_runtime(adev))
4509                 tmo = msecs_to_jiffies(8000);
4510         else
4511                 tmo = msecs_to_jiffies(100);
4512
4513         dev_info(adev->dev, "recover vram bo from shadow start\n");
4514         mutex_lock(&adev->shadow_list_lock);
4515         list_for_each_entry(vmbo, &adev->shadow_list, shadow_list) {
4516                 shadow = &vmbo->bo;
4517                 /* No need to recover an evicted BO */
4518                 if (shadow->tbo.resource->mem_type != TTM_PL_TT ||
4519                     shadow->tbo.resource->start == AMDGPU_BO_INVALID_OFFSET ||
4520                     shadow->parent->tbo.resource->mem_type != TTM_PL_VRAM)
4521                         continue;
4522
4523                 r = amdgpu_bo_restore_shadow(shadow, &next);
4524                 if (r)
4525                         break;
4526
4527                 if (fence) {
4528                         tmo = dma_fence_wait_timeout(fence, false, tmo);
4529                         dma_fence_put(fence);
4530                         fence = next;
4531                         if (tmo == 0) {
4532                                 r = -ETIMEDOUT;
4533                                 break;
4534                         } else if (tmo < 0) {
4535                                 r = tmo;
4536                                 break;
4537                         }
4538                 } else {
4539                         fence = next;
4540                 }
4541         }
4542         mutex_unlock(&adev->shadow_list_lock);
4543
4544         if (fence)
4545                 tmo = dma_fence_wait_timeout(fence, false, tmo);
4546         dma_fence_put(fence);
4547
4548         if (r < 0 || tmo <= 0) {
4549                 dev_err(adev->dev, "recover vram bo from shadow failed, r is %ld, tmo is %ld\n", r, tmo);
4550                 return -EIO;
4551         }
4552
4553         dev_info(adev->dev, "recover vram bo from shadow done\n");
4554         return 0;
4555 }
4556
4557
4558 /**
4559  * amdgpu_device_reset_sriov - reset ASIC for SR-IOV vf
4560  *
4561  * @adev: amdgpu_device pointer
4562  * @from_hypervisor: request from hypervisor
4563  *
4564  * do VF FLR and reinitialize Asic
4565  * return 0 means succeeded otherwise failed
4566  */
4567 static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
4568                                      bool from_hypervisor)
4569 {
4570         int r;
4571         struct amdgpu_hive_info *hive = NULL;
4572         int retry_limit = 0;
4573
4574 retry:
4575         amdgpu_amdkfd_pre_reset(adev);
4576
4577         if (from_hypervisor)
4578                 r = amdgpu_virt_request_full_gpu(adev, true);
4579         else
4580                 r = amdgpu_virt_reset_gpu(adev);
4581         if (r)
4582                 return r;
4583
4584         /* Resume IP prior to SMC */
4585         r = amdgpu_device_ip_reinit_early_sriov(adev);
4586         if (r)
4587                 goto error;
4588
4589         amdgpu_virt_init_data_exchange(adev);
4590
4591         r = amdgpu_device_fw_loading(adev);
4592         if (r)
4593                 return r;
4594
4595         /* now we are okay to resume SMC/CP/SDMA */
4596         r = amdgpu_device_ip_reinit_late_sriov(adev);
4597         if (r)
4598                 goto error;
4599
4600         hive = amdgpu_get_xgmi_hive(adev);
4601         /* Update PSP FW topology after reset */
4602         if (hive && adev->gmc.xgmi.num_physical_nodes > 1)
4603                 r = amdgpu_xgmi_update_topology(hive, adev);
4604
4605         if (hive)
4606                 amdgpu_put_xgmi_hive(hive);
4607
4608         if (!r) {
4609                 amdgpu_irq_gpu_reset_resume_helper(adev);
4610                 r = amdgpu_ib_ring_tests(adev);
4611
4612                 amdgpu_amdkfd_post_reset(adev);
4613         }
4614
4615 error:
4616         if (!r && adev->virt.gim_feature & AMDGIM_FEATURE_GIM_FLR_VRAMLOST) {
4617                 amdgpu_inc_vram_lost(adev);
4618                 r = amdgpu_device_recover_vram(adev);
4619         }
4620         amdgpu_virt_release_full_gpu(adev, true);
4621
4622         if (AMDGPU_RETRY_SRIOV_RESET(r)) {
4623                 if (retry_limit < AMDGPU_MAX_RETRY_LIMIT) {
4624                         retry_limit++;
4625                         goto retry;
4626                 } else
4627                         DRM_ERROR("GPU reset retry is beyond the retry limit\n");
4628         }
4629
4630         return r;
4631 }
4632
4633 /**
4634  * amdgpu_device_has_job_running - check if there is any job in mirror list
4635  *
4636  * @adev: amdgpu_device pointer
4637  *
4638  * check if there is any job in mirror list
4639  */
4640 bool amdgpu_device_has_job_running(struct amdgpu_device *adev)
4641 {
4642         int i;
4643         struct drm_sched_job *job;
4644
4645         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4646                 struct amdgpu_ring *ring = adev->rings[i];
4647
4648                 if (!ring || !ring->sched.thread)
4649                         continue;
4650
4651                 spin_lock(&ring->sched.job_list_lock);
4652                 job = list_first_entry_or_null(&ring->sched.pending_list,
4653                                                struct drm_sched_job, list);
4654                 spin_unlock(&ring->sched.job_list_lock);
4655                 if (job)
4656                         return true;
4657         }
4658         return false;
4659 }
4660
4661 /**
4662  * amdgpu_device_should_recover_gpu - check if we should try GPU recovery
4663  *
4664  * @adev: amdgpu_device pointer
4665  *
4666  * Check amdgpu_gpu_recovery and SRIOV status to see if we should try to recover
4667  * a hung GPU.
