drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c

   1 /*
   2  * Copyright 2014 Advanced Micro Devices, Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  */
  23 #include <linux/firmware.h>
  24 #include "drmP.h"
  25 #include "amdgpu.h"
  26 #include "cikd.h"
  27 #include "cik.h"
  28 #include "gmc_v7_0.h"
  29 #include "amdgpu_ucode.h"
  30
  31 #include "bif/bif_4_1_d.h"
  32 #include "bif/bif_4_1_sh_mask.h"
  33
  34 #include "gmc/gmc_7_1_d.h"
  35 #include "gmc/gmc_7_1_sh_mask.h"
  36
  37 #include "oss/oss_2_0_d.h"
  38 #include "oss/oss_2_0_sh_mask.h"
  39
  40 static void gmc_v7_0_set_gart_funcs(struct amdgpu_device *adev);
  41 static void gmc_v7_0_set_irq_funcs(struct amdgpu_device *adev);
  42
  43 MODULE_FIRMWARE("radeon/bonaire_mc.bin");
  44 MODULE_FIRMWARE("radeon/hawaii_mc.bin");
  45
  46 /**
  47  * gmc8_mc_wait_for_idle - wait for MC idle callback.
  48  *
  49  * @adev: amdgpu_device pointer
  50  *
  51  * Wait for the MC (memory controller) to be idle.
  52  * (evergreen+).
  53  * Returns 0 if the MC is idle, -1 if not.
  54  */
  55 int gmc_v7_0_mc_wait_for_idle(struct amdgpu_device *adev)
  56 {
  57         unsigned i;
  58         u32 tmp;
  59
  60         for (i = 0; i < adev->usec_timeout; i++) {
  61                 /* read MC_STATUS */
  62                 tmp = RREG32(mmSRBM_STATUS) & 0x1F00;
  63                 if (!tmp)
  64                         return 0;
  65                 udelay(1);
  66         }
  67         return -1;
  68 }
  69
  70 void gmc_v7_0_mc_stop(struct amdgpu_device *adev,
  71                       struct amdgpu_mode_mc_save *save)
  72 {
  73         u32 blackout;
  74
  75         if (adev->mode_info.num_crtc)
  76                 amdgpu_display_stop_mc_access(adev, save);
  77
  78         amdgpu_asic_wait_for_mc_idle(adev);
  79
  80         blackout = RREG32(mmMC_SHARED_BLACKOUT_CNTL);
  81         if (REG_GET_FIELD(blackout, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE) != 1) {
  82                 /* Block CPU access */
  83                 WREG32(mmBIF_FB_EN, 0);
  84                 /* blackout the MC */
  85                 blackout = REG_SET_FIELD(blackout,
  86                                          MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 0);
  87                 WREG32(mmMC_SHARED_BLACKOUT_CNTL, blackout | 1);
  88         }
  89         /* wait for the MC to settle */
  90         udelay(100);
  91 }
  92
  93 void gmc_v7_0_mc_resume(struct amdgpu_device *adev,
  94                         struct amdgpu_mode_mc_save *save)
  95 {
  96         u32 tmp;
  97
  98         /* unblackout the MC */
  99         tmp = RREG32(mmMC_SHARED_BLACKOUT_CNTL);
 100         tmp = REG_SET_FIELD(tmp, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 0);
 101         WREG32(mmMC_SHARED_BLACKOUT_CNTL, tmp);
 102         /* allow CPU access */
 103         tmp = REG_SET_FIELD(0, BIF_FB_EN, FB_READ_EN, 1);
 104         tmp = REG_SET_FIELD(tmp, BIF_FB_EN, FB_WRITE_EN, 1);
 105         WREG32(mmBIF_FB_EN, tmp);
 106
 107         if (adev->mode_info.num_crtc)
 108                 amdgpu_display_resume_mc_access(adev, save);
 109 }
 110
 111 /**
 112  * gmc_v7_0_init_microcode - load ucode images from disk
 113  *
 114  * @adev: amdgpu_device pointer
 115  *
 116  * Use the firmware interface to load the ucode images into
 117  * the driver (not loaded into hw).
 118  * Returns 0 on success, error on failure.
 119  */
 120 static int gmc_v7_0_init_microcode(struct amdgpu_device *adev)
 121 {
 122         const char *chip_name;
 123         char fw_name[30];
 124         int err;
 125
 126         DRM_DEBUG("\n");
 127
 128         switch (adev->asic_type) {
 129         case CHIP_BONAIRE:
 130                 chip_name = "bonaire";
 131                 break;
 132         case CHIP_HAWAII:
 133                 chip_name = "hawaii";
 134                 break;
 135         case CHIP_KAVERI:
 136         case CHIP_KABINI:
 137                 return 0;
 138         default: BUG();
 139         }
 140
 141         snprintf(fw_name, sizeof(fw_name), "radeon/%s_mc.bin", chip_name);
 142         err = request_firmware(&adev->mc.fw, fw_name, adev->dev);
 143         if (err)
 144                 goto out;
 145         err = amdgpu_ucode_validate(adev->mc.fw);
 146
 147 out:
 148         if (err) {
 149                 printk(KERN_ERR
 150                        "cik_mc: Failed to load firmware \"%s\"\n",
 151                        fw_name);
 152                 release_firmware(adev->mc.fw);
 153                 adev->mc.fw = NULL;
 154         }
 155         return err;
 156 }
 157
 158 /**
 159  * gmc_v7_0_mc_load_microcode - load MC ucode into the hw
 160  *
 161  * @adev: amdgpu_device pointer
 162  *
 163  * Load the GDDR MC ucode into the hw (CIK).
 164  * Returns 0 on success, error on failure.
