drivers/gpu/drm/etnaviv/etnaviv_gpu.c

   1 /*
   2  * Copyright (C) 2015 Etnaviv Project
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms of the GNU General Public License version 2 as published by
   6  * the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  11  * more details.
  12  *
  13  * You should have received a copy of the GNU General Public License along with
  14  * this program.  If not, see <http://www.gnu.org/licenses/>.
  15  */
  16
  17 #include <linux/component.h>
  18 #include <linux/fence.h>
  19 #include <linux/moduleparam.h>
  20 #include <linux/of_device.h>
  21 #include "etnaviv_dump.h"
  22 #include "etnaviv_gpu.h"
  23 #include "etnaviv_gem.h"
  24 #include "etnaviv_mmu.h"
  25 #include "etnaviv_iommu.h"
  26 #include "etnaviv_iommu_v2.h"
  27 #include "common.xml.h"
  28 #include "state.xml.h"
  29 #include "state_hi.xml.h"
  30 #include "cmdstream.xml.h"
  31
  32 static const struct platform_device_id gpu_ids[] = {
  33         { .name = "etnaviv-gpu,2d" },
  34         { },
  35 };
  36
  37 static bool etnaviv_dump_core = true;
  38 module_param_named(dump_core, etnaviv_dump_core, bool, 0600);
  39
  40 /*
  41  * Driver functions:
  42  */
  43
  44 int etnaviv_gpu_get_param(struct etnaviv_gpu *gpu, u32 param, u64 *value)
  45 {
  46         switch (param) {
  47         case ETNAVIV_PARAM_GPU_MODEL:
  48                 *value = gpu->identity.model;
  49                 break;
  50
  51         case ETNAVIV_PARAM_GPU_REVISION:
  52                 *value = gpu->identity.revision;
  53                 break;
  54
  55         case ETNAVIV_PARAM_GPU_FEATURES_0:
  56                 *value = gpu->identity.features;
  57                 break;
  58
  59         case ETNAVIV_PARAM_GPU_FEATURES_1:
  60                 *value = gpu->identity.minor_features0;
  61                 break;
  62
  63         case ETNAVIV_PARAM_GPU_FEATURES_2:
  64                 *value = gpu->identity.minor_features1;
  65                 break;
  66
  67         case ETNAVIV_PARAM_GPU_FEATURES_3:
  68                 *value = gpu->identity.minor_features2;
  69                 break;
  70
  71         case ETNAVIV_PARAM_GPU_FEATURES_4:
  72                 *value = gpu->identity.minor_features3;
  73                 break;
  74
  75         case ETNAVIV_PARAM_GPU_FEATURES_5:
  76                 *value = gpu->identity.minor_features4;
  77                 break;
  78
  79         case ETNAVIV_PARAM_GPU_FEATURES_6:
  80                 *value = gpu->identity.minor_features5;
  81                 break;
  82
  83         case ETNAVIV_PARAM_GPU_STREAM_COUNT:
  84                 *value = gpu->identity.stream_count;
  85                 break;
  86
  87         case ETNAVIV_PARAM_GPU_REGISTER_MAX:
  88                 *value = gpu->identity.register_max;
  89                 break;
  90
  91         case ETNAVIV_PARAM_GPU_THREAD_COUNT:
  92                 *value = gpu->identity.thread_count;
  93                 break;
  94
  95         case ETNAVIV_PARAM_GPU_VERTEX_CACHE_SIZE:
  96                 *value = gpu->identity.vertex_cache_size;
  97                 break;
  98
  99         case ETNAVIV_PARAM_GPU_SHADER_CORE_COUNT:
 100                 *value = gpu->identity.shader_core_count;
 101                 break;
 102
 103         case ETNAVIV_PARAM_GPU_PIXEL_PIPES:
 104                 *value = gpu->identity.pixel_pipes;
 105                 break;
 106
 107         case ETNAVIV_PARAM_GPU_VERTEX_OUTPUT_BUFFER_SIZE:
 108                 *value = gpu->identity.vertex_output_buffer_size;
 109                 break;
 110
 111         case ETNAVIV_PARAM_GPU_BUFFER_SIZE:
 112                 *value = gpu->identity.buffer_size;
 113                 break;
 114
 115         case ETNAVIV_PARAM_GPU_INSTRUCTION_COUNT:
 116                 *value = gpu->identity.instruction_count;
 117                 break;
 118
 119         case ETNAVIV_PARAM_GPU_NUM_CONSTANTS:
 120                 *value = gpu->identity.num_constants;
 121                 break;
 122
 123         case ETNAVIV_PARAM_GPU_NUM_VARYINGS:
 124                 *value = gpu->identity.varyings_count;
 125                 break;
 126
 127         default:
 128                 DBG("%s: invalid param: %u", dev_name(gpu->dev), param);
 129                 return -EINVAL;
 130         }
 131
 132         return 0;
 133 }
 134
 135
 136 #define etnaviv_is_model_rev(gpu, mod, rev) \
 137         ((gpu)->identity.model == chipModel_##mod && \
 138          (gpu)->identity.revision == rev)
 139 #define etnaviv_field(val, field) \
 140         (((val) & field##__MASK) >> field##__SHIFT)
 141
 142 static void etnaviv_hw_specs(struct etnaviv_gpu *gpu)
 143 {
 144         if (gpu->identity.minor_features0 &
 145             chipMinorFeatures0_MORE_MINOR_FEATURES) {
 146                 u32 specs[4];
 147                 unsigned int streams;
 148
 149                 specs[0] = gpu_read(gpu, VIVS_HI_CHIP_SPECS);
 150                 specs[1] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_2);
 151                 specs[2] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_3);
 152                 specs[3] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_4);
 153
 154                 gpu->identity.stream_count = etnaviv_field(specs[0],
 155                                         VIVS_HI_CHIP_SPECS_STREAM_COUNT);
 156                 gpu->identity.register_max = etnaviv_field(specs[0],
 157                                         VIVS_HI_CHIP_SPECS_REGISTER_MAX);
 158                 gpu->identity.thread_count = etnaviv_field(specs[0],
 159                                         VIVS_HI_CHIP_SPECS_THREAD_COUNT);
 160                 gpu->identity.vertex_cache_size = etnaviv_field(specs[0],
 161                                         VIVS_HI_CHIP_SPECS_VERTEX_CACHE_SIZE);
 162                 gpu->identity.shader_core_count = etnaviv_field(specs[0],
 163                                         VIVS_HI_CHIP_SPECS_SHADER_CORE_COUNT);
 164                 gpu->identity.pixel_pipes = etnaviv_field(specs[0],
 165                                         VIVS_HI_CHIP_SPECS_PIXEL_PIPES);
 166                 gpu->identity.vertex_output_buffer_size =
 167                         etnaviv_field(specs[0],
 168                                 VIVS_HI_CHIP_SPECS_VERTEX_OUTPUT_BUFFER_SIZE);
 169
 170                 gpu->identity.buffer_size = etnaviv_field(specs[1],
 171                                         VIVS_HI_CHIP_SPECS_2_BUFFER_SIZE);
 172                 gpu->identity.instruction_count = etnaviv_field(specs[1],
 173                                         VIVS_HI_CHIP_SPECS_2_INSTRUCTION_COUNT);
 174                 gpu->identity.num_constants = etnaviv_field(specs[1],
 175                                         VIVS_HI_CHIP_SPECS_2_NUM_CONSTANTS);
 176
 177                 gpu->identity.varyings_count = etnaviv_field(specs[2],
 178                                         VIVS_HI_CHIP_SPECS_3_VARYINGS_COUNT);
 179
 180                 /* This overrides the value from older register if non-zero */
 181                 streams = etnaviv_field(specs[3],
 182                                         VIVS_HI_CHIP_SPECS_4_STREAM_COUNT);
 183                 if (streams)
 184                         gpu->identity.stream_count = streams;
 185         }
 186
 187         /* Fill in the stream count if not specified */