4668  */
4669 bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev)
4670 {
4671
4672         if (amdgpu_gpu_recovery == 0)
4673                 goto disabled;
4674
4675         /* Skip soft reset check in fatal error mode */
4676         if (!amdgpu_ras_is_poison_mode_supported(adev))
4677                 return true;
4678
4679         if (amdgpu_sriov_vf(adev))
4680                 return true;
4681
4682         if (amdgpu_gpu_recovery == -1) {
4683                 switch (adev->asic_type) {
4684 #ifdef CONFIG_DRM_AMDGPU_SI
4685                 case CHIP_VERDE:
4686                 case CHIP_TAHITI:
4687                 case CHIP_PITCAIRN:
4688                 case CHIP_OLAND:
4689                 case CHIP_HAINAN:
4690 #endif
4691 #ifdef CONFIG_DRM_AMDGPU_CIK
4692                 case CHIP_KAVERI:
4693                 case CHIP_KABINI:
4694                 case CHIP_MULLINS:
4695 #endif
4696                 case CHIP_CARRIZO:
4697                 case CHIP_STONEY:
4698                 case CHIP_CYAN_SKILLFISH:
4699                         goto disabled;
4700                 default:
4701                         break;
4702                 }
4703         }
4704
4705         return true;
4706
4707 disabled:
4708                 dev_info(adev->dev, "GPU recovery disabled.\n");
4709                 return false;
4710 }
4711
4712 int amdgpu_device_mode1_reset(struct amdgpu_device *adev)
4713 {
4714         u32 i;
4715         int ret = 0;
4716
4717         amdgpu_atombios_scratch_regs_engine_hung(adev, true);
4718
4719         dev_info(adev->dev, "GPU mode1 reset\n");
4720
4721         /* disable BM */
4722         pci_clear_master(adev->pdev);
4723
4724         amdgpu_device_cache_pci_state(adev->pdev);
4725
4726         if (amdgpu_dpm_is_mode1_reset_supported(adev)) {
4727                 dev_info(adev->dev, "GPU smu mode1 reset\n");
4728                 ret = amdgpu_dpm_mode1_reset(adev);
4729         } else {
4730                 dev_info(adev->dev, "GPU psp mode1 reset\n");
4731                 ret = psp_gpu_reset(adev);
4732         }
4733
4734         if (ret)
4735                 dev_err(adev->dev, "GPU mode1 reset failed\n");
4736
4737         amdgpu_device_load_pci_state(adev->pdev);
4738
4739         /* wait for asic to come out of reset */
4740         for (i = 0; i < adev->usec_timeout; i++) {
4741                 u32 memsize = adev->nbio.funcs->get_memsize(adev);
4742
4743                 if (memsize != 0xffffffff)
4744                         break;
4745                 udelay(1);
4746         }
4747
4748         amdgpu_atombios_scratch_regs_engine_hung(adev, false);
4749         return ret;
4750 }
4751
4752 int amdgpu_device_pre_asic_reset(struct amdgpu_device *adev,
4753                                  struct amdgpu_reset_context *reset_context)
4754 {
4755         int i, r = 0;
4756         struct amdgpu_job *job = NULL;
4757         bool need_full_reset =
4758                 test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
4759
4760         if (reset_context->reset_req_dev == adev)
4761                 job = reset_context->job;
4762
4763         if (amdgpu_sriov_vf(adev)) {
4764                 /* stop the data exchange thread */
4765                 amdgpu_virt_fini_data_exchange(adev);
4766         }
4767
4768         amdgpu_fence_driver_isr_toggle(adev, true);
4769
4770         /* block all schedulers and reset given job's ring */
4771         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
4772                 struct amdgpu_ring *ring = adev->rings[i];
4773
4774                 if (!ring || !ring->sched.thread)
4775                         continue;
4776
4777                 /*clear job fence from fence drv to avoid force_completion
4778                  *leave NULL and vm flush fence in fence drv */
4779                 amdgpu_fence_driver_clear_job_fences(ring);
4780
4781                 /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
4782                 amdgpu_fence_driver_force_completion(ring);
4783         }
4784
4785         amdgpu_fence_driver_isr_toggle(adev, false);
4786
4787         if (job && job->vm)
4788                 drm_sched_increase_karma(&job->base);
4789
4790         r = amdgpu_reset_prepare_hwcontext(adev, reset_context);
4791         /* If reset handler not implemented, continue; otherwise return */
4792         if (r == -ENOSYS)
4793                 r = 0;
4794         else
4795                 return r;
4796
4797         /* Don't suspend on bare metal if we are not going to HW reset the ASIC */
4798         if (!amdgpu_sriov_vf(adev)) {
4799
4800                 if (!need_full_reset)
4801                         need_full_reset = amdgpu_device_ip_need_full_reset(adev);
4802
4803                 if (!need_full_reset && amdgpu_gpu_recovery &&
4804                     amdgpu_device_ip_check_soft_reset(adev)) {
4805                         amdgpu_device_ip_pre_soft_reset(adev);
4806                         r = amdgpu_device_ip_soft_reset(adev);
4807                         amdgpu_device_ip_post_soft_reset(adev);
4808                         if (r || amdgpu_device_ip_check_soft_reset(adev)) {
4809                                 dev_info(adev->dev, "soft reset failed, will fallback to full reset!\n");
4810                                 need_full_reset = true;
4811                         }
4812                 }
4813
4814                 if (need_full_reset)
4815                         r = amdgpu_device_ip_suspend(adev);
4816                 if (need_full_reset)
4817                         set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
4818                 else
4819                         clear_bit(AMDGPU_NEED_FULL_RESET,
4820                                   &reset_context->flags);
4821         }
4822
4823         return r;
4824 }
4825
4826 static int amdgpu_reset_reg_dumps(struct amdgpu_device *adev)
4827 {
4828         int i;
4829
4830         lockdep_assert_held(&adev->reset_domain->sem);
4831
4832         for (i = 0; i < adev->num_regs; i++) {
4833                 adev->reset_dump_reg_value[i] = RREG32(adev->reset_dump_reg_list[i]);
4834                 trace_amdgpu_reset_reg_dumps(adev->reset_dump_reg_list[i],
4835                                              adev->reset_dump_reg_value[i]);
4836         }
4837
4838         return 0;
4839 }
4840
4841 #ifdef CONFIG_DEV_COREDUMP
4842 static ssize_t amdgpu_devcoredump_read(char *buffer, loff_t offset,
4843                 size_t count, void *data, size_t datalen)
4844 {
4845         struct drm_printer p;
4846         struct amdgpu_device *adev = data;
4847         struct drm_print_iterator iter;
4848         int i;
4849
4850         iter.data = buffer;
4851         iter.offset = 0;
4852         iter.start = offset;
4853         iter.remain = count;
4854
4855         p = drm_coredump_printer(&iter);
4856
4857         drm_printf(&p, "**** AMDGPU Device Coredump ****\n");
4858         drm_printf(&p, "kernel: " UTS_RELEASE "\n");
4859         drm_printf(&p, "module: " KBUILD_MODNAME "\n");
4860         drm_printf(&p, "time: %lld.%09ld\n", adev->reset_time.tv_sec, adev->reset_time.tv_nsec);
4861         if (adev->reset_task_info.pid)
4862                 drm_printf(&p, "process_name: %s PID: %d\n",
4863                            adev->reset_task_info.process_name,
4864                            adev->reset_task_info.pid);
4865
4866         if (adev->reset_vram_lost)
4867                 drm_printf(&p, "VRAM is lost due to GPU reset!\n");
4868         if (adev->num_regs) {
4869                 drm_printf(&p, "AMDGPU register dumps:\nOffset:     Value:\n");
4870
4871                 for (i = 0; i < adev->num_regs; i++)
4872                         drm_printf(&p, "0x%08x: 0x%08x\n",
4873                                    adev->reset_dump_reg_list[i],
4874                                    adev->reset_dump_reg_value[i]);
4875         }
4876
4877         return count - iter.remain;
4878 }
4879
4880 static void amdgpu_devcoredump_free(void *data)
4881 {
4882 }
4883
4884 static void amdgpu_reset_capture_coredumpm(struct amdgpu_device *adev)
4885 {
4886         struct drm_device *dev = adev_to_drm(adev);
4887
4888         ktime_get_ts64(&adev->reset_time);
4889         dev_coredumpm(dev->dev, THIS_MODULE, adev, 0, GFP_KERNEL,
4890                       amdgpu_devcoredump_read, amdgpu_devcoredump_free);
4891 }
4892 #endif
4893
4894 int amdgpu_do_asic_reset(struct list_head *device_list_handle,
4895                          struct amdgpu_reset_context *reset_context)
4896 {
4897         struct amdgpu_device *tmp_adev = NULL;
4898         bool need_full_reset, skip_hw_reset, vram_lost = false;
4899         int r = 0;
4900         bool gpu_reset_for_dev_remove = 0;
4901
4902         /* Try reset handler method first */
4903         tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,
4904                                     reset_list);
4905         amdgpu_reset_reg_dumps(tmp_adev);
4906
4907         reset_context->reset_device_list = device_list_handle;
4908         r = amdgpu_reset_perform_reset(tmp_adev, reset_context);
4909         /* If reset handler not implemented, continue; otherwise return */
4910         if (r == -ENOSYS)
4911                 r = 0;
4912         else
4913                 return r;
4914
4915         /* Reset handler not implemented, use the default method */
4916         need_full_reset =
4917                 test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
4918         skip_hw_reset = test_bit(AMDGPU_SKIP_HW_RESET, &reset_context->flags);
4919
4920         gpu_reset_for_dev_remove =
4921                 test_bit(AMDGPU_RESET_FOR_DEVICE_REMOVE, &reset_context->flags) &&
4922                         test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
4923
4924         /*
4925          * ASIC reset has to be done on all XGMI hive nodes ASAP
4926          * to allow proper links negotiation in FW (within 1 sec)
4927          */
4928         if (!skip_hw_reset && need_full_reset) {
