drivers/gpu/drm/msm/adreno/adreno_gpu.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2013 Red Hat
   4  * Author: Rob Clark <[email protected]>
   5  *
   6  * Copyright (c) 2014 The Linux Foundation. All rights reserved.
   7  */
   8
   9 #include <linux/ascii85.h>
  10 #include <linux/interconnect.h>
  11 #include <linux/firmware/qcom/qcom_scm.h>
  12 #include <linux/kernel.h>
  13 #include <linux/of_address.h>
  14 #include <linux/pm_opp.h>
  15 #include <linux/slab.h>
  16 #include <linux/soc/qcom/mdt_loader.h>
  17 #include <linux/nvmem-consumer.h>
  18 #include <soc/qcom/ocmem.h>
  19 #include "adreno_gpu.h"
  20 #include "a6xx_gpu.h"
  21 #include "msm_gem.h"
  22 #include "msm_mmu.h"
  23
  24 static u64 address_space_size = 0;
  25 MODULE_PARM_DESC(address_space_size, "Override for size of processes private GPU address space");
  26 module_param(address_space_size, ullong, 0600);
  27
  28 static bool zap_available = true;
  29
  30 static int zap_shader_load_mdt(struct msm_gpu *gpu, const char *fwname,
  31                 u32 pasid)
  32 {
  33         struct device *dev = &gpu->pdev->dev;
  34         const struct firmware *fw;
  35         const char *signed_fwname = NULL;
  36         struct device_node *np, *mem_np;
  37         struct resource r;
  38         phys_addr_t mem_phys;
  39         ssize_t mem_size;
  40         void *mem_region = NULL;
  41         int ret;
  42
  43         if (!IS_ENABLED(CONFIG_ARCH_QCOM)) {
  44                 zap_available = false;
  45                 return -EINVAL;
  46         }
  47
  48         np = of_get_child_by_name(dev->of_node, "zap-shader");
  49         if (!np) {
  50                 zap_available = false;
  51                 return -ENODEV;
  52         }
  53
  54         mem_np = of_parse_phandle(np, "memory-region", 0);
  55         of_node_put(np);
  56         if (!mem_np) {
  57                 zap_available = false;
  58                 return -EINVAL;
  59         }
  60
  61         ret = of_address_to_resource(mem_np, 0, &r);
  62         of_node_put(mem_np);
  63         if (ret)
  64                 return ret;
  65
  66         mem_phys = r.start;
  67
  68         /*
  69          * Check for a firmware-name property.  This is the new scheme
  70          * to handle firmware that may be signed with device specific
  71          * keys, allowing us to have a different zap fw path for different
  72          * devices.
  73          *
  74          * If the firmware-name property is found, we bypass the
  75          * adreno_request_fw() mechanism, because we don't need to handle
  76          * the /lib/firmware/qcom/... vs /lib/firmware/... case.
  77          *
  78          * If the firmware-name property is not found, for backwards
  79          * compatibility we fall back to the fwname from the gpulist
  80          * table.
  81          */
  82         of_property_read_string_index(np, "firmware-name", 0, &signed_fwname);
  83         if (signed_fwname) {
  84                 fwname = signed_fwname;
  85                 ret = request_firmware_direct(&fw, fwname, gpu->dev->dev);
  86                 if (ret)
  87                         fw = ERR_PTR(ret);
  88         } else if (fwname) {
  89                 /* Request the MDT file from the default location: */
  90                 fw = adreno_request_fw(to_adreno_gpu(gpu), fwname);
  91         } else {
  92                 /*
  93                  * For new targets, we require the firmware-name property,
  94                  * if a zap-shader is required, rather than falling back
  95                  * to a firmware name specified in gpulist.
  96                  *
  97                  * Because the firmware is signed with a (potentially)
  98                  * device specific key, having the name come from gpulist
  99                  * was a bad idea, and is only provided for backwards
 100                  * compatibility for older targets.