 188         if (gpu->identity.stream_count == 0) {
 189                 if (gpu->identity.model >= 0x1000)
 190                         gpu->identity.stream_count = 4;
 191                 else
 192                         gpu->identity.stream_count = 1;
 193         }
 194
 195         /* Convert the register max value */
 196         if (gpu->identity.register_max)
 197                 gpu->identity.register_max = 1 << gpu->identity.register_max;
 198         else if (gpu->identity.model == chipModel_GC400)
 199                 gpu->identity.register_max = 32;
 200         else
 201                 gpu->identity.register_max = 64;
 202
 203         /* Convert thread count */
 204         if (gpu->identity.thread_count)
 205                 gpu->identity.thread_count = 1 << gpu->identity.thread_count;
 206         else if (gpu->identity.model == chipModel_GC400)
 207                 gpu->identity.thread_count = 64;
 208         else if (gpu->identity.model == chipModel_GC500 ||
 209                  gpu->identity.model == chipModel_GC530)
 210                 gpu->identity.thread_count = 128;
 211         else
 212                 gpu->identity.thread_count = 256;
 213
 214         if (gpu->identity.vertex_cache_size == 0)
 215                 gpu->identity.vertex_cache_size = 8;
 216
 217         if (gpu->identity.shader_core_count == 0) {
 218                 if (gpu->identity.model >= 0x1000)
 219                         gpu->identity.shader_core_count = 2;
 220                 else
 221                         gpu->identity.shader_core_count = 1;
 222         }
 223
 224         if (gpu->identity.pixel_pipes == 0)
 225                 gpu->identity.pixel_pipes = 1;
 226
 227         /* Convert virtex buffer size */
 228         if (gpu->identity.vertex_output_buffer_size) {
 229                 gpu->identity.vertex_output_buffer_size =
 230                         1 << gpu->identity.vertex_output_buffer_size;
 231         } else if (gpu->identity.model == chipModel_GC400) {
 232                 if (gpu->identity.revision < 0x4000)
 233                         gpu->identity.vertex_output_buffer_size = 512;
 234                 else if (gpu->identity.revision < 0x4200)
 235                         gpu->identity.vertex_output_buffer_size = 256;
 236                 else
 237                         gpu->identity.vertex_output_buffer_size = 128;
 238         } else {
 239                 gpu->identity.vertex_output_buffer_size = 512;
 240         }
 241
 242         switch (gpu->identity.instruction_count) {
 243         case 0:
 244                 if (etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
 245                     gpu->identity.model == chipModel_GC880)
 246                         gpu->identity.instruction_count = 512;
 247                 else
 248                         gpu->identity.instruction_count = 256;
 249                 break;
 250
 251         case 1:
 252                 gpu->identity.instruction_count = 1024;
 253                 break;
 254
 255         case 2:
 256                 gpu->identity.instruction_count = 2048;
 257                 break;
 258
 259         default:
 260                 gpu->identity.instruction_count = 256;
 261                 break;
 262         }
 263
 264         if (gpu->identity.num_constants == 0)
 265                 gpu->identity.num_constants = 168;
 266
 267         if (gpu->identity.varyings_count == 0) {
 268                 if (gpu->identity.minor_features1 & chipMinorFeatures1_HALTI0)
 269                         gpu->identity.varyings_count = 12;
 270                 else
 271                         gpu->identity.varyings_count = 8;
 272         }
 273
 274         /*
 275          * For some cores, two varyings are consumed for position, so the
 276          * maximum varying count needs to be reduced by one.
 277          */
 278         if (etnaviv_is_model_rev(gpu, GC5000, 0x5434) ||
 279             etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
 280             etnaviv_is_model_rev(gpu, GC4000, 0x5245) ||
 281             etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
 282             etnaviv_is_model_rev(gpu, GC3000, 0x5435) ||
 283             etnaviv_is_model_rev(gpu, GC2200, 0x5244) ||
 284             etnaviv_is_model_rev(gpu, GC2100, 0x5108) ||
 285             etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
 286             etnaviv_is_model_rev(gpu, GC1500, 0x5246) ||
 287             etnaviv_is_model_rev(gpu, GC880, 0x5107) ||
 288             etnaviv_is_model_rev(gpu, GC880, 0x5106))
 289                 gpu->identity.varyings_count -= 1;
 290 }
 291
 292 static void etnaviv_hw_identify(struct etnaviv_gpu *gpu)
 293 {
 294         u32 chipIdentity;
 295
 296         chipIdentity = gpu_read(gpu, VIVS_HI_CHIP_IDENTITY);
 297
 298         /* Special case for older graphic cores. */
 299         if (etnaviv_field(chipIdentity, VIVS_HI_CHIP_IDENTITY_FAMILY) == 0x01) {
 300                 gpu->identity.model    = chipModel_GC500;
 301                 gpu->identity.revision = etnaviv_field(chipIdentity,
 302                                          VIVS_HI_CHIP_IDENTITY_REVISION);
 303         } else {
 304
 305                 gpu->identity.model = gpu_read(gpu, VIVS_HI_CHIP_MODEL);
 306                 gpu->identity.revision = gpu_read(gpu, VIVS_HI_CHIP_REV);
 307
 308                 /*
 309                  * !!!! HACK ALERT !!!!
 310                  * Because people change device IDs without letting software
 311                  * know about it - here is the hack to make it all look the
 312                  * same.  Only for GC400 family.
 313                  */
 314                 if ((gpu->identity.model & 0xff00) == 0x0400 &&
 315                     gpu->identity.model != chipModel_GC420) {
 316                         gpu->identity.model = gpu->identity.model & 0x0400;
 317                 }
 318
 319                 /* Another special case */
 320                 if (etnaviv_is_model_rev(gpu, GC300, 0x2201)) {
 321                         u32 chipDate = gpu_read(gpu, VIVS_HI_CHIP_DATE);
 322                         u32 chipTime = gpu_read(gpu, VIVS_HI_CHIP_TIME);
 323
 324                         if (chipDate == 0x20080814 && chipTime == 0x12051100) {
 325                                 /*
 326                                  * This IP has an ECO; put the correct
 327                                  * revision in it.
 328                                  */
 329                                 gpu->identity.revision = 0x1051;
 330                         }
 331                 }
 332         }
 333
 334         dev_info(gpu->dev, "model: GC%x, revision: %x\n",
 335                  gpu->identity.model, gpu->identity.revision);
 336
 337         gpu->identity.features = gpu_read(gpu, VIVS_HI_CHIP_FEATURE);
 338
 339         /* Disable fast clear on GC700. */
 340         if (gpu->identity.model == chipModel_GC700)
 341                 gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
 342
 343         if ((gpu->identity.model == chipModel_GC500 &&
 344              gpu->identity.revision < 2) ||
 345             (gpu->identity.model == chipModel_GC300 &&
 346              gpu->identity.revision < 0x2000)) {
 347
 348                 /*
 349                  * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these
 350                  * registers.