4929                 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
4930                         /* For XGMI run all resets in parallel to speed up the process */
4931                         if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
4932                                 tmp_adev->gmc.xgmi.pending_reset = false;
4933                                 if (!queue_work(system_unbound_wq, &tmp_adev->xgmi_reset_work))
4934                                         r = -EALREADY;
4935                         } else
4936                                 r = amdgpu_asic_reset(tmp_adev);
4937
4938                         if (r) {
4939                                 dev_err(tmp_adev->dev, "ASIC reset failed with error, %d for drm dev, %s",
4940                                          r, adev_to_drm(tmp_adev)->unique);
4941                                 break;
4942                         }
4943                 }
4944
4945                 /* For XGMI wait for all resets to complete before proceed */
4946                 if (!r) {
4947                         list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
4948                                 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) {
4949                                         flush_work(&tmp_adev->xgmi_reset_work);
4950                                         r = tmp_adev->asic_reset_res;
4951                                         if (r)
4952                                                 break;
4953                                 }
4954                         }
4955                 }
4956         }
4957
4958         if (!r && amdgpu_ras_intr_triggered()) {
4959                 list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
4960                         if (tmp_adev->mmhub.ras && tmp_adev->mmhub.ras->ras_block.hw_ops &&
4961                             tmp_adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count)
4962                                 tmp_adev->mmhub.ras->ras_block.hw_ops->reset_ras_error_count(tmp_adev);
4963                 }
4964
4965                 amdgpu_ras_intr_cleared();
4966         }
4967
4968         /* Since the mode1 reset affects base ip blocks, the
4969          * phase1 ip blocks need to be resumed. Otherwise there
4970          * will be a BIOS signature error and the psp bootloader
4971          * can't load kdb on the next amdgpu install.
4972          */
4973         if (gpu_reset_for_dev_remove) {
4974                 list_for_each_entry(tmp_adev, device_list_handle, reset_list)
4975                         amdgpu_device_ip_resume_phase1(tmp_adev);
4976
4977                 goto end;
4978         }
4979
4980         list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
4981                 if (need_full_reset) {
4982                         /* post card */
4983                         r = amdgpu_device_asic_init(tmp_adev);
4984                         if (r) {
4985                                 dev_warn(tmp_adev->dev, "asic atom init failed!");
4986                         } else {
4987                                 dev_info(tmp_adev->dev, "GPU reset succeeded, trying to resume\n");
4988                                 r = amdgpu_amdkfd_resume_iommu(tmp_adev);
4989                                 if (r)
4990                                         goto out;
4991
4992                                 r = amdgpu_device_ip_resume_phase1(tmp_adev);
4993                                 if (r)
4994                                         goto out;
4995
4996                                 vram_lost = amdgpu_device_check_vram_lost(tmp_adev);
4997 #ifdef CONFIG_DEV_COREDUMP
4998                                 tmp_adev->reset_vram_lost = vram_lost;
4999                                 memset(&tmp_adev->reset_task_info, 0,
5000                                                 sizeof(tmp_adev->reset_task_info));
5001                                 if (reset_context->job && reset_context->job->vm)
5002                                         tmp_adev->reset_task_info =
5003                                                 reset_context->job->vm->task_info;
5004                                 amdgpu_reset_capture_coredumpm(tmp_adev);
5005 #endif
5006                                 if (vram_lost) {
5007                                         DRM_INFO("VRAM is lost due to GPU reset!\n");
5008                                         amdgpu_inc_vram_lost(tmp_adev);
5009                                 }
5010
5011                                 r = amdgpu_device_fw_loading(tmp_adev);
5012                                 if (r)
5013                                         return r;
5014
5015                                 r = amdgpu_device_ip_resume_phase2(tmp_adev);
5016                                 if (r)
5017                                         goto out;
5018
5019                                 if (vram_lost)
5020                                         amdgpu_device_fill_reset_magic(tmp_adev);
5021
5022                                 /*
5023                                  * Add this ASIC as tracked as reset was already
5024                                  * complete successfully.
5025                                  */
5026                                 amdgpu_register_gpu_instance(tmp_adev);
5027
5028                                 if (!reset_context->hive &&
5029                                     tmp_adev->gmc.xgmi.num_physical_nodes > 1)
5030                                         amdgpu_xgmi_add_device(tmp_adev);
5031
5032                                 r = amdgpu_device_ip_late_init(tmp_adev);
5033                                 if (r)
5034                                         goto out;
5035
5036                                 drm_fb_helper_set_suspend_unlocked(adev_to_drm(tmp_adev)->fb_helper, false);
5037
5038                                 /*
5039                                  * The GPU enters bad state once faulty pages
5040                                  * by ECC has reached the threshold, and ras
5041                                  * recovery is scheduled next. So add one check
5042                                  * here to break recovery if it indeed exceeds
5043                                  * bad page threshold, and remind user to
5044                                  * retire this GPU or setting one bigger
5045                                  * bad_page_threshold value to fix this once
5046                                  * probing driver again.
5047                                  */
5048                                 if (!amdgpu_ras_eeprom_check_err_threshold(tmp_adev)) {
5049                                         /* must succeed. */
5050                                         amdgpu_ras_resume(tmp_adev);
5051                                 } else {
5052                                         r = -EINVAL;
5053                                         goto out;
5054                                 }
5055
5056                                 /* Update PSP FW topology after reset */
5057                                 if (reset_context->hive &&
5058                                     tmp_adev->gmc.xgmi.num_physical_nodes > 1)
5059                                         r = amdgpu_xgmi_update_topology(
5060                                                 reset_context->hive, tmp_adev);
5061                         }
5062                 }
5063
5064 out:
5065                 if (!r) {
5066                         amdgpu_irq_gpu_reset_resume_helper(tmp_adev);
5067                         r = amdgpu_ib_ring_tests(tmp_adev);
5068                         if (r) {
5069                                 dev_err(tmp_adev->dev, "ib ring test failed (%d).\n", r);
5070                                 need_full_reset = true;
5071                                 r = -EAGAIN;
5072                                 goto end;
5073                         }
5074                 }
5075
5076                 if (!r)
5077                         r = amdgpu_device_recover_vram(tmp_adev);
5078                 else
5079                         tmp_adev->asic_reset_res = r;
5080         }
5081
5082 end:
5083         if (need_full_reset)
5084                 set_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5085         else
5086                 clear_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5087         return r;
5088 }
5089
5090 static void amdgpu_device_set_mp1_state(struct amdgpu_device *adev)
5091 {
5092
5093         switch (amdgpu_asic_reset_method(adev)) {
5094         case AMD_RESET_METHOD_MODE1:
5095                 adev->mp1_state = PP_MP1_STATE_SHUTDOWN;
5096                 break;
5097         case AMD_RESET_METHOD_MODE2:
5098                 adev->mp1_state = PP_MP1_STATE_RESET;
5099                 break;
5100         default:
5101                 adev->mp1_state = PP_MP1_STATE_NONE;
5102                 break;
5103         }
5104 }
5105
5106 static void amdgpu_device_unset_mp1_state(struct amdgpu_device *adev)
5107 {
5108         amdgpu_vf_error_trans_all(adev);
5109         adev->mp1_state = PP_MP1_STATE_NONE;
5110 }
5111
5112 static void amdgpu_device_resume_display_audio(struct amdgpu_device *adev)
5113 {
5114         struct pci_dev *p = NULL;
5115
5116         p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus),
5117                         adev->pdev->bus->number, 1);
5118         if (p) {
5119                 pm_runtime_enable(&(p->dev));
5120                 pm_runtime_resume(&(p->dev));
5121         }
5122
5123         pci_dev_put(p);
5124 }
5125
5126 static int amdgpu_device_suspend_display_audio(struct amdgpu_device *adev)
5127 {
5128         enum amd_reset_method reset_method;
5129         struct pci_dev *p = NULL;
5130         u64 expires;
5131
5132         /*
5133          * For now, only BACO and mode1 reset are confirmed
5134          * to suffer the audio issue without proper suspended.