 101                  */
 102                 return -ENODEV;
 103         }
 104
 105         if (IS_ERR(fw)) {
 106                 DRM_DEV_ERROR(dev, "Unable to load %s\n", fwname);
 107                 return PTR_ERR(fw);
 108         }
 109
 110         /* Figure out how much memory we need */
 111         mem_size = qcom_mdt_get_size(fw);
 112         if (mem_size < 0) {
 113                 ret = mem_size;
 114                 goto out;
 115         }
 116
 117         if (mem_size > resource_size(&r)) {
 118                 DRM_DEV_ERROR(dev,
 119                         "memory region is too small to load the MDT\n");
 120                 ret = -E2BIG;
 121                 goto out;
 122         }
 123
 124         /* Allocate memory for the firmware image */
 125         mem_region = memremap(mem_phys, mem_size,  MEMREMAP_WC);
 126         if (!mem_region) {
 127                 ret = -ENOMEM;
 128                 goto out;
 129         }
 130
 131         /*
 132          * Load the rest of the MDT
 133          *
 134          * Note that we could be dealing with two different paths, since
 135          * with upstream linux-firmware it would be in a qcom/ subdir..
 136          * adreno_request_fw() handles this, but qcom_mdt_load() does
 137          * not.  But since we've already gotten through adreno_request_fw()
 138          * we know which of the two cases it is:
 139          */
 140         if (signed_fwname || (to_adreno_gpu(gpu)->fwloc == FW_LOCATION_LEGACY)) {
 141                 ret = qcom_mdt_load(dev, fw, fwname, pasid,
 142                                 mem_region, mem_phys, mem_size, NULL);
 143         } else {
 144                 char *newname;
 145
 146                 newname = kasprintf(GFP_KERNEL, "qcom/%s", fwname);
 147
 148                 ret = qcom_mdt_load(dev, fw, newname, pasid,
 149                                 mem_region, mem_phys, mem_size, NULL);
 150                 kfree(newname);
 151         }
 152         if (ret)
 153                 goto out;
 154
 155         /* Send the image to the secure world */
 156         ret = qcom_scm_pas_auth_and_reset(pasid);
 157
 158         /*
 159          * If the scm call returns -EOPNOTSUPP we assume that this target
 160          * doesn't need/support the zap shader so quietly fail
 161          */
 162         if (ret == -EOPNOTSUPP)
 163                 zap_available = false;
 164         else if (ret)
 165                 DRM_DEV_ERROR(dev, "Unable to authorize the image\n");
 166
 167 out:
 168         if (mem_region)
 169                 memunmap(mem_region);
 170
 171         release_firmware(fw);
 172
 173         return ret;
 174 }
 175
 176 int adreno_zap_shader_load(struct msm_gpu *gpu, u32 pasid)
 177 {
 178         struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
 179         struct platform_device *pdev = gpu->pdev;
 180
 181         /* Short cut if we determine the zap shader isn't available/needed */
 182         if (!zap_available)
 183                 return -ENODEV;
 184
 185         /* We need SCM to be able to load the firmware */
 186         if (!qcom_scm_is_available()) {
 187                 DRM_DEV_ERROR(&pdev->dev, "SCM is not available\n");
 188                 return -EPROBE_DEFER;
 189         }
 190
 191         return zap_shader_load_mdt(gpu, adreno_gpu->info->zapfw, pasid);
 192 }
 193
 194 struct msm_gem_address_space *
 195 adreno_create_address_space(struct msm_gpu *gpu,
 196                             struct platform_device *pdev)
 197 {
 198         return adreno_iommu_create_address_space(gpu, pdev, 0);
 199 }
 200
 201 struct msm_gem_address_space *
 202 adreno_iommu_create_address_space(struct msm_gpu *gpu,
 203                                   struct platform_device *pdev,
 204                                   unsigned long quirks)
 205 {
 206         struct iommu_domain_geometry *geometry;
 207         struct msm_mmu *mmu;
 208         struct msm_gem_address_space *aspace;
 209         u64 start, size;
 210
 211         mmu = msm_iommu_gpu_new(&pdev->dev, gpu, quirks);
 212         if (IS_ERR_OR_NULL(mmu))
 213                 return ERR_CAST(mmu);
 214
 215         geometry = msm_iommu_get_geometry(mmu);
 216         if (IS_ERR(geometry))
 217                 return ERR_CAST(geometry);
 218
 219         /*
 220          * Use the aperture start or SZ_16M, whichever is greater. This will
 221          * ensure that we align with the allocated pagetable range while still
 222          * allowing room in the lower 32 bits for GMEM and whatnot
 223          */
 224         start = max_t(u64, SZ_16M, geometry->aperture_start);
 225         size = geometry->aperture_end - start + 1;
 226
 227         aspace = msm_gem_address_space_create(mmu, "gpu",
 228                 start & GENMASK_ULL(48, 0), size);
 229
 230         if (IS_ERR(aspace) && !IS_ERR(mmu))
 231                 mmu->funcs->destroy(mmu);
 232
 233         return aspace;
 234 }
 235
 236 u64 adreno_private_address_space_size(struct msm_gpu *gpu)
 237 {
 238         struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
 239
 240         if (address_space_size)
 241                 return address_space_size;
 242
 243         if (adreno_gpu->info->address_space_size)
 244                 return adreno_gpu->info->address_space_size;
 245
 246         return SZ_4G;
 247 }
 248
 249 #define ARM_SMMU_FSR_TF                 BIT(1)
 250 #define ARM_SMMU_FSR_PF                 BIT(3)
 251 #define ARM_SMMU_FSR_EF                 BIT(4)
 252
 253 int adreno_fault_handler(struct msm_gpu *gpu, unsigned long iova, int flags,
 254                          struct adreno_smmu_fault_info *info, const char *block,
 255                          u32 scratch[4])
 256 {
 257         const char *type = "UNKNOWN";
 258         bool do_devcoredump = info && !READ_ONCE(gpu->crashstate);
 259
 260         /*
 261          * If we aren't going to be resuming later from fault_worker, then do
 262          * it now.