 165  */
 166 static int gmc_v7_0_mc_load_microcode(struct amdgpu_device *adev)
 167 {
 168         const struct mc_firmware_header_v1_0 *hdr;
 169         const __le32 *fw_data = NULL;
 170         const __le32 *io_mc_regs = NULL;
 171         u32 running, blackout = 0;
 172         int i, ucode_size, regs_size;
 173
 174         if (!adev->mc.fw)
 175                 return -EINVAL;
 176
 177         hdr = (const struct mc_firmware_header_v1_0 *)adev->mc.fw->data;
 178         amdgpu_ucode_print_mc_hdr(&hdr->header);
 179
 180         adev->mc.fw_version = le32_to_cpu(hdr->header.ucode_version);
 181         regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
 182         io_mc_regs = (const __le32 *)
 183                 (adev->mc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
 184         ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
 185         fw_data = (const __le32 *)
 186                 (adev->mc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
 187
 188         running = REG_GET_FIELD(RREG32(mmMC_SEQ_SUP_CNTL), MC_SEQ_SUP_CNTL, RUN);
 189
 190         if (running == 0) {
 191                 if (running) {
 192                         blackout = RREG32(mmMC_SHARED_BLACKOUT_CNTL);
 193                         WREG32(mmMC_SHARED_BLACKOUT_CNTL, blackout | 1);
 194                 }
 195
 196                 /* reset the engine and set to writable */
 197                 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008);
 198                 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000010);
 199
 200                 /* load mc io regs */
 201                 for (i = 0; i < regs_size; i++) {
 202                         WREG32(mmMC_SEQ_IO_DEBUG_INDEX, le32_to_cpup(io_mc_regs++));
 203                         WREG32(mmMC_SEQ_IO_DEBUG_DATA, le32_to_cpup(io_mc_regs++));
 204                 }
 205                 /* load the MC ucode */
 206                 for (i = 0; i < ucode_size; i++)
 207                         WREG32(mmMC_SEQ_SUP_PGM, le32_to_cpup(fw_data++));
 208
 209                 /* put the engine back into the active state */
 210                 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008);
 211                 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000004);
 212                 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000001);
 213
 214                 /* wait for training to complete */
 215                 for (i = 0; i < adev->usec_timeout; i++) {
 216                         if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL),
 217                                           MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D0))
 218                                 break;
 219                         udelay(1);
 220                 }
 221                 for (i = 0; i < adev->usec_timeout; i++) {
 222                         if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL),
 223                                           MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D1))
 224                                 break;
 225                         udelay(1);
 226                 }
 227
 228                 if (running)
 229                         WREG32(mmMC_SHARED_BLACKOUT_CNTL, blackout);
 230         }
 231
 232         return 0;
 233 }
 234
 235 static void gmc_v7_0_vram_gtt_location(struct amdgpu_device *adev,
 236                                        struct amdgpu_mc *mc)
 237 {
 238         if (mc->mc_vram_size > 0xFFC0000000ULL) {
 239                 /* leave room for at least 1024M GTT */
 240                 dev_warn(adev->dev, "limiting VRAM\n");
 241                 mc->real_vram_size = 0xFFC0000000ULL;
 242                 mc->mc_vram_size = 0xFFC0000000ULL;
 243         }
 244         amdgpu_vram_location(adev, &adev->mc, 0);
 245         adev->mc.gtt_base_align = 0;
 246         amdgpu_gtt_location(adev, mc);
 247 }
 248
 249 /**
 250  * gmc_v7_0_mc_program - program the GPU memory controller
 251  *
 252  * @adev: amdgpu_device pointer
 253  *
 254  * Set the location of vram, gart, and AGP in the GPU's
 255  * physical address space (CIK).
 256  */
 257 static void gmc_v7_0_mc_program(struct amdgpu_device *adev)
 258 {
 259         struct amdgpu_mode_mc_save save;
 260         u32 tmp;
 261         int i, j;
 262
 263         /* Initialize HDP */
 264         for (i = 0, j = 0; i < 32; i++, j += 0x6) {
 265                 WREG32((0xb05 + j), 0x00000000);
 266                 WREG32((0xb06 + j), 0x00000000);
 267                 WREG32((0xb07 + j), 0x00000000);
 268                 WREG32((0xb08 + j), 0x00000000);
 269                 WREG32((0xb09 + j), 0x00000000);
 270         }
 271         WREG32(mmHDP_REG_COHERENCY_FLUSH_CNTL, 0);
 272
 273         if (adev->mode_info.num_crtc)
 274                 amdgpu_display_set_vga_render_state(adev, false);
 275
 276         gmc_v7_0_mc_stop(adev, &save);
 277         if (amdgpu_asic_wait_for_mc_idle(adev)) {
 278                 dev_warn(adev->dev, "Wait for MC idle timedout !\n");
 279         }
 280         /* Update configuration */
 281         WREG32(mmMC_VM_SYSTEM_APERTURE_LOW_ADDR,
 282                adev->mc.vram_start >> 12);
 283         WREG32(mmMC_VM_SYSTEM_APERTURE_HIGH_ADDR,
 284                adev->mc.vram_end >> 12);
 285         WREG32(mmMC_VM_SYSTEM_APERTURE_DEFAULT_ADDR,
 286                adev->vram_scratch.gpu_addr >> 12);
 287         tmp = ((adev->mc.vram_end >> 24) & 0xFFFF) << 16;
 288         tmp |= ((adev->mc.vram_start >> 24) & 0xFFFF);
 289         WREG32(mmMC_VM_FB_LOCATION, tmp);
 290         /* XXX double check these! */
 291         WREG32(mmHDP_NONSURFACE_BASE, (adev->mc.vram_start >> 8));
 292         WREG32(mmHDP_NONSURFACE_INFO, (2 << 7) | (1 << 30));
 293         WREG32(mmHDP_NONSURFACE_SIZE, 0x3FFFFFFF);
 294         WREG32(mmMC_VM_AGP_BASE, 0);
 295         WREG32(mmMC_VM_AGP_TOP, 0x0FFFFFFF);
 296         WREG32(mmMC_VM_AGP_BOT, 0x0FFFFFFF);
 297         if (amdgpu_asic_wait_for_mc_idle(adev)) {
 298                 dev_warn(adev->dev, "Wait for MC idle timedout !\n");
 299         }
 300         gmc_v7_0_mc_resume(adev, &save);
 301
 302         WREG32(mmBIF_FB_EN, BIF_FB_EN__FB_READ_EN_MASK | BIF_FB_EN__FB_WRITE_EN_MASK);
 303
 304         tmp = RREG32(mmHDP_MISC_CNTL);
 305         tmp = REG_SET_FIELD(tmp, HDP_MISC_CNTL, FLUSH_INVALIDATE_CACHE, 1);
 306         WREG32(mmHDP_MISC_CNTL, tmp);
 307
 308         tmp = RREG32(mmHDP_HOST_PATH_CNTL);
 309         WREG32(mmHDP_HOST_PATH_CNTL, tmp);
 310 }
 311
 312 /**
 313  * gmc_v7_0_mc_init - initialize the memory controller driver params
 314  *
 315  * @adev: amdgpu_device pointer
 316  *
 317  * Look up the amount of vram, vram width, and decide how to place
 318  * vram and gart within the GPU's physical address space (CIK).
 319  * Returns 0 for success.
 320  */
 321 static int gmc_v7_0_mc_init(struct amdgpu_device *adev)
 322 {
 323         u32 tmp;
 324         int chansize, numchan;
 325
 326         /* Get VRAM informations */
 327         tmp = RREG32(mmMC_ARB_RAMCFG);
 328         if (REG_GET_FIELD(tmp, MC_ARB_RAMCFG, CHANSIZE)) {
 329                 chansize = 64;
 330         } else {
 331                 chansize = 32;
 332         }
 333         tmp = RREG32(mmMC_SHARED_CHMAP);
 334         switch (REG_GET_FIELD(tmp, MC_SHARED_CHMAP, NOOFCHAN)) {
 335         case 0:
 336         default:
 337                 numchan = 1;
 338                 break;
 339         case 1:
 340                 numchan = 2;
 341                 break;
 342         case 2:
 343                 numchan = 4;
 344                 break;
 345         case 3:
 346                 numchan = 8;
 347                 break;
 348         case 4:
 349                 numchan = 3;
 350                 break;
 351         case 5:
 352                 numchan = 6;
 353                 break;
 354         case 6:
 355                 numchan = 10;
 356                 break;
 357         case 7:
 358                 numchan = 12;
 359                 break;
 360         case 8:
 361                 numchan = 16;
 362                 break;
 363         }
 364         adev->mc.vram_width = numchan * chansize;
 365         /* Could aper size report 0 ? */
 366         adev->mc.aper_base = pci_resource_start(adev->pdev, 0);
 367         adev->mc.aper_size = pci_resource_len(adev->pdev, 0);
 368         /* size in MB on si */
 369         adev->mc.mc_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL;
 370         adev->mc.real_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL;
 371         adev->mc.visible_vram_size = adev->mc.aper_size;
 372
 373         /* In case the PCI BAR is larger than the actual amount of vram */
 374         if (adev->mc.visible_vram_size > adev->mc.real_vram_size)
 375                 adev->mc.visible_vram_size = adev->mc.real_vram_size;
 376
 377         /* unless the user had overridden it, set the gart
 378          * size equal to the 1024 or vram, whichever is larger.
