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

   1 /*
   2  * Copyright 2008 Advanced Micro Devices, Inc.
   3  * Copyright 2008 Red Hat Inc.
   4  * Copyright 2009 Jerome Glisse.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice shall be included in
  14  * all copies or substantial portions of the Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  22  * OTHER DEALINGS IN THE SOFTWARE.
  23  *
  24  * Authors: Dave Airlie
  25  *          Alex Deucher
  26  *          Jerome Glisse
  27  */
  28 #include <linux/kthread.h>
  29 #include <linux/console.h>
  30 #include <linux/slab.h>
  31 #include <linux/debugfs.h>
  32 #include <drm/drmP.h>
  33 #include <drm/drm_crtc_helper.h>
  34 #include <drm/amdgpu_drm.h>
  35 #include <linux/vgaarb.h>
  36 #include <linux/vga_switcheroo.h>
  37 #include <linux/efi.h>
  38 #include "amdgpu.h"
  39 #include "amdgpu_trace.h"
  40 #include "amdgpu_i2c.h"
  41 #include "atom.h"
  42 #include "amdgpu_atombios.h"
  43 #include "amd_pcie.h"
  44 #ifdef CONFIG_DRM_AMDGPU_SI
  45 #include "si.h"
  46 #endif
  47 #ifdef CONFIG_DRM_AMDGPU_CIK
  48 #include "cik.h"
  49 #endif
  50 #include "vi.h"
  51 #include "bif/bif_4_1_d.h"
  52 #include <linux/pci.h>
  53 #include <linux/firmware.h>
  54
  55 static int amdgpu_debugfs_regs_init(struct amdgpu_device *adev);
  56 static void amdgpu_debugfs_regs_cleanup(struct amdgpu_device *adev);
  57
  58 static const char *amdgpu_asic_name[] = {
  59         "TAHITI",
  60         "PITCAIRN",
  61         "VERDE",
  62         "OLAND",
  63         "HAINAN",
  64         "BONAIRE",
  65         "KAVERI",
  66         "KABINI",
  67         "HAWAII",
  68         "MULLINS",
  69         "TOPAZ",
  70         "TONGA",
  71         "FIJI",
  72         "CARRIZO",
  73         "STONEY",
  74         "POLARIS10",
  75         "POLARIS11",
  76         "LAST",
  77 };
  78
  79 bool amdgpu_device_is_px(struct drm_device *dev)
  80 {
  81         struct amdgpu_device *adev = dev->dev_private;
  82
  83         if (adev->flags & AMD_IS_PX)
  84                 return true;
  85         return false;
  86 }
  87
  88 /*
  89  * MMIO register access helper functions.
  90  */
  91 uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg,
  92                         bool always_indirect)
  93 {
  94         uint32_t ret;
  95
  96         if ((reg * 4) < adev->rmmio_size && !always_indirect)
  97                 ret = readl(((void __iomem *)adev->rmmio) + (reg * 4));
  98         else {
  99                 unsigned long flags;
 100
 101                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
 102                 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
 103                 ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
 104                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
 105         }
 106         trace_amdgpu_mm_rreg(adev->pdev->device, reg, ret);
 107         return ret;
 108 }
 109
 110 void amdgpu_mm_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v,
 111                     bool always_indirect)
 112 {
 113         trace_amdgpu_mm_wreg(adev->pdev->device, reg, v);
 114
 115         if ((reg * 4) < adev->rmmio_size && !always_indirect)
 116                 writel(v, ((void __iomem *)adev->rmmio) + (reg * 4));
 117         else {
 118                 unsigned long flags;
 119
 120                 spin_lock_irqsave(&adev->mmio_idx_lock, flags);
 121                 writel((reg * 4), ((void __iomem *)adev->rmmio) + (mmMM_INDEX * 4));
 122                 writel(v, ((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));
 123                 spin_unlock_irqrestore(&adev->mmio_idx_lock, flags);
 124         }
 125 }
 126
 127 u32 amdgpu_io_rreg(struct amdgpu_device *adev, u32 reg)
 128 {
 129         if ((reg * 4) < adev->rio_mem_size)
 130                 return ioread32(adev->rio_mem + (reg * 4));
 131         else {
 132                 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
 133                 return ioread32(adev->rio_mem + (mmMM_DATA * 4));
 134         }
 135 }
 136
 137 void amdgpu_io_wreg(struct amdgpu_device *adev, u32 reg, u32 v)
 138 {
 139
 140         if ((reg * 4) < adev->rio_mem_size)
 141                 iowrite32(v, adev->rio_mem + (reg * 4));
 142         else {
 143                 iowrite32((reg * 4), adev->rio_mem + (mmMM_INDEX * 4));
 144                 iowrite32(v, adev->rio_mem + (mmMM_DATA * 4));
 145         }
 146 }
 147
 148 /**
 149  * amdgpu_mm_rdoorbell - read a doorbell dword
 150  *
 151  * @adev: amdgpu_device pointer
 152  * @index: doorbell index
 153  *
 154  * Returns the value in the doorbell aperture at the
 155  * requested doorbell index (CIK).
 156  */
 157 u32 amdgpu_mm_rdoorbell(struct amdgpu_device *adev, u32 index)
 158 {
 159         if (index < adev->doorbell.num_doorbells) {
 160                 return readl(adev->doorbell.ptr + index);
 161         } else {
 162                 DRM_ERROR("reading beyond doorbell aperture: 0x%08x!\n", index);
 163                 return 0;
 164         }
 165 }
 166
 167 /**
 168  * amdgpu_mm_wdoorbell - write a doorbell dword
 169  *
 170  * @adev: amdgpu_device pointer
 171  * @index: doorbell index
 172  * @v: value to write
 173  *
 174  * Writes @v to the doorbell aperture at the
 175  * requested doorbell index (CIK).
 176  */
 177 void amdgpu_mm_wdoorbell(struct amdgpu_device *adev, u32 index, u32 v)
 178 {
 179         if (index < adev->doorbell.num_doorbells) {
 180                 writel(v, adev->doorbell.ptr + index);
 181         } else {
 182                 DRM_ERROR("writing beyond doorbell aperture: 0x%08x!\n", index);
 183         }
 184 }
 185
 186 /**
 187  * amdgpu_invalid_rreg - dummy reg read function
 188  *
 189  * @adev: amdgpu device pointer
 190  * @reg: offset of register
 191  *
 192  * Dummy register read function.  Used for register blocks
 193  * that certain asics don't have (all asics).
 194  * Returns the value in the register.
 195  */
 196 static uint32_t amdgpu_invalid_rreg(struct amdgpu_device *adev, uint32_t reg)
 197 {
 198         DRM_ERROR("Invalid callback to read register 0x%04X\n", reg);
 199         BUG();
 200         return 0;
 201 }
 202
 203 /**
 204  * amdgpu_invalid_wreg - dummy reg write function
 205  *
 206  * @adev: amdgpu device pointer
 207  * @reg: offset of register
 208  * @v: value to write to the register
 209  *
 210  * Dummy register read function.  Used for register blocks
 211  * that certain asics don't have (all asics).
 212  */
 213 static void amdgpu_invalid_wreg(struct amdgpu_device *adev, uint32_t reg, uint32_t v)
 214 {
 215         DRM_ERROR("Invalid callback to write register 0x%04X with 0x%08X\n",
 216                   reg, v);
 217         BUG();
 218 }
 219
 220 /**
 221  * amdgpu_block_invalid_rreg - dummy reg read function
 222  *
 223  * @adev: amdgpu device pointer
 224  * @block: offset of instance
 225  * @reg: offset of register
 226  *
 227  * Dummy register read function.  Used for register blocks
 228  * that certain asics don't have (all asics).
 229  * Returns the value in the register.
 230  */
 231 static uint32_t amdgpu_block_invalid_rreg(struct amdgpu_device *adev,
 232                                           uint32_t block, uint32_t reg)
 233 {
 234         DRM_ERROR("Invalid callback to read register 0x%04X in block 0x%04X\n",
 235                   reg, block);
 236         BUG();
 237         return 0;
 238 }
 239
 240 /**
 241  * amdgpu_block_invalid_wreg - dummy reg write function
 242  *
 243  * @adev: amdgpu device pointer
 244  * @block: offset of instance
 245  * @reg: offset of register
 246  * @v: value to write to the register
 247  *
 248  * Dummy register read function.  Used for register blocks
 249  * that certain asics don't have (all asics).
 250  */
 251 static void amdgpu_block_invalid_wreg(struct amdgpu_device *adev,
 252                                       uint32_t block,
 253                                       uint32_t reg, uint32_t v)
 254 {
 255         DRM_ERROR("Invalid block callback to write register 0x%04X in block 0x%04X with 0x%08X\n",
 256                   reg, block, v);
 257         BUG();
 258 }
 259
 260 static int amdgpu_vram_scratch_init(struct amdgpu_device *adev)
 261 {
 262         int r;
 263
 264         if (adev->vram_scratch.robj == NULL) {
 265                 r = amdgpu_bo_create(adev, AMDGPU_GPU_PAGE_SIZE,
 266                                      PAGE_SIZE, true, AMDGPU_GEM_DOMAIN_VRAM,
 267                                      AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED,
 268                                      NULL, NULL, &adev->vram_scratch.robj);
 269                 if (r) {
 270                         return r;
 271                 }
 272         }
 273
 274         r = amdgpu_bo_reserve(adev->vram_scratch.robj, false);
 275         if (unlikely(r != 0))
 276                 return r;
 277         r = amdgpu_bo_pin(adev->vram_scratch.robj,
 278                           AMDGPU_GEM_DOMAIN_VRAM, &adev->vram_scratch.gpu_addr);
 279         if (r) {
 280                 amdgpu_bo_unreserve(adev->vram_scratch.robj);
 281                 return r;
 282         }
 283         r = amdgpu_bo_kmap(adev->vram_scratch.robj,
 284                                 (void **)&adev->vram_scratch.ptr);
 285         if (r)
 286                 amdgpu_bo_unpin(adev->vram_scratch.robj);
 287         amdgpu_bo_unreserve(adev->vram_scratch.robj);
 288
 289         return r;
 290 }
 291
 292 static void amdgpu_vram_scratch_fini(struct amdgpu_device *adev)
 293 {
 294         int r;
 295
 296         if (adev->vram_scratch.robj == NULL) {
 297                 return;
 298         }
 299         r = amdgpu_bo_reserve(adev->vram_scratch.robj, false);
 300         if (likely(r == 0)) {
 301                 amdgpu_bo_kunmap(adev->vram_scratch.robj);
 302                 amdgpu_bo_unpin(adev->vram_scratch.robj);
 303                 amdgpu_bo_unreserve(adev->vram_scratch.robj);
 304         }
 305         amdgpu_bo_unref(&adev->vram_scratch.robj);
 306 }
 307
 308 /**
 309  * amdgpu_program_register_sequence - program an array of registers.
 310  *
 311  * @adev: amdgpu_device pointer
 312  * @registers: pointer to the register array
 313  * @array_size: size of the register array
 314  *
 315  * Programs an array or registers with and and or masks.
 316  * This is a helper for setting golden registers.
 317  */
 318 void amdgpu_program_register_sequence(struct amdgpu_device *adev,
 319                                       const u32 *registers,
 320                                       const u32 array_size)
 321 {
 322         u32 tmp, reg, and_mask, or_mask;
 323         int i;
 324
 325         if (array_size % 3)
 326                 return;
 327
 328         for (i = 0; i < array_size; i +=3) {
 329                 reg = registers[i + 0];
 330                 and_mask = registers[i + 1];
 331                 or_mask = registers[i + 2];
 332
 333                 if (and_mask == 0xffffffff) {
 334                         tmp = or_mask;
 335                 } else {
 336                         tmp = RREG32(reg);
 337                         tmp &= ~and_mask;
 338                         tmp |= or_mask;
 339                 }
 340                 WREG32(reg, tmp);
 341         }
 342 }
 343
 344 void amdgpu_pci_config_reset(struct amdgpu_device *adev)
 345 {
 346         pci_write_config_dword(adev->pdev, 0x7c, AMDGPU_ASIC_RESET_DATA);
 347 }
 348
 349 /*
 350  * GPU doorbell aperture helpers function.
 351  */
 352 /**
 353  * amdgpu_doorbell_init - Init doorbell driver information.
 354  *
 355  * @adev: amdgpu_device pointer
 356  *
 357  * Init doorbell driver information (CIK)
 358  * Returns 0 on success, error on failure.
 359  */
 360 static int amdgpu_doorbell_init(struct amdgpu_device *adev)
 361 {
 362         /* doorbell bar mapping */
 363         adev->doorbell.base = pci_resource_start(adev->pdev, 2);
 364         adev->doorbell.size = pci_resource_len(adev->pdev, 2);
 365
 366         adev->doorbell.num_doorbells = min_t(u32, adev->doorbell.size / sizeof(u32),
 367                                              AMDGPU_DOORBELL_MAX_ASSIGNMENT+1);
 368         if (adev->doorbell.num_doorbells == 0)
 369                 return -EINVAL;
 370
 371         adev->doorbell.ptr = ioremap(adev->doorbell.base, adev->doorbell.num_doorbells * sizeof(u32));
 372         if (adev->doorbell.ptr == NULL) {
 373                 return -ENOMEM;
 374         }
 375         DRM_INFO("doorbell mmio base: 0x%08X\n", (uint32_t)adev->doorbell.base);
 376         DRM_INFO("doorbell mmio size: %u\n", (unsigned)adev->doorbell.size);
 377
 378         return 0;
 379 }
 380
 381 /**
 382  * amdgpu_doorbell_fini - Tear down doorbell driver information.
 383  *
 384  * @adev: amdgpu_device pointer
 385  *
 386  * Tear down doorbell driver information (CIK)
 387  */
 388 static void amdgpu_doorbell_fini(struct amdgpu_device *adev)
 389 {
 390         iounmap(adev->doorbell.ptr);
 391         adev->doorbell.ptr = NULL;
 392 }
 393
 394 /**
 395  * amdgpu_doorbell_get_kfd_info - Report doorbell configuration required to
 396  *                                setup amdkfd
 397  *
 398  * @adev: amdgpu_device pointer
 399  * @aperture_base: output returning doorbell aperture base physical address
 400  * @aperture_size: output returning doorbell aperture size in bytes
 401  * @start_offset: output returning # of doorbell bytes reserved for amdgpu.