 263          */
 264         if (!do_devcoredump) {
 265                 gpu->aspace->mmu->funcs->resume_translation(gpu->aspace->mmu);
 266         }
 267
 268         /*
 269          * Print a default message if we couldn't get the data from the
 270          * adreno-smmu-priv
 271          */
 272         if (!info) {
 273                 pr_warn_ratelimited("*** gpu fault: iova=%.16lx flags=%d (%u,%u,%u,%u)\n",
 274                         iova, flags,
 275                         scratch[0], scratch[1], scratch[2], scratch[3]);
 276
 277                 return 0;
 278         }
 279
 280         if (info->fsr & ARM_SMMU_FSR_TF)
 281                 type = "TRANSLATION";
 282         else if (info->fsr & ARM_SMMU_FSR_PF)
 283                 type = "PERMISSION";
 284         else if (info->fsr & ARM_SMMU_FSR_EF)
 285                 type = "EXTERNAL";
 286
 287         pr_warn_ratelimited("*** gpu fault: ttbr0=%.16llx iova=%.16lx dir=%s type=%s source=%s (%u,%u,%u,%u)\n",
 288                         info->ttbr0, iova,
 289                         flags & IOMMU_FAULT_WRITE ? "WRITE" : "READ",
 290                         type, block,
 291                         scratch[0], scratch[1], scratch[2], scratch[3]);
 292
 293         if (do_devcoredump) {
 294                 /* Turn off the hangcheck timer to keep it from bothering us */
 295                 del_timer(&gpu->hangcheck_timer);
 296
 297                 gpu->fault_info.ttbr0 = info->ttbr0;
 298                 gpu->fault_info.iova  = iova;
 299                 gpu->fault_info.flags = flags;
 300                 gpu->fault_info.type  = type;
 301                 gpu->fault_info.block = block;
 302
 303                 kthread_queue_work(gpu->worker, &gpu->fault_work);
 304         }
 305
 306         return 0;
 307 }
 308
 309 int adreno_get_param(struct msm_gpu *gpu, struct msm_file_private *ctx,
 310                      uint32_t param, uint64_t *value, uint32_t *len)
 311 {
 312         struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
 313
 314         /* No pointer params yet */
 315         if (*len != 0)
 316                 return -EINVAL;
 317
 318         switch (param) {
 319         case MSM_PARAM_GPU_ID:
 320                 *value = adreno_gpu->info->revn;
 321                 return 0;
 322         case MSM_PARAM_GMEM_SIZE:
 323                 *value = adreno_gpu->gmem;
 324                 return 0;
 325         case MSM_PARAM_GMEM_BASE:
 326                 *value = !adreno_is_a650_family(adreno_gpu) ? 0x100000 : 0;
 327                 return 0;
 328         case MSM_PARAM_CHIP_ID:
 329                 *value =  (uint64_t)adreno_gpu->rev.patchid |
 330                          ((uint64_t)adreno_gpu->rev.minor << 8) |
 331                          ((uint64_t)adreno_gpu->rev.major << 16) |
 332                          ((uint64_t)adreno_gpu->rev.core  << 24);
 333                 if (!adreno_gpu->info->revn)
 334                         *value |= ((uint64_t) adreno_gpu->speedbin) << 32;
 335                 return 0;
 336         case MSM_PARAM_MAX_FREQ:
 337                 *value = adreno_gpu->base.fast_rate;
 338                 return 0;
 339         case MSM_PARAM_TIMESTAMP:
 340                 if (adreno_gpu->funcs->get_timestamp) {
 341                         int ret;
 342
 343                         pm_runtime_get_sync(&gpu->pdev->dev);
 344                         ret = adreno_gpu->funcs->get_timestamp(gpu, value);
 345                         pm_runtime_put_autosuspend(&gpu->pdev->dev);
 346
 347                         return ret;
 348                 }
 349                 return -EINVAL;
 350         case MSM_PARAM_PRIORITIES:
 351                 *value = gpu->nr_rings * NR_SCHED_PRIORITIES;
 352                 return 0;
 353         case MSM_PARAM_PP_PGTABLE:
 354                 *value = 0;
 355                 return 0;
 356         case MSM_PARAM_FAULTS:
 357                 if (ctx->aspace)
 358                         *value = gpu->global_faults + ctx->aspace->faults;
 359                 else
 360                         *value = gpu->global_faults;