 351                  */
 352                 gpu->identity.minor_features0 = 0;
 353                 gpu->identity.minor_features1 = 0;
 354                 gpu->identity.minor_features2 = 0;
 355                 gpu->identity.minor_features3 = 0;
 356                 gpu->identity.minor_features4 = 0;
 357                 gpu->identity.minor_features5 = 0;
 358         } else
 359                 gpu->identity.minor_features0 =
 360                                 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0);
 361
 362         if (gpu->identity.minor_features0 &
 363             chipMinorFeatures0_MORE_MINOR_FEATURES) {
 364                 gpu->identity.minor_features1 =
 365                                 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1);
 366                 gpu->identity.minor_features2 =
 367                                 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2);
 368                 gpu->identity.minor_features3 =
 369                                 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3);
 370                 gpu->identity.minor_features4 =
 371                                 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_4);
 372                 gpu->identity.minor_features5 =
 373                                 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_5);
 374         }
 375
 376         /* GC600 idle register reports zero bits where modules aren't present */
 377         if (gpu->identity.model == chipModel_GC600) {
 378                 gpu->idle_mask = VIVS_HI_IDLE_STATE_TX |
 379                                  VIVS_HI_IDLE_STATE_RA |
 380                                  VIVS_HI_IDLE_STATE_SE |
 381                                  VIVS_HI_IDLE_STATE_PA |
 382                                  VIVS_HI_IDLE_STATE_SH |
 383                                  VIVS_HI_IDLE_STATE_PE |
 384                                  VIVS_HI_IDLE_STATE_DE |
 385                                  VIVS_HI_IDLE_STATE_FE;
 386         } else {
 387                 gpu->idle_mask = ~VIVS_HI_IDLE_STATE_AXI_LP;
 388         }
 389
 390         etnaviv_hw_specs(gpu);
 391 }
 392
 393 static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock)
 394 {
 395         gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock |
 396                   VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD);
 397         gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock);
 398 }
 399
 400 static int etnaviv_hw_reset(struct etnaviv_gpu *gpu)
 401 {
 402         u32 control, idle;
 403         unsigned long timeout;
 404         bool failed = true;
 405
 406         /* TODO
 407          *
 408          * - clock gating
 409          * - puls eater
 410          * - what about VG?
 411          */
 412
 413         /* We hope that the GPU resets in under one second */
 414         timeout = jiffies + msecs_to_jiffies(1000);
 415
 416         while (time_is_after_jiffies(timeout)) {
 417                 control = VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS |
 418                           VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(0x40);
 419
 420                 /* enable clock */
 421                 etnaviv_gpu_load_clock(gpu, control);
 422
 423                 /* Wait for stable clock.  Vivante's code waited for 1ms */
 424                 usleep_range(1000, 10000);
 425
 426                 /* isolate the GPU. */
 427                 control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
 428                 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
 429
 430                 /* set soft reset. */
 431                 control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
 432                 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
 433
 434                 /* wait for reset. */
 435                 msleep(1);
 436
 437                 /* reset soft reset bit. */
 438                 control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
 439                 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
 440
 441                 /* reset GPU isolation. */
 442                 control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
 443                 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
 444
 445                 /* read idle register. */
 446                 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
 447
 448                 /* try reseting again if FE it not idle */
 449                 if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) {
 450                         dev_dbg(gpu->dev, "FE is not idle\n");
 451                         continue;
 452                 }
 453
 454                 /* read reset register. */
 455                 control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
 456
 457                 /* is the GPU idle? */
 458                 if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) ||
 459                     ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) {
 460                         dev_dbg(gpu->dev, "GPU is not idle\n");
 461                         continue;
 462                 }
 463
 464                 failed = false;
 465                 break;
 466         }
 467
 468         if (failed) {
 469                 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
 470                 control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
 471
 472                 dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n",
 473                         idle & VIVS_HI_IDLE_STATE_FE ? "" : "not ",
 474                         control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not ",
 475                         control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not ");
 476
 477                 return -EBUSY;
 478         }
 479
 480         /* We rely on the GPU running, so program the clock */
 481         control = VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS |
 482                   VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(0x40);
 483
 484         /* enable clock */
 485         etnaviv_gpu_load_clock(gpu, control);
 486
 487         return 0;
 488 }
 489
 490 static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
 491 {
 492         u16 prefetch;
 493
 494         if ((etnaviv_is_model_rev(gpu, GC320, 0x5007) ||
 495              etnaviv_is_model_rev(gpu, GC320, 0x5220)) &&
 496             gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) {
 497                 u32 mc_memory_debug;
 498
 499                 mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff;
 500
 501                 if (gpu->identity.revision == 0x5007)
 502                         mc_memory_debug |= 0x0c;
 503                 else
 504                         mc_memory_debug |= 0x08;
 505
 506                 gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug);
 507         }
 508
 509         /*
 510          * Update GPU AXI cache atttribute to "cacheable, no allocate".
 511          * This is necessary to prevent the iMX6 SoC locking up.
 512          */
 513         gpu_write(gpu, VIVS_HI_AXI_CONFIG,
 514                   VIVS_HI_AXI_CONFIG_AWCACHE(2) |
 515                   VIVS_HI_AXI_CONFIG_ARCACHE(2));
 516
 517         /* GC2000 rev 5108 needs a special bus config */
 518         if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) {
 519                 u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG);
 520                 bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK |
 521                                 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK);
 522                 bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
 523                               VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
 524                 gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config);
 525         }
 526
 527         /* set base addresses */
 528         gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_RA, gpu->memory_base);
 529         gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_FE, gpu->memory_base);
 530         gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_TX, gpu->memory_base);
 531         gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_PEZ, gpu->memory_base);
 532         gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_PE, gpu->memory_base);
 533
 534         /* setup the MMU page table pointers */
 535         etnaviv_iommu_domain_restore(gpu, gpu->mmu->domain);
 536
 537         /* Start command processor */
 538         prefetch = etnaviv_buffer_init(gpu);
 539
 540         gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
 541         gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS,
 542                   gpu->buffer->paddr - gpu->memory_base);