 402  *
 403  * amdgpu and amdkfd share the doorbell aperture. amdgpu sets it up,
 404  * takes doorbells required for its own rings and reports the setup to amdkfd.
 405  * amdgpu reserved doorbells are at the start of the doorbell aperture.
 406  */
 407 void amdgpu_doorbell_get_kfd_info(struct amdgpu_device *adev,
 408                                 phys_addr_t *aperture_base,
 409                                 size_t *aperture_size,
 410                                 size_t *start_offset)
 411 {
 412         /*
 413          * The first num_doorbells are used by amdgpu.
 414          * amdkfd takes whatever's left in the aperture.
 415          */
 416         if (adev->doorbell.size > adev->doorbell.num_doorbells * sizeof(u32)) {
 417                 *aperture_base = adev->doorbell.base;
 418                 *aperture_size = adev->doorbell.size;
 419                 *start_offset = adev->doorbell.num_doorbells * sizeof(u32);
 420         } else {
 421                 *aperture_base = 0;
 422                 *aperture_size = 0;
 423                 *start_offset = 0;
 424         }
 425 }
 426
 427 /*
 428  * amdgpu_wb_*()
 429  * Writeback is the the method by which the the GPU updates special pages
 430  * in memory with the status of certain GPU events (fences, ring pointers,
 431  * etc.).
 432  */
 433
 434 /**
 435  * amdgpu_wb_fini - Disable Writeback and free memory
 436  *
 437  * @adev: amdgpu_device pointer
 438  *
 439  * Disables Writeback and frees the Writeback memory (all asics).
 440  * Used at driver shutdown.
 441  */
 442 static void amdgpu_wb_fini(struct amdgpu_device *adev)
 443 {
 444         if (adev->wb.wb_obj) {
 445                 if (!amdgpu_bo_reserve(adev->wb.wb_obj, false)) {
 446                         amdgpu_bo_kunmap(adev->wb.wb_obj);
 447                         amdgpu_bo_unpin(adev->wb.wb_obj);
 448                         amdgpu_bo_unreserve(adev->wb.wb_obj);
 449                 }
 450                 amdgpu_bo_unref(&adev->wb.wb_obj);
 451                 adev->wb.wb = NULL;
 452                 adev->wb.wb_obj = NULL;
 453         }
 454 }
 455
 456 /**
 457  * amdgpu_wb_init- Init Writeback driver info and allocate memory
 458  *
 459  * @adev: amdgpu_device pointer
 460  *
 461  * Disables Writeback and frees the Writeback memory (all asics).
 462  * Used at driver startup.
 463  * Returns 0 on success or an -error on failure.
 464  */
 465 static int amdgpu_wb_init(struct amdgpu_device *adev)
 466 {
 467         int r;
 468
 469         if (adev->wb.wb_obj == NULL) {
 470                 r = amdgpu_bo_create(adev, AMDGPU_MAX_WB * 4, PAGE_SIZE, true,
 471                                      AMDGPU_GEM_DOMAIN_GTT, 0,  NULL, NULL,
 472                                      &adev->wb.wb_obj);
 473                 if (r) {
 474                         dev_warn(adev->dev, "(%d) create WB bo failed\n", r);
 475                         return r;
 476                 }
 477                 r = amdgpu_bo_reserve(adev->wb.wb_obj, false);
 478                 if (unlikely(r != 0)) {
 479                         amdgpu_wb_fini(adev);
 480                         return r;
 481                 }
 482                 r = amdgpu_bo_pin(adev->wb.wb_obj, AMDGPU_GEM_DOMAIN_GTT,
 483                                 &adev->wb.gpu_addr);
 484                 if (r) {
 485                         amdgpu_bo_unreserve(adev->wb.wb_obj);
 486                         dev_warn(adev->dev, "(%d) pin WB bo failed\n", r);
 487                         amdgpu_wb_fini(adev);
 488                         return r;
 489                 }
 490                 r = amdgpu_bo_kmap(adev->wb.wb_obj, (void **)&adev->wb.wb);
 491                 amdgpu_bo_unreserve(adev->wb.wb_obj);
 492                 if (r) {
 493                         dev_warn(adev->dev, "(%d) map WB bo failed\n", r);
 494                         amdgpu_wb_fini(adev);
 495                         return r;
 496                 }
 497
 498                 adev->wb.num_wb = AMDGPU_MAX_WB;
 499                 memset(&adev->wb.used, 0, sizeof(adev->wb.used));
 500
 501                 /* clear wb memory */
 502                 memset((char *)adev->wb.wb, 0, AMDGPU_GPU_PAGE_SIZE);
 503         }
 504
 505         return 0;
 506 }
 507
 508 /**
 509  * amdgpu_wb_get - Allocate a wb entry
 510  *
 511  * @adev: amdgpu_device pointer
 512  * @wb: wb index
 513  *
 514  * Allocate a wb slot for use by the driver (all asics).
 515  * Returns 0 on success or -EINVAL on failure.
 516  */
 517 int amdgpu_wb_get(struct amdgpu_device *adev, u32 *wb)
 518 {
 519         unsigned long offset = find_first_zero_bit(adev->wb.used, adev->wb.num_wb);
 520         if (offset < adev->wb.num_wb) {
 521                 __set_bit(offset, adev->wb.used);
 522                 *wb = offset;
 523                 return 0;
 524         } else {
 525                 return -EINVAL;
 526         }
 527 }
 528
 529 /**
 530  * amdgpu_wb_free - Free a wb entry
 531  *
 532  * @adev: amdgpu_device pointer
 533  * @wb: wb index
 534  *
 535  * Free a wb slot allocated for use by the driver (all asics)
 536  */
 537 void amdgpu_wb_free(struct amdgpu_device *adev, u32 wb)
 538 {
 539         if (wb < adev->wb.num_wb)
 540                 __clear_bit(wb, adev->wb.used);
 541 }
 542
 543 /**
 544  * amdgpu_vram_location - try to find VRAM location
 545  * @adev: amdgpu device structure holding all necessary informations
 546  * @mc: memory controller structure holding memory informations
 547  * @base: base address at which to put VRAM
 548  *
 549  * Function will place try to place VRAM at base address provided
 550  * as parameter (which is so far either PCI aperture address or
 551  * for IGP TOM base address).
 552  *
 553  * If there is not enough space to fit the unvisible VRAM in the 32bits
 554  * address space then we limit the VRAM size to the aperture.
 555  *
 556  * Note: We don't explicitly enforce VRAM start to be aligned on VRAM size,
 557  * this shouldn't be a problem as we are using the PCI aperture as a reference.
 558  * Otherwise this would be needed for rv280, all r3xx, and all r4xx, but
 559  * not IGP.
 560  *
 561  * Note: we use mc_vram_size as on some board we need to program the mc to
 562  * cover the whole aperture even if VRAM size is inferior to aperture size
 563  * Novell bug 204882 + along with lots of ubuntu ones
 564  *
 565  * Note: when limiting vram it's safe to overwritte real_vram_size because
 566  * we are not in case where real_vram_size is inferior to mc_vram_size (ie
 567  * note afected by bogus hw of Novell bug 204882 + along with lots of ubuntu
 568  * ones)
 569  *
 570  * Note: IGP TOM addr should be the same as the aperture addr, we don't
 571  * explicitly check for that thought.
 572  *
 573  * FIXME: when reducing VRAM size align new size on power of 2.
 574  */
 575 void amdgpu_vram_location(struct amdgpu_device *adev, struct amdgpu_mc *mc, u64 base)
 576 {
 577         uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
 578
 579         mc->vram_start = base;
 580         if (mc->mc_vram_size > (adev->mc.mc_mask - base + 1)) {
 581                 dev_warn(adev->dev, "limiting VRAM to PCI aperture size\n");
 582                 mc->real_vram_size = mc->aper_size;
 583                 mc->mc_vram_size = mc->aper_size;
 584         }
 585         mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
 586         if (limit && limit < mc->real_vram_size)
 587                 mc->real_vram_size = limit;
 588         dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
 589                         mc->mc_vram_size >> 20, mc->vram_start,
 590                         mc->vram_end, mc->real_vram_size >> 20);
 591 }
 592
 593 /**
 594  * amdgpu_gtt_location - try to find GTT location
 595  * @adev: amdgpu device structure holding all necessary informations
 596  * @mc: memory controller structure holding memory informations
 597  *
 598  * Function will place try to place GTT before or after VRAM.
 599  *
 600  * If GTT size is bigger than space left then we ajust GTT size.
 601  * Thus function will never fails.
 602  *
 603  * FIXME: when reducing GTT size align new size on power of 2.
 604  */
 605 void amdgpu_gtt_location(struct amdgpu_device *adev, struct amdgpu_mc *mc)
 606 {
 607         u64 size_af, size_bf;
 608
 609         size_af = ((adev->mc.mc_mask - mc->vram_end) + mc->gtt_base_align) & ~mc->gtt_base_align;
 610         size_bf = mc->vram_start & ~mc->gtt_base_align;
 611         if (size_bf > size_af) {
 612                 if (mc->gtt_size > size_bf) {
 613                         dev_warn(adev->dev, "limiting GTT\n");
 614                         mc->gtt_size = size_bf;
 615                 }
 616                 mc->gtt_start = (mc->vram_start & ~mc->gtt_base_align) - mc->gtt_size;
 617         } else {
 618                 if (mc->gtt_size > size_af) {
 619                         dev_warn(adev->dev, "limiting GTT\n");
 620                         mc->gtt_size = size_af;
 621                 }
 622                 mc->gtt_start = (mc->vram_end + 1 + mc->gtt_base_align) & ~mc->gtt_base_align;
 623         }
 624         mc->gtt_end = mc->gtt_start + mc->gtt_size - 1;
 625         dev_info(adev->dev, "GTT: %lluM 0x%016llX - 0x%016llX\n",
 626                         mc->gtt_size >> 20, mc->gtt_start, mc->gtt_end);
 627 }
 628
 629 /*
 630  * GPU helpers function.
 631  */
 632 /**
 633  * amdgpu_card_posted - check if the hw has already been initialized
 634  *
 635  * @adev: amdgpu_device pointer
 636  *
 637  * Check if the asic has been initialized (all asics).
 638  * Used at driver startup.
 639  * Returns true if initialized or false if not.
 640  */
 641 bool amdgpu_card_posted(struct amdgpu_device *adev)
 642 {
 643         uint32_t reg;
 644
 645         /* then check MEM_SIZE, in case the crtcs are off */
 646         reg = RREG32(mmCONFIG_MEMSIZE);
 647
 648         if (reg)
 649                 return true;
 650
 651         return false;
 652
 653 }
 654
 655 static bool amdgpu_vpost_needed(struct amdgpu_device *adev)
 656 {
 657         if (amdgpu_sriov_vf(adev))
 658                 return false;
 659
 660         if (amdgpu_passthrough(adev)) {
 661                 /* for FIJI: In whole GPU pass-through virtualization case, after VM reboot
 662                  * some old smc fw still need driver do vPost otherwise gpu hang, while
 663                  * those smc fw version above 22.15 doesn't have this flaw, so we force
 664                  * vpost executed for smc version below 22.15
 665                  */
 666                 if (adev->asic_type == CHIP_FIJI) {
 667                         int err;
 668                         uint32_t fw_ver;
 669                         err = reject_firmware(&adev->pm.fw, "/*(DEBLOBBED)*/", adev->dev);
 670                         /* force vPost if error occured */
 671                         if (err)
 672                                 return true;
 673
 674                         fw_ver = *((uint32_t *)adev->pm.fw->data + 69);
 675                         if (fw_ver < 0x00160e00)
 676                                 return true;
 677                 }
 678         }
 679         return !amdgpu_card_posted(adev);
 680 }
 681
 682 /**
 683  * amdgpu_dummy_page_init - init dummy page used by the driver
 684  *
 685  * @adev: amdgpu_device pointer
 686  *
 687  * Allocate the dummy page used by the driver (all asics).
 688  * This dummy page is used by the driver as a filler for gart entries
 689  * when pages are taken out of the GART
 690  * Returns 0 on sucess, -ENOMEM on failure.
 691  */
 692 int amdgpu_dummy_page_init(struct amdgpu_device *adev)
 693 {
 694         if (adev->dummy_page.page)
 695                 return 0;
 696         adev->dummy_page.page = alloc_page(GFP_DMA32 | GFP_KERNEL | __GFP_ZERO);
 697         if (adev->dummy_page.page == NULL)
 698                 return -ENOMEM;
 699         adev->dummy_page.addr = pci_map_page(adev->pdev, adev->dummy_page.page,
 700                                         0, PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 701         if (pci_dma_mapping_error(adev->pdev, adev->dummy_page.addr)) {
 702                 dev_err(&adev->pdev->dev, "Failed to DMA MAP the dummy page\n");
 703                 __free_page(adev->dummy_page.page);
 704                 adev->dummy_page.page = NULL;
 705                 return -ENOMEM;
 706         }
 707         return 0;
 708 }
 709
 710 /**
 711  * amdgpu_dummy_page_fini - free dummy page used by the driver
 712  *
 713  * @adev: amdgpu_device pointer
 714  *
 715  * Frees the dummy page used by the driver (all asics).
 716  */
 717 void amdgpu_dummy_page_fini(struct amdgpu_device *adev)
 718 {
 719         if (adev->dummy_page.page == NULL)
 720                 return;
 721         pci_unmap_page(adev->pdev, adev->dummy_page.addr,
 722                         PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 723         __free_page(adev->dummy_page.page);
 724         adev->dummy_page.page = NULL;
 725 }
 726
 727
 728 /* ATOM accessor methods */
 729 /*
 730  * ATOM is an interpreted byte code stored in tables in the vbios.  The
 731  * driver registers callbacks to access registers and the interpreter
 732  * in the driver parses the tables and executes then to program specific
 733  * actions (set display modes, asic init, etc.).  See amdgpu_atombios.c,
 734  * atombios.h, and atom.c
 735  */
 736
 737 /**
 738  * cail_pll_read - read PLL register
 739  *
 740  * @info: atom card_info pointer
 741  * @reg: PLL register offset
 742  *
 743  * Provides a PLL register accessor for the atom interpreter (r4xx+).