5135          */
5136         reset_method = amdgpu_asic_reset_method(adev);
5137         if ((reset_method != AMD_RESET_METHOD_BACO) &&
5138              (reset_method != AMD_RESET_METHOD_MODE1))
5139                 return -EINVAL;
5140
5141         p = pci_get_domain_bus_and_slot(pci_domain_nr(adev->pdev->bus),
5142                         adev->pdev->bus->number, 1);
5143         if (!p)
5144                 return -ENODEV;
5145
5146         expires = pm_runtime_autosuspend_expiration(&(p->dev));
5147         if (!expires)
5148                 /*
5149                  * If we cannot get the audio device autosuspend delay,
5150                  * a fixed 4S interval will be used. Considering 3S is
5151                  * the audio controller default autosuspend delay setting.
5152                  * 4S used here is guaranteed to cover that.
5153                  */
5154                 expires = ktime_get_mono_fast_ns() + NSEC_PER_SEC * 4ULL;
5155
5156         while (!pm_runtime_status_suspended(&(p->dev))) {
5157                 if (!pm_runtime_suspend(&(p->dev)))
5158                         break;
5159
5160                 if (expires < ktime_get_mono_fast_ns()) {
5161                         dev_warn(adev->dev, "failed to suspend display audio\n");
5162                         pci_dev_put(p);
5163                         /* TODO: abort the succeeding gpu reset? */
5164                         return -ETIMEDOUT;
5165                 }
5166         }
5167
5168         pm_runtime_disable(&(p->dev));
5169
5170         pci_dev_put(p);
5171         return 0;
5172 }
5173
5174 static inline void amdgpu_device_stop_pending_resets(struct amdgpu_device *adev)
5175 {
5176         struct amdgpu_ras *con = amdgpu_ras_get_context(adev);
5177
5178 #if defined(CONFIG_DEBUG_FS)
5179         if (!amdgpu_sriov_vf(adev))
5180                 cancel_work(&adev->reset_work);
5181 #endif
5182
5183         if (adev->kfd.dev)
5184                 cancel_work(&adev->kfd.reset_work);
5185
5186         if (amdgpu_sriov_vf(adev))
5187                 cancel_work(&adev->virt.flr_work);
5188
5189         if (con && adev->ras_enabled)
5190                 cancel_work(&con->recovery_work);
5191
5192 }
5193
5194 /**
5195  * amdgpu_device_gpu_recover - reset the asic and recover scheduler
5196  *
5197  * @adev: amdgpu_device pointer
5198  * @job: which job trigger hang
5199  * @reset_context: amdgpu reset context pointer
5200  *
5201  * Attempt to reset the GPU if it has hung (all asics).
5202  * Attempt to do soft-reset or full-reset and reinitialize Asic
5203  * Returns 0 for success or an error on failure.
5204  */
5205
5206 int amdgpu_device_gpu_recover(struct amdgpu_device *adev,
5207                               struct amdgpu_job *job,
5208                               struct amdgpu_reset_context *reset_context)
5209 {
5210         struct list_head device_list, *device_list_handle =  NULL;
5211         bool job_signaled = false;
5212         struct amdgpu_hive_info *hive = NULL;
5213         struct amdgpu_device *tmp_adev = NULL;
5214         int i, r = 0;
5215         bool need_emergency_restart = false;
5216         bool audio_suspended = false;
5217         bool gpu_reset_for_dev_remove = false;
5218
5219         gpu_reset_for_dev_remove =
5220                         test_bit(AMDGPU_RESET_FOR_DEVICE_REMOVE, &reset_context->flags) &&
5221                                 test_bit(AMDGPU_NEED_FULL_RESET, &reset_context->flags);
5222
5223         /*
5224          * Special case: RAS triggered and full reset isn't supported
5225          */
5226         need_emergency_restart = amdgpu_ras_need_emergency_restart(adev);
5227
5228         /*
5229          * Flush RAM to disk so that after reboot
5230          * the user can read log and see why the system rebooted.
5231          */
5232         if (need_emergency_restart && amdgpu_ras_get_context(adev)->reboot) {
5233                 DRM_WARN("Emergency reboot.");
5234
5235                 ksys_sync_helper();
5236                 emergency_restart();
5237         }
5238
5239         dev_info(adev->dev, "GPU %s begin!\n",
5240                 need_emergency_restart ? "jobs stop":"reset");
5241
5242         if (!amdgpu_sriov_vf(adev))
5243                 hive = amdgpu_get_xgmi_hive(adev);
5244         if (hive)
5245                 mutex_lock(&hive->hive_lock);
5246
5247         reset_context->job = job;
5248         reset_context->hive = hive;
5249         /*
5250          * Build list of devices to reset.
5251          * In case we are in XGMI hive mode, resort the device list
5252          * to put adev in the 1st position.
5253          */
5254         INIT_LIST_HEAD(&device_list);
5255         if (!amdgpu_sriov_vf(adev) && (adev->gmc.xgmi.num_physical_nodes > 1)) {
5256                 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) {
5257                         list_add_tail(&tmp_adev->reset_list, &device_list);
5258                         if (gpu_reset_for_dev_remove && adev->shutdown)
5259                                 tmp_adev->shutdown = true;
5260                 }
5261                 if (!list_is_first(&adev->reset_list, &device_list))
5262                         list_rotate_to_front(&adev->reset_list, &device_list);
5263                 device_list_handle = &device_list;
5264         } else {
5265                 list_add_tail(&adev->reset_list, &device_list);
5266                 device_list_handle = &device_list;
5267         }
5268
5269         /* We need to lock reset domain only once both for XGMI and single device */
5270         tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,
5271                                     reset_list);
5272         amdgpu_device_lock_reset_domain(tmp_adev->reset_domain);
5273
5274         /* block all schedulers and reset given job's ring */
5275         list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5276
5277                 amdgpu_device_set_mp1_state(tmp_adev);
5278
5279                 /*
5280                  * Try to put the audio codec into suspend state
5281                  * before gpu reset started.
5282                  *
5283                  * Due to the power domain of the graphics device
5284                  * is shared with AZ power domain. Without this,
5285                  * we may change the audio hardware from behind
5286                  * the audio driver's back. That will trigger
5287                  * some audio codec errors.