 379          */
 380         if (amdgpu_gart_size == -1)
 381                 adev->mc.gtt_size = max((1024ULL << 20), adev->mc.mc_vram_size);
 382         else
 383                 adev->mc.gtt_size = (uint64_t)amdgpu_gart_size << 20;
 384
 385         gmc_v7_0_vram_gtt_location(adev, &adev->mc);
 386
 387         return 0;
 388 }
 389
 390 /*
 391  * GART
 392  * VMID 0 is the physical GPU addresses as used by the kernel.
 393  * VMIDs 1-15 are used for userspace clients and are handled
 394  * by the amdgpu vm/hsa code.
 395  */
 396
 397 /**
 398  * gmc_v7_0_gart_flush_gpu_tlb - gart tlb flush callback
 399  *
 400  * @adev: amdgpu_device pointer
 401  * @vmid: vm instance to flush
 402  *
 403  * Flush the TLB for the requested page table (CIK).
 404  */
 405 static void gmc_v7_0_gart_flush_gpu_tlb(struct amdgpu_device *adev,
 406                                         uint32_t vmid)
 407 {
 408         /* flush hdp cache */
 409         WREG32(mmHDP_MEM_COHERENCY_FLUSH_CNTL, 0);
 410
 411         /* bits 0-15 are the VM contexts0-15 */
 412         WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
 413 }
 414
 415 /**
 416  * gmc_v7_0_gart_set_pte_pde - update the page tables using MMIO
 417  *
 418  * @adev: amdgpu_device pointer
 419  * @cpu_pt_addr: cpu address of the page table
 420  * @gpu_page_idx: entry in the page table to update
 421  * @addr: dst addr to write into pte/pde
 422  * @flags: access flags
 423  *
 424  * Update the page tables using the CPU.
 425  */
 426 static int gmc_v7_0_gart_set_pte_pde(struct amdgpu_device *adev,
 427                                      void *cpu_pt_addr,
 428                                      uint32_t gpu_page_idx,
 429                                      uint64_t addr,
 430                                      uint32_t flags)
 431 {
 432         void __iomem *ptr = (void *)cpu_pt_addr;
 433         uint64_t value;
 434
 435         value = addr & 0xFFFFFFFFFFFFF000ULL;
 436         value |= flags;
 437         writeq(value, ptr + (gpu_page_idx * 8));
 438
 439         return 0;
 440 }
 441
 442 /**
 443  * gmc_v8_0_set_fault_enable_default - update VM fault handling
 444  *
 445  * @adev: amdgpu_device pointer
 446  * @value: true redirects VM faults to the default page
 447  */
 448 static void gmc_v7_0_set_fault_enable_default(struct amdgpu_device *adev,
 449                                               bool value)
 450 {
 451         u32 tmp;
 452
 453         tmp = RREG32(mmVM_CONTEXT1_CNTL);
 454         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
 455                             RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
 456         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
 457                             DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
 458         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
 459                             PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, value);
 460         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
 461                             VALID_PROTECTION_FAULT_ENABLE_DEFAULT, value);
 462         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
 463                             READ_PROTECTION_FAULT_ENABLE_DEFAULT, value);
 464         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
 465                             WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
 466         WREG32(mmVM_CONTEXT1_CNTL, tmp);
 467 }
 468
 469 /**
 470  * gmc_v7_0_gart_enable - gart enable
 471  *
 472  * @adev: amdgpu_device pointer
 473  *
 474  * This sets up the TLBs, programs the page tables for VMID0,
 475  * sets up the hw for VMIDs 1-15 which are allocated on
 476  * demand, and sets up the global locations for the LDS, GDS,
 477  * and GPUVM for FSA64 clients (CIK).
 478  * Returns 0 for success, errors for failure.
 479  */
 480 static int gmc_v7_0_gart_enable(struct amdgpu_device *adev)
 481 {
 482         int r, i;
 483         u32 tmp;
 484
 485         if (adev->gart.robj == NULL) {
 486                 dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
 487                 return -EINVAL;
 488         }
 489         r = amdgpu_gart_table_vram_pin(adev);
 490         if (r)
 491                 return r;
 492         /* Setup TLB control */
 493         tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL);
 494         tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 1);
 495         tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 1);
 496         tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE, 3);
 497         tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 1);
 498         tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_APERTURE_UNMAPPED_ACCESS, 0);
 499         WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp);
 500         /* Setup L2 cache */
 501         tmp = RREG32(mmVM_L2_CNTL);
 502         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 1);
 503         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING, 1);
 504         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE, 1);
 505         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE, 1);
 506         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, EFFECTIVE_L2_QUEUE_SIZE, 7);
 507         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, CONTEXT1_IDENTITY_ACCESS_MODE, 1);
 508         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_DEFAULT_PAGE_OUT_TO_SYSTEM_MEMORY, 1);
 509         WREG32(mmVM_L2_CNTL, tmp);
 510         tmp = REG_SET_FIELD(0, VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS, 1);
 511         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_L2_CACHE, 1);
 512         WREG32(mmVM_L2_CNTL2, tmp);
 513         tmp = RREG32(mmVM_L2_CNTL3);
 514         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY, 1);
 515         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, BANK_SELECT, 4);
 516         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_FRAGMENT_SIZE, 4);
 517         WREG32(mmVM_L2_CNTL3, tmp);
 518         /* setup context0 */
 519         WREG32(mmVM_CONTEXT0_PAGE_TABLE_START_ADDR, adev->mc.gtt_start >> 12);
 520         WREG32(mmVM_CONTEXT0_PAGE_TABLE_END_ADDR, (adev->mc.gtt_end >> 12) - 1);
 521         WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR, adev->gart.table_addr >> 12);
 522         WREG32(mmVM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
 523                         (u32)(adev->dummy_page.addr >> 12));
 524         WREG32(mmVM_CONTEXT0_CNTL2, 0);
 525         tmp = RREG32(mmVM_CONTEXT0_CNTL);
 526         tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, ENABLE_CONTEXT, 1);
 527         tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, PAGE_TABLE_DEPTH, 0);