 744  * Returns the value of the PLL register.
 745  */
 746 static uint32_t cail_pll_read(struct card_info *info, uint32_t reg)
 747 {
 748         return 0;
 749 }
 750
 751 /**
 752  * cail_pll_write - write PLL register
 753  *
 754  * @info: atom card_info pointer
 755  * @reg: PLL register offset
 756  * @val: value to write to the pll register
 757  *
 758  * Provides a PLL register accessor for the atom interpreter (r4xx+).
 759  */
 760 static void cail_pll_write(struct card_info *info, uint32_t reg, uint32_t val)
 761 {
 762
 763 }
 764
 765 /**
 766  * cail_mc_read - read MC (Memory Controller) register
 767  *
 768  * @info: atom card_info pointer
 769  * @reg: MC register offset
 770  *
 771  * Provides an MC register accessor for the atom interpreter (r4xx+).
 772  * Returns the value of the MC register.
 773  */
 774 static uint32_t cail_mc_read(struct card_info *info, uint32_t reg)
 775 {
 776         return 0;
 777 }
 778
 779 /**
 780  * cail_mc_write - write MC (Memory Controller) register
 781  *
 782  * @info: atom card_info pointer
 783  * @reg: MC register offset
 784  * @val: value to write to the pll register
 785  *
 786  * Provides a MC register accessor for the atom interpreter (r4xx+).
 787  */
 788 static void cail_mc_write(struct card_info *info, uint32_t reg, uint32_t val)
 789 {
 790
 791 }
 792
 793 /**
 794  * cail_reg_write - write MMIO register
 795  *
 796  * @info: atom card_info pointer
 797  * @reg: MMIO register offset
 798  * @val: value to write to the pll register
 799  *
 800  * Provides a MMIO register accessor for the atom interpreter (r4xx+).
 801  */
 802 static void cail_reg_write(struct card_info *info, uint32_t reg, uint32_t val)
 803 {
 804         struct amdgpu_device *adev = info->dev->dev_private;
 805
 806         WREG32(reg, val);
 807 }
 808
 809 /**
 810  * cail_reg_read - read MMIO register
 811  *
 812  * @info: atom card_info pointer
 813  * @reg: MMIO register offset
 814  *
 815  * Provides an MMIO register accessor for the atom interpreter (r4xx+).
 816  * Returns the value of the MMIO register.
 817  */
 818 static uint32_t cail_reg_read(struct card_info *info, uint32_t reg)
 819 {
 820         struct amdgpu_device *adev = info->dev->dev_private;
 821         uint32_t r;
 822
 823         r = RREG32(reg);
 824         return r;
 825 }
 826
 827 /**
 828  * cail_ioreg_write - write IO register
 829  *
 830  * @info: atom card_info pointer
 831  * @reg: IO register offset
 832  * @val: value to write to the pll register
 833  *
 834  * Provides a IO register accessor for the atom interpreter (r4xx+).
 835  */
 836 static void cail_ioreg_write(struct card_info *info, uint32_t reg, uint32_t val)
 837 {
 838         struct amdgpu_device *adev = info->dev->dev_private;
 839
 840         WREG32_IO(reg, val);
 841 }
 842
 843 /**
 844  * cail_ioreg_read - read IO register
 845  *
 846  * @info: atom card_info pointer
 847  * @reg: IO register offset
 848  *
 849  * Provides an IO register accessor for the atom interpreter (r4xx+).
 850  * Returns the value of the IO register.
 851  */
 852 static uint32_t cail_ioreg_read(struct card_info *info, uint32_t reg)
 853 {
 854         struct amdgpu_device *adev = info->dev->dev_private;
 855         uint32_t r;
 856
 857         r = RREG32_IO(reg);
 858         return r;
 859 }
 860
 861 /**
 862  * amdgpu_atombios_fini - free the driver info and callbacks for atombios
 863  *
 864  * @adev: amdgpu_device pointer
 865  *
 866  * Frees the driver info and register access callbacks for the ATOM
 867  * interpreter (r4xx+).
 868  * Called at driver shutdown.
 869  */
 870 static void amdgpu_atombios_fini(struct amdgpu_device *adev)
 871 {
 872         if (adev->mode_info.atom_context) {
 873                 kfree(adev->mode_info.atom_context->scratch);
 874                 kfree(adev->mode_info.atom_context->iio);
 875         }
 876         kfree(adev->mode_info.atom_context);
 877         adev->mode_info.atom_context = NULL;
 878         kfree(adev->mode_info.atom_card_info);
 879         adev->mode_info.atom_card_info = NULL;
 880 }
 881
 882 /**
 883  * amdgpu_atombios_init - init the driver info and callbacks for atombios
 884  *
 885  * @adev: amdgpu_device pointer
 886  *
 887  * Initializes the driver info and register access callbacks for the
 888  * ATOM interpreter (r4xx+).
 889  * Returns 0 on sucess, -ENOMEM on failure.
 890  * Called at driver startup.
 891  */
 892 static int amdgpu_atombios_init(struct amdgpu_device *adev)
 893 {
 894         struct card_info *atom_card_info =
 895             kzalloc(sizeof(struct card_info), GFP_KERNEL);
 896
 897         if (!atom_card_info)
 898                 return -ENOMEM;
 899
 900         adev->mode_info.atom_card_info = atom_card_info;
 901         atom_card_info->dev = adev->ddev;
 902         atom_card_info->reg_read = cail_reg_read;
 903         atom_card_info->reg_write = cail_reg_write;
 904         /* needed for iio ops */
 905         if (adev->rio_mem) {
 906                 atom_card_info->ioreg_read = cail_ioreg_read;
 907                 atom_card_info->ioreg_write = cail_ioreg_write;
 908         } else {
 909                 DRM_ERROR("Unable to find PCI I/O BAR; using MMIO for ATOM IIO\n");
 910                 atom_card_info->ioreg_read = cail_reg_read;
 911                 atom_card_info->ioreg_write = cail_reg_write;
 912         }
 913         atom_card_info->mc_read = cail_mc_read;
 914         atom_card_info->mc_write = cail_mc_write;
 915         atom_card_info->pll_read = cail_pll_read;
 916         atom_card_info->pll_write = cail_pll_write;
 917
 918         adev->mode_info.atom_context = amdgpu_atom_parse(atom_card_info, adev->bios);
 919         if (!adev->mode_info.atom_context) {
 920                 amdgpu_atombios_fini(adev);
 921                 return -ENOMEM;
 922         }
 923
 924         mutex_init(&adev->mode_info.atom_context->mutex);
 925         amdgpu_atombios_scratch_regs_init(adev);
 926         amdgpu_atom_allocate_fb_scratch(adev->mode_info.atom_context);
 927         return 0;
 928 }
 929
 930 /* if we get transitioned to only one device, take VGA back */
 931 /**
 932  * amdgpu_vga_set_decode - enable/disable vga decode
 933  *
 934  * @cookie: amdgpu_device pointer
 935  * @state: enable/disable vga decode
 936  *
 937  * Enable/disable vga decode (all asics).
 938  * Returns VGA resource flags.
 939  */
 940 static unsigned int amdgpu_vga_set_decode(void *cookie, bool state)
 941 {
 942         struct amdgpu_device *adev = cookie;
 943         amdgpu_asic_set_vga_state(adev, state);
 944         if (state)
 945                 return VGA_RSRC_LEGACY_IO | VGA_RSRC_LEGACY_MEM |
 946                        VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
 947         else
 948                 return VGA_RSRC_NORMAL_IO | VGA_RSRC_NORMAL_MEM;
 949 }
 950
 951 /**
 952  * amdgpu_check_pot_argument - check that argument is a power of two
 953  *
 954  * @arg: value to check
 955  *
 956  * Validates that a certain argument is a power of two (all asics).
 957  * Returns true if argument is valid.
 958  */
 959 static bool amdgpu_check_pot_argument(int arg)
 960 {
 961         return (arg & (arg - 1)) == 0;
 962 }
 963
 964 /**
 965  * amdgpu_check_arguments - validate module params
 966  *
 967  * @adev: amdgpu_device pointer
 968  *
 969  * Validates certain module parameters and updates
 970  * the associated values used by the driver (all asics).
 971  */
 972 static void amdgpu_check_arguments(struct amdgpu_device *adev)
 973 {
 974         if (amdgpu_sched_jobs < 4) {
 975                 dev_warn(adev->dev, "sched jobs (%d) must be at least 4\n",
 976                          amdgpu_sched_jobs);
 977                 amdgpu_sched_jobs = 4;
 978         } else if (!amdgpu_check_pot_argument(amdgpu_sched_jobs)){
 979                 dev_warn(adev->dev, "sched jobs (%d) must be a power of 2\n",
 980                          amdgpu_sched_jobs);
 981                 amdgpu_sched_jobs = roundup_pow_of_two(amdgpu_sched_jobs);
 982         }
 983
 984         if (amdgpu_gart_size != -1) {
 985                 /* gtt size must be greater or equal to 32M */
 986                 if (amdgpu_gart_size < 32) {
 987                         dev_warn(adev->dev, "gart size (%d) too small\n",
 988                                  amdgpu_gart_size);
 989                         amdgpu_gart_size = -1;
 990                 }
 991         }
 992
 993         if (!amdgpu_check_pot_argument(amdgpu_vm_size)) {
 994                 dev_warn(adev->dev, "VM size (%d) must be a power of 2\n",
 995                          amdgpu_vm_size);
 996                 amdgpu_vm_size = 8;
 997         }
 998
 999         if (amdgpu_vm_size < 1) {
1000                 dev_warn(adev->dev, "VM size (%d) too small, min is 1GB\n",
1001                          amdgpu_vm_size);
1002                 amdgpu_vm_size = 8;
1003         }
1004
1005         /*
1006          * Max GPUVM size for Cayman, SI and CI are 40 bits.
1007          */
1008         if (amdgpu_vm_size > 1024) {
1009                 dev_warn(adev->dev, "VM size (%d) too large, max is 1TB\n",
1010                          amdgpu_vm_size);
1011                 amdgpu_vm_size = 8;
1012         }
1013
1014         /* defines number of bits in page table versus page directory,
1015          * a page is 4KB so we have 12 bits offset, minimum 9 bits in the
1016          * page table and the remaining bits are in the page directory */
1017         if (amdgpu_vm_block_size == -1) {
1018
1019                 /* Total bits covered by PD + PTs */
1020                 unsigned bits = ilog2(amdgpu_vm_size) + 18;
1021
1022                 /* Make sure the PD is 4K in size up to 8GB address space.
1023                    Above that split equal between PD and PTs */
1024                 if (amdgpu_vm_size <= 8)
1025                         amdgpu_vm_block_size = bits - 9;
1026                 else
1027                         amdgpu_vm_block_size = (bits + 3) / 2;
1028
1029         } else if (amdgpu_vm_block_size < 9) {
1030                 dev_warn(adev->dev, "VM page table size (%d) too small\n",
1031                          amdgpu_vm_block_size);
1032                 amdgpu_vm_block_size = 9;
1033         }
1034
1035         if (amdgpu_vm_block_size > 24 ||
1036             (amdgpu_vm_size * 1024) < (1ull << amdgpu_vm_block_size)) {
1037                 dev_warn(adev->dev, "VM page table size (%d) too large\n",
1038                          amdgpu_vm_block_size);
1039                 amdgpu_vm_block_size = 9;
1040         }
1041 }
1042
1043 /**
1044  * amdgpu_switcheroo_set_state - set switcheroo state
1045  *
1046  * @pdev: pci dev pointer
1047  * @state: vga_switcheroo state
1048  *
1049  * Callback for the switcheroo driver.  Suspends or resumes the
1050  * the asics before or after it is powered up using ACPI methods.
1051  */
1052 static void amdgpu_switcheroo_set_state(struct pci_dev *pdev, enum vga_switcheroo_state state)
1053 {
1054         struct drm_device *dev = pci_get_drvdata(pdev);
1055
1056         if (amdgpu_device_is_px(dev) && state == VGA_SWITCHEROO_OFF)
1057                 return;
1058
1059         if (state == VGA_SWITCHEROO_ON) {
1060                 unsigned d3_delay = dev->pdev->d3_delay;
1061
1062                 printk(KERN_INFO "amdgpu: switched on\n");
1063                 /* don't suspend or resume card normally */
1064                 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1065
1066                 amdgpu_device_resume(dev, true, true);
1067
1068                 dev->pdev->d3_delay = d3_delay;
1069
1070                 dev->switch_power_state = DRM_SWITCH_POWER_ON;
1071                 drm_kms_helper_poll_enable(dev);
1072         } else {
1073                 printk(KERN_INFO "amdgpu: switched off\n");
1074                 drm_kms_helper_poll_disable(dev);
1075                 dev->switch_power_state = DRM_SWITCH_POWER_CHANGING;
1076                 amdgpu_device_suspend(dev, true, true);
1077                 dev->switch_power_state = DRM_SWITCH_POWER_OFF;
1078         }
1079 }
1080
1081 /**
1082  * amdgpu_switcheroo_can_switch - see if switcheroo state can change
1083  *
1084  * @pdev: pci dev pointer
1085  *
1086  * Callback for the switcheroo driver.  Check of the switcheroo
1087  * state can be changed.
1088  * Returns true if the state can be changed, false if not.
1089  */
1090 static bool amdgpu_switcheroo_can_switch(struct pci_dev *pdev)
1091 {
1092         struct drm_device *dev = pci_get_drvdata(pdev);
1093
1094         /*
1095         * FIXME: open_count is protected by drm_global_mutex but that would lead to
1096         * locking inversion with the driver load path. And the access here is
1097         * completely racy anyway. So don't bother with locking for now.