 361                 return 0;
 362         case MSM_PARAM_SUSPENDS:
 363                 *value = gpu->suspend_count;
 364                 return 0;
 365         case MSM_PARAM_VA_START:
 366                 if (ctx->aspace == gpu->aspace)
 367                         return -EINVAL;
 368                 *value = ctx->aspace->va_start;
 369                 return 0;
 370         case MSM_PARAM_VA_SIZE:
 371                 if (ctx->aspace == gpu->aspace)
 372                         return -EINVAL;
 373                 *value = ctx->aspace->va_size;
 374                 return 0;
 375         default:
 376                 DBG("%s: invalid param: %u", gpu->name, param);
 377                 return -EINVAL;
 378         }
 379 }
 380
 381 int adreno_set_param(struct msm_gpu *gpu, struct msm_file_private *ctx,
 382                      uint32_t param, uint64_t value, uint32_t len)
 383 {
 384         switch (param) {
 385         case MSM_PARAM_COMM:
 386         case MSM_PARAM_CMDLINE:
 387                 /* kstrdup_quotable_cmdline() limits to PAGE_SIZE, so
 388                  * that should be a reasonable upper bound
 389                  */
 390                 if (len > PAGE_SIZE)
 391                         return -EINVAL;
 392                 break;
 393         default:
 394                 if (len != 0)
 395                         return -EINVAL;
 396         }
 397
 398         switch (param) {
 399         case MSM_PARAM_COMM:
 400         case MSM_PARAM_CMDLINE: {
 401                 char *str, **paramp;
 402
 403                 str = kmalloc(len + 1, GFP_KERNEL);
 404                 if (!str)
 405                         return -ENOMEM;
 406
 407                 if (copy_from_user(str, u64_to_user_ptr(value), len)) {
 408                         kfree(str);
 409                         return -EFAULT;
 410                 }
 411
 412                 /* Ensure string is null terminated: */
 413                 str[len] = '\0';
 414
 415                 mutex_lock(&gpu->lock);
 416
 417                 if (param == MSM_PARAM_COMM) {
 418                         paramp = &ctx->comm;
 419                 } else {
 420                         paramp = &ctx->cmdline;
 421                 }
 422
 423                 kfree(*paramp);
 424                 *paramp = str;
 425
 426                 mutex_unlock(&gpu->lock);
 427
 428                 return 0;
 429         }
 430         case MSM_PARAM_SYSPROF:
 431                 if (!capable(CAP_SYS_ADMIN))
 432                         return -EPERM;
 433                 return msm_file_private_set_sysprof(ctx, gpu, value);
 434         default:
 435                 DBG("%s: invalid param: %u", gpu->name, param);
 436                 return -EINVAL;
 437         }
 438 }
 439
 440 const struct firmware *
 441 adreno_request_fw(struct adreno_gpu *adreno_gpu, const char *fwname)
 442 {
 443         struct drm_device *drm = adreno_gpu->base.dev;
 444         const struct firmware *fw = NULL;
 445         char *newname;
 446         int ret;
 447
 448         newname = kasprintf(GFP_KERNEL, "qcom/%s", fwname);
 449         if (!newname)
 450                 return ERR_PTR(-ENOMEM);
 451
 452         /*
 453          * Try first to load from qcom/$fwfile using a direct load (to avoid
 454          * a potential timeout waiting for usermode helper)
 455          */
 456         if ((adreno_gpu->fwloc == FW_LOCATION_UNKNOWN) ||
 457             (adreno_gpu->fwloc == FW_LOCATION_NEW)) {
 458
 459                 ret = request_firmware_direct(&fw, newname, drm->dev);
 460                 if (!ret) {
 461                         DRM_DEV_INFO(drm->dev, "loaded %s from new location\n",
 462                                 newname);
 463                         adreno_gpu->fwloc = FW_LOCATION_NEW;
 464                         goto out;
 465                 } else if (adreno_gpu->fwloc != FW_LOCATION_UNKNOWN) {
 466                         DRM_DEV_ERROR(drm->dev, "failed to load %s: %d\n",
 467                                 newname, ret);
 468                         fw = ERR_PTR(ret);