 528         tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
 529         WREG32(mmVM_CONTEXT0_CNTL, tmp);
 530
 531         WREG32(0x575, 0);
 532         WREG32(0x576, 0);
 533         WREG32(0x577, 0);
 534
 535         /* empty context1-15 */
 536         /* FIXME start with 4G, once using 2 level pt switch to full
 537          * vm size space
 538          */
 539         /* set vm size, must be a multiple of 4 */
 540         WREG32(mmVM_CONTEXT1_PAGE_TABLE_START_ADDR, 0);
 541         WREG32(mmVM_CONTEXT1_PAGE_TABLE_END_ADDR, adev->vm_manager.max_pfn - 1);
 542         for (i = 1; i < 16; i++) {
 543                 if (i < 8)
 544                         WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + i,
 545                                adev->gart.table_addr >> 12);
 546                 else
 547                         WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + i - 8,
 548                                adev->gart.table_addr >> 12);
 549         }
 550
 551         /* enable context1-15 */
 552         WREG32(mmVM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR,
 553                (u32)(adev->dummy_page.addr >> 12));
 554         WREG32(mmVM_CONTEXT1_CNTL2, 4);
 555         tmp = RREG32(mmVM_CONTEXT1_CNTL);
 556         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, ENABLE_CONTEXT, 1);
 557         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_DEPTH, 1);
 558         tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_BLOCK_SIZE,
 559                             amdgpu_vm_block_size - 9);
 560         WREG32(mmVM_CONTEXT1_CNTL, tmp);
 561         if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS)
 562                 gmc_v7_0_set_fault_enable_default(adev, false);
 563         else
 564                 gmc_v7_0_set_fault_enable_default(adev, true);
 565
 566         if (adev->asic_type == CHIP_KAVERI) {
 567                 tmp = RREG32(mmCHUB_CONTROL);
 568                 tmp &= ~BYPASS_VM;
 569                 WREG32(mmCHUB_CONTROL, tmp);
 570         }
 571
 572         gmc_v7_0_gart_flush_gpu_tlb(adev, 0);
 573         DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
 574                  (unsigned)(adev->mc.gtt_size >> 20),
 575                  (unsigned long long)adev->gart.table_addr);
 576         adev->gart.ready = true;
 577         return 0;
 578 }
 579
 580 static int gmc_v7_0_gart_init(struct amdgpu_device *adev)
 581 {
 582         int r;
 583
 584         if (adev->gart.robj) {
 585                 WARN(1, "R600 PCIE GART already initialized\n");
 586                 return 0;
 587         }
 588         /* Initialize common gart structure */
 589         r = amdgpu_gart_init(adev);
 590         if (r)
 591                 return r;
 592         adev->gart.table_size = adev->gart.num_gpu_pages * 8;
 593         return amdgpu_gart_table_vram_alloc(adev);
 594 }
 595
 596 /**
 597  * gmc_v7_0_gart_disable - gart disable
 598  *
 599  * @adev: amdgpu_device pointer
 600  *
 601  * This disables all VM page table (CIK).
 602  */
 603 static void gmc_v7_0_gart_disable(struct amdgpu_device *adev)
 604 {
 605         u32 tmp;
 606
 607         /* Disable all tables */
 608         WREG32(mmVM_CONTEXT0_CNTL, 0);
 609         WREG32(mmVM_CONTEXT1_CNTL, 0);
 610         /* Setup TLB control */
 611         tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL);
 612         tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 0);
 613         tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 0);
 614         tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 0);
 615         WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp);
 616         /* Setup L2 cache */
 617         tmp = RREG32(mmVM_L2_CNTL);
 618         tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 0);
 619         WREG32(mmVM_L2_CNTL, tmp);
 620         WREG32(mmVM_L2_CNTL2, 0);
 621         amdgpu_gart_table_vram_unpin(adev);
 622 }
 623
 624 /**
 625  * gmc_v7_0_gart_fini - vm fini callback
 626  *
 627  * @adev: amdgpu_device pointer
 628  *
 629  * Tears down the driver GART/VM setup (CIK).
 630  */
 631 static void gmc_v7_0_gart_fini(struct amdgpu_device *adev)
 632 {
 633         amdgpu_gart_table_vram_free(adev);
 634         amdgpu_gart_fini(adev);
 635 }
 636
 637 /*
 638  * vm
 639  * VMID 0 is the physical GPU addresses as used by the kernel.
 640  * VMIDs 1-15 are used for userspace clients and are handled
 641  * by the amdgpu vm/hsa code.
 642  */
 643 /**
 644  * gmc_v7_0_vm_init - cik vm init callback
 645  *
 646  * @adev: amdgpu_device pointer
 647  *
 648  * Inits cik specific vm parameters (number of VMs, base of vram for
 649  * VMIDs 1-15) (CIK).
 650  * Returns 0 for success.
 651  */
 652 static int gmc_v7_0_vm_init(struct amdgpu_device *adev)
 653 {
 654         /*
 655          * number of VMs
 656          * VMID 0 is reserved for System
 657          * amdgpu graphics/compute will use VMIDs 1-7
 658          * amdkfd will use VMIDs 8-15
 659          */
 660         adev->vm_manager.nvm = AMDGPU_NUM_OF_VMIDS;
 661
 662         /* base offset of vram pages */
 663         if (adev->flags & AMD_IS_APU) {
 664                 u64 tmp = RREG32(mmMC_VM_FB_OFFSET);
 665                 tmp <<= 22;
 666                 adev->vm_manager.vram_base_offset = tmp;
 667         } else
 668                 adev->vm_manager.vram_base_offset = 0;
 669
 670         return 0;
 671 }
 672
 673 /**
 674  * gmc_v7_0_vm_fini - cik vm fini callback
 675  *
 676  * @adev: amdgpu_device pointer
 677  *
 678  * Tear down any asic specific VM setup (CIK).
 679  */
 680 static void gmc_v7_0_vm_fini(struct amdgpu_device *adev)
 681 {
 682 }
 683
 684 /**
 685  * gmc_v7_0_vm_decode_fault - print human readable fault info
 686  *
 687  * @adev: amdgpu_device pointer
 688  * @status: VM_CONTEXT1_PROTECTION_FAULT_STATUS register value
 689  * @addr: VM_CONTEXT1_PROTECTION_FAULT_ADDR register value
 690  *
 691  * Print human readable fault information (CIK).
 692  */
 693 static void gmc_v7_0_vm_decode_fault(struct amdgpu_device *adev,
 694                                      u32 status, u32 addr, u32 mc_client)
 695 {
 696         u32 mc_id;
 697         u32 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, VMID);
 698         u32 protections = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
 699                                         PROTECTIONS);
 700         char block[5] = { mc_client >> 24, (mc_client >> 16) & 0xff,
 701                 (mc_client >> 8) & 0xff, mc_client & 0xff, 0 };
 702
 703         mc_id = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
 704                               MEMORY_CLIENT_ID);
 705
 706         printk("VM fault (0x%02x, vmid %d) at page %u, %s from '%s' (0x%08x) (%d)\n",