1098         */
1099         return dev->open_count == 0;
1100 }
1101
1102 static const struct vga_switcheroo_client_ops amdgpu_switcheroo_ops = {
1103         .set_gpu_state = amdgpu_switcheroo_set_state,
1104         .reprobe = NULL,
1105         .can_switch = amdgpu_switcheroo_can_switch,
1106 };
1107
1108 int amdgpu_set_clockgating_state(struct amdgpu_device *adev,
1109                                   enum amd_ip_block_type block_type,
1110                                   enum amd_clockgating_state state)
1111 {
1112         int i, r = 0;
1113
1114         for (i = 0; i < adev->num_ip_blocks; i++) {
1115                 if (!adev->ip_block_status[i].valid)
1116                         continue;
1117                 if (adev->ip_blocks[i].type == block_type) {
1118                         r = adev->ip_blocks[i].funcs->set_clockgating_state((void *)adev,
1119                                                                             state);
1120                         if (r)
1121                                 return r;
1122                         break;
1123                 }
1124         }
1125         return r;
1126 }
1127
1128 int amdgpu_set_powergating_state(struct amdgpu_device *adev,
1129                                   enum amd_ip_block_type block_type,
1130                                   enum amd_powergating_state state)
1131 {
1132         int i, r = 0;
1133
1134         for (i = 0; i < adev->num_ip_blocks; i++) {
1135                 if (!adev->ip_block_status[i].valid)
1136                         continue;
1137                 if (adev->ip_blocks[i].type == block_type) {
1138                         r = adev->ip_blocks[i].funcs->set_powergating_state((void *)adev,
1139                                                                             state);
1140                         if (r)
1141                                 return r;
1142                         break;
1143                 }
1144         }
1145         return r;
1146 }
1147
1148 int amdgpu_wait_for_idle(struct amdgpu_device *adev,
1149                          enum amd_ip_block_type block_type)
1150 {
1151         int i, r;
1152
1153         for (i = 0; i < adev->num_ip_blocks; i++) {
1154                 if (!adev->ip_block_status[i].valid)
1155                         continue;
1156                 if (adev->ip_blocks[i].type == block_type) {
1157                         r = adev->ip_blocks[i].funcs->wait_for_idle((void *)adev);
1158                         if (r)
1159                                 return r;
1160                         break;
1161                 }
1162         }
1163         return 0;
1164
1165 }
1166
1167 bool amdgpu_is_idle(struct amdgpu_device *adev,
1168                     enum amd_ip_block_type block_type)
1169 {
1170         int i;
1171
1172         for (i = 0; i < adev->num_ip_blocks; i++) {
1173                 if (!adev->ip_block_status[i].valid)
1174                         continue;
1175                 if (adev->ip_blocks[i].type == block_type)
1176                         return adev->ip_blocks[i].funcs->is_idle((void *)adev);
1177         }
1178         return true;
1179
1180 }
1181
1182 const struct amdgpu_ip_block_version * amdgpu_get_ip_block(
1183                                         struct amdgpu_device *adev,
1184                                         enum amd_ip_block_type type)
1185 {
1186         int i;
1187
1188         for (i = 0; i < adev->num_ip_blocks; i++)
1189                 if (adev->ip_blocks[i].type == type)
1190                         return &adev->ip_blocks[i];
1191
1192         return NULL;
1193 }
1194
1195 /**
1196  * amdgpu_ip_block_version_cmp
1197  *
1198  * @adev: amdgpu_device pointer
1199  * @type: enum amd_ip_block_type
1200  * @major: major version
1201  * @minor: minor version
1202  *
1203  * return 0 if equal or greater
1204  * return 1 if smaller or the ip_block doesn't exist
1205  */
1206 int amdgpu_ip_block_version_cmp(struct amdgpu_device *adev,
1207                                 enum amd_ip_block_type type,
1208                                 u32 major, u32 minor)
1209 {
1210         const struct amdgpu_ip_block_version *ip_block;
1211         ip_block = amdgpu_get_ip_block(adev, type);
1212
1213         if (ip_block && ((ip_block->major > major) ||
1214                         ((ip_block->major == major) &&
1215                         (ip_block->minor >= minor))))
1216                 return 0;
1217
1218         return 1;
1219 }
1220
1221 static void amdgpu_whether_enable_virtual_display(struct amdgpu_device *adev)
1222 {
1223         adev->enable_virtual_display = false;
1224
1225         if (amdgpu_virtual_display) {
1226                 struct drm_device *ddev = adev->ddev;
1227                 const char *pci_address_name = pci_name(ddev->pdev);
1228                 char *pciaddstr, *pciaddstr_tmp, *pciaddname;
1229
1230                 pciaddstr = kstrdup(amdgpu_virtual_display, GFP_KERNEL);
1231                 pciaddstr_tmp = pciaddstr;
1232                 while ((pciaddname = strsep(&pciaddstr_tmp, ";"))) {
1233                         if (!strcmp(pci_address_name, pciaddname)) {
1234                                 adev->enable_virtual_display = true;
1235                                 break;
1236                         }
1237                 }
1238
1239                 DRM_INFO("virtual display string:%s, %s:virtual_display:%d\n",
1240                                  amdgpu_virtual_display, pci_address_name,
1241                                  adev->enable_virtual_display);
1242
1243                 kfree(pciaddstr);
1244         }
1245 }
1246
1247 static int amdgpu_early_init(struct amdgpu_device *adev)
1248 {
1249         int i, r;
1250
1251         amdgpu_whether_enable_virtual_display(adev);
1252
1253         switch (adev->asic_type) {
1254         case CHIP_TOPAZ:
1255         case CHIP_TONGA:
1256         case CHIP_FIJI:
1257         case CHIP_POLARIS11:
1258         case CHIP_POLARIS10:
1259         case CHIP_CARRIZO:
1260         case CHIP_STONEY:
1261                 if (adev->asic_type == CHIP_CARRIZO || adev->asic_type == CHIP_STONEY)
1262                         adev->family = AMDGPU_FAMILY_CZ;
1263                 else
1264                         adev->family = AMDGPU_FAMILY_VI;
1265
1266                 r = vi_set_ip_blocks(adev);
1267                 if (r)
1268                         return r;
1269                 break;
1270 #ifdef CONFIG_DRM_AMDGPU_SI
1271         case CHIP_VERDE:
1272         case CHIP_TAHITI:
1273         case CHIP_PITCAIRN:
1274         case CHIP_OLAND:
1275         case CHIP_HAINAN:
1276                 adev->family = AMDGPU_FAMILY_SI;
1277                 r = si_set_ip_blocks(adev);
1278                 if (r)
1279                         return r;
1280                 break;
1281 #endif
1282 #ifdef CONFIG_DRM_AMDGPU_CIK
1283         case CHIP_BONAIRE:
1284         case CHIP_HAWAII:
1285         case CHIP_KAVERI:
1286         case CHIP_KABINI:
1287         case CHIP_MULLINS:
1288                 if ((adev->asic_type == CHIP_BONAIRE) || (adev->asic_type == CHIP_HAWAII))
1289                         adev->family = AMDGPU_FAMILY_CI;
1290                 else
1291                         adev->family = AMDGPU_FAMILY_KV;
1292
1293                 r = cik_set_ip_blocks(adev);
1294                 if (r)
1295                         return r;
1296                 break;
1297 #endif
1298         default:
1299                 /* FIXME: not supported yet */
1300                 return -EINVAL;
1301         }
1302
1303         adev->ip_block_status = kcalloc(adev->num_ip_blocks,
1304                                         sizeof(struct amdgpu_ip_block_status), GFP_KERNEL);
1305         if (adev->ip_block_status == NULL)
1306                 return -ENOMEM;
1307
1308         if (adev->ip_blocks == NULL) {
1309                 DRM_ERROR("No IP blocks found!\n");
1310                 return r;
1311         }
1312
1313         for (i = 0; i < adev->num_ip_blocks; i++) {
1314                 if ((amdgpu_ip_block_mask & (1 << i)) == 0) {
1315                         DRM_ERROR("disabled ip block: %d\n", i);
1316                         adev->ip_block_status[i].valid = false;
1317                 } else {
1318                         if (adev->ip_blocks[i].funcs->early_init) {
1319                                 r = adev->ip_blocks[i].funcs->early_init((void *)adev);
1320                                 if (r == -ENOENT) {
1321                                         adev->ip_block_status[i].valid = false;
1322                                 } else if (r) {
1323                                         DRM_ERROR("early_init of IP block <%s> failed %d\n", adev->ip_blocks[i].funcs->name, r);
1324                                         return r;
1325                                 } else {
1326                                         adev->ip_block_status[i].valid = true;
1327                                 }
1328                         } else {
1329                                 adev->ip_block_status[i].valid = true;
1330                         }
1331                 }
1332         }
1333
1334         adev->cg_flags &= amdgpu_cg_mask;
1335         adev->pg_flags &= amdgpu_pg_mask;
1336
1337         return 0;
1338 }
1339
1340 static int amdgpu_init(struct amdgpu_device *adev)
1341 {
1342         int i, r;
1343
1344         for (i = 0; i < adev->num_ip_blocks; i++) {
1345                 if (!adev->ip_block_status[i].valid)
1346                         continue;
1347                 r = adev->ip_blocks[i].funcs->sw_init((void *)adev);
1348                 if (r) {
1349                         DRM_ERROR("sw_init of IP block <%s> failed %d\n", adev->ip_blocks[i].funcs->name, r);
1350                         return r;
1351                 }
1352                 adev->ip_block_status[i].sw = true;
1353                 /* need to do gmc hw init early so we can allocate gpu mem */
1354                 if (adev->ip_blocks[i].type == AMD_IP_BLOCK_TYPE_GMC) {
1355                         r = amdgpu_vram_scratch_init(adev);
1356                         if (r) {
1357                                 DRM_ERROR("amdgpu_vram_scratch_init failed %d\n", r);
1358                                 return r;
1359                         }
1360                         r = adev->ip_blocks[i].funcs->hw_init((void *)adev);
1361                         if (r) {
1362                                 DRM_ERROR("hw_init %d failed %d\n", i, r);
1363                                 return r;
1364                         }
1365                         r = amdgpu_wb_init(adev);
1366                         if (r) {
1367                                 DRM_ERROR("amdgpu_wb_init failed %d\n", r);
1368                                 return r;
1369                         }
1370                         adev->ip_block_status[i].hw = true;
1371                 }
1372         }
1373
1374         for (i = 0; i < adev->num_ip_blocks; i++) {
1375                 if (!adev->ip_block_status[i].sw)
1376                         continue;
1377                 /* gmc hw init is done early */
1378                 if (adev->ip_blocks[i].type == AMD_IP_BLOCK_TYPE_GMC)
1379                         continue;
1380                 r = adev->ip_blocks[i].funcs->hw_init((void *)adev);
1381                 if (r) {
1382                         DRM_ERROR("hw_init of IP block <%s> failed %d\n", adev->ip_blocks[i].funcs->name, r);
1383                         return r;
1384                 }
1385                 adev->ip_block_status[i].hw = true;
1386         }
1387
1388         return 0;
1389 }
1390
1391 static int amdgpu_late_init(struct amdgpu_device *adev)
1392 {
1393         int i = 0, r;
1394
1395         for (i = 0; i < adev->num_ip_blocks; i++) {
1396                 if (!adev->ip_block_status[i].valid)
1397                         continue;
1398                 if (adev->ip_blocks[i].funcs->late_init) {
1399                         r = adev->ip_blocks[i].funcs->late_init((void *)adev);
1400                         if (r) {
1401                                 DRM_ERROR("late_init of IP block <%s> failed %d\n", adev->ip_blocks[i].funcs->name, r);
1402                                 return r;
1403                         }
1404                         adev->ip_block_status[i].late_initialized = true;
1405                 }
1406                 /* skip CG for VCE/UVD, it's handled specially */
1407                 if (adev->ip_blocks[i].type != AMD_IP_BLOCK_TYPE_UVD &&
1408                     adev->ip_blocks[i].type != AMD_IP_BLOCK_TYPE_VCE) {
1409                         /* enable clockgating to save power */
1410                         r = adev->ip_blocks[i].funcs->set_clockgating_state((void *)adev,
1411                                                                             AMD_CG_STATE_GATE);
1412                         if (r) {
1413                                 DRM_ERROR("set_clockgating_state(gate) of IP block <%s> failed %d\n",
1414                                           adev->ip_blocks[i].funcs->name, r);
1415                                 return r;
1416                         }
1417                 }
1418         }
1419
1420         return 0;
1421 }
1422
1423 static int amdgpu_fini(struct amdgpu_device *adev)
1424 {
1425         int i, r;
1426
1427         /* need to disable SMC first */
1428         for (i = 0; i < adev->num_ip_blocks; i++) {
1429                 if (!adev->ip_block_status[i].hw)
1430                         continue;
1431                 if (adev->ip_blocks[i].type == AMD_IP_BLOCK_TYPE_SMC) {
1432                         /* ungate blocks before hw fini so that we can shutdown the blocks safely */
1433                         r = adev->ip_blocks[i].funcs->set_clockgating_state((void *)adev,
1434                                                                             AMD_CG_STATE_UNGATE);
1435                         if (r) {
1436                                 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n",
1437                                           adev->ip_blocks[i].funcs->name, r);
1438                                 return r;
1439                         }
1440                         r = adev->ip_blocks[i].funcs->hw_fini((void *)adev);
1441                         /* XXX handle errors */
1442                         if (r) {
1443                                 DRM_DEBUG("hw_fini of IP block <%s> failed %d\n",
1444                                           adev->ip_blocks[i].funcs->name, r);
1445                         }
1446                         adev->ip_block_status[i].hw = false;
1447                         break;
1448                 }
1449         }
1450
1451         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1452                 if (!adev->ip_block_status[i].hw)
1453                         continue;
1454                 if (adev->ip_blocks[i].type == AMD_IP_BLOCK_TYPE_GMC) {
1455                         amdgpu_wb_fini(adev);
1456                         amdgpu_vram_scratch_fini(adev);
1457                 }
1458                 /* ungate blocks before hw fini so that we can shutdown the blocks safely */
1459                 r = adev->ip_blocks[i].funcs->set_clockgating_state((void *)adev,
1460                                                                     AMD_CG_STATE_UNGATE);
1461                 if (r) {
1462                         DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n", adev->ip_blocks[i].funcs->name, r);
1463                         return r;
1464                 }
1465                 r = adev->ip_blocks[i].funcs->hw_fini((void *)adev);
1466                 /* XXX handle errors */
1467                 if (r) {
1468                         DRM_DEBUG("hw_fini of IP block <%s> failed %d\n", adev->ip_blocks[i].funcs->name, r);
1469                 }
1470                 adev->ip_block_status[i].hw = false;
1471         }
1472
1473         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1474                 if (!adev->ip_block_status[i].sw)
1475                         continue;
1476                 r = adev->ip_blocks[i].funcs->sw_fini((void *)adev);
1477                 /* XXX handle errors */
1478                 if (r) {
1479                         DRM_DEBUG("sw_fini of IP block <%s> failed %d\n", adev->ip_blocks[i].funcs->name, r);
1480                 }
1481                 adev->ip_block_status[i].sw = false;
1482                 adev->ip_block_status[i].valid = false;
1483         }
1484
1485         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1486                 if (!adev->ip_block_status[i].late_initialized)
1487                         continue;
1488                 if (adev->ip_blocks[i].funcs->late_fini)
1489                         adev->ip_blocks[i].funcs->late_fini((void *)adev);
1490                 adev->ip_block_status[i].late_initialized = false;
1491         }
1492
1493         return 0;
1494 }
1495
1496 int amdgpu_suspend(struct amdgpu_device *adev)
1497 {
1498         int i, r;
1499
1500         /* ungate SMC block first */
1501         r = amdgpu_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_SMC,
1502                                          AMD_CG_STATE_UNGATE);
1503         if (r) {
1504                 DRM_ERROR("set_clockgating_state(ungate) SMC failed %d\n",r);
1505         }
1506
1507         for (i = adev->num_ip_blocks - 1; i >= 0; i--) {
1508                 if (!adev->ip_block_status[i].valid)
1509                         continue;
1510                 /* ungate blocks so that suspend can properly shut them down */
1511                 if (i != AMD_IP_BLOCK_TYPE_SMC) {
1512                         r = adev->ip_blocks[i].funcs->set_clockgating_state((void *)adev,
1513                                                                             AMD_CG_STATE_UNGATE);
1514                         if (r) {
1515                                 DRM_ERROR("set_clockgating_state(ungate) of IP block <%s> failed %d\n", adev->ip_blocks[i].funcs->name, r);
1516                         }
1517                 }
1518                 /* XXX handle errors */
1519                 r = adev->ip_blocks[i].funcs->suspend(adev);
1520                 /* XXX handle errors */
1521                 if (r) {
1522                         DRM_ERROR("suspend of IP block <%s> failed %d\n", adev->ip_blocks[i].funcs->name, r);
1523                 }
1524         }
1525
1526         return 0;
1527 }
1528
1529 static int amdgpu_resume(struct amdgpu_device *adev)
1530 {
1531         int i, r;
1532
1533         for (i = 0; i < adev->num_ip_blocks; i++) {
1534                 if (!adev->ip_block_status[i].valid)
1535                         continue;
1536                 r = adev->ip_blocks[i].funcs->resume(adev);
1537                 if (r) {
1538                         DRM_ERROR("resume of IP block <%s> failed %d\n", adev->ip_blocks[i].funcs->name, r);
1539                         return r;
1540                 }
1541         }
1542
1543         return 0;
1544 }
1545
1546 static void amdgpu_device_detect_sriov_bios(struct amdgpu_device *adev)
1547 {
1548         if (amdgpu_atombios_has_gpu_virtualization_table(adev))
1549                 adev->virtualization.virtual_caps |= AMDGPU_SRIOV_CAPS_SRIOV_VBIOS;
1550 }
1551
1552 /**
1553  * amdgpu_device_init - initialize the driver
1554  *
1555  * @adev: amdgpu_device pointer
1556  * @pdev: drm dev pointer
1557  * @pdev: pci dev pointer
1558  * @flags: driver flags
1559  *
1560  * Initializes the driver info and hw (all asics).