 543         gpu_write(gpu, VIVS_FE_COMMAND_CONTROL,
 544                   VIVS_FE_COMMAND_CONTROL_ENABLE |
 545                   VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch));
 546 }
 547
 548 int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
 549 {
 550         int ret, i;
 551         struct iommu_domain *iommu;
 552         enum etnaviv_iommu_version version;
 553         bool mmuv2;
 554
 555         ret = pm_runtime_get_sync(gpu->dev);
 556         if (ret < 0)
 557                 return ret;
 558
 559         etnaviv_hw_identify(gpu);
 560
 561         if (gpu->identity.model == 0) {
 562                 dev_err(gpu->dev, "Unknown GPU model\n");
 563                 ret = -ENXIO;
 564                 goto fail;
 565         }
 566
 567         /* Exclude VG cores with FE2.0 */
 568         if (gpu->identity.features & chipFeatures_PIPE_VG &&
 569             gpu->identity.features & chipFeatures_FE20) {
 570                 dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n");
 571                 ret = -ENXIO;
 572                 goto fail;
 573         }
 574
 575         ret = etnaviv_hw_reset(gpu);
 576         if (ret)
 577                 goto fail;
 578
 579         /* Setup IOMMU.. eventually we will (I think) do this once per context
 580          * and have separate page tables per context.  For now, to keep things
 581          * simple and to get something working, just use a single address space:
 582          */
 583         mmuv2 = gpu->identity.minor_features1 & chipMinorFeatures1_MMU_VERSION;
 584         dev_dbg(gpu->dev, "mmuv2: %d\n", mmuv2);
 585
 586         if (!mmuv2) {
 587                 iommu = etnaviv_iommu_domain_alloc(gpu);
 588                 version = ETNAVIV_IOMMU_V1;
 589         } else {
 590                 iommu = etnaviv_iommu_v2_domain_alloc(gpu);
 591                 version = ETNAVIV_IOMMU_V2;
 592         }
 593
 594         if (!iommu) {
 595                 ret = -ENOMEM;
 596                 goto fail;
 597         }
 598
 599         gpu->mmu = etnaviv_iommu_new(gpu, iommu, version);
 600         if (!gpu->mmu) {
 601                 iommu_domain_free(iommu);
 602                 ret = -ENOMEM;
 603                 goto fail;
 604         }
 605
 606         /* Create buffer: */
 607         gpu->buffer = etnaviv_gpu_cmdbuf_new(gpu, PAGE_SIZE, 0);
 608         if (!gpu->buffer) {
 609                 ret = -ENOMEM;
 610                 dev_err(gpu->dev, "could not create command buffer\n");
 611                 goto destroy_iommu;
 612         }
 613         if (gpu->buffer->paddr - gpu->memory_base > 0x80000000) {
 614                 ret = -EINVAL;
 615                 dev_err(gpu->dev,
 616                         "command buffer outside valid memory window\n");
 617                 goto free_buffer;
 618         }
 619
 620         /* Setup event management */
 621         spin_lock_init(&gpu->event_spinlock);
 622         init_completion(&gpu->event_free);
 623         for (i = 0; i < ARRAY_SIZE(gpu->event); i++) {
 624                 gpu->event[i].used = false;
 625                 complete(&gpu->event_free);
 626         }
 627
 628         /* Now program the hardware */
 629         mutex_lock(&gpu->lock);
 630         etnaviv_gpu_hw_init(gpu);
 631         gpu->exec_state = -1;
 632         mutex_unlock(&gpu->lock);
 633
 634         pm_runtime_mark_last_busy(gpu->dev);
 635         pm_runtime_put_autosuspend(gpu->dev);
 636
 637         return 0;
 638
 639 free_buffer:
 640         etnaviv_gpu_cmdbuf_free(gpu->buffer);
 641         gpu->buffer = NULL;
 642 destroy_iommu:
 643         etnaviv_iommu_destroy(gpu->mmu);
 644         gpu->mmu = NULL;
 645 fail:
 646         pm_runtime_mark_last_busy(gpu->dev);
 647         pm_runtime_put_autosuspend(gpu->dev);
 648
 649         return ret;
 650 }
 651
 652 #ifdef CONFIG_DEBUG_FS
 653 struct dma_debug {
 654         u32 address[2];
 655         u32 state[2];
 656 };
 657
 658 static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug)
 659 {
 660         u32 i;
 661
 662         debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
 663         debug->state[0]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
 664
 665         for (i = 0; i < 500; i++) {
 666                 debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
 667                 debug->state[1]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
 668
 669                 if (debug->address[0] != debug->address[1])
 670                         break;
 671
 672                 if (debug->state[0] != debug->state[1])
 673                         break;
 674         }
 675 }
 676
 677 int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m)
 678 {
 679         struct dma_debug debug;
 680         u32 dma_lo, dma_hi, axi, idle;
 681         int ret;
 682
 683         seq_printf(m, "%s Status:\n", dev_name(gpu->dev));
 684
 685         ret = pm_runtime_get_sync(gpu->dev);
 686         if (ret < 0)
 687                 return ret;
 688
 689         dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW);
 690         dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH);
 691         axi = gpu_read(gpu, VIVS_HI_AXI_STATUS);
 692         idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
 693
 694         verify_dma(gpu, &debug);
 695
 696         seq_puts(m, "\tfeatures\n");
 697         seq_printf(m, "\t minor_features0: 0x%08x\n",
 698                    gpu->identity.minor_features0);
 699         seq_printf(m, "\t minor_features1: 0x%08x\n",
 700                    gpu->identity.minor_features1);
 701         seq_printf(m, "\t minor_features2: 0x%08x\n",
 702                    gpu->identity.minor_features2);
 703         seq_printf(m, "\t minor_features3: 0x%08x\n",
 704                    gpu->identity.minor_features3);
 705         seq_printf(m, "\t minor_features4: 0x%08x\n",
 706                    gpu->identity.minor_features4);
 707         seq_printf(m, "\t minor_features5: 0x%08x\n",
 708                    gpu->identity.minor_features5);
 709
 710         seq_puts(m, "\tspecs\n");
 711         seq_printf(m, "\t stream_count:  %d\n",
 712                         gpu->identity.stream_count);
 713         seq_printf(m, "\t register_max: %d\n",
 714                         gpu->identity.register_max);
 715         seq_printf(m, "\t thread_count: %d\n",
 716                         gpu->identity.thread_count);
 717         seq_printf(m, "\t vertex_cache_size: %d\n",
 718                         gpu->identity.vertex_cache_size);
 719         seq_printf(m, "\t shader_core_count: %d\n",
 720                         gpu->identity.shader_core_count);
 721         seq_printf(m, "\t pixel_pipes: %d\n",
 722                         gpu->identity.pixel_pipes);
 723         seq_printf(m, "\t vertex_output_buffer_size: %d\n",
 724                         gpu->identity.vertex_output_buffer_size);
 725         seq_printf(m, "\t buffer_size: %d\n",
 726                         gpu->identity.buffer_size);
 727         seq_printf(m, "\t instruction_count: %d\n",
 728                         gpu->identity.instruction_count);
 729         seq_printf(m, "\t num_constants: %d\n",
 730                         gpu->identity.num_constants);
 731         seq_printf(m, "\t varyings_count: %d\n",
 732                         gpu->identity.varyings_count);
 733
 734         seq_printf(m, "\taxi: 0x%08x\n", axi);
 735         seq_printf(m, "\tidle: 0x%08x\n", idle);
 736         idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP;
 737         if ((idle & VIVS_HI_IDLE_STATE_FE) == 0)
 738                 seq_puts(m, "\t FE is not idle\n");
 739         if ((idle & VIVS_HI_IDLE_STATE_DE) == 0)
 740                 seq_puts(m, "\t DE is not idle\n");
 741         if ((idle & VIVS_HI_IDLE_STATE_PE) == 0)
 742                 seq_puts(m, "\t PE is not idle\n");
 743         if ((idle & VIVS_HI_IDLE_STATE_SH) == 0)
 744                 seq_puts(m, "\t SH is not idle\n");
 745         if ((idle & VIVS_HI_IDLE_STATE_PA) == 0)
 746                 seq_puts(m, "\t PA is not idle\n");
 747         if ((idle & VIVS_HI_IDLE_STATE_SE) == 0)
 748                 seq_puts(m, "\t SE is not idle\n");
 749         if ((idle & VIVS_HI_IDLE_STATE_RA) == 0)
 750                 seq_puts(m, "\t RA is not idle\n");
 751         if ((idle & VIVS_HI_IDLE_STATE_TX) == 0)
 752                 seq_puts(m, "\t TX is not idle\n");
 753         if ((idle & VIVS_HI_IDLE_STATE_VG) == 0)
 754                 seq_puts(m, "\t VG is not idle\n");
 755         if ((idle & VIVS_HI_IDLE_STATE_IM) == 0)
 756                 seq_puts(m, "\t IM is not idle\n");
 757         if ((idle & VIVS_HI_IDLE_STATE_FP) == 0)
 758                 seq_puts(m, "\t FP is not idle\n");
 759         if ((idle & VIVS_HI_IDLE_STATE_TS) == 0)
 760                 seq_puts(m, "\t TS is not idle\n");
 761         if (idle & VIVS_HI_IDLE_STATE_AXI_LP)