5288                  */
5289                 if (!amdgpu_device_suspend_display_audio(tmp_adev))
5290                         audio_suspended = true;
5291
5292                 amdgpu_ras_set_error_query_ready(tmp_adev, false);
5293
5294                 cancel_delayed_work_sync(&tmp_adev->delayed_init_work);
5295
5296                 if (!amdgpu_sriov_vf(tmp_adev))
5297                         amdgpu_amdkfd_pre_reset(tmp_adev);
5298
5299                 /*
5300                  * Mark these ASICs to be reseted as untracked first
5301                  * And add them back after reset completed
5302                  */
5303                 amdgpu_unregister_gpu_instance(tmp_adev);
5304
5305                 drm_fb_helper_set_suspend_unlocked(adev_to_drm(tmp_adev)->fb_helper, true);
5306
5307                 /* disable ras on ALL IPs */
5308                 if (!need_emergency_restart &&
5309                       amdgpu_device_ip_need_full_reset(tmp_adev))
5310                         amdgpu_ras_suspend(tmp_adev);
5311
5312                 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5313                         struct amdgpu_ring *ring = tmp_adev->rings[i];
5314
5315                         if (!ring || !ring->sched.thread)
5316                                 continue;
5317
5318                         drm_sched_stop(&ring->sched, job ? &job->base : NULL);
5319
5320                         if (need_emergency_restart)
5321                                 amdgpu_job_stop_all_jobs_on_sched(&ring->sched);
5322                 }
5323                 atomic_inc(&tmp_adev->gpu_reset_counter);
5324         }
5325
5326         if (need_emergency_restart)
5327                 goto skip_sched_resume;
5328
5329         /*
5330          * Must check guilty signal here since after this point all old
5331          * HW fences are force signaled.
5332          *
5333          * job->base holds a reference to parent fence
5334          */
5335         if (job && dma_fence_is_signaled(&job->hw_fence)) {
5336                 job_signaled = true;
5337                 dev_info(adev->dev, "Guilty job already signaled, skipping HW reset");
5338                 goto skip_hw_reset;
5339         }
5340
5341 retry:  /* Rest of adevs pre asic reset from XGMI hive. */
5342         list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5343                 if (gpu_reset_for_dev_remove) {
5344                         /* Workaroud for ASICs need to disable SMC first */
5345                         amdgpu_device_smu_fini_early(tmp_adev);
5346                 }
5347                 r = amdgpu_device_pre_asic_reset(tmp_adev, reset_context);
5348                 /*TODO Should we stop ?*/
5349                 if (r) {
5350                         dev_err(tmp_adev->dev, "GPU pre asic reset failed with err, %d for drm dev, %s ",
5351                                   r, adev_to_drm(tmp_adev)->unique);
5352                         tmp_adev->asic_reset_res = r;
5353                 }
5354
5355                 /*
5356                  * Drop all pending non scheduler resets. Scheduler resets
5357                  * were already dropped during drm_sched_stop
5358                  */
5359                 amdgpu_device_stop_pending_resets(tmp_adev);
5360         }
5361
5362         /* Actual ASIC resets if needed.*/
5363         /* Host driver will handle XGMI hive reset for SRIOV */
5364         if (amdgpu_sriov_vf(adev)) {
5365                 r = amdgpu_device_reset_sriov(adev, job ? false : true);
5366                 if (r)
5367                         adev->asic_reset_res = r;
5368
5369                 /* Aldebaran and gfx_11_0_3 support ras in SRIOV, so need resume ras during reset */
5370                 if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 2) ||
5371                     adev->ip_versions[GC_HWIP][0] == IP_VERSION(11, 0, 3))
5372                         amdgpu_ras_resume(adev);
5373         } else {
5374                 r = amdgpu_do_asic_reset(device_list_handle, reset_context);
5375                 if (r && r == -EAGAIN)
5376                         goto retry;
5377
5378                 if (!r && gpu_reset_for_dev_remove)
5379                         goto recover_end;
5380         }
5381
5382 skip_hw_reset:
5383
5384         /* Post ASIC reset for all devs .*/
5385         list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5386
5387                 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5388                         struct amdgpu_ring *ring = tmp_adev->rings[i];
5389
5390                         if (!ring || !ring->sched.thread)
5391                                 continue;
5392
5393                         drm_sched_start(&ring->sched, true);
5394                 }
5395
5396                 if (adev->enable_mes && adev->ip_versions[GC_HWIP][0] != IP_VERSION(11, 0, 3))
5397                         amdgpu_mes_self_test(tmp_adev);
5398
5399                 if (!drm_drv_uses_atomic_modeset(adev_to_drm(tmp_adev)) && !job_signaled) {
5400                         drm_helper_resume_force_mode(adev_to_drm(tmp_adev));
5401                 }
5402
5403                 if (tmp_adev->asic_reset_res)
5404                         r = tmp_adev->asic_reset_res;
5405
5406                 tmp_adev->asic_reset_res = 0;
5407
5408                 if (r) {
5409                         /* bad news, how to tell it to userspace ? */
5410                         dev_info(tmp_adev->dev, "GPU reset(%d) failed\n", atomic_read(&tmp_adev->gpu_reset_counter));
5411                         amdgpu_vf_error_put(tmp_adev, AMDGIM_ERROR_VF_GPU_RESET_FAIL, 0, r);
5412                 } else {
5413                         dev_info(tmp_adev->dev, "GPU reset(%d) succeeded!\n", atomic_read(&tmp_adev->gpu_reset_counter));
5414                         if (amdgpu_acpi_smart_shift_update(adev_to_drm(tmp_adev), AMDGPU_SS_DEV_D0))
5415                                 DRM_WARN("smart shift update failed\n");
5416                 }
5417         }
5418
5419 skip_sched_resume:
5420         list_for_each_entry(tmp_adev, device_list_handle, reset_list) {
5421                 /* unlock kfd: SRIOV would do it separately */
5422                 if (!need_emergency_restart && !amdgpu_sriov_vf(tmp_adev))
5423                         amdgpu_amdkfd_post_reset(tmp_adev);
5424
5425                 /* kfd_post_reset will do nothing if kfd device is not initialized,
5426                  * need to bring up kfd here if it's not be initialized before
5427                  */
5428                 if (!adev->kfd.init_complete)
5429                         amdgpu_amdkfd_device_init(adev);
5430
5431                 if (audio_suspended)
5432                         amdgpu_device_resume_display_audio(tmp_adev);
5433
5434                 amdgpu_device_unset_mp1_state(tmp_adev);
5435
5436                 amdgpu_ras_set_error_query_ready(tmp_adev, true);
5437         }
5438
5439 recover_end:
5440         tmp_adev = list_first_entry(device_list_handle, struct amdgpu_device,
5441                                             reset_list);
5442         amdgpu_device_unlock_reset_domain(tmp_adev->reset_domain);
5443
5444         if (hive) {
5445                 mutex_unlock(&hive->hive_lock);
5446                 amdgpu_put_xgmi_hive(hive);
5447         }
5448
5449         if (r)
5450                 dev_info(adev->dev, "GPU reset end with ret = %d\n", r);
5451
5452         atomic_set(&adev->reset_domain->reset_res, r);
5453         return r;
5454 }
5455
5456 /**
5457  * amdgpu_device_get_pcie_info - fence pcie info about the PCIE slot
5458  *
5459  * @adev: amdgpu_device pointer
5460  *
5461  * Fetchs and stores in the driver the PCIE capabilities (gen speed
5462  * and lanes) of the slot the device is in. Handles APUs and
5463  * virtualized environments where PCIE config space may not be available.