 707                protections, vmid, addr,
 708                REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
 709                              MEMORY_CLIENT_RW) ?
 710                "write" : "read", block, mc_client, mc_id);
 711 }
 712
 713
 714 static const u32 mc_cg_registers[] = {
 715         mmMC_HUB_MISC_HUB_CG,
 716         mmMC_HUB_MISC_SIP_CG,
 717         mmMC_HUB_MISC_VM_CG,
 718         mmMC_XPB_CLK_GAT,
 719         mmATC_MISC_CG,
 720         mmMC_CITF_MISC_WR_CG,
 721         mmMC_CITF_MISC_RD_CG,
 722         mmMC_CITF_MISC_VM_CG,
 723         mmVM_L2_CG,
 724 };
 725
 726 static const u32 mc_cg_ls_en[] = {
 727         MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK,
 728         MC_HUB_MISC_SIP_CG__MEM_LS_ENABLE_MASK,
 729         MC_HUB_MISC_VM_CG__MEM_LS_ENABLE_MASK,
 730         MC_XPB_CLK_GAT__MEM_LS_ENABLE_MASK,
 731         ATC_MISC_CG__MEM_LS_ENABLE_MASK,
 732         MC_CITF_MISC_WR_CG__MEM_LS_ENABLE_MASK,
 733         MC_CITF_MISC_RD_CG__MEM_LS_ENABLE_MASK,
 734         MC_CITF_MISC_VM_CG__MEM_LS_ENABLE_MASK,
 735         VM_L2_CG__MEM_LS_ENABLE_MASK,
 736 };
 737
 738 static const u32 mc_cg_en[] = {
 739         MC_HUB_MISC_HUB_CG__ENABLE_MASK,
 740         MC_HUB_MISC_SIP_CG__ENABLE_MASK,
 741         MC_HUB_MISC_VM_CG__ENABLE_MASK,
 742         MC_XPB_CLK_GAT__ENABLE_MASK,
 743         ATC_MISC_CG__ENABLE_MASK,
 744         MC_CITF_MISC_WR_CG__ENABLE_MASK,
 745         MC_CITF_MISC_RD_CG__ENABLE_MASK,
 746         MC_CITF_MISC_VM_CG__ENABLE_MASK,
 747         VM_L2_CG__ENABLE_MASK,
 748 };
 749
 750 static void gmc_v7_0_enable_mc_ls(struct amdgpu_device *adev,
 751                                   bool enable)
 752 {
 753         int i;
 754         u32 orig, data;
 755
 756         for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) {
 757                 orig = data = RREG32(mc_cg_registers[i]);
 758                 if (enable && (adev->cg_flags & AMDGPU_CG_SUPPORT_MC_LS))
 759                         data |= mc_cg_ls_en[i];
 760                 else
 761                         data &= ~mc_cg_ls_en[i];
 762                 if (data != orig)
 763                         WREG32(mc_cg_registers[i], data);
 764         }
 765 }
 766
 767 static void gmc_v7_0_enable_mc_mgcg(struct amdgpu_device *adev,
 768                                     bool enable)
 769 {
 770         int i;
 771         u32 orig, data;
 772
 773         for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) {
 774                 orig = data = RREG32(mc_cg_registers[i]);
 775                 if (enable && (adev->cg_flags & AMDGPU_CG_SUPPORT_MC_MGCG))
 776                         data |= mc_cg_en[i];
 777                 else
 778                         data &= ~mc_cg_en[i];
 779                 if (data != orig)
 780                         WREG32(mc_cg_registers[i], data);
 781         }
 782 }
 783
 784 static void gmc_v7_0_enable_bif_mgls(struct amdgpu_device *adev,
 785                                      bool enable)
 786 {
 787         u32 orig, data;
 788
 789         orig = data = RREG32_PCIE(ixPCIE_CNTL2);
 790
 791         if (enable && (adev->cg_flags & AMDGPU_CG_SUPPORT_BIF_LS)) {
 792                 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_LS_EN, 1);
 793                 data = REG_SET_FIELD(data, PCIE_CNTL2, MST_MEM_LS_EN, 1);
 794                 data = REG_SET_FIELD(data, PCIE_CNTL2, REPLAY_MEM_LS_EN, 1);
 795                 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_AGGRESSIVE_LS_EN, 1);
 796         } else {
 797                 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_LS_EN, 0);
 798                 data = REG_SET_FIELD(data, PCIE_CNTL2, MST_MEM_LS_EN, 0);
 799                 data = REG_SET_FIELD(data, PCIE_CNTL2, REPLAY_MEM_LS_EN, 0);
 800                 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_AGGRESSIVE_LS_EN, 0);
 801         }
 802
 803         if (orig != data)
 804                 WREG32_PCIE(ixPCIE_CNTL2, data);
 805 }
 806
 807 static void gmc_v7_0_enable_hdp_mgcg(struct amdgpu_device *adev,
 808                                      bool enable)
 809 {
 810         u32 orig, data;
 811
 812         orig = data = RREG32(mmHDP_HOST_PATH_CNTL);
 813
 814         if (enable && (adev->cg_flags & AMDGPU_CG_SUPPORT_HDP_MGCG))
 815                 data = REG_SET_FIELD(data, HDP_HOST_PATH_CNTL, CLOCK_GATING_DIS, 0);
 816         else
 817                 data = REG_SET_FIELD(data, HDP_HOST_PATH_CNTL, CLOCK_GATING_DIS, 1);
 818
 819         if (orig != data)
 820                 WREG32(mmHDP_HOST_PATH_CNTL, data);
 821 }
 822
 823 static void gmc_v7_0_enable_hdp_ls(struct amdgpu_device *adev,
 824                                    bool enable)
 825 {
 826         u32 orig, data;
 827
 828         orig = data = RREG32(mmHDP_MEM_POWER_LS);
 829
 830         if (enable && (adev->cg_flags & AMDGPU_CG_SUPPORT_HDP_LS))
 831                 data = REG_SET_FIELD(data, HDP_MEM_POWER_LS, LS_ENABLE, 1);
 832         else
 833                 data = REG_SET_FIELD(data, HDP_MEM_POWER_LS, LS_ENABLE, 0);
 834
 835         if (orig != data)
 836                 WREG32(mmHDP_MEM_POWER_LS, data);
 837 }
 838
 839 static int gmc_v7_0_convert_vram_type(int mc_seq_vram_type)
 840 {
 841         switch (mc_seq_vram_type) {
 842         case MC_SEQ_MISC0__MT__GDDR1:
 843                 return AMDGPU_VRAM_TYPE_GDDR1;
 844         case MC_SEQ_MISC0__MT__DDR2:
 845                 return AMDGPU_VRAM_TYPE_DDR2;
 846         case MC_SEQ_MISC0__MT__GDDR3:
 847                 return AMDGPU_VRAM_TYPE_GDDR3;
 848         case MC_SEQ_MISC0__MT__GDDR4:
 849                 return AMDGPU_VRAM_TYPE_GDDR4;
 850         case MC_SEQ_MISC0__MT__GDDR5:
 851                 return AMDGPU_VRAM_TYPE_GDDR5;
 852         case MC_SEQ_MISC0__MT__HBM:
 853                 return AMDGPU_VRAM_TYPE_HBM;
 854         case MC_SEQ_MISC0__MT__DDR3:
 855                 return AMDGPU_VRAM_TYPE_DDR3;
 856         default:
 857                 return AMDGPU_VRAM_TYPE_UNKNOWN;
 858         }
 859 }
 860
 861 static int gmc_v7_0_early_init(void *handle)
 862 {
 863         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 864
 865         gmc_v7_0_set_gart_funcs(adev);
 866         gmc_v7_0_set_irq_funcs(adev);
 867
 868         if (adev->flags & AMD_IS_APU) {
 869                 adev->mc.vram_type = AMDGPU_VRAM_TYPE_UNKNOWN;
 870         } else {
 871                 u32 tmp = RREG32(mmMC_SEQ_MISC0);
 872                 tmp &= MC_SEQ_MISC0__MT__MASK;
 873                 adev->mc.vram_type = gmc_v7_0_convert_vram_type(tmp);
 874         }
 875
 876         return 0;
 877 }
 878
 879 static int gmc_v7_0_late_init(void *handle)
 880 {
 881         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 882
 883         return amdgpu_irq_get(adev, &adev->mc.vm_fault, 0);
 884 }
 885
 886 static int gmc_v7_0_sw_init(void *handle)
 887 {
 888         int r;
 889         int dma_bits;
 890         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 891
 892         r = amdgpu_gem_init(adev);
 893         if (r)
 894                 return r;
 895
 896         r = amdgpu_irq_add_id(adev, 146, &adev->mc.vm_fault);
 897         if (r)
 898                 return r;
 899
 900         r = amdgpu_irq_add_id(adev, 147, &adev->mc.vm_fault);
 901         if (r)
 902                 return r;
 903
 904         /* Adjust VM size here.