 762                 seq_puts(m, "\t AXI low power mode\n");
 763
 764         if (gpu->identity.features & chipFeatures_DEBUG_MODE) {
 765                 u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0);
 766                 u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1);
 767                 u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE);
 768
 769                 seq_puts(m, "\tMC\n");
 770                 seq_printf(m, "\t read0: 0x%08x\n", read0);
 771                 seq_printf(m, "\t read1: 0x%08x\n", read1);
 772                 seq_printf(m, "\t write: 0x%08x\n", write);
 773         }
 774
 775         seq_puts(m, "\tDMA ");
 776
 777         if (debug.address[0] == debug.address[1] &&
 778             debug.state[0] == debug.state[1]) {
 779                 seq_puts(m, "seems to be stuck\n");
 780         } else if (debug.address[0] == debug.address[1]) {
 781                 seq_puts(m, "adress is constant\n");
 782         } else {
 783                 seq_puts(m, "is runing\n");
 784         }
 785
 786         seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]);
 787         seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]);
 788         seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]);
 789         seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]);
 790         seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
 791                    dma_lo, dma_hi);
 792
 793         ret = 0;
 794
 795         pm_runtime_mark_last_busy(gpu->dev);
 796         pm_runtime_put_autosuspend(gpu->dev);
 797
 798         return ret;
 799 }
 800 #endif
 801
 802 /*
 803  * Power Management:
 804  */
 805 static int enable_clk(struct etnaviv_gpu *gpu)
 806 {
 807         if (gpu->clk_core)
 808                 clk_prepare_enable(gpu->clk_core);
 809         if (gpu->clk_shader)
 810                 clk_prepare_enable(gpu->clk_shader);
 811
 812         return 0;
 813 }
 814
 815 static int disable_clk(struct etnaviv_gpu *gpu)
 816 {
 817         if (gpu->clk_core)
 818                 clk_disable_unprepare(gpu->clk_core);
 819         if (gpu->clk_shader)
 820                 clk_disable_unprepare(gpu->clk_shader);
 821
 822         return 0;
 823 }
 824
 825 static int enable_axi(struct etnaviv_gpu *gpu)
 826 {
 827         if (gpu->clk_bus)
 828                 clk_prepare_enable(gpu->clk_bus);
 829
 830         return 0;
 831 }
 832
 833 static int disable_axi(struct etnaviv_gpu *gpu)
 834 {
 835         if (gpu->clk_bus)
 836                 clk_disable_unprepare(gpu->clk_bus);
 837
 838         return 0;
 839 }
 840
 841 /*
 842  * Hangcheck detection for locked gpu:
 843  */
 844 static void recover_worker(struct work_struct *work)
 845 {
 846         struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
 847                                                recover_work);
 848         unsigned long flags;
 849         unsigned int i;
 850
 851         dev_err(gpu->dev, "hangcheck recover!\n");
 852
 853         if (pm_runtime_get_sync(gpu->dev) < 0)
 854                 return;
 855
 856         mutex_lock(&gpu->lock);
 857
 858         /* Only catch the first event, or when manually re-armed */
 859         if (etnaviv_dump_core) {
 860                 etnaviv_core_dump(gpu);
 861                 etnaviv_dump_core = false;
 862         }
 863
 864         etnaviv_hw_reset(gpu);
 865
 866         /* complete all events, the GPU won't do it after the reset */
 867         spin_lock_irqsave(&gpu->event_spinlock, flags);
 868         for (i = 0; i < ARRAY_SIZE(gpu->event); i++) {
 869                 if (!gpu->event[i].used)
 870                         continue;
 871                 fence_signal(gpu->event[i].fence);
 872                 gpu->event[i].fence = NULL;
 873                 gpu->event[i].used = false;
 874                 complete(&gpu->event_free);
 875         }
 876         spin_unlock_irqrestore(&gpu->event_spinlock, flags);
 877         gpu->completed_fence = gpu->active_fence;
 878
 879         etnaviv_gpu_hw_init(gpu);
 880         gpu->switch_context = true;
 881         gpu->exec_state = -1;
 882
 883         mutex_unlock(&gpu->lock);
 884         pm_runtime_mark_last_busy(gpu->dev);
 885         pm_runtime_put_autosuspend(gpu->dev);
 886
 887         /* Retire the buffer objects in a work */
 888         etnaviv_queue_work(gpu->drm, &gpu->retire_work);
 889 }
 890
 891 static void hangcheck_timer_reset(struct etnaviv_gpu *gpu)
 892 {
 893         DBG("%s", dev_name(gpu->dev));
 894         mod_timer(&gpu->hangcheck_timer,
 895                   round_jiffies_up(jiffies + DRM_ETNAVIV_HANGCHECK_JIFFIES));
 896 }
 897
 898 static void hangcheck_handler(unsigned long data)
 899 {
 900         struct etnaviv_gpu *gpu = (struct etnaviv_gpu *)data;
 901         u32 fence = gpu->completed_fence;
 902         bool progress = false;
 903
 904         if (fence != gpu->hangcheck_fence) {
 905                 gpu->hangcheck_fence = fence;
 906                 progress = true;
 907         }
 908
 909         if (!progress) {
 910                 u32 dma_addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
 911                 int change = dma_addr - gpu->hangcheck_dma_addr;
 912
 913                 if (change < 0 || change > 16) {
 914                         gpu->hangcheck_dma_addr = dma_addr;
 915                         progress = true;
 916                 }
 917         }
 918
 919         if (!progress && fence_after(gpu->active_fence, fence)) {
 920                 dev_err(gpu->dev, "hangcheck detected gpu lockup!\n");
 921                 dev_err(gpu->dev, "     completed fence: %u\n", fence);
 922                 dev_err(gpu->dev, "     active fence: %u\n",
 923                         gpu->active_fence);
 924                 etnaviv_queue_work(gpu->drm, &gpu->recover_work);
 925         }
 926
 927         /* if still more pending work, reset the hangcheck timer: */
 928         if (fence_after(gpu->active_fence, gpu->hangcheck_fence))
 929                 hangcheck_timer_reset(gpu);
 930 }
 931
 932 static void hangcheck_disable(struct etnaviv_gpu *gpu)
 933 {
 934         del_timer_sync(&gpu->hangcheck_timer);
 935         cancel_work_sync(&gpu->recover_work);
 936 }
 937
 938 /* fence object management */
 939 struct etnaviv_fence {
 940         struct etnaviv_gpu *gpu;
 941         struct fence base;
 942 };
 943
 944 static inline struct etnaviv_fence *to_etnaviv_fence(struct fence *fence)
 945 {
 946         return container_of(fence, struct etnaviv_fence, base);
 947 }
 948
 949 static const char *etnaviv_fence_get_driver_name(struct fence *fence)
 950 {
 951         return "etnaviv";
 952 }
 953
 954 static const char *etnaviv_fence_get_timeline_name(struct fence *fence)
 955 {
 956         struct etnaviv_fence *f = to_etnaviv_fence(fence);
 957
 958         return dev_name(f->gpu->dev);
 959 }
 960
 961 static bool etnaviv_fence_enable_signaling(struct fence *fence)
 962 {
 963         return true;
 964 }
 965
 966 static bool etnaviv_fence_signaled(struct fence *fence)
 967 {
 968         struct etnaviv_fence *f = to_etnaviv_fence(fence);
 969
 970         return fence_completed(f->gpu, f->base.seqno);
 971 }
 972
 973 static void etnaviv_fence_release(struct fence *fence)
 974 {
 975         struct etnaviv_fence *f = to_etnaviv_fence(fence);
 976
 977         kfree_rcu(f, base.rcu);
 978 }
 979
 980 static const struct fence_ops etnaviv_fence_ops = {
 981         .get_driver_name = etnaviv_fence_get_driver_name,
 982         .get_timeline_name = etnaviv_fence_get_timeline_name,
 983         .enable_signaling = etnaviv_fence_enable_signaling,
 984         .signaled = etnaviv_fence_signaled,
 985         .wait = fence_default_wait,
 986         .release = etnaviv_fence_release,
 987 };
 988
 989 static struct fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu)
 990 {
 991         struct etnaviv_fence *f;
 992
 993         f = kzalloc(sizeof(*f), GFP_KERNEL);
 994         if (!f)
 995                 return NULL;
 996
 997         f->gpu = gpu;
 998
 999         fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock,
1000                    gpu->fence_context, ++gpu->next_fence);
1001
1002         return &f->base;
1003 }
1004
1005 int etnaviv_gpu_fence_sync_obj(struct etnaviv_gem_object *etnaviv_obj,
1006         unsigned int context, bool exclusive)
1007 {
1008         struct reservation_object *robj = etnaviv_obj->resv;
1009         struct reservation_object_list *fobj;
1010         struct fence *fence;
1011         int i, ret;
1012
1013         if (!exclusive) {
1014                 ret = reservation_object_reserve_shared(robj);
1015                 if (ret)
1016                         return ret;
1017         }
1018
1019         /*
1020          * If we have any shared fences, then the exclusive fence
1021          * should be ignored as it will already have been signalled.