1561  * Returns 0 for success or an error on failure.
1562  * Called at driver startup.
1563  */
1564 int amdgpu_device_init(struct amdgpu_device *adev,
1565                        struct drm_device *ddev,
1566                        struct pci_dev *pdev,
1567                        uint32_t flags)
1568 {
1569         int r, i;
1570         bool runtime = false;
1571         u32 max_MBps;
1572
1573         adev->shutdown = false;
1574         adev->dev = &pdev->dev;
1575         adev->ddev = ddev;
1576         adev->pdev = pdev;
1577         adev->flags = flags;
1578         adev->asic_type = flags & AMD_ASIC_MASK;
1579         adev->is_atom_bios = false;
1580         adev->usec_timeout = AMDGPU_MAX_USEC_TIMEOUT;
1581         adev->mc.gtt_size = 512 * 1024 * 1024;
1582         adev->accel_working = false;
1583         adev->num_rings = 0;
1584         adev->mman.buffer_funcs = NULL;
1585         adev->mman.buffer_funcs_ring = NULL;
1586         adev->vm_manager.vm_pte_funcs = NULL;
1587         adev->vm_manager.vm_pte_num_rings = 0;
1588         adev->gart.gart_funcs = NULL;
1589         adev->fence_context = fence_context_alloc(AMDGPU_MAX_RINGS);
1590
1591         adev->smc_rreg = &amdgpu_invalid_rreg;
1592         adev->smc_wreg = &amdgpu_invalid_wreg;
1593         adev->pcie_rreg = &amdgpu_invalid_rreg;
1594         adev->pcie_wreg = &amdgpu_invalid_wreg;
1595         adev->pciep_rreg = &amdgpu_invalid_rreg;
1596         adev->pciep_wreg = &amdgpu_invalid_wreg;
1597         adev->uvd_ctx_rreg = &amdgpu_invalid_rreg;
1598         adev->uvd_ctx_wreg = &amdgpu_invalid_wreg;
1599         adev->didt_rreg = &amdgpu_invalid_rreg;
1600         adev->didt_wreg = &amdgpu_invalid_wreg;
1601         adev->gc_cac_rreg = &amdgpu_invalid_rreg;
1602         adev->gc_cac_wreg = &amdgpu_invalid_wreg;
1603         adev->audio_endpt_rreg = &amdgpu_block_invalid_rreg;
1604         adev->audio_endpt_wreg = &amdgpu_block_invalid_wreg;
1605
1606
1607         DRM_INFO("initializing kernel modesetting (%s 0x%04X:0x%04X 0x%04X:0x%04X 0x%02X).\n",
1608                  amdgpu_asic_name[adev->asic_type], pdev->vendor, pdev->device,
1609                  pdev->subsystem_vendor, pdev->subsystem_device, pdev->revision);
1610
1611         /* mutex initialization are all done here so we
1612          * can recall function without having locking issues */
1613         mutex_init(&adev->vm_manager.lock);
1614         atomic_set(&adev->irq.ih.lock, 0);
1615         mutex_init(&adev->pm.mutex);
1616         mutex_init(&adev->gfx.gpu_clock_mutex);
1617         mutex_init(&adev->srbm_mutex);
1618         mutex_init(&adev->grbm_idx_mutex);
1619         mutex_init(&adev->mn_lock);
1620         hash_init(adev->mn_hash);
1621
1622         amdgpu_check_arguments(adev);
1623
1624         /* Registers mapping */
1625         /* TODO: block userspace mapping of io register */
1626         spin_lock_init(&adev->mmio_idx_lock);
1627         spin_lock_init(&adev->smc_idx_lock);
1628         spin_lock_init(&adev->pcie_idx_lock);
1629         spin_lock_init(&adev->uvd_ctx_idx_lock);
1630         spin_lock_init(&adev->didt_idx_lock);
1631         spin_lock_init(&adev->gc_cac_idx_lock);
1632         spin_lock_init(&adev->audio_endpt_idx_lock);
1633         spin_lock_init(&adev->mm_stats.lock);
1634
1635         INIT_LIST_HEAD(&adev->shadow_list);
1636         mutex_init(&adev->shadow_list_lock);
1637
1638         INIT_LIST_HEAD(&adev->gtt_list);
1639         spin_lock_init(&adev->gtt_list_lock);
1640
1641         if (adev->asic_type >= CHIP_BONAIRE) {
1642                 adev->rmmio_base = pci_resource_start(adev->pdev, 5);
1643                 adev->rmmio_size = pci_resource_len(adev->pdev, 5);
1644         } else {
1645                 adev->rmmio_base = pci_resource_start(adev->pdev, 2);
1646                 adev->rmmio_size = pci_resource_len(adev->pdev, 2);
1647         }
1648
1649         adev->rmmio = ioremap(adev->rmmio_base, adev->rmmio_size);
1650         if (adev->rmmio == NULL) {
1651                 return -ENOMEM;
1652         }
1653         DRM_INFO("register mmio base: 0x%08X\n", (uint32_t)adev->rmmio_base);
1654         DRM_INFO("register mmio size: %u\n", (unsigned)adev->rmmio_size);
1655
1656         if (adev->asic_type >= CHIP_BONAIRE)
1657                 /* doorbell bar mapping */
1658                 amdgpu_doorbell_init(adev);
1659
1660         /* io port mapping */
1661         for (i = 0; i < DEVICE_COUNT_RESOURCE; i++) {
1662                 if (pci_resource_flags(adev->pdev, i) & IORESOURCE_IO) {
1663                         adev->rio_mem_size = pci_resource_len(adev->pdev, i);
1664                         adev->rio_mem = pci_iomap(adev->pdev, i, adev->rio_mem_size);
1665                         break;
1666                 }
1667         }
1668         if (adev->rio_mem == NULL)
1669                 DRM_ERROR("Unable to find PCI I/O BAR\n");
1670
1671         /* early init functions */
1672         r = amdgpu_early_init(adev);
1673         if (r)
1674                 return r;
1675
1676         /* if we have > 1 VGA cards, then disable the amdgpu VGA resources */
1677         /* this will fail for cards that aren't VGA class devices, just
1678          * ignore it */
1679         vga_client_register(adev->pdev, adev, NULL, amdgpu_vga_set_decode);
1680
1681         if (amdgpu_runtime_pm == 1)
1682                 runtime = true;
1683         if (amdgpu_device_is_px(ddev))
1684                 runtime = true;
1685         vga_switcheroo_register_client(adev->pdev, &amdgpu_switcheroo_ops, runtime);
1686         if (runtime)
1687                 vga_switcheroo_init_domain_pm_ops(adev->dev, &adev->vga_pm_domain);
1688
1689         /* Read BIOS */
1690         if (!amdgpu_get_bios(adev)) {
1691                 r = -EINVAL;
1692                 goto failed;
1693         }
1694         /* Must be an ATOMBIOS */
1695         if (!adev->is_atom_bios) {
1696                 dev_err(adev->dev, "Expecting atombios for GPU\n");
1697                 r = -EINVAL;
1698                 goto failed;
1699         }
1700         r = amdgpu_atombios_init(adev);
1701         if (r) {
1702                 dev_err(adev->dev, "amdgpu_atombios_init failed\n");
1703                 goto failed;
1704         }
1705
1706         /* detect if we are with an SRIOV vbios */
1707         amdgpu_device_detect_sriov_bios(adev);
1708
1709         /* Post card if necessary */
1710         if (amdgpu_vpost_needed(adev)) {
1711                 if (!adev->bios) {
1712                         dev_err(adev->dev, "no vBIOS found\n");
1713                         r = -EINVAL;
1714                         goto failed;
1715                 }
1716                 DRM_INFO("GPU posting now...\n");
1717                 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
1718                 if (r) {
1719                         dev_err(adev->dev, "gpu post error!\n");
1720                         goto failed;
1721                 }
1722         } else {
1723                 DRM_INFO("GPU post is not needed\n");
1724         }
1725
1726         /* Initialize clocks */
1727         r = amdgpu_atombios_get_clock_info(adev);
1728         if (r) {
1729                 dev_err(adev->dev, "amdgpu_atombios_get_clock_info failed\n");
1730                 goto failed;
1731         }
1732         /* init i2c buses */
1733         amdgpu_atombios_i2c_init(adev);
1734
1735         /* Fence driver */
1736         r = amdgpu_fence_driver_init(adev);
1737         if (r) {
1738                 dev_err(adev->dev, "amdgpu_fence_driver_init failed\n");
1739                 goto failed;
1740         }
1741
1742         /* init the mode config */
1743         drm_mode_config_init(adev->ddev);
1744
1745         r = amdgpu_init(adev);
1746         if (r) {
1747                 dev_err(adev->dev, "amdgpu_init failed\n");
1748                 amdgpu_fini(adev);
1749                 goto failed;
1750         }
1751
1752         adev->accel_working = true;
1753
1754         /* Initialize the buffer migration limit. */
1755         if (amdgpu_moverate >= 0)
1756                 max_MBps = amdgpu_moverate;
1757         else
1758                 max_MBps = 8; /* Allow 8 MB/s. */
1759         /* Get a log2 for easy divisions. */
1760         adev->mm_stats.log2_max_MBps = ilog2(max(1u, max_MBps));
1761
1762         amdgpu_fbdev_init(adev);
1763
1764         r = amdgpu_ib_pool_init(adev);
1765         if (r) {
1766                 dev_err(adev->dev, "IB initialization failed (%d).\n", r);
1767                 goto failed;
1768         }
1769
1770         r = amdgpu_ib_ring_tests(adev);
1771         if (r)
1772                 DRM_ERROR("ib ring test failed (%d).\n", r);
1773
1774         r = amdgpu_gem_debugfs_init(adev);
1775         if (r) {
1776                 DRM_ERROR("registering gem debugfs failed (%d).\n", r);
1777         }
1778
1779         r = amdgpu_debugfs_regs_init(adev);
1780         if (r) {
1781                 DRM_ERROR("registering register debugfs failed (%d).\n", r);
1782         }
1783
1784         r = amdgpu_debugfs_firmware_init(adev);
1785         if (r) {
1786                 DRM_ERROR("registering firmware debugfs failed (%d).\n", r);
1787                 return r;
1788         }
1789
1790         if ((amdgpu_testing & 1)) {
1791                 if (adev->accel_working)
1792                         amdgpu_test_moves(adev);
1793                 else
1794                         DRM_INFO("amdgpu: acceleration disabled, skipping move tests\n");
1795         }
1796         if ((amdgpu_testing & 2)) {
1797                 if (adev->accel_working)
1798                         amdgpu_test_syncing(adev);
1799                 else
1800                         DRM_INFO("amdgpu: acceleration disabled, skipping sync tests\n");
1801         }
1802         if (amdgpu_benchmarking) {
1803                 if (adev->accel_working)
1804                         amdgpu_benchmark(adev, amdgpu_benchmarking);
1805                 else
1806                         DRM_INFO("amdgpu: acceleration disabled, skipping benchmarks\n");
1807         }
1808
1809         /* enable clockgating, etc. after ib tests, etc. since some blocks require
1810          * explicit gating rather than handling it automatically.