5464  */
5465 static void amdgpu_device_get_pcie_info(struct amdgpu_device *adev)
5466 {
5467         struct pci_dev *pdev;
5468         enum pci_bus_speed speed_cap, platform_speed_cap;
5469         enum pcie_link_width platform_link_width;
5470
5471         if (amdgpu_pcie_gen_cap)
5472                 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
5473
5474         if (amdgpu_pcie_lane_cap)
5475                 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
5476
5477         /* covers APUs as well */
5478         if (pci_is_root_bus(adev->pdev->bus)) {
5479                 if (adev->pm.pcie_gen_mask == 0)
5480                         adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
5481                 if (adev->pm.pcie_mlw_mask == 0)
5482                         adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
5483                 return;
5484         }
5485
5486         if (adev->pm.pcie_gen_mask && adev->pm.pcie_mlw_mask)
5487                 return;
5488
5489         pcie_bandwidth_available(adev->pdev, NULL,
5490                                  &platform_speed_cap, &platform_link_width);
5491
5492         if (adev->pm.pcie_gen_mask == 0) {
5493                 /* asic caps */
5494                 pdev = adev->pdev;
5495                 speed_cap = pcie_get_speed_cap(pdev);
5496                 if (speed_cap == PCI_SPEED_UNKNOWN) {
5497                         adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5498                                                   CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5499                                                   CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
5500                 } else {
5501                         if (speed_cap == PCIE_SPEED_32_0GT)
5502                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5503                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5504                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5505                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4 |
5506                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN5);
5507                         else if (speed_cap == PCIE_SPEED_16_0GT)
5508                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5509                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5510                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5511                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN4);
5512                         else if (speed_cap == PCIE_SPEED_8_0GT)
5513                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5514                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5515                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
5516                         else if (speed_cap == PCIE_SPEED_5_0GT)
5517                                 adev->pm.pcie_gen_mask |= (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5518                                                           CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2);
5519                         else
5520                                 adev->pm.pcie_gen_mask |= CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1;
5521                 }
5522                 /* platform caps */
5523                 if (platform_speed_cap == PCI_SPEED_UNKNOWN) {
5524                         adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5525                                                    CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
5526                 } else {
5527                         if (platform_speed_cap == PCIE_SPEED_32_0GT)
5528                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5529                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5530                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5531                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4 |
5532                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN5);
5533                         else if (platform_speed_cap == PCIE_SPEED_16_0GT)
5534                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5535                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5536                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3 |
5537                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN4);
5538                         else if (platform_speed_cap == PCIE_SPEED_8_0GT)
5539                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5540                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2 |
5541                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3);
5542                         else if (platform_speed_cap == PCIE_SPEED_5_0GT)
5543                                 adev->pm.pcie_gen_mask |= (CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1 |
5544                                                            CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2);
5545                         else
5546                                 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
5547
5548                 }
5549         }
5550         if (adev->pm.pcie_mlw_mask == 0) {
5551                 if (platform_link_width == PCIE_LNK_WIDTH_UNKNOWN) {
5552                         adev->pm.pcie_mlw_mask |= AMDGPU_DEFAULT_PCIE_MLW_MASK;
5553                 } else {
5554                         switch (platform_link_width) {
5555                         case PCIE_LNK_X32:
5556                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
5557                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
5558                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
5559                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5560                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5561                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5562                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5563                                 break;
5564                         case PCIE_LNK_X16:
5565                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
5566                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
5567                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5568                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5569                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5570                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5571                                 break;
5572                         case PCIE_LNK_X12:
5573                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
5574                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5575                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5576                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5577                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5578                                 break;
5579                         case PCIE_LNK_X8:
5580                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
5581                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5582                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5583                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5584                                 break;
5585                         case PCIE_LNK_X4:
5586                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
5587                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5588                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5589                                 break;
5590                         case PCIE_LNK_X2:
5591                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
5592                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
5593                                 break;
5594                         case PCIE_LNK_X1:
5595                                 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
5596                                 break;
5597                         default:
5598                                 break;
5599                         }
5600                 }
5601         }
5602 }
5603
5604 /**
5605  * amdgpu_device_is_peer_accessible - Check peer access through PCIe BAR
5606  *
5607  * @adev: amdgpu_device pointer
5608  * @peer_adev: amdgpu_device pointer for peer device trying to access @adev
5609  *
5610  * Return true if @peer_adev can access (DMA) @adev through the PCIe
5611  * BAR, i.e. @adev is "large BAR" and the BAR matches the DMA mask of
5612  * @peer_adev.
5613  */
5614 bool amdgpu_device_is_peer_accessible(struct amdgpu_device *adev,
5615                                       struct amdgpu_device *peer_adev)
5616 {
5617 #ifdef CONFIG_HSA_AMD_P2P
5618         uint64_t address_mask = peer_adev->dev->dma_mask ?
5619                 ~*peer_adev->dev->dma_mask : ~((1ULL << 32) - 1);
5620         resource_size_t aper_limit =
5621                 adev->gmc.aper_base + adev->gmc.aper_size - 1;
5622         bool p2p_access =
5623                 !adev->gmc.xgmi.connected_to_cpu &&
5624                 !(pci_p2pdma_distance(adev->pdev, peer_adev->dev, false) < 0);
5625
5626         return pcie_p2p && p2p_access && (adev->gmc.visible_vram_size &&
5627                 adev->gmc.real_vram_size == adev->gmc.visible_vram_size &&
5628                 !(adev->gmc.aper_base & address_mask ||
5629                   aper_limit & address_mask));
5630 #else
5631         return false;
5632 #endif
5633 }
5634
5635 int amdgpu_device_baco_enter(struct drm_device *dev)
5636 {
5637         struct amdgpu_device *adev = drm_to_adev(dev);
5638         struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
5639
5640         if (!amdgpu_device_supports_baco(dev))
5641                 return -ENOTSUPP;
5642
5643         if (ras && adev->ras_enabled &&
5644             adev->nbio.funcs->enable_doorbell_interrupt)
5645                 adev->nbio.funcs->enable_doorbell_interrupt(adev, false);
5646
5647         return amdgpu_dpm_baco_enter(adev);
5648 }
5649
5650 int amdgpu_device_baco_exit(struct drm_device *dev)
5651 {
5652         struct amdgpu_device *adev = drm_to_adev(dev);
5653         struct amdgpu_ras *ras = amdgpu_ras_get_context(adev);
5654         int ret = 0;
5655
5656         if (!amdgpu_device_supports_baco(dev))
5657                 return -ENOTSUPP;
5658
5659         ret = amdgpu_dpm_baco_exit(adev);
5660         if (ret)
5661                 return ret;
5662
5663         if (ras && adev->ras_enabled &&
5664             adev->nbio.funcs->enable_doorbell_interrupt)
5665                 adev->nbio.funcs->enable_doorbell_interrupt(adev, true);
5666
5667         if (amdgpu_passthrough(adev) &&
5668             adev->nbio.funcs->clear_doorbell_interrupt)
5669                 adev->nbio.funcs->clear_doorbell_interrupt(adev);
5670
5671         return 0;
5672 }
5673
5674 /**
5675  * amdgpu_pci_error_detected - Called when a PCI error is detected.
5676  * @pdev: PCI device struct
5677  * @state: PCI channel state
5678  *
5679  * Description: Called when a PCI error is detected.