 469                         goto out;
 470                 }
 471         }
 472
 473         /*
 474          * Then try the legacy location without qcom/ prefix
 475          */
 476         if ((adreno_gpu->fwloc == FW_LOCATION_UNKNOWN) ||
 477             (adreno_gpu->fwloc == FW_LOCATION_LEGACY)) {
 478
 479                 ret = request_firmware_direct(&fw, fwname, drm->dev);
 480                 if (!ret) {
 481                         DRM_DEV_INFO(drm->dev, "loaded %s from legacy location\n",
 482                                 newname);
 483                         adreno_gpu->fwloc = FW_LOCATION_LEGACY;
 484                         goto out;
 485                 } else if (adreno_gpu->fwloc != FW_LOCATION_UNKNOWN) {
 486                         DRM_DEV_ERROR(drm->dev, "failed to load %s: %d\n",
 487                                 fwname, ret);
 488                         fw = ERR_PTR(ret);
 489                         goto out;
 490                 }
 491         }
 492
 493         /*
 494          * Finally fall back to request_firmware() for cases where the
 495          * usermode helper is needed (I think mainly android)
 496          */
 497         if ((adreno_gpu->fwloc == FW_LOCATION_UNKNOWN) ||
 498             (adreno_gpu->fwloc == FW_LOCATION_HELPER)) {
 499
 500                 ret = request_firmware(&fw, newname, drm->dev);
 501                 if (!ret) {
 502                         DRM_DEV_INFO(drm->dev, "loaded %s with helper\n",
 503                                 newname);
 504                         adreno_gpu->fwloc = FW_LOCATION_HELPER;
 505                         goto out;
 506                 } else if (adreno_gpu->fwloc != FW_LOCATION_UNKNOWN) {
 507                         DRM_DEV_ERROR(drm->dev, "failed to load %s: %d\n",
 508                                 newname, ret);
 509                         fw = ERR_PTR(ret);
 510                         goto out;
 511                 }
 512         }
 513
 514         DRM_DEV_ERROR(drm->dev, "failed to load %s\n", fwname);
 515         fw = ERR_PTR(-ENOENT);
 516 out:
 517         kfree(newname);
 518         return fw;
 519 }
 520
 521 int adreno_load_fw(struct adreno_gpu *adreno_gpu)
 522 {
 523         int i;
 524
 525         for (i = 0; i < ARRAY_SIZE(adreno_gpu->info->fw); i++) {
 526                 const struct firmware *fw;
 527
 528                 if (!adreno_gpu->info->fw[i])
 529                         continue;
 530
 531                 /* Skip if the firmware has already been loaded */
 532                 if (adreno_gpu->fw[i])
 533                         continue;
 534
 535                 fw = adreno_request_fw(adreno_gpu, adreno_gpu->info->fw[i]);
 536                 if (IS_ERR(fw))
 537                         return PTR_ERR(fw);
 538
 539                 adreno_gpu->fw[i] = fw;
 540         }
 541
 542         return 0;
 543 }
 544
 545 struct drm_gem_object *adreno_fw_create_bo(struct msm_gpu *gpu,
 546                 const struct firmware *fw, u64 *iova)
 547 {
 548         struct drm_gem_object *bo;
 549         void *ptr;
 550
 551         ptr = msm_gem_kernel_new(gpu->dev, fw->size - 4,
 552                 MSM_BO_WC | MSM_BO_GPU_READONLY, gpu->aspace, &bo, iova);
 553
 554         if (IS_ERR(ptr))
 555                 return ERR_CAST(ptr);
 556
 557         memcpy(ptr, &fw->data[4], fw->size - 4);
 558
 559         msm_gem_put_vaddr(bo);
 560
 561         return bo;
 562 }
 563
 564 int adreno_hw_init(struct msm_gpu *gpu)
 565 {
 566         VERB("%s", gpu->name);
 567
 568         for (int i = 0; i < gpu->nr_rings; i++) {
 569                 struct msm_ringbuffer *ring = gpu->rb[i];
 570
 571                 if (!ring)
 572                         continue;
 573
 574                 ring->cur = ring->start;
 575                 ring->next = ring->start;
 576                 ring->memptrs->rptr = 0;
 577
 578                 /* Detect and clean up an impossible fence, ie. if GPU managed
 579                  * to scribble something invalid, we don't want that to confuse
 580                  * us into mistakingly believing that submits have completed.