1811          */
1812         r = amdgpu_late_init(adev);
1813         if (r) {
1814                 dev_err(adev->dev, "amdgpu_late_init failed\n");
1815                 goto failed;
1816         }
1817
1818         return 0;
1819
1820 failed:
1821         if (runtime)
1822                 vga_switcheroo_fini_domain_pm_ops(adev->dev);
1823         return r;
1824 }
1825
1826 static void amdgpu_debugfs_remove_files(struct amdgpu_device *adev);
1827
1828 /**
1829  * amdgpu_device_fini - tear down the driver
1830  *
1831  * @adev: amdgpu_device pointer
1832  *
1833  * Tear down the driver info (all asics).
1834  * Called at driver shutdown.
1835  */
1836 void amdgpu_device_fini(struct amdgpu_device *adev)
1837 {
1838         int r;
1839
1840         DRM_INFO("amdgpu: finishing device.\n");
1841         adev->shutdown = true;
1842         drm_crtc_force_disable_all(adev->ddev);
1843         /* evict vram memory */
1844         amdgpu_bo_evict_vram(adev);
1845         amdgpu_ib_pool_fini(adev);
1846         amdgpu_fence_driver_fini(adev);
1847         amdgpu_fbdev_fini(adev);
1848         r = amdgpu_fini(adev);
1849         kfree(adev->ip_block_status);
1850         adev->ip_block_status = NULL;
1851         adev->accel_working = false;
1852         /* free i2c buses */
1853         amdgpu_i2c_fini(adev);
1854         amdgpu_atombios_fini(adev);
1855         kfree(adev->bios);
1856         adev->bios = NULL;
1857         vga_switcheroo_unregister_client(adev->pdev);
1858         if (adev->flags & AMD_IS_PX)
1859                 vga_switcheroo_fini_domain_pm_ops(adev->dev);
1860         vga_client_register(adev->pdev, NULL, NULL, NULL);
1861         if (adev->rio_mem)
1862                 pci_iounmap(adev->pdev, adev->rio_mem);
1863         adev->rio_mem = NULL;
1864         iounmap(adev->rmmio);
1865         adev->rmmio = NULL;
1866         if (adev->asic_type >= CHIP_BONAIRE)
1867                 amdgpu_doorbell_fini(adev);
1868         amdgpu_debugfs_regs_cleanup(adev);
1869         amdgpu_debugfs_remove_files(adev);
1870 }
1871
1872
1873 /*
1874  * Suspend & resume.
1875  */
1876 /**
1877  * amdgpu_device_suspend - initiate device suspend
1878  *
1879  * @pdev: drm dev pointer
1880  * @state: suspend state
1881  *
1882  * Puts the hw in the suspend state (all asics).
1883  * Returns 0 for success or an error on failure.
1884  * Called at driver suspend.
1885  */
1886 int amdgpu_device_suspend(struct drm_device *dev, bool suspend, bool fbcon)
1887 {
1888         struct amdgpu_device *adev;
1889         struct drm_crtc *crtc;
1890         struct drm_connector *connector;
1891         int r;
1892
1893         if (dev == NULL || dev->dev_private == NULL) {
1894                 return -ENODEV;
1895         }
1896
1897         adev = dev->dev_private;
1898
1899         if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
1900                 return 0;
1901
1902         drm_kms_helper_poll_disable(dev);
1903
1904         /* turn off display hw */
1905         drm_modeset_lock_all(dev);
1906         list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
1907                 drm_helper_connector_dpms(connector, DRM_MODE_DPMS_OFF);
1908         }
1909         drm_modeset_unlock_all(dev);
1910
1911         /* unpin the front buffers and cursors */
1912         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
1913                 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
1914                 struct amdgpu_framebuffer *rfb = to_amdgpu_framebuffer(crtc->primary->fb);
1915                 struct amdgpu_bo *robj;
1916
1917                 if (amdgpu_crtc->cursor_bo) {
1918                         struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
1919                         r = amdgpu_bo_reserve(aobj, false);
1920                         if (r == 0) {
1921                                 amdgpu_bo_unpin(aobj);
1922                                 amdgpu_bo_unreserve(aobj);
1923                         }
1924                 }
1925
1926                 if (rfb == NULL || rfb->obj == NULL) {
1927                         continue;
1928                 }
1929                 robj = gem_to_amdgpu_bo(rfb->obj);
1930                 /* don't unpin kernel fb objects */
1931                 if (!amdgpu_fbdev_robj_is_fb(adev, robj)) {
1932                         r = amdgpu_bo_reserve(robj, false);
1933                         if (r == 0) {
1934                                 amdgpu_bo_unpin(robj);
1935                                 amdgpu_bo_unreserve(robj);
1936                         }
1937                 }
1938         }
1939         /* evict vram memory */
1940         amdgpu_bo_evict_vram(adev);
1941
1942         amdgpu_fence_driver_suspend(adev);
1943
1944         r = amdgpu_suspend(adev);
1945
1946         /* evict remaining vram memory */
1947         amdgpu_bo_evict_vram(adev);
1948
1949         amdgpu_atombios_scratch_regs_save(adev);
1950         pci_save_state(dev->pdev);
1951         if (suspend) {
1952                 /* Shut down the device */
1953                 pci_disable_device(dev->pdev);
1954                 pci_set_power_state(dev->pdev, PCI_D3hot);
1955         } else {
1956                 r = amdgpu_asic_reset(adev);
1957                 if (r)
1958                         DRM_ERROR("amdgpu asic reset failed\n");
1959         }
1960
1961         if (fbcon) {
1962                 console_lock();
1963                 amdgpu_fbdev_set_suspend(adev, 1);
1964                 console_unlock();
1965         }
1966         return 0;
1967 }
1968
1969 /**
1970  * amdgpu_device_resume - initiate device resume
1971  *
1972  * @pdev: drm dev pointer
1973  *
1974  * Bring the hw back to operating state (all asics).
1975  * Returns 0 for success or an error on failure.
1976  * Called at driver resume.
1977  */
1978 int amdgpu_device_resume(struct drm_device *dev, bool resume, bool fbcon)
1979 {
1980         struct drm_connector *connector;
1981         struct amdgpu_device *adev = dev->dev_private;
1982         struct drm_crtc *crtc;
1983         int r;
1984
1985         if (dev->switch_power_state == DRM_SWITCH_POWER_OFF)
1986                 return 0;
1987
1988         if (fbcon)
1989                 console_lock();
1990
1991         if (resume) {
1992                 pci_set_power_state(dev->pdev, PCI_D0);
1993                 pci_restore_state(dev->pdev);
1994                 r = pci_enable_device(dev->pdev);
1995                 if (r) {
1996                         if (fbcon)
1997                                 console_unlock();
1998                         return r;
1999                 }
2000         }
2001         amdgpu_atombios_scratch_regs_restore(adev);
2002
2003         /* post card */
2004         if (!amdgpu_card_posted(adev) || !resume) {
2005                 r = amdgpu_atom_asic_init(adev->mode_info.atom_context);
2006                 if (r)
2007                         DRM_ERROR("amdgpu asic init failed\n");
2008         }
2009
2010         r = amdgpu_resume(adev);
2011         if (r)
2012                 DRM_ERROR("amdgpu_resume failed (%d).\n", r);
2013
2014         amdgpu_fence_driver_resume(adev);
2015
2016         if (resume) {
2017                 r = amdgpu_ib_ring_tests(adev);
2018                 if (r)
2019                         DRM_ERROR("ib ring test failed (%d).\n", r);
2020         }
2021
2022         r = amdgpu_late_init(adev);
2023         if (r)
2024                 return r;
2025
2026         /* pin cursors */
2027         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2028                 struct amdgpu_crtc *amdgpu_crtc = to_amdgpu_crtc(crtc);
2029
2030                 if (amdgpu_crtc->cursor_bo) {
2031                         struct amdgpu_bo *aobj = gem_to_amdgpu_bo(amdgpu_crtc->cursor_bo);
2032                         r = amdgpu_bo_reserve(aobj, false);
2033                         if (r == 0) {
2034                                 r = amdgpu_bo_pin(aobj,
2035                                                   AMDGPU_GEM_DOMAIN_VRAM,
2036                                                   &amdgpu_crtc->cursor_addr);
2037                                 if (r != 0)
2038                                         DRM_ERROR("Failed to pin cursor BO (%d)\n", r);
2039                                 amdgpu_bo_unreserve(aobj);
2040                         }
2041                 }
2042         }
2043
2044         /* blat the mode back in */
2045         if (fbcon) {
2046                 drm_helper_resume_force_mode(dev);
2047                 /* turn on display hw */
2048                 drm_modeset_lock_all(dev);
2049                 list_for_each_entry(connector, &dev->mode_config.connector_list, head) {
2050                         drm_helper_connector_dpms(connector, DRM_MODE_DPMS_ON);
2051                 }
2052                 drm_modeset_unlock_all(dev);
2053         }
2054
2055         drm_kms_helper_poll_enable(dev);
2056
2057         /*
2058          * Most of the connector probing functions try to acquire runtime pm
2059          * refs to ensure that the GPU is powered on when connector polling is
2060          * performed. Since we're calling this from a runtime PM callback,
2061          * trying to acquire rpm refs will cause us to deadlock.
2062          *
2063          * Since we're guaranteed to be holding the rpm lock, it's safe to
2064          * temporarily disable the rpm helpers so this doesn't deadlock us.
2065          */
2066 #ifdef CONFIG_PM
2067         dev->dev->power.disable_depth++;
2068 #endif
2069         drm_helper_hpd_irq_event(dev);
2070 #ifdef CONFIG_PM
2071         dev->dev->power.disable_depth--;
2072 #endif
2073
2074         if (fbcon) {
2075                 amdgpu_fbdev_set_suspend(adev, 0);
2076                 console_unlock();
2077         }
2078
2079         return 0;
2080 }
2081
2082 static bool amdgpu_check_soft_reset(struct amdgpu_device *adev)
2083 {
2084         int i;
2085         bool asic_hang = false;
2086
2087         for (i = 0; i < adev->num_ip_blocks; i++) {
2088                 if (!adev->ip_block_status[i].valid)
2089                         continue;
2090                 if (adev->ip_blocks[i].funcs->check_soft_reset)
2091                         adev->ip_block_status[i].hang =
2092                                 adev->ip_blocks[i].funcs->check_soft_reset(adev);
2093                 if (adev->ip_block_status[i].hang) {
2094                         DRM_INFO("IP block:%d is hang!\n", i);
2095                         asic_hang = true;
2096                 }
2097         }
2098         return asic_hang;
2099 }
2100
2101 static int amdgpu_pre_soft_reset(struct amdgpu_device *adev)
2102 {
2103         int i, r = 0;
2104
2105         for (i = 0; i < adev->num_ip_blocks; i++) {
2106                 if (!adev->ip_block_status[i].valid)
2107                         continue;
2108                 if (adev->ip_block_status[i].hang &&
2109                     adev->ip_blocks[i].funcs->pre_soft_reset) {
2110                         r = adev->ip_blocks[i].funcs->pre_soft_reset(adev);
2111                         if (r)
2112                                 return r;
2113                 }
2114         }
2115
2116         return 0;
2117 }
2118
2119 static bool amdgpu_need_full_reset(struct amdgpu_device *adev)
2120 {
2121         int i;
2122
2123         for (i = 0; i < adev->num_ip_blocks; i++) {
2124                 if (!adev->ip_block_status[i].valid)
2125                         continue;
2126                 if ((adev->ip_blocks[i].type == AMD_IP_BLOCK_TYPE_GMC) ||
2127                     (adev->ip_blocks[i].type == AMD_IP_BLOCK_TYPE_SMC) ||
2128                     (adev->ip_blocks[i].type == AMD_IP_BLOCK_TYPE_ACP) ||
2129                     (adev->ip_blocks[i].type == AMD_IP_BLOCK_TYPE_DCE)) {
2130                         if (adev->ip_block_status[i].hang) {
2131                                 DRM_INFO("Some block need full reset!\n");
2132                                 return true;
2133                         }
2134                 }
2135         }
2136         return false;
2137 }
2138
2139 static int amdgpu_soft_reset(struct amdgpu_device *adev)
2140 {
2141         int i, r = 0;
2142
2143         for (i = 0; i < adev->num_ip_blocks; i++) {
2144                 if (!adev->ip_block_status[i].valid)
2145                         continue;
2146                 if (adev->ip_block_status[i].hang &&
2147                     adev->ip_blocks[i].funcs->soft_reset) {
2148                         r = adev->ip_blocks[i].funcs->soft_reset(adev);
2149                         if (r)
2150                                 return r;
2151                 }
2152         }
2153
2154         return 0;
2155 }
2156
2157 static int amdgpu_post_soft_reset(struct amdgpu_device *adev)
2158 {
2159         int i, r = 0;
2160
2161         for (i = 0; i < adev->num_ip_blocks; i++) {
2162                 if (!adev->ip_block_status[i].valid)
2163                         continue;
2164                 if (adev->ip_block_status[i].hang &&
2165                     adev->ip_blocks[i].funcs->post_soft_reset)
2166                         r = adev->ip_blocks[i].funcs->post_soft_reset(adev);
2167                 if (r)
2168                         return r;
2169         }
2170
2171         return 0;
2172 }
2173
2174 bool amdgpu_need_backup(struct amdgpu_device *adev)
2175 {
2176         if (adev->flags & AMD_IS_APU)
2177                 return false;
2178
2179         return amdgpu_lockup_timeout > 0 ? true : false;
2180 }
2181
2182 static int amdgpu_recover_vram_from_shadow(struct amdgpu_device *adev,
2183                                            struct amdgpu_ring *ring,
2184                                            struct amdgpu_bo *bo,
2185                                            struct fence **fence)
2186 {
2187         uint32_t domain;
2188         int r;
2189
2190        if (!bo->shadow)
2191                return 0;
2192
2193        r = amdgpu_bo_reserve(bo, false);
2194        if (r)
2195                return r;
2196        domain = amdgpu_mem_type_to_domain(bo->tbo.mem.mem_type);
2197        /* if bo has been evicted, then no need to recover */
2198        if (domain == AMDGPU_GEM_DOMAIN_VRAM) {
2199                r = amdgpu_bo_restore_from_shadow(adev, ring, bo,
2200                                                  NULL, fence, true);
2201                if (r) {
2202                        DRM_ERROR("recover page table failed!\n");
2203                        goto err;
2204                }
2205        }
2206 err:
2207        amdgpu_bo_unreserve(bo);
2208        return r;
2209 }
2210
2211 /**
2212  * amdgpu_gpu_reset - reset the asic
2213  *
2214  * @adev: amdgpu device pointer
2215  *
2216  * Attempt the reset the GPU if it has hung (all asics).