5680  *
5681  * Return: PCI_ERS_RESULT_NEED_RESET or PCI_ERS_RESULT_DISCONNECT.
5682  */
5683 pci_ers_result_t amdgpu_pci_error_detected(struct pci_dev *pdev, pci_channel_state_t state)
5684 {
5685         struct drm_device *dev = pci_get_drvdata(pdev);
5686         struct amdgpu_device *adev = drm_to_adev(dev);
5687         int i;
5688
5689         DRM_INFO("PCI error: detected callback, state(%d)!!\n", state);
5690
5691         if (adev->gmc.xgmi.num_physical_nodes > 1) {
5692                 DRM_WARN("No support for XGMI hive yet...");
5693                 return PCI_ERS_RESULT_DISCONNECT;
5694         }
5695
5696         adev->pci_channel_state = state;
5697
5698         switch (state) {
5699         case pci_channel_io_normal:
5700                 return PCI_ERS_RESULT_CAN_RECOVER;
5701         /* Fatal error, prepare for slot reset */
5702         case pci_channel_io_frozen:
5703                 /*
5704                  * Locking adev->reset_domain->sem will prevent any external access
5705                  * to GPU during PCI error recovery
5706                  */
5707                 amdgpu_device_lock_reset_domain(adev->reset_domain);
5708                 amdgpu_device_set_mp1_state(adev);
5709
5710                 /*
5711                  * Block any work scheduling as we do for regular GPU reset
5712                  * for the duration of the recovery
5713                  */
5714                 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5715                         struct amdgpu_ring *ring = adev->rings[i];
5716
5717                         if (!ring || !ring->sched.thread)
5718                                 continue;
5719
5720                         drm_sched_stop(&ring->sched, NULL);
5721                 }
5722                 atomic_inc(&adev->gpu_reset_counter);
5723                 return PCI_ERS_RESULT_NEED_RESET;
5724         case pci_channel_io_perm_failure:
5725                 /* Permanent error, prepare for device removal */
5726                 return PCI_ERS_RESULT_DISCONNECT;
5727         }
5728
5729         return PCI_ERS_RESULT_NEED_RESET;
5730 }
5731
5732 /**
5733  * amdgpu_pci_mmio_enabled - Enable MMIO and dump debug registers
5734  * @pdev: pointer to PCI device
5735  */
5736 pci_ers_result_t amdgpu_pci_mmio_enabled(struct pci_dev *pdev)
5737 {
5738
5739         DRM_INFO("PCI error: mmio enabled callback!!\n");
5740
5741         /* TODO - dump whatever for debugging purposes */
5742
5743         /* This called only if amdgpu_pci_error_detected returns
5744          * PCI_ERS_RESULT_CAN_RECOVER. Read/write to the device still
5745          * works, no need to reset slot.
5746          */
5747
5748         return PCI_ERS_RESULT_RECOVERED;
5749 }
5750
5751 /**
5752  * amdgpu_pci_slot_reset - Called when PCI slot has been reset.
5753  * @pdev: PCI device struct
5754  *
5755  * Description: This routine is called by the pci error recovery
5756  * code after the PCI slot has been reset, just before we
5757  * should resume normal operations.
5758  */
5759 pci_ers_result_t amdgpu_pci_slot_reset(struct pci_dev *pdev)
5760 {
5761         struct drm_device *dev = pci_get_drvdata(pdev);
5762         struct amdgpu_device *adev = drm_to_adev(dev);
5763         int r, i;
5764         struct amdgpu_reset_context reset_context;
5765         u32 memsize;
5766         struct list_head device_list;
5767
5768         DRM_INFO("PCI error: slot reset callback!!\n");
5769
5770         memset(&reset_context, 0, sizeof(reset_context));
5771
5772         INIT_LIST_HEAD(&device_list);
5773         list_add_tail(&adev->reset_list, &device_list);
5774
5775         /* wait for asic to come out of reset */
5776         msleep(500);
5777
5778         /* Restore PCI confspace */
5779         amdgpu_device_load_pci_state(pdev);
5780
5781         /* confirm  ASIC came out of reset */
5782         for (i = 0; i < adev->usec_timeout; i++) {
5783                 memsize = amdgpu_asic_get_config_memsize(adev);
5784
5785                 if (memsize != 0xffffffff)
5786                         break;
5787                 udelay(1);
5788         }
5789         if (memsize == 0xffffffff) {
5790                 r = -ETIME;
5791                 goto out;
5792         }
5793
5794         reset_context.method = AMD_RESET_METHOD_NONE;
5795         reset_context.reset_req_dev = adev;
5796         set_bit(AMDGPU_NEED_FULL_RESET, &reset_context.flags);
5797         set_bit(AMDGPU_SKIP_HW_RESET, &reset_context.flags);
5798
5799         adev->no_hw_access = true;
5800         r = amdgpu_device_pre_asic_reset(adev, &reset_context);
5801         adev->no_hw_access = false;
5802         if (r)
5803                 goto out;
5804
5805         r = amdgpu_do_asic_reset(&device_list, &reset_context);
5806
5807 out:
5808         if (!r) {
5809                 if (amdgpu_device_cache_pci_state(adev->pdev))
5810                         pci_restore_state(adev->pdev);
5811
5812                 DRM_INFO("PCIe error recovery succeeded\n");
5813         } else {
5814                 DRM_ERROR("PCIe error recovery failed, err:%d", r);
5815                 amdgpu_device_unset_mp1_state(adev);
5816                 amdgpu_device_unlock_reset_domain(adev->reset_domain);
5817         }
5818
5819         return r ? PCI_ERS_RESULT_DISCONNECT : PCI_ERS_RESULT_RECOVERED;
5820 }
5821
5822 /**
5823  * amdgpu_pci_resume() - resume normal ops after PCI reset
5824  * @pdev: pointer to PCI device
5825  *
5826  * Called when the error recovery driver tells us that its
5827  * OK to resume normal operation.