 581                  */
 582                 if (fence_before(ring->fctx->last_fence, ring->memptrs->fence)) {
 583                         ring->memptrs->fence = ring->fctx->last_fence;
 584                 }
 585         }
 586
 587         return 0;
 588 }
 589
 590 /* Use this helper to read rptr, since a430 doesn't update rptr in memory */
 591 static uint32_t get_rptr(struct adreno_gpu *adreno_gpu,
 592                 struct msm_ringbuffer *ring)
 593 {
 594         struct msm_gpu *gpu = &adreno_gpu->base;
 595
 596         return gpu->funcs->get_rptr(gpu, ring);
 597 }
 598
 599 struct msm_ringbuffer *adreno_active_ring(struct msm_gpu *gpu)
 600 {
 601         return gpu->rb[0];
 602 }
 603
 604 void adreno_recover(struct msm_gpu *gpu)
 605 {
 606         struct drm_device *dev = gpu->dev;
 607         int ret;
 608
 609         // XXX pm-runtime??  we *need* the device to be off after this
 610         // so maybe continuing to call ->pm_suspend/resume() is better?
 611
 612         gpu->funcs->pm_suspend(gpu);
 613         gpu->funcs->pm_resume(gpu);
 614
 615         ret = msm_gpu_hw_init(gpu);
 616         if (ret) {
 617                 DRM_DEV_ERROR(dev->dev, "gpu hw init failed: %d\n", ret);
 618                 /* hmm, oh well? */
 619         }
 620 }
 621
 622 void adreno_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring, u32 reg)
 623 {
 624         uint32_t wptr;
 625
 626         /* Copy the shadow to the actual register */
 627         ring->cur = ring->next;
 628
 629         /*
 630          * Mask wptr value that we calculate to fit in the HW range. This is
 631          * to account for the possibility that the last command fit exactly into
 632          * the ringbuffer and rb->next hasn't wrapped to zero yet
 633          */
 634         wptr = get_wptr(ring);
 635
 636         /* ensure writes to ringbuffer have hit system memory: */
 637         mb();
 638
 639         gpu_write(gpu, reg, wptr);
 640 }
 641
 642 bool adreno_idle(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
 643 {
 644         struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
 645         uint32_t wptr = get_wptr(ring);
 646
 647         /* wait for CP to drain ringbuffer: */
 648         if (!spin_until(get_rptr(adreno_gpu, ring) == wptr))
 649                 return true;
 650
 651         /* TODO maybe we need to reset GPU here to recover from hang? */
 652         DRM_ERROR("%s: timeout waiting to drain ringbuffer %d rptr/wptr = %X/%X\n",
 653                 gpu->name, ring->id, get_rptr(adreno_gpu, ring), wptr);
 654
 655         return false;
 656 }
 657
 658 int adreno_gpu_state_get(struct msm_gpu *gpu, struct msm_gpu_state *state)
 659 {
 660         struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
 661         int i, count = 0;
 662
 663         WARN_ON(!mutex_is_locked(&gpu->lock));
 664
 665         kref_init(&state->ref);
 666
 667         ktime_get_real_ts64(&state->time);
 668
 669         for (i = 0; i < gpu->nr_rings; i++) {
 670                 int size = 0, j;
 671
 672                 state->ring[i].fence = gpu->rb[i]->memptrs->fence;
 673                 state->ring[i].iova = gpu->rb[i]->iova;
 674                 state->ring[i].seqno = gpu->rb[i]->fctx->last_fence;
 675                 state->ring[i].rptr = get_rptr(adreno_gpu, gpu->rb[i]);
 676                 state->ring[i].wptr = get_wptr(gpu->rb[i]);
 677
 678                 /* Copy at least 'wptr' dwords of the data */
 679                 size = state->ring[i].wptr;
 680
 681                 /* After wptr find the last non zero dword to save space */
 682                 for (j = state->ring[i].wptr; j < MSM_GPU_RINGBUFFER_SZ >> 2; j++)
 683                         if (gpu->rb[i]->start[j])