 905          * Currently set to 4GB ((1 << 20) 4k pages).
 906          * Max GPUVM size for cayman and SI is 40 bits.
 907          */
 908         adev->vm_manager.max_pfn = amdgpu_vm_size << 18;
 909
 910         /* Set the internal MC address mask
 911          * This is the max address of the GPU's
 912          * internal address space.
 913          */
 914         adev->mc.mc_mask = 0xffffffffffULL; /* 40 bit MC */
 915
 916         /* set DMA mask + need_dma32 flags.
 917          * PCIE - can handle 40-bits.
 918          * IGP - can handle 40-bits
 919          * PCI - dma32 for legacy pci gart, 40 bits on newer asics
 920          */
 921         adev->need_dma32 = false;
 922         dma_bits = adev->need_dma32 ? 32 : 40;
 923         r = pci_set_dma_mask(adev->pdev, DMA_BIT_MASK(dma_bits));
 924         if (r) {
 925                 adev->need_dma32 = true;
 926                 dma_bits = 32;
 927                 printk(KERN_WARNING "amdgpu: No suitable DMA available.\n");
 928         }
 929         r = pci_set_consistent_dma_mask(adev->pdev, DMA_BIT_MASK(dma_bits));
 930         if (r) {
 931                 pci_set_consistent_dma_mask(adev->pdev, DMA_BIT_MASK(32));
 932                 printk(KERN_WARNING "amdgpu: No coherent DMA available.\n");
 933         }
 934
 935         r = gmc_v7_0_init_microcode(adev);
 936         if (r) {
 937                 DRM_ERROR("Failed to load mc firmware!\n");
 938                 return r;
 939         }
 940
 941         r = gmc_v7_0_mc_init(adev);
 942         if (r)
 943                 return r;
 944
 945         /* Memory manager */
 946         r = amdgpu_bo_init(adev);
 947         if (r)
 948                 return r;
 949
 950         r = gmc_v7_0_gart_init(adev);
 951         if (r)
 952                 return r;
 953
 954         if (!adev->vm_manager.enabled) {
 955                 r = gmc_v7_0_vm_init(adev);
 956                 if (r) {
 957                         dev_err(adev->dev, "vm manager initialization failed (%d).\n", r);
 958                         return r;
 959                 }
 960                 adev->vm_manager.enabled = true;
 961         }
 962
 963         return r;
 964 }
 965
 966 static int gmc_v7_0_sw_fini(void *handle)
 967 {
 968         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 969
 970         if (adev->vm_manager.enabled) {
 971                 amdgpu_vm_manager_fini(adev);
 972                 gmc_v7_0_vm_fini(adev);
 973                 adev->vm_manager.enabled = false;
 974         }
 975         gmc_v7_0_gart_fini(adev);
 976         amdgpu_gem_fini(adev);
 977         amdgpu_bo_fini(adev);
 978
 979         return 0;
 980 }
 981
 982 static int gmc_v7_0_hw_init(void *handle)
 983 {
 984         int r;
 985         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 986
 987         gmc_v7_0_mc_program(adev);
 988
 989         if (!(adev->flags & AMD_IS_APU)) {
 990                 r = gmc_v7_0_mc_load_microcode(adev);
 991                 if (r) {
 992                         DRM_ERROR("Failed to load MC firmware!\n");
 993                         return r;
 994                 }
 995         }
 996
 997         r = gmc_v7_0_gart_enable(adev);
 998         if (r)
 999                 return r;
1000
1001         return r;
1002 }
1003
1004 static int gmc_v7_0_hw_fini(void *handle)
1005 {
1006         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1007
1008         amdgpu_irq_put(adev, &adev->mc.vm_fault, 0);
1009         gmc_v7_0_gart_disable(adev);
1010
1011         return 0;
1012 }
1013
1014 static int gmc_v7_0_suspend(void *handle)
1015 {
1016         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1017
1018         if (adev->vm_manager.enabled) {
1019                 amdgpu_vm_manager_fini(adev);
1020                 gmc_v7_0_vm_fini(adev);
1021                 adev->vm_manager.enabled = false;
1022         }
1023         gmc_v7_0_hw_fini(adev);
1024
1025         return 0;
1026 }
1027
1028 static int gmc_v7_0_resume(void *handle)
1029 {
1030         int r;
1031         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1032
1033         r = gmc_v7_0_hw_init(adev);
1034         if (r)
1035                 return r;
1036
1037         if (!adev->vm_manager.enabled) {
1038                 r = gmc_v7_0_vm_init(adev);
1039                 if (r) {
1040                         dev_err(adev->dev, "vm manager initialization failed (%d).\n", r);
1041                         return r;
1042                 }
1043                 adev->vm_manager.enabled = true;
1044         }
1045
1046         return r;
1047 }
1048
1049 static bool gmc_v7_0_is_idle(void *handle)
1050 {
1051         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1052         u32 tmp = RREG32(mmSRBM_STATUS);
1053
1054         if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1055                    SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK | SRBM_STATUS__VMC_BUSY_MASK))
1056                 return false;
1057
1058         return true;
1059 }
1060
1061 static int gmc_v7_0_wait_for_idle(void *handle)
1062 {
1063         unsigned i;
1064         u32 tmp;
1065         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1066
1067         for (i = 0; i < adev->usec_timeout; i++) {
1068                 /* read MC_STATUS */
1069                 tmp = RREG32(mmSRBM_STATUS) & (SRBM_STATUS__MCB_BUSY_MASK |
1070                                                SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1071                                                SRBM_STATUS__MCC_BUSY_MASK |
1072                                                SRBM_STATUS__MCD_BUSY_MASK |
1073                                                SRBM_STATUS__VMC_BUSY_MASK);
1074                 if (!tmp)
1075                         return 0;
1076                 udelay(1);
1077         }
1078         return -ETIMEDOUT;
1079
1080 }
1081
1082 static void gmc_v7_0_print_status(void *handle)
1083 {
1084         int i, j;
1085         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1086
1087         dev_info(adev->dev, "GMC 8.x registers\n");
1088         dev_info(adev->dev, "  SRBM_STATUS=0x%08X\n",
1089                 RREG32(mmSRBM_STATUS));
1090         dev_info(adev->dev, "  SRBM_STATUS2=0x%08X\n",
1091                 RREG32(mmSRBM_STATUS2));
1092
1093         dev_info(adev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_ADDR   0x%08X\n",
1094                  RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_ADDR));