5828  */
5829 void amdgpu_pci_resume(struct pci_dev *pdev)
5830 {
5831         struct drm_device *dev = pci_get_drvdata(pdev);
5832         struct amdgpu_device *adev = drm_to_adev(dev);
5833         int i;
5834
5835
5836         DRM_INFO("PCI error: resume callback!!\n");
5837
5838         /* Only continue execution for the case of pci_channel_io_frozen */
5839         if (adev->pci_channel_state != pci_channel_io_frozen)
5840                 return;
5841
5842         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
5843                 struct amdgpu_ring *ring = adev->rings[i];
5844
5845                 if (!ring || !ring->sched.thread)
5846                         continue;
5847
5848                 drm_sched_start(&ring->sched, true);
5849         }
5850
5851         amdgpu_device_unset_mp1_state(adev);
5852         amdgpu_device_unlock_reset_domain(adev->reset_domain);
5853 }
5854
5855 bool amdgpu_device_cache_pci_state(struct pci_dev *pdev)
5856 {
5857         struct drm_device *dev = pci_get_drvdata(pdev);
5858         struct amdgpu_device *adev = drm_to_adev(dev);
5859         int r;
5860
5861         r = pci_save_state(pdev);
5862         if (!r) {
5863                 kfree(adev->pci_state);
5864
5865                 adev->pci_state = pci_store_saved_state(pdev);
5866
5867                 if (!adev->pci_state) {
5868                         DRM_ERROR("Failed to store PCI saved state");
5869                         return false;
5870                 }
5871         } else {
5872                 DRM_WARN("Failed to save PCI state, err:%d\n", r);
5873                 return false;
5874         }
5875
5876         return true;
5877 }
5878
5879 bool amdgpu_device_load_pci_state(struct pci_dev *pdev)
5880 {
5881         struct drm_device *dev = pci_get_drvdata(pdev);
5882         struct amdgpu_device *adev = drm_to_adev(dev);
5883         int r;
5884
5885         if (!adev->pci_state)
5886                 return false;
5887
5888         r = pci_load_saved_state(pdev, adev->pci_state);
5889
5890         if (!r) {
5891                 pci_restore_state(pdev);
5892         } else {
5893                 DRM_WARN("Failed to load PCI state, err:%d\n", r);
5894                 return false;
5895         }
5896
5897         return true;
5898 }
5899
5900 void amdgpu_device_flush_hdp(struct amdgpu_device *adev,
5901                 struct amdgpu_ring *ring)
5902 {
5903 #ifdef CONFIG_X86_64
5904         if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev))
5905                 return;
5906 #endif
5907         if (adev->gmc.xgmi.connected_to_cpu)
5908                 return;
5909
5910         if (ring && ring->funcs->emit_hdp_flush)
5911                 amdgpu_ring_emit_hdp_flush(ring);
5912         else
5913                 amdgpu_asic_flush_hdp(adev, ring);
5914 }
5915
5916 void amdgpu_device_invalidate_hdp(struct amdgpu_device *adev,
5917                 struct amdgpu_ring *ring)
5918 {
5919 #ifdef CONFIG_X86_64
5920         if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev))
5921                 return;
5922 #endif
5923         if (adev->gmc.xgmi.connected_to_cpu)
5924                 return;
5925
5926         amdgpu_asic_invalidate_hdp(adev, ring);
5927 }
5928
5929 int amdgpu_in_reset(struct amdgpu_device *adev)
5930 {
5931         return atomic_read(&adev->reset_domain->in_gpu_reset);
5932 }
5933
5934 /**
5935  * amdgpu_device_halt() - bring hardware to some kind of halt state
5936  *
5937  * @adev: amdgpu_device pointer
5938  *
5939  * Bring hardware to some kind of halt state so that no one can touch it
5940  * any more. It will help to maintain error context when error occurred.
5941  * Compare to a simple hang, the system will keep stable at least for SSH
5942  * access. Then it should be trivial to inspect the hardware state and
5943  * see what's going on. Implemented as following:
5944  *
5945  * 1. drm_dev_unplug() makes device inaccessible to user space(IOCTLs, etc),
5946  *    clears all CPU mappings to device, disallows remappings through page faults
5947  * 2. amdgpu_irq_disable_all() disables all interrupts
5948  * 3. amdgpu_fence_driver_hw_fini() signals all HW fences
5949  * 4. set adev->no_hw_access to avoid potential crashes after setp 5
5950  * 5. amdgpu_device_unmap_mmio() clears all MMIO mappings
5951  * 6. pci_disable_device() and pci_wait_for_pending_transaction()
5952  *    flush any in flight DMA operations
5953  */
5954 void amdgpu_device_halt(struct amdgpu_device *adev)
5955 {
5956         struct pci_dev *pdev = adev->pdev;
5957         struct drm_device *ddev = adev_to_drm(adev);
5958
5959         drm_dev_unplug(ddev);
5960
5961         amdgpu_irq_disable_all(adev);
5962
5963         amdgpu_fence_driver_hw_fini(adev);
5964
5965         adev->no_hw_access = true;
5966
5967         amdgpu_device_unmap_mmio(adev);
5968
5969         pci_disable_device(pdev);
5970         pci_wait_for_pending_transaction(pdev);
5971 }
5972
5973 u32 amdgpu_device_pcie_port_rreg(struct amdgpu_device *adev,
5974                                 u32 reg)
5975 {
5976         unsigned long flags, address, data;
5977         u32 r;
5978
5979         address = adev->nbio.funcs->get_pcie_port_index_offset(adev);
5980         data = adev->nbio.funcs->get_pcie_port_data_offset(adev);
5981
5982         spin_lock_irqsave(&adev->pcie_idx_lock, flags);
5983         WREG32(address, reg * 4);
5984         (void)RREG32(address);
5985         r = RREG32(data);
5986         spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
5987         return r;
5988 }
5989
5990 void amdgpu_device_pcie_port_wreg(struct amdgpu_device *adev,
5991                                 u32 reg, u32 v)
5992 {
5993         unsigned long flags, address, data;
5994
5995         address = adev->nbio.funcs->get_pcie_port_index_offset(adev);
5996         data = adev->nbio.funcs->get_pcie_port_data_offset(adev);
5997
5998         spin_lock_irqsave(&adev->pcie_idx_lock, flags);
5999         WREG32(address, reg * 4);
6000         (void)RREG32(address);
6001         WREG32(data, v);
6002         (void)RREG32(data);
6003         spin_unlock_irqrestore(&adev->pcie_idx_lock, flags);
6004 }
6005
6006 /**
6007  * amdgpu_device_switch_gang - switch to a new gang
6008  * @adev: amdgpu_device pointer
6009  * @gang: the gang to switch to
6010  *
6011  * Try to switch to a new gang.
6012  * Returns: NULL if we switched to the new gang or a reference to the current
6013  * gang leader.
6014  */
6015 struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev,
6016                                             struct dma_fence *gang)
6017 {
6018         struct dma_fence *old = NULL;
6019
6020         do {
6021                 dma_fence_put(old);
6022                 rcu_read_lock();
6023                 old = dma_fence_get_rcu_safe(&adev->gang_submit);
6024                 rcu_read_unlock();
6025
6026                 if (old == gang)
6027                         break;
6028
6029                 if (!dma_fence_is_signaled(old))
6030                         return old;
6031
6032         } while (cmpxchg((struct dma_fence __force **)&adev->gang_submit,
6033                          old, gang) != old);
6034
6035         dma_fence_put(old);
6036         return NULL;
6037 }
6038
6039 bool amdgpu_device_has_display_hardware(struct amdgpu_device *adev)
6040 {
6041         switch (adev->asic_type) {
6042 #ifdef CONFIG_DRM_AMDGPU_SI
6043         case CHIP_HAINAN:
6044 #endif
6045         case CHIP_TOPAZ:
6046                 /* chips with no display hardware */
6047                 return false;
6048 #ifdef CONFIG_DRM_AMDGPU_SI
6049         case CHIP_TAHITI:
6050         case CHIP_PITCAIRN:
6051         case CHIP_VERDE:
6052         case CHIP_OLAND:
6053 #endif
6054 #ifdef CONFIG_DRM_AMDGPU_CIK
6055         case CHIP_BONAIRE:
6056         case CHIP_HAWAII:
6057         case CHIP_KAVERI:
6058         case CHIP_KABINI:
6059         case CHIP_MULLINS:
6060 #endif
6061         case CHIP_TONGA:
6062         case CHIP_FIJI:
6063         case CHIP_POLARIS10:
6064         case CHIP_POLARIS11:
6065         case CHIP_POLARIS12:
6066         case CHIP_VEGAM:
6067         case CHIP_CARRIZO:
6068         case CHIP_STONEY:
6069                 /* chips with display hardware */
6070                 return true;
6071         default:
6072                 /* IP discovery */
6073                 if (!adev->ip_versions[DCE_HWIP][0] ||
6074                     (adev->harvest_ip_mask & AMD_HARVEST_IP_DMU_MASK))
6075                         return false;
6076                 return true;
6077         }
6078 }