2217  * Returns 0 for success or an error on failure.
2218  */
2219 int amdgpu_gpu_reset(struct amdgpu_device *adev)
2220 {
2221         int i, r;
2222         int resched;
2223         bool need_full_reset;
2224
2225         if (!amdgpu_check_soft_reset(adev)) {
2226                 DRM_INFO("No hardware hang detected. Did some blocks stall?\n");
2227                 return 0;
2228         }
2229
2230         atomic_inc(&adev->gpu_reset_counter);
2231
2232         /* block TTM */
2233         resched = ttm_bo_lock_delayed_workqueue(&adev->mman.bdev);
2234
2235         /* block scheduler */
2236         for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2237                 struct amdgpu_ring *ring = adev->rings[i];
2238
2239                 if (!ring)
2240                         continue;
2241                 kthread_park(ring->sched.thread);
2242                 amd_sched_hw_job_reset(&ring->sched);
2243         }
2244         /* after all hw jobs are reset, hw fence is meaningless, so force_completion */
2245         amdgpu_fence_driver_force_completion(adev);
2246
2247         need_full_reset = amdgpu_need_full_reset(adev);
2248
2249         if (!need_full_reset) {
2250                 amdgpu_pre_soft_reset(adev);
2251                 r = amdgpu_soft_reset(adev);
2252                 amdgpu_post_soft_reset(adev);
2253                 if (r || amdgpu_check_soft_reset(adev)) {
2254                         DRM_INFO("soft reset failed, will fallback to full reset!\n");
2255                         need_full_reset = true;
2256                 }
2257         }
2258
2259         if (need_full_reset) {
2260                 r = amdgpu_suspend(adev);
2261
2262 retry:
2263                 /* Disable fb access */
2264                 if (adev->mode_info.num_crtc) {
2265                         struct amdgpu_mode_mc_save save;
2266                         amdgpu_display_stop_mc_access(adev, &save);
2267                         amdgpu_wait_for_idle(adev, AMD_IP_BLOCK_TYPE_GMC);
2268                 }
2269                 amdgpu_atombios_scratch_regs_save(adev);
2270                 r = amdgpu_asic_reset(adev);
2271                 amdgpu_atombios_scratch_regs_restore(adev);
2272                 /* post card */
2273                 amdgpu_atom_asic_init(adev->mode_info.atom_context);
2274
2275                 if (!r) {
2276                         dev_info(adev->dev, "GPU reset succeeded, trying to resume\n");
2277                         r = amdgpu_resume(adev);
2278                 }
2279         }
2280         if (!r) {
2281                 amdgpu_irq_gpu_reset_resume_helper(adev);
2282                 if (need_full_reset && amdgpu_need_backup(adev)) {
2283                         r = amdgpu_ttm_recover_gart(adev);
2284                         if (r)
2285                                 DRM_ERROR("gart recovery failed!!!\n");
2286                 }
2287                 r = amdgpu_ib_ring_tests(adev);
2288                 if (r) {
2289                         dev_err(adev->dev, "ib ring test failed (%d).\n", r);
2290                         r = amdgpu_suspend(adev);
2291                         need_full_reset = true;
2292                         goto retry;
2293                 }
2294                 /**
2295                  * recovery vm page tables, since we cannot depend on VRAM is
2296                  * consistent after gpu full reset.
2297                  */
2298                 if (need_full_reset && amdgpu_need_backup(adev)) {
2299                         struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
2300                         struct amdgpu_bo *bo, *tmp;
2301                         struct fence *fence = NULL, *next = NULL;
2302
2303                         DRM_INFO("recover vram bo from shadow\n");
2304                         mutex_lock(&adev->shadow_list_lock);
2305                         list_for_each_entry_safe(bo, tmp, &adev->shadow_list, shadow_list) {
2306                                 amdgpu_recover_vram_from_shadow(adev, ring, bo, &next);
2307                                 if (fence) {
2308                                         r = fence_wait(fence, false);
2309                                         if (r) {
2310                                                 WARN(r, "recovery from shadow isn't comleted\n");
2311                                                 break;
2312                                         }
2313                                 }
2314
2315                                 fence_put(fence);
2316                                 fence = next;
2317                         }
2318                         mutex_unlock(&adev->shadow_list_lock);
2319                         if (fence) {
2320                                 r = fence_wait(fence, false);
2321                                 if (r)
2322                                         WARN(r, "recovery from shadow isn't comleted\n");
2323                         }
2324                         fence_put(fence);
2325                 }
2326                 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2327                         struct amdgpu_ring *ring = adev->rings[i];
2328                         if (!ring)
2329                                 continue;
2330
2331                         amd_sched_job_recovery(&ring->sched);
2332                         kthread_unpark(ring->sched.thread);
2333                 }
2334         } else {
2335                 dev_err(adev->dev, "asic resume failed (%d).\n", r);
2336                 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
2337                         if (adev->rings[i]) {
2338                                 kthread_unpark(adev->rings[i]->sched.thread);
2339                         }
2340                 }
2341         }
2342
2343         drm_helper_resume_force_mode(adev->ddev);
2344
2345         ttm_bo_unlock_delayed_workqueue(&adev->mman.bdev, resched);
2346         if (r) {
2347                 /* bad news, how to tell it to userspace ? */
2348                 dev_info(adev->dev, "GPU reset failed\n");
2349         }
2350
2351         return r;
2352 }
2353
2354 void amdgpu_get_pcie_info(struct amdgpu_device *adev)
2355 {
2356         u32 mask;
2357         int ret;
2358
2359         if (amdgpu_pcie_gen_cap)
2360                 adev->pm.pcie_gen_mask = amdgpu_pcie_gen_cap;
2361
2362         if (amdgpu_pcie_lane_cap)
2363                 adev->pm.pcie_mlw_mask = amdgpu_pcie_lane_cap;
2364
2365         /* covers APUs as well */
2366         if (pci_is_root_bus(adev->pdev->bus)) {
2367                 if (adev->pm.pcie_gen_mask == 0)
2368                         adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
2369                 if (adev->pm.pcie_mlw_mask == 0)
2370                         adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
2371                 return;
2372         }
2373
2374         if (adev->pm.pcie_gen_mask == 0) {
2375                 ret = drm_pcie_get_speed_cap_mask(adev->ddev, &mask);
2376                 if (!ret) {
2377                         adev->pm.pcie_gen_mask = (CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN1 |
2378                                                   CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN2 |
2379                                                   CAIL_ASIC_PCIE_LINK_SPEED_SUPPORT_GEN3);
2380
2381                         if (mask & DRM_PCIE_SPEED_25)
2382                                 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN1;
2383                         if (mask & DRM_PCIE_SPEED_50)
2384                                 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN2;
2385                         if (mask & DRM_PCIE_SPEED_80)
2386                                 adev->pm.pcie_gen_mask |= CAIL_PCIE_LINK_SPEED_SUPPORT_GEN3;
2387                 } else {
2388                         adev->pm.pcie_gen_mask = AMDGPU_DEFAULT_PCIE_GEN_MASK;
2389                 }
2390         }
2391         if (adev->pm.pcie_mlw_mask == 0) {
2392                 ret = drm_pcie_get_max_link_width(adev->ddev, &mask);
2393                 if (!ret) {
2394                         switch (mask) {
2395                         case 32:
2396                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X32 |
2397                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
2398                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
2399                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2400                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2401                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2402                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2403                                 break;
2404                         case 16:
2405                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X16 |
2406                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
2407                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2408                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2409                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2410                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2411                                 break;
2412                         case 12:
2413                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X12 |
2414                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2415                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2416                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2417                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2418                                 break;
2419                         case 8:
2420                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X8 |
2421                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2422                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2423                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2424                                 break;
2425                         case 4:
2426                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X4 |
2427                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2428                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2429                                 break;
2430                         case 2:
2431                                 adev->pm.pcie_mlw_mask = (CAIL_PCIE_LINK_WIDTH_SUPPORT_X2 |
2432                                                           CAIL_PCIE_LINK_WIDTH_SUPPORT_X1);
2433                                 break;
2434                         case 1:
2435                                 adev->pm.pcie_mlw_mask = CAIL_PCIE_LINK_WIDTH_SUPPORT_X1;
2436                                 break;
2437                         default:
2438                                 break;
2439                         }
2440                 } else {
2441                         adev->pm.pcie_mlw_mask = AMDGPU_DEFAULT_PCIE_MLW_MASK;
2442                 }
2443         }
2444 }
2445
2446 /*
2447  * Debugfs
2448  */
2449 int amdgpu_debugfs_add_files(struct amdgpu_device *adev,
2450                              const struct drm_info_list *files,
2451                              unsigned nfiles)
2452 {
2453         unsigned i;
2454
2455         for (i = 0; i < adev->debugfs_count; i++) {
2456                 if (adev->debugfs[i].files == files) {
2457                         /* Already registered */
2458                         return 0;
2459                 }
2460         }
2461
2462         i = adev->debugfs_count + 1;
2463         if (i > AMDGPU_DEBUGFS_MAX_COMPONENTS) {
2464                 DRM_ERROR("Reached maximum number of debugfs components.\n");
2465                 DRM_ERROR("Report so we increase "
2466                           "AMDGPU_DEBUGFS_MAX_COMPONENTS.\n");
2467                 return -EINVAL;
2468         }
2469         adev->debugfs[adev->debugfs_count].files = files;
2470         adev->debugfs[adev->debugfs_count].num_files = nfiles;
2471         adev->debugfs_count = i;
2472 #if defined(CONFIG_DEBUG_FS)
2473         drm_debugfs_create_files(files, nfiles,
2474                                  adev->ddev->control->debugfs_root,
2475                                  adev->ddev->control);
2476         drm_debugfs_create_files(files, nfiles,
2477                                  adev->ddev->primary->debugfs_root,
2478                                  adev->ddev->primary);
2479 #endif
2480         return 0;
2481 }
2482
2483 static void amdgpu_debugfs_remove_files(struct amdgpu_device *adev)
2484 {
2485 #if defined(CONFIG_DEBUG_FS)
2486         unsigned i;
2487
2488         for (i = 0; i < adev->debugfs_count; i++) {
2489                 drm_debugfs_remove_files(adev->debugfs[i].files,
2490                                          adev->debugfs[i].num_files,
2491                                          adev->ddev->control);
2492                 drm_debugfs_remove_files(adev->debugfs[i].files,
2493                                          adev->debugfs[i].num_files,
2494                                          adev->ddev->primary);
2495         }
2496 #endif
2497 }
2498
2499 #if defined(CONFIG_DEBUG_FS)
2500
2501 static ssize_t amdgpu_debugfs_regs_read(struct file *f, char __user *buf,
2502                                         size_t size, loff_t *pos)
2503 {
2504         struct amdgpu_device *adev = f->f_inode->i_private;
2505         ssize_t result = 0;
2506         int r;
2507         bool pm_pg_lock, use_bank;
2508         unsigned instance_bank, sh_bank, se_bank;
2509
2510         if (size & 0x3 || *pos & 0x3)
2511                 return -EINVAL;
2512
2513         /* are we reading registers for which a PG lock is necessary? */
2514         pm_pg_lock = (*pos >> 23) & 1;
2515
2516         if (*pos & (1ULL << 62)) {
2517                 se_bank = (*pos >> 24) & 0x3FF;
2518                 sh_bank = (*pos >> 34) & 0x3FF;
2519                 instance_bank = (*pos >> 44) & 0x3FF;
2520                 use_bank = 1;
2521         } else {
2522                 use_bank = 0;
2523         }
2524
2525         *pos &= 0x3FFFF;
2526
2527         if (use_bank) {
2528                 if (sh_bank >= adev->gfx.config.max_sh_per_se ||
2529                     se_bank >= adev->gfx.config.max_shader_engines)
2530                         return -EINVAL;
2531                 mutex_lock(&adev->grbm_idx_mutex);
2532                 amdgpu_gfx_select_se_sh(adev, se_bank,