1095         dev_info(adev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
1096                  RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_STATUS));
1097         dev_info(adev->dev, "  MC_VM_MX_L1_TLB_CNTL=0x%08X\n",
1098                  RREG32(mmMC_VM_MX_L1_TLB_CNTL));
1099         dev_info(adev->dev, "  VM_L2_CNTL=0x%08X\n",
1100                  RREG32(mmVM_L2_CNTL));
1101         dev_info(adev->dev, "  VM_L2_CNTL2=0x%08X\n",
1102                  RREG32(mmVM_L2_CNTL2));
1103         dev_info(adev->dev, "  VM_L2_CNTL3=0x%08X\n",
1104                  RREG32(mmVM_L2_CNTL3));
1105         dev_info(adev->dev, "  VM_CONTEXT0_PAGE_TABLE_START_ADDR=0x%08X\n",
1106                  RREG32(mmVM_CONTEXT0_PAGE_TABLE_START_ADDR));
1107         dev_info(adev->dev, "  VM_CONTEXT0_PAGE_TABLE_END_ADDR=0x%08X\n",
1108                  RREG32(mmVM_CONTEXT0_PAGE_TABLE_END_ADDR));
1109         dev_info(adev->dev, "  VM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR=0x%08X\n",
1110                  RREG32(mmVM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR));
1111         dev_info(adev->dev, "  VM_CONTEXT0_CNTL2=0x%08X\n",
1112                  RREG32(mmVM_CONTEXT0_CNTL2));
1113         dev_info(adev->dev, "  VM_CONTEXT0_CNTL=0x%08X\n",
1114                  RREG32(mmVM_CONTEXT0_CNTL));
1115         dev_info(adev->dev, "  0x15D4=0x%08X\n",
1116                  RREG32(0x575));
1117         dev_info(adev->dev, "  0x15D8=0x%08X\n",
1118                  RREG32(0x576));
1119         dev_info(adev->dev, "  0x15DC=0x%08X\n",
1120                  RREG32(0x577));
1121         dev_info(adev->dev, "  VM_CONTEXT1_PAGE_TABLE_START_ADDR=0x%08X\n",
1122                  RREG32(mmVM_CONTEXT1_PAGE_TABLE_START_ADDR));
1123         dev_info(adev->dev, "  VM_CONTEXT1_PAGE_TABLE_END_ADDR=0x%08X\n",
1124                  RREG32(mmVM_CONTEXT1_PAGE_TABLE_END_ADDR));
1125         dev_info(adev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR=0x%08X\n",
1126                  RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR));
1127         dev_info(adev->dev, "  VM_CONTEXT1_CNTL2=0x%08X\n",
1128                  RREG32(mmVM_CONTEXT1_CNTL2));
1129         dev_info(adev->dev, "  VM_CONTEXT1_CNTL=0x%08X\n",
1130                  RREG32(mmVM_CONTEXT1_CNTL));
1131         for (i = 0; i < 16; i++) {
1132                 if (i < 8)
1133                         dev_info(adev->dev, "  VM_CONTEXT%d_PAGE_TABLE_BASE_ADDR=0x%08X\n",
1134                                  i, RREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + i));
1135                 else
1136                         dev_info(adev->dev, "  VM_CONTEXT%d_PAGE_TABLE_BASE_ADDR=0x%08X\n",
1137                                  i, RREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + i - 8));
1138         }
1139         dev_info(adev->dev, "  MC_VM_SYSTEM_APERTURE_LOW_ADDR=0x%08X\n",
1140                  RREG32(mmMC_VM_SYSTEM_APERTURE_LOW_ADDR));
1141         dev_info(adev->dev, "  MC_VM_SYSTEM_APERTURE_HIGH_ADDR=0x%08X\n",
1142                  RREG32(mmMC_VM_SYSTEM_APERTURE_HIGH_ADDR));
1143         dev_info(adev->dev, "  MC_VM_SYSTEM_APERTURE_DEFAULT_ADDR=0x%08X\n",
1144                  RREG32(mmMC_VM_SYSTEM_APERTURE_DEFAULT_ADDR));
1145         dev_info(adev->dev, "  MC_VM_FB_LOCATION=0x%08X\n",
1146                  RREG32(mmMC_VM_FB_LOCATION));
1147         dev_info(adev->dev, "  MC_VM_AGP_BASE=0x%08X\n",
1148                  RREG32(mmMC_VM_AGP_BASE));
1149         dev_info(adev->dev, "  MC_VM_AGP_TOP=0x%08X\n",
1150                  RREG32(mmMC_VM_AGP_TOP));
1151         dev_info(adev->dev, "  MC_VM_AGP_BOT=0x%08X\n",
1152                  RREG32(mmMC_VM_AGP_BOT));
1153
1154         if (adev->asic_type == CHIP_KAVERI) {
1155                 dev_info(adev->dev, "  CHUB_CONTROL=0x%08X\n",
1156                          RREG32(mmCHUB_CONTROL));
1157         }
1158
1159         dev_info(adev->dev, "  HDP_REG_COHERENCY_FLUSH_CNTL=0x%08X\n",
1160                  RREG32(mmHDP_REG_COHERENCY_FLUSH_CNTL));
1161         dev_info(adev->dev, "  HDP_NONSURFACE_BASE=0x%08X\n",
1162                  RREG32(mmHDP_NONSURFACE_BASE));
1163         dev_info(adev->dev, "  HDP_NONSURFACE_INFO=0x%08X\n",
1164                  RREG32(mmHDP_NONSURFACE_INFO));
1165         dev_info(adev->dev, "  HDP_NONSURFACE_SIZE=0x%08X\n",
1166                  RREG32(mmHDP_NONSURFACE_SIZE));
1167         dev_info(adev->dev, "  HDP_MISC_CNTL=0x%08X\n",
1168                  RREG32(mmHDP_MISC_CNTL));
1169         dev_info(adev->dev, "  HDP_HOST_PATH_CNTL=0x%08X\n",
1170                  RREG32(mmHDP_HOST_PATH_CNTL));
1171
1172         for (i = 0, j = 0; i < 32; i++, j += 0x6) {
1173                 dev_info(adev->dev, "  %d:\n", i);
1174                 dev_info(adev->dev, "  0x%04X=0x%08X\n",
1175                          0xb05 + j, RREG32(0xb05 + j));
1176                 dev_info(adev->dev, "  0x%04X=0x%08X\n",
1177                          0xb06 + j, RREG32(0xb06 + j));
1178                 dev_info(adev->dev, "  0x%04X=0x%08X\n",
1179                          0xb07 + j, RREG32(0xb07 + j));
1180                 dev_info(adev->dev, "  0x%04X=0x%08X\n",
1181                          0xb08 + j, RREG32(0xb08 + j));
1182                 dev_info(adev->dev, "  0x%04X=0x%08X\n",
1183                          0xb09 + j, RREG32(0xb09 + j));
1184         }
1185
1186         dev_info(adev->dev, "  BIF_FB_EN=0x%08X\n",
1187                  RREG32(mmBIF_FB_EN));
1188 }
1189
1190 static int gmc_v7_0_soft_reset(void *handle)
1191 {
1192         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1193         struct amdgpu_mode_mc_save save;
1194         u32 srbm_soft_reset = 0;
1195         u32 tmp = RREG32(mmSRBM_STATUS);
1196
1197         if (tmp & SRBM_STATUS__VMC_BUSY_MASK)
1198                 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset,
1199                                                 SRBM_SOFT_RESET, SOFT_RESET_VMC, 1);
1200
1201         if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1202                    SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK)) {
1203                 if (!(adev->flags & AMD_IS_APU))
1204                         srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset,
1205                                                         SRBM_SOFT_RESET, SOFT_RESET_MC, 1);
1206         }
1207
1208         if (srbm_soft_reset) {
1209                 gmc_v7_0_print_status((void *)adev);
1210
1211                 gmc_v7_0_mc_stop(adev, &save);