1022          */
1023         fobj = reservation_object_get_list(robj);
1024         if (!fobj || fobj->shared_count == 0) {
1025                 /* Wait on any existing exclusive fence which isn't our own */
1026                 fence = reservation_object_get_excl(robj);
1027                 if (fence && fence->context != context) {
1028                         ret = fence_wait(fence, true);
1029                         if (ret)
1030                                 return ret;
1031                 }
1032         }
1033
1034         if (!exclusive || !fobj)
1035                 return 0;
1036
1037         for (i = 0; i < fobj->shared_count; i++) {
1038                 fence = rcu_dereference_protected(fobj->shared[i],
1039                                                 reservation_object_held(robj));
1040                 if (fence->context != context) {
1041                         ret = fence_wait(fence, true);
1042                         if (ret)
1043                                 return ret;
1044                 }
1045         }
1046
1047         return 0;
1048 }
1049
1050 /*
1051  * event management:
1052  */
1053
1054 static unsigned int event_alloc(struct etnaviv_gpu *gpu)
1055 {
1056         unsigned long ret, flags;
1057         unsigned int i, event = ~0U;
1058
1059         ret = wait_for_completion_timeout(&gpu->event_free,
1060                                           msecs_to_jiffies(10 * 10000));
1061         if (!ret)
1062                 dev_err(gpu->dev, "wait_for_completion_timeout failed");
1063
1064         spin_lock_irqsave(&gpu->event_spinlock, flags);
1065
1066         /* find first free event */
1067         for (i = 0; i < ARRAY_SIZE(gpu->event); i++) {
1068                 if (gpu->event[i].used == false) {
1069                         gpu->event[i].used = true;
1070                         event = i;
1071                         break;
1072                 }
1073         }
1074
1075         spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1076
1077         return event;
1078 }
1079
1080 static void event_free(struct etnaviv_gpu *gpu, unsigned int event)
1081 {
1082         unsigned long flags;
1083
1084         spin_lock_irqsave(&gpu->event_spinlock, flags);
1085
1086         if (gpu->event[event].used == false) {
1087                 dev_warn(gpu->dev, "event %u is already marked as free",
1088                          event);
1089                 spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1090         } else {
1091                 gpu->event[event].used = false;
1092                 spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1093
1094                 complete(&gpu->event_free);
1095         }
1096 }
1097
1098 /*
1099  * Cmdstream submission/retirement:
1100  */
1101
1102 struct etnaviv_cmdbuf *etnaviv_gpu_cmdbuf_new(struct etnaviv_gpu *gpu, u32 size,
1103         size_t nr_bos)
1104 {
1105         struct etnaviv_cmdbuf *cmdbuf;
1106         size_t sz = size_vstruct(nr_bos, sizeof(cmdbuf->bo_map[0]),
1107                                  sizeof(*cmdbuf));
1108
1109         cmdbuf = kzalloc(sz, GFP_KERNEL);
1110         if (!cmdbuf)
1111                 return NULL;
1112
1113         cmdbuf->vaddr = dma_alloc_wc(gpu->dev, size, &cmdbuf->paddr,
1114                                      GFP_KERNEL);
1115         if (!cmdbuf->vaddr) {
1116                 kfree(cmdbuf);
1117                 return NULL;
1118         }
1119
1120         cmdbuf->gpu = gpu;
1121         cmdbuf->size = size;
1122
1123         return cmdbuf;
1124 }
1125
1126 void etnaviv_gpu_cmdbuf_free(struct etnaviv_cmdbuf *cmdbuf)
1127 {
1128         dma_free_wc(cmdbuf->gpu->dev, cmdbuf->size, cmdbuf->vaddr,
1129                     cmdbuf->paddr);
1130         kfree(cmdbuf);
1131 }
1132
1133 static void retire_worker(struct work_struct *work)
1134 {
1135         struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
1136                                                retire_work);
1137         u32 fence = gpu->completed_fence;
1138         struct etnaviv_cmdbuf *cmdbuf, *tmp;
1139         unsigned int i;
1140
1141         mutex_lock(&gpu->lock);
1142         list_for_each_entry_safe(cmdbuf, tmp, &gpu->active_cmd_list, node) {
1143                 if (!fence_is_signaled(cmdbuf->fence))
1144                         break;
1145
1146                 list_del(&cmdbuf->node);
1147                 fence_put(cmdbuf->fence);
1148
1149                 for (i = 0; i < cmdbuf->nr_bos; i++) {
1150                         struct etnaviv_vram_mapping *mapping = cmdbuf->bo_map[i];
1151                         struct etnaviv_gem_object *etnaviv_obj = mapping->object;
1152
1153                         atomic_dec(&etnaviv_obj->gpu_active);
1154                         /* drop the refcount taken in etnaviv_gpu_submit */
1155                         etnaviv_gem_mapping_unreference(mapping);
1156                 }
1157
1158                 etnaviv_gpu_cmdbuf_free(cmdbuf);
1159                 /*
1160                  * We need to balance the runtime PM count caused by
1161                  * each submission.  Upon submission, we increment
1162                  * the runtime PM counter, and allocate one event.
1163                  * So here, we put the runtime PM count for each
1164                  * completed event.
1165                  */
1166                 pm_runtime_put_autosuspend(gpu->dev);
1167         }
1168
1169         gpu->retired_fence = fence;
1170
1171         mutex_unlock(&gpu->lock);
1172
1173         wake_up_all(&gpu->fence_event);
1174 }
1175
1176 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
1177         u32 fence, struct timespec *timeout)
1178 {
1179         int ret;
1180
1181         if (fence_after(fence, gpu->next_fence)) {
1182                 DRM_ERROR("waiting on invalid fence: %u (of %u)\n",
1183                                 fence, gpu->next_fence);
1184                 return -EINVAL;
1185         }
1186
1187         if (!timeout) {
1188                 /* No timeout was requested: just test for completion */
1189                 ret = fence_completed(gpu, fence) ? 0 : -EBUSY;
1190         } else {
1191                 unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
1192
1193                 ret = wait_event_interruptible_timeout(gpu->fence_event,
1194                                                 fence_completed(gpu, fence),
1195                                                 remaining);
1196                 if (ret == 0) {
1197                         DBG("timeout waiting for fence: %u (retired: %u completed: %u)",
1198                                 fence, gpu->retired_fence,
1199                                 gpu->completed_fence);
1200                         ret = -ETIMEDOUT;
1201                 } else if (ret != -ERESTARTSYS) {
1202                         ret = 0;
1203                 }
1204         }
1205
1206         return ret;
1207 }
1208
1209 /*
1210  * Wait for an object to become inactive.  This, on it's own, is not race
1211  * free: the object is moved by the retire worker off the active list, and
1212  * then the iova is put.  Moreover, the object could be re-submitted just
1213  * after we notice that it's become inactive.
1214  *
1215  * Although the retirement happens under the gpu lock, we don't want to hold
1216  * that lock in this function while waiting.