2533                                         sh_bank, instance_bank);
2534         }
2535
2536         if (pm_pg_lock)
2537                 mutex_lock(&adev->pm.mutex);
2538
2539         while (size) {
2540                 uint32_t value;
2541
2542                 if (*pos > adev->rmmio_size)
2543                         goto end;
2544
2545                 value = RREG32(*pos >> 2);
2546                 r = put_user(value, (uint32_t *)buf);
2547                 if (r) {
2548                         result = r;
2549                         goto end;
2550                 }
2551
2552                 result += 4;
2553                 buf += 4;
2554                 *pos += 4;
2555                 size -= 4;
2556         }
2557
2558 end:
2559         if (use_bank) {
2560                 amdgpu_gfx_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff);
2561                 mutex_unlock(&adev->grbm_idx_mutex);
2562         }
2563
2564         if (pm_pg_lock)
2565                 mutex_unlock(&adev->pm.mutex);
2566
2567         return result;
2568 }
2569
2570 static ssize_t amdgpu_debugfs_regs_write(struct file *f, const char __user *buf,
2571                                          size_t size, loff_t *pos)
2572 {
2573         struct amdgpu_device *adev = f->f_inode->i_private;
2574         ssize_t result = 0;
2575         int r;
2576
2577         if (size & 0x3 || *pos & 0x3)
2578                 return -EINVAL;
2579
2580         while (size) {
2581                 uint32_t value;
2582
2583                 if (*pos > adev->rmmio_size)
2584                         return result;
2585
2586                 r = get_user(value, (uint32_t *)buf);
2587                 if (r)
2588                         return r;
2589
2590                 WREG32(*pos >> 2, value);
2591
2592                 result += 4;
2593                 buf += 4;
2594                 *pos += 4;
2595                 size -= 4;
2596         }
2597
2598         return result;
2599 }
2600
2601 static ssize_t amdgpu_debugfs_regs_pcie_read(struct file *f, char __user *buf,
2602                                         size_t size, loff_t *pos)
2603 {
2604         struct amdgpu_device *adev = f->f_inode->i_private;
2605         ssize_t result = 0;
2606         int r;
2607
2608         if (size & 0x3 || *pos & 0x3)
2609                 return -EINVAL;
2610
2611         while (size) {
2612                 uint32_t value;
2613
2614                 value = RREG32_PCIE(*pos >> 2);
2615                 r = put_user(value, (uint32_t *)buf);
2616                 if (r)
2617                         return r;
2618
2619                 result += 4;
2620                 buf += 4;
2621                 *pos += 4;
2622                 size -= 4;
2623         }
2624
2625         return result;
2626 }
2627
2628 static ssize_t amdgpu_debugfs_regs_pcie_write(struct file *f, const char __user *buf,
2629                                          size_t size, loff_t *pos)
2630 {
2631         struct amdgpu_device *adev = f->f_inode->i_private;
2632         ssize_t result = 0;
2633         int r;
2634
2635         if (size & 0x3 || *pos & 0x3)
2636                 return -EINVAL;
2637
2638         while (size) {
2639                 uint32_t value;
2640
2641                 r = get_user(value, (uint32_t *)buf);
2642                 if (r)
2643                         return r;
2644
2645                 WREG32_PCIE(*pos >> 2, value);
2646
2647                 result += 4;
2648                 buf += 4;
2649                 *pos += 4;
2650                 size -= 4;
2651         }
2652
2653         return result;
2654 }
2655
2656 static ssize_t amdgpu_debugfs_regs_didt_read(struct file *f, char __user *buf,
2657                                         size_t size, loff_t *pos)
2658 {
2659         struct amdgpu_device *adev = f->f_inode->i_private;
2660         ssize_t result = 0;
2661         int r;
2662
2663         if (size & 0x3 || *pos & 0x3)
2664                 return -EINVAL;
2665
2666         while (size) {
2667                 uint32_t value;
2668
2669                 value = RREG32_DIDT(*pos >> 2);
2670                 r = put_user(value, (uint32_t *)buf);
2671                 if (r)
2672                         return r;
2673
2674                 result += 4;
2675                 buf += 4;
2676                 *pos += 4;
2677                 size -= 4;
2678         }
2679
2680         return result;
2681 }
2682
2683 static ssize_t amdgpu_debugfs_regs_didt_write(struct file *f, const char __user *buf,
2684                                          size_t size, loff_t *pos)
2685 {
2686         struct amdgpu_device *adev = f->f_inode->i_private;
2687         ssize_t result = 0;
2688         int r;
2689
2690         if (size & 0x3 || *pos & 0x3)
2691                 return -EINVAL;
2692
2693         while (size) {
2694                 uint32_t value;
2695
2696                 r = get_user(value, (uint32_t *)buf);
2697                 if (r)
2698                         return r;
2699
2700                 WREG32_DIDT(*pos >> 2, value);
2701
2702                 result += 4;
2703                 buf += 4;
2704                 *pos += 4;
2705                 size -= 4;
2706         }
2707
2708         return result;
2709 }
2710
2711 static ssize_t amdgpu_debugfs_regs_smc_read(struct file *f, char __user *buf,
2712                                         size_t size, loff_t *pos)
2713 {
2714         struct amdgpu_device *adev = f->f_inode->i_private;
2715         ssize_t result = 0;
2716         int r;
2717
2718         if (size & 0x3 || *pos & 0x3)
2719                 return -EINVAL;
2720
2721         while (size) {
2722                 uint32_t value;
2723
2724                 value = RREG32_SMC(*pos);
2725                 r = put_user(value, (uint32_t *)buf);
2726                 if (r)
2727                         return r;
2728
2729                 result += 4;
2730                 buf += 4;
2731                 *pos += 4;
2732                 size -= 4;
2733         }
2734
2735         return result;
2736 }
2737
2738 static ssize_t amdgpu_debugfs_regs_smc_write(struct file *f, const char __user *buf,
2739                                          size_t size, loff_t *pos)
2740 {
2741         struct amdgpu_device *adev = f->f_inode->i_private;
2742         ssize_t result = 0;
2743         int r;
2744
2745         if (size & 0x3 || *pos & 0x3)
2746                 return -EINVAL;
2747
2748         while (size) {
2749                 uint32_t value;
2750
2751                 r = get_user(value, (uint32_t *)buf);
2752                 if (r)
2753                         return r;
2754
2755                 WREG32_SMC(*pos, value);
2756
2757                 result += 4;
2758                 buf += 4;
2759                 *pos += 4;
2760                 size -= 4;
2761         }
2762
2763         return result;
2764 }
2765
2766 static ssize_t amdgpu_debugfs_gca_config_read(struct file *f, char __user *buf,
2767                                         size_t size, loff_t *pos)
2768 {
2769         struct amdgpu_device *adev = f->f_inode->i_private;
2770         ssize_t result = 0;
2771         int r;
2772         uint32_t *config, no_regs = 0;
2773
2774         if (size & 0x3 || *pos & 0x3)
2775                 return -EINVAL;
2776
2777         config = kmalloc_array(256, sizeof(*config), GFP_KERNEL);
2778         if (!config)
2779                 return -ENOMEM;
2780
2781         /* version, increment each time something is added */
2782         config[no_regs++] = 2;
2783         config[no_regs++] = adev->gfx.config.max_shader_engines;
2784         config[no_regs++] = adev->gfx.config.max_tile_pipes;
2785         config[no_regs++] = adev->gfx.config.max_cu_per_sh;
2786         config[no_regs++] = adev->gfx.config.max_sh_per_se;
2787         config[no_regs++] = adev->gfx.config.max_backends_per_se;
2788         config[no_regs++] = adev->gfx.config.max_texture_channel_caches;
2789         config[no_regs++] = adev->gfx.config.max_gprs;
2790         config[no_regs++] = adev->gfx.config.max_gs_threads;
2791         config[no_regs++] = adev->gfx.config.max_hw_contexts;
2792         config[no_regs++] = adev->gfx.config.sc_prim_fifo_size_frontend;
2793         config[no_regs++] = adev->gfx.config.sc_prim_fifo_size_backend;
2794         config[no_regs++] = adev->gfx.config.sc_hiz_tile_fifo_size;
2795         config[no_regs++] = adev->gfx.config.sc_earlyz_tile_fifo_size;
2796         config[no_regs++] = adev->gfx.config.num_tile_pipes;
2797         config[no_regs++] = adev->gfx.config.backend_enable_mask;
2798         config[no_regs++] = adev->gfx.config.mem_max_burst_length_bytes;
2799         config[no_regs++] = adev->gfx.config.mem_row_size_in_kb;
2800         config[no_regs++] = adev->gfx.config.shader_engine_tile_size;
2801         config[no_regs++] = adev->gfx.config.num_gpus;
2802         config[no_regs++] = adev->gfx.config.multi_gpu_tile_size;
2803         config[no_regs++] = adev->gfx.config.mc_arb_ramcfg;
2804         config[no_regs++] = adev->gfx.config.gb_addr_config;
2805         config[no_regs++] = adev->gfx.config.num_rbs;
2806
2807         /* rev==1 */
2808         config[no_regs++] = adev->rev_id;
2809         config[no_regs++] = adev->pg_flags;
2810         config[no_regs++] = adev->cg_flags;
2811
2812         /* rev==2 */
2813         config[no_regs++] = adev->family;
2814         config[no_regs++] = adev->external_rev_id;
2815
2816         while (size && (*pos < no_regs * 4)) {
2817                 uint32_t value;
2818
2819                 value = config[*pos >> 2];
2820                 r = put_user(value, (uint32_t *)buf);
2821                 if (r) {
2822                         kfree(config);
2823                         return r;
2824                 }
2825
2826                 result += 4;
2827                 buf += 4;
2828                 *pos += 4;
2829                 size -= 4;
2830         }
2831
2832         kfree(config);
2833         return result;
2834 }
2835
2836 static ssize_t amdgpu_debugfs_sensor_read(struct file *f, char __user *buf,
2837                                         size_t size, loff_t *pos)
2838 {
2839         struct amdgpu_device *adev = f->f_inode->i_private;
2840         int idx, r;
2841         int32_t value;
2842
2843         if (size != 4 || *pos & 0x3)
2844                 return -EINVAL;
2845
2846         /* convert offset to sensor number */
2847         idx = *pos >> 2;
2848
2849         if (adev->powerplay.pp_funcs && adev->powerplay.pp_funcs->read_sensor)
2850                 r = adev->powerplay.pp_funcs->read_sensor(adev->powerplay.pp_handle, idx, &value);
2851         else
2852                 return -EINVAL;
2853
2854         if (!r)
2855                 r = put_user(value, (int32_t *)buf);
2856
2857         return !r ? 4 : r;
2858 }
2859
2860 static const struct file_operations amdgpu_debugfs_regs_fops = {
2861         .owner = THIS_MODULE,
2862         .read = amdgpu_debugfs_regs_read,
2863         .write = amdgpu_debugfs_regs_write,
2864         .llseek = default_llseek
2865 };
2866 static const struct file_operations amdgpu_debugfs_regs_didt_fops = {
2867         .owner = THIS_MODULE,
2868         .read = amdgpu_debugfs_regs_didt_read,
2869         .write = amdgpu_debugfs_regs_didt_write,
2870         .llseek = default_llseek
2871 };
2872 static const struct file_operations amdgpu_debugfs_regs_pcie_fops = {
2873         .owner = THIS_MODULE,
2874         .read = amdgpu_debugfs_regs_pcie_read,
2875         .write = amdgpu_debugfs_regs_pcie_write,
2876         .llseek = default_llseek
2877 };
2878 static const struct file_operations amdgpu_debugfs_regs_smc_fops = {
2879         .owner = THIS_MODULE,
2880         .read = amdgpu_debugfs_regs_smc_read,
2881         .write = amdgpu_debugfs_regs_smc_write,
2882         .llseek = default_llseek
2883 };
2884
2885 static const struct file_operations amdgpu_debugfs_gca_config_fops = {
2886         .owner = THIS_MODULE,
2887         .read = amdgpu_debugfs_gca_config_read,
2888         .llseek = default_llseek
2889 };
2890
2891 static const struct file_operations amdgpu_debugfs_sensors_fops = {
2892         .owner = THIS_MODULE,
2893         .read = amdgpu_debugfs_sensor_read,
2894         .llseek = default_llseek
2895 };
2896
2897 static const struct file_operations *debugfs_regs[] = {
2898         &amdgpu_debugfs_regs_fops,
2899         &amdgpu_debugfs_regs_didt_fops,
2900         &amdgpu_debugfs_regs_pcie_fops,
2901         &amdgpu_debugfs_regs_smc_fops,
2902         &amdgpu_debugfs_gca_config_fops,
2903         &amdgpu_debugfs_sensors_fops,
2904 };
2905
2906 static const char *debugfs_regs_names[] = {
2907         "amdgpu_regs",
2908         "amdgpu_regs_didt",
2909         "amdgpu_regs_pcie",
2910         "amdgpu_regs_smc",
2911         "amdgpu_gca_config",
2912         "amdgpu_sensors",
2913 };
2914
2915 static int amdgpu_debugfs_regs_init(struct amdgpu_device *adev)
2916 {
2917         struct drm_minor *minor = adev->ddev->primary;
2918         struct dentry *ent, *root = minor->debugfs_root;
2919         unsigned i, j;
2920
2921         for (i = 0; i < ARRAY_SIZE(debugfs_regs); i++) {
2922                 ent = debugfs_create_file(debugfs_regs_names[i],
2923                                           S_IFREG | S_IRUGO, root,
2924                                           adev, debugfs_regs[i]);
2925                 if (IS_ERR(ent)) {
2926                         for (j = 0; j < i; j++) {
2927                                 debugfs_remove(adev->debugfs_regs[i]);
2928                                 adev->debugfs_regs[i] = NULL;
2929                         }
2930                         return PTR_ERR(ent);
2931                 }
2932
2933                 if (!i)
2934                         i_size_write(ent->d_inode, adev->rmmio_size);
2935                 adev->debugfs_regs[i] = ent;
2936         }
2937
2938         return 0;
2939 }
2940
2941 static void amdgpu_debugfs_regs_cleanup(struct amdgpu_device *adev)
2942 {
2943         unsigned i;
2944
2945         for (i = 0; i < ARRAY_SIZE(debugfs_regs); i++) {
2946                 if (adev->debugfs_regs[i]) {
2947                         debugfs_remove(adev->debugfs_regs[i]);
2948                         adev->debugfs_regs[i] = NULL;
2949                 }
2950         }
2951 }
2952
2953 int amdgpu_debugfs_init(struct drm_minor *minor)
2954 {
2955         return 0;
2956 }
2957
2958 void amdgpu_debugfs_cleanup(struct drm_minor *minor)
2959 {
2960 }
2961 #else
2962 static int amdgpu_debugfs_regs_init(struct amdgpu_device *adev)
2963 {
2964         return 0;
2965 }
2966 static void amdgpu_debugfs_regs_cleanup(struct amdgpu_device *adev) { }
2967 #endif