1212                 if (gmc_v7_0_wait_for_idle(adev)) {
1213                         dev_warn(adev->dev, "Wait for GMC idle timed out !\n");
1214                 }
1215
1216
1217                 tmp = RREG32(mmSRBM_SOFT_RESET);
1218                 tmp |= srbm_soft_reset;
1219                 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1220                 WREG32(mmSRBM_SOFT_RESET, tmp);
1221                 tmp = RREG32(mmSRBM_SOFT_RESET);
1222
1223                 udelay(50);
1224
1225                 tmp &= ~srbm_soft_reset;
1226                 WREG32(mmSRBM_SOFT_RESET, tmp);
1227                 tmp = RREG32(mmSRBM_SOFT_RESET);
1228
1229                 /* Wait a little for things to settle down */
1230                 udelay(50);
1231
1232                 gmc_v7_0_mc_resume(adev, &save);
1233                 udelay(50);
1234
1235                 gmc_v7_0_print_status((void *)adev);
1236         }
1237
1238         return 0;
1239 }
1240
1241 static int gmc_v7_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
1242                                              struct amdgpu_irq_src *src,
1243                                              unsigned type,
1244                                              enum amdgpu_interrupt_state state)
1245 {
1246         u32 tmp;
1247         u32 bits = (VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1248                     VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1249                     VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1250                     VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1251                     VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1252                     VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK);
1253
1254         switch (state) {
1255         case AMDGPU_IRQ_STATE_DISABLE:
1256                 /* system context */
1257                 tmp = RREG32(mmVM_CONTEXT0_CNTL);
1258                 tmp &= ~bits;
1259                 WREG32(mmVM_CONTEXT0_CNTL, tmp);
1260                 /* VMs */
1261                 tmp = RREG32(mmVM_CONTEXT1_CNTL);
1262                 tmp &= ~bits;
1263                 WREG32(mmVM_CONTEXT1_CNTL, tmp);
1264                 break;
1265         case AMDGPU_IRQ_STATE_ENABLE:
1266                 /* system context */
1267                 tmp = RREG32(mmVM_CONTEXT0_CNTL);
1268                 tmp |= bits;
1269                 WREG32(mmVM_CONTEXT0_CNTL, tmp);
1270                 /* VMs */
1271                 tmp = RREG32(mmVM_CONTEXT1_CNTL);
1272                 tmp |= bits;
1273                 WREG32(mmVM_CONTEXT1_CNTL, tmp);
1274                 break;
1275         default:
1276                 break;
1277         }
1278
1279         return 0;
1280 }
1281
1282 static int gmc_v7_0_process_interrupt(struct amdgpu_device *adev,
1283                                       struct amdgpu_irq_src *source,
1284                                       struct amdgpu_iv_entry *entry)
1285 {
1286         u32 addr, status, mc_client;
1287
1288         addr = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_ADDR);
1289         status = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_STATUS);
1290         mc_client = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_MCCLIENT);
1291         /* reset addr and status */
1292         WREG32_P(mmVM_CONTEXT1_CNTL2, 1, ~1);
1293
1294         if (!addr && !status)
1295                 return 0;
1296
1297         if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_FIRST)
1298                 gmc_v7_0_set_fault_enable_default(adev, false);
1299
1300         dev_err(adev->dev, "GPU fault detected: %d 0x%08x\n",
1301                 entry->src_id, entry->src_data);
1302         dev_err(adev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_ADDR   0x%08X\n",
1303                 addr);
1304         dev_err(adev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
1305                 status);
1306         gmc_v7_0_vm_decode_fault(adev, status, addr, mc_client);
1307
1308         return 0;
1309 }
1310
1311 static int gmc_v7_0_set_clockgating_state(void *handle,
1312                                           enum amd_clockgating_state state)
1313 {
1314         bool gate = false;
1315         struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1316
1317         if (state == AMD_CG_STATE_GATE)
1318                 gate = true;
1319
1320         if (!(adev->flags & AMD_IS_APU)) {
1321                 gmc_v7_0_enable_mc_mgcg(adev, gate);
1322                 gmc_v7_0_enable_mc_ls(adev, gate);
1323         }
1324         gmc_v7_0_enable_bif_mgls(adev, gate);
1325         gmc_v7_0_enable_hdp_mgcg(adev, gate);
1326         gmc_v7_0_enable_hdp_ls(adev, gate);
1327
1328         return 0;
1329 }
1330
1331 static int gmc_v7_0_set_powergating_state(void *handle,
1332                                           enum amd_powergating_state state)
1333 {
1334         return 0;
1335 }
1336
1337 const struct amd_ip_funcs gmc_v7_0_ip_funcs = {
1338         .early_init = gmc_v7_0_early_init,
1339         .late_init = gmc_v7_0_late_init,
1340         .sw_init = gmc_v7_0_sw_init,
1341         .sw_fini = gmc_v7_0_sw_fini,
1342         .hw_init = gmc_v7_0_hw_init,
1343         .hw_fini = gmc_v7_0_hw_fini,
1344         .suspend = gmc_v7_0_suspend,
1345         .resume = gmc_v7_0_resume,
1346         .is_idle = gmc_v7_0_is_idle,
1347         .wait_for_idle = gmc_v7_0_wait_for_idle,
1348         .soft_reset = gmc_v7_0_soft_reset,
1349         .print_status = gmc_v7_0_print_status,
1350         .set_clockgating_state = gmc_v7_0_set_clockgating_state,
1351         .set_powergating_state = gmc_v7_0_set_powergating_state,
1352 };
1353
1354 static const struct amdgpu_gart_funcs gmc_v7_0_gart_funcs = {
1355         .flush_gpu_tlb = gmc_v7_0_gart_flush_gpu_tlb,
1356         .set_pte_pde = gmc_v7_0_gart_set_pte_pde,
1357 };
1358
1359 static const struct amdgpu_irq_src_funcs gmc_v7_0_irq_funcs = {
1360         .set = gmc_v7_0_vm_fault_interrupt_state,
1361         .process = gmc_v7_0_process_interrupt,
1362 };
1363
1364 static void gmc_v7_0_set_gart_funcs(struct amdgpu_device *adev)
1365 {
1366         if (adev->gart.gart_funcs == NULL)
1367                 adev->gart.gart_funcs = &gmc_v7_0_gart_funcs;
1368 }
1369
1370 static void gmc_v7_0_set_irq_funcs(struct amdgpu_device *adev)
1371 {
1372         adev->mc.vm_fault.num_types = 1;
1373         adev->mc.vm_fault.funcs = &gmc_v7_0_irq_funcs;
1374 }