1217  */
1218 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
1219         struct etnaviv_gem_object *etnaviv_obj, struct timespec *timeout)
1220 {
1221         unsigned long remaining;
1222         long ret;
1223
1224         if (!timeout)
1225                 return !is_active(etnaviv_obj) ? 0 : -EBUSY;
1226
1227         remaining = etnaviv_timeout_to_jiffies(timeout);
1228
1229         ret = wait_event_interruptible_timeout(gpu->fence_event,
1230                                                !is_active(etnaviv_obj),
1231                                                remaining);
1232         if (ret > 0) {
1233                 struct etnaviv_drm_private *priv = gpu->drm->dev_private;
1234
1235                 /* Synchronise with the retire worker */
1236                 flush_workqueue(priv->wq);
1237                 return 0;
1238         } else if (ret == -ERESTARTSYS) {
1239                 return -ERESTARTSYS;
1240         } else {
1241                 return -ETIMEDOUT;
1242         }
1243 }
1244
1245 int etnaviv_gpu_pm_get_sync(struct etnaviv_gpu *gpu)
1246 {
1247         return pm_runtime_get_sync(gpu->dev);
1248 }
1249
1250 void etnaviv_gpu_pm_put(struct etnaviv_gpu *gpu)
1251 {
1252         pm_runtime_mark_last_busy(gpu->dev);
1253         pm_runtime_put_autosuspend(gpu->dev);
1254 }
1255
1256 /* add bo's to gpu's ring, and kick gpu: */
1257 int etnaviv_gpu_submit(struct etnaviv_gpu *gpu,
1258         struct etnaviv_gem_submit *submit, struct etnaviv_cmdbuf *cmdbuf)
1259 {
1260         struct fence *fence;
1261         unsigned int event, i;
1262         int ret;
1263
1264         ret = etnaviv_gpu_pm_get_sync(gpu);
1265         if (ret < 0)
1266                 return ret;
1267
1268         mutex_lock(&gpu->lock);
1269
1270         /*
1271          * TODO
1272          *
1273          * - flush
1274          * - data endian
1275          * - prefetch
1276          *
1277          */
1278
1279         event = event_alloc(gpu);
1280         if (unlikely(event == ~0U)) {
1281                 DRM_ERROR("no free event\n");
1282                 ret = -EBUSY;
1283                 goto out_unlock;
1284         }
1285
1286         fence = etnaviv_gpu_fence_alloc(gpu);
1287         if (!fence) {
1288                 event_free(gpu, event);
1289                 ret = -ENOMEM;
1290                 goto out_unlock;
1291         }
1292
1293         gpu->event[event].fence = fence;
1294         submit->fence = fence->seqno;
1295         gpu->active_fence = submit->fence;
1296
1297         if (gpu->lastctx != cmdbuf->ctx) {
1298                 gpu->mmu->need_flush = true;
1299                 gpu->switch_context = true;
1300                 gpu->lastctx = cmdbuf->ctx;
1301         }
1302
1303         etnaviv_buffer_queue(gpu, event, cmdbuf);
1304
1305         cmdbuf->fence = fence;
1306         list_add_tail(&cmdbuf->node, &gpu->active_cmd_list);
1307
1308         /* We're committed to adding this command buffer, hold a PM reference */
1309         pm_runtime_get_noresume(gpu->dev);
1310
1311         for (i = 0; i < submit->nr_bos; i++) {
1312                 struct etnaviv_gem_object *etnaviv_obj = submit->bos[i].obj;
1313
1314                 /* Each cmdbuf takes a refcount on the mapping */
1315                 etnaviv_gem_mapping_reference(submit->bos[i].mapping);
1316                 cmdbuf->bo_map[i] = submit->bos[i].mapping;
1317                 atomic_inc(&etnaviv_obj->gpu_active);
1318
1319                 if (submit->bos[i].flags & ETNA_SUBMIT_BO_WRITE)
1320                         reservation_object_add_excl_fence(etnaviv_obj->resv,
1321                                                           fence);
1322                 else
1323                         reservation_object_add_shared_fence(etnaviv_obj->resv,
1324                                                             fence);
1325         }
1326         cmdbuf->nr_bos = submit->nr_bos;
1327         hangcheck_timer_reset(gpu);
1328         ret = 0;
1329
1330 out_unlock:
1331         mutex_unlock(&gpu->lock);
1332
1333         etnaviv_gpu_pm_put(gpu);
1334
1335         return ret;
1336 }
1337
1338 /*
1339  * Init/Cleanup:
1340  */
1341 static irqreturn_t irq_handler(int irq, void *data)
1342 {
1343         struct etnaviv_gpu *gpu = data;
1344         irqreturn_t ret = IRQ_NONE;
1345
1346         u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE);
1347
1348         if (intr != 0) {
1349                 int event;
1350
1351                 pm_runtime_mark_last_busy(gpu->dev);
1352
1353                 dev_dbg(gpu->dev, "intr 0x%08x\n", intr);
1354
1355                 if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) {
1356                         dev_err(gpu->dev, "AXI bus error\n");
1357                         intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR;
1358                 }
1359
1360                 while ((event = ffs(intr)) != 0) {
1361                         struct fence *fence;
1362
1363                         event -= 1;
1364
1365                         intr &= ~(1 << event);
1366
1367                         dev_dbg(gpu->dev, "event %u\n", event);
1368
1369                         fence = gpu->event[event].fence;
1370                         gpu->event[event].fence = NULL;
1371                         fence_signal(fence);
1372
1373                         /*
1374                          * Events can be processed out of order.  Eg,
1375                          * - allocate and queue event 0
1376                          * - allocate event 1
1377                          * - event 0 completes, we process it
1378                          * - allocate and queue event 0
1379                          * - event 1 and event 0 complete
1380                          * we can end up processing event 0 first, then 1.
1381                          */
1382                         if (fence_after(fence->seqno, gpu->completed_fence))
1383                                 gpu->completed_fence = fence->seqno;
1384
1385                         event_free(gpu, event);
1386                 }
1387
1388                 /* Retire the buffer objects in a work */
1389                 etnaviv_queue_work(gpu->drm, &gpu->retire_work);
1390
1391                 ret = IRQ_HANDLED;
1392         }
1393
1394         return ret;
1395 }
1396
1397 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu)
1398 {
1399         int ret;
1400
1401         ret = enable_clk(gpu);
1402         if (ret)
1403                 return ret;
1404
1405         ret = enable_axi(gpu);
1406         if (ret) {
1407                 disable_clk(gpu);
1408                 return ret;
1409         }
1410
1411         return 0;
1412 }
1413
1414 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu)
1415 {
1416         int ret;
1417
1418         ret = disable_axi(gpu);
1419         if (ret)
1420                 return ret;
1421
1422         ret = disable_clk(gpu);
1423         if (ret)
1424                 return ret;
1425
1426         return 0;
1427 }
1428
1429 static int etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu)
1430 {
1431         if (gpu->buffer) {
1432                 unsigned long timeout;
1433
1434                 /* Replace the last WAIT with END */
1435                 etnaviv_buffer_end(gpu);
1436
1437                 /*
1438                  * We know that only the FE is busy here, this should
1439                  * happen quickly (as the WAIT is only 200 cycles).  If
1440                  * we fail, just warn and continue.
1441                  */
1442                 timeout = jiffies + msecs_to_jiffies(100);
1443                 do {
1444                         u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
1445
1446                         if ((idle & gpu->idle_mask) == gpu->idle_mask)
1447                                 break;
1448
1449                         if (time_is_before_jiffies(timeout)) {
1450                                 dev_warn(gpu->dev,
1451                                          "timed out waiting for idle: idle=0x%x\n",
1452                                          idle);
1453                                 break;
1454                         }
1455
1456                         udelay(5);
1457                 } while (1);
1458         }
1459
1460         return etnaviv_gpu_clk_disable(gpu);
1461 }
1462
1463 #ifdef CONFIG_PM
1464 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
1465 {
1466         u32 clock;
1467         int ret;
1468
1469         ret = mutex_lock_killable(&gpu->lock);
1470         if (ret)
1471                 return ret;
1472
1473         clock = VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS |
1474                 VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(0x40);
1475
1476         etnaviv_gpu_load_clock(gpu, clock);
1477         etnaviv_gpu_hw_init(gpu);
1478
1479         gpu->switch_context = true;
1480         gpu->exec_state = -1;
1481
1482         mutex_unlock(&gpu->lock);
1483
1484         return 0;
1485 }
1486 #endif
1487
1488 static int etnaviv_gpu_bind(struct device *dev, struct device *master,
1489         void *data)
1490 {
1491         struct drm_device *drm = data;
1492         struct etnaviv_drm_private *priv = drm->dev_private;
1493         struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1494         int ret;
1495
1496 #ifdef CONFIG_PM
1497         ret = pm_runtime_get_sync(gpu->dev);
1498 #else
1499         ret = etnaviv_gpu_clk_enable(gpu);
1500 #endif
1501         if (ret < 0)
1502                 return ret;
1503
1504         gpu->drm = drm;
1505         gpu->fence_context = fence_context_alloc(1);
1506         spin_lock_init(&gpu->fence_spinlock);
1507
1508         INIT_LIST_HEAD(&gpu->active_cmd_list);
1509         INIT_WORK(&gpu->retire_work, retire_worker);
1510         INIT_WORK(&gpu->recover_work, recover_worker);
1511         init_waitqueue_head(&gpu->fence_event);
1512
1513         setup_timer(&gpu->hangcheck_timer, hangcheck_handler,
1514                         (unsigned long)gpu);
1515
1516         priv->gpu[priv->num_gpus++] = gpu;
1517
1518         pm_runtime_mark_last_busy(gpu->dev);
1519         pm_runtime_put_autosuspend(gpu->dev);
1520
1521         return 0;
1522 }
1523
1524 static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
1525         void *data)
1526 {
1527         struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1528
1529         DBG("%s", dev_name(gpu->dev));
1530
1531         hangcheck_disable(gpu);
1532
1533 #ifdef CONFIG_PM
1534         pm_runtime_get_sync(gpu->dev);
1535         pm_runtime_put_sync_suspend(gpu->dev);
1536 #else
1537         etnaviv_gpu_hw_suspend(gpu);
1538 #endif
1539
1540         if (gpu->buffer) {
1541                 etnaviv_gpu_cmdbuf_free(gpu->buffer);
1542                 gpu->buffer = NULL;
1543         }
1544
1545         if (gpu->mmu) {
1546                 etnaviv_iommu_destroy(gpu->mmu);
1547                 gpu->mmu = NULL;
1548         }
1549
1550         gpu->drm = NULL;
1551 }
1552
1553 static const struct component_ops gpu_ops = {
1554         .bind = etnaviv_gpu_bind,
1555         .unbind = etnaviv_gpu_unbind,
1556 };
1557
1558 static const struct of_device_id etnaviv_gpu_match[] = {
1559         {
1560                 .compatible = "vivante,gc"
1561         },
1562         { /* sentinel */ }
1563 };
1564
1565 static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
1566 {
1567         struct device *dev = &pdev->dev;
1568         struct etnaviv_gpu *gpu;
1569         u32 dma_mask;
1570         int err = 0;
1571
1572         gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
1573         if (!gpu)
1574                 return -ENOMEM;
1575
1576         gpu->dev = &pdev->dev;
1577         mutex_init(&gpu->lock);
1578
1579         /*
1580          * Set the GPU linear window to be at the end of the DMA window, where
1581          * the CMA area is likely to reside. This ensures that we are able to
1582          * map the command buffers while having the linear window overlap as
1583          * much RAM as possible, so we can optimize mappings for other buffers.
1584          */
1585         dma_mask = (u32)dma_get_required_mask(dev);
1586         if (dma_mask < PHYS_OFFSET + SZ_2G)
1587                 gpu->memory_base = PHYS_OFFSET;
1588         else
1589                 gpu->memory_base = dma_mask - SZ_2G + 1;
1590
1591         /* Map registers: */
1592         gpu->mmio = etnaviv_ioremap(pdev, NULL, dev_name(gpu->dev));
1593         if (IS_ERR(gpu->mmio))
1594                 return PTR_ERR(gpu->mmio);
1595
1596         /* Get Interrupt: */
1597         gpu->irq = platform_get_irq(pdev, 0);
1598         if (gpu->irq < 0) {
1599                 err = gpu->irq;
1600                 dev_err(dev, "failed to get irq: %d\n", err);
1601                 goto fail;
1602         }
1603
1604         err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0,
1605                                dev_name(gpu->dev), gpu);
1606         if (err) {
1607                 dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err);
1608                 goto fail;
1609         }
1610
1611         /* Get Clocks: */
1612         gpu->clk_bus = devm_clk_get(&pdev->dev, "bus");
1613         DBG("clk_bus: %p", gpu->clk_bus);
1614         if (IS_ERR(gpu->clk_bus))
1615                 gpu->clk_bus = NULL;
1616
1617         gpu->clk_core = devm_clk_get(&pdev->dev, "core");
1618         DBG("clk_core: %p", gpu->clk_core);
1619         if (IS_ERR(gpu->clk_core))
1620                 gpu->clk_core = NULL;
1621
1622         gpu->clk_shader = devm_clk_get(&pdev->dev, "shader");
1623         DBG("clk_shader: %p", gpu->clk_shader);
1624         if (IS_ERR(gpu->clk_shader))
1625                 gpu->clk_shader = NULL;
1626
1627         /* TODO: figure out max mapped size */
1628         dev_set_drvdata(dev, gpu);
1629
1630         /*
1631          * We treat the device as initially suspended.  The runtime PM
1632          * autosuspend delay is rather arbitary: no measurements have
1633          * yet been performed to determine an appropriate value.
1634          */
1635         pm_runtime_use_autosuspend(gpu->dev);
1636         pm_runtime_set_autosuspend_delay(gpu->dev, 200);
1637         pm_runtime_enable(gpu->dev);
1638
1639         err = component_add(&pdev->dev, &gpu_ops);
1640         if (err < 0) {
1641                 dev_err(&pdev->dev, "failed to register component: %d\n", err);
1642                 goto fail;
1643         }
1644
1645         return 0;
1646
1647 fail:
1648         return err;
1649 }
1650
1651 static int etnaviv_gpu_platform_remove(struct platform_device *pdev)
1652 {
1653         component_del(&pdev->dev, &gpu_ops);
1654         pm_runtime_disable(&pdev->dev);
1655         return 0;
1656 }
1657
1658 #ifdef CONFIG_PM
1659 static int etnaviv_gpu_rpm_suspend(struct device *dev)
1660 {
1661         struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1662         u32 idle, mask;
1663
1664         /* If we have outstanding fences, we're not idle */
1665         if (gpu->completed_fence != gpu->active_fence)
1666                 return -EBUSY;
1667
1668         /* Check whether the hardware (except FE) is idle */
1669         mask = gpu->idle_mask & ~VIVS_HI_IDLE_STATE_FE;
1670         idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask;
1671         if (idle != mask)
1672                 return -EBUSY;
1673
1674         return etnaviv_gpu_hw_suspend(gpu);
1675 }
1676
1677 static int etnaviv_gpu_rpm_resume(struct device *dev)
1678 {
1679         struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1680         int ret;
1681
1682         ret = etnaviv_gpu_clk_enable(gpu);
1683         if (ret)
1684                 return ret;
1685
1686         /* Re-initialise the basic hardware state */
1687         if (gpu->drm && gpu->buffer) {
1688                 ret = etnaviv_gpu_hw_resume(gpu);
1689                 if (ret) {
1690                         etnaviv_gpu_clk_disable(gpu);
1691                         return ret;
1692                 }
1693         }
1694
1695         return 0;
1696 }
1697 #endif
1698
1699 static const struct dev_pm_ops etnaviv_gpu_pm_ops = {
1700         SET_RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume,
1701                            NULL)
1702 };
1703
1704 struct platform_driver etnaviv_gpu_driver = {
1705         .driver = {
1706                 .name = "etnaviv-gpu",
1707                 .owner = THIS_MODULE,
1708                 .pm = &etnaviv_gpu_pm_ops,
1709                 .of_match_table = etnaviv_gpu_match,
1710         },
1711         .probe = etnaviv_gpu_platform_probe,
1712         .remove = etnaviv_gpu_platform_remove,
1713         .id_table = gpu_ids,
1714 };