 684                                 size = j + 1;
 685
 686                 if (size) {
 687                         state->ring[i].data = kvmalloc(size << 2, GFP_KERNEL);
 688                         if (state->ring[i].data) {
 689                                 memcpy(state->ring[i].data, gpu->rb[i]->start, size << 2);
 690                                 state->ring[i].data_size = size << 2;
 691                         }
 692                 }
 693         }
 694
 695         /* Some targets prefer to collect their own registers */
 696         if (!adreno_gpu->registers)
 697                 return 0;
 698
 699         /* Count the number of registers */
 700         for (i = 0; adreno_gpu->registers[i] != ~0; i += 2)
 701                 count += adreno_gpu->registers[i + 1] -
 702                         adreno_gpu->registers[i] + 1;
 703
 704         state->registers = kcalloc(count * 2, sizeof(u32), GFP_KERNEL);
 705         if (state->registers) {
 706                 int pos = 0;
 707
 708                 for (i = 0; adreno_gpu->registers[i] != ~0; i += 2) {
 709                         u32 start = adreno_gpu->registers[i];
 710                         u32 end   = adreno_gpu->registers[i + 1];
 711                         u32 addr;
 712
 713                         for (addr = start; addr <= end; addr++) {
 714                                 state->registers[pos++] = addr;
 715                                 state->registers[pos++] = gpu_read(gpu, addr);
 716                         }
 717                 }
 718
 719                 state->nr_registers = count;
 720         }
 721
 722         return 0;
 723 }
 724
 725 void adreno_gpu_state_destroy(struct msm_gpu_state *state)
 726 {
 727         int i;
 728
 729         for (i = 0; i < ARRAY_SIZE(state->ring); i++)
 730                 kvfree(state->ring[i].data);
 731
 732         for (i = 0; state->bos && i < state->nr_bos; i++)
 733                 kvfree(state->bos[i].data);
 734
 735         kfree(state->bos);
 736         kfree(state->comm);
 737         kfree(state->cmd);
 738         kfree(state->registers);
 739 }
 740
 741 static void adreno_gpu_state_kref_destroy(struct kref *kref)
 742 {
 743         struct msm_gpu_state *state = container_of(kref,
 744                 struct msm_gpu_state, ref);
 745
 746         adreno_gpu_state_destroy(state);
 747         kfree(state);
 748 }
 749
 750 int adreno_gpu_state_put(struct msm_gpu_state *state)
 751 {
 752         if (IS_ERR_OR_NULL(state))
 753                 return 1;
 754
 755         return kref_put(&state->ref, adreno_gpu_state_kref_destroy);
 756 }
 757
 758 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
 759
 760 static char *adreno_gpu_ascii85_encode(u32 *src, size_t len)
 761 {
 762         void *buf;
 763         size_t buf_itr = 0, buffer_size;
 764         char out[ASCII85_BUFSZ];
 765         long l;
 766         int i;
 767
 768         if (!src || !len)
 769                 return NULL;
 770
 771         l = ascii85_encode_len(len);
 772
 773         /*
 774          * Ascii85 outputs either a 5 byte string or a 1 byte string. So we
 775          * account for the worst case of 5 bytes per dword plus the 1 for '\0'
 776          */
 777         buffer_size = (l * 5) + 1;
 778
 779         buf = kvmalloc(buffer_size, GFP_KERNEL);
 780         if (!buf)
 781                 return NULL;
 782
 783         for (i = 0; i < l; i++)
 784                 buf_itr += scnprintf(buf + buf_itr, buffer_size - buf_itr, "%s",
 785                                 ascii85_encode(src[i], out));
 786
 787         return buf;
 788 }
 789
 790 /* len is expected to be in bytes
 791  *
 792  * WARNING: *ptr should be allocated with kvmalloc or friends.  It can be free'd
 793  * with kvfree() and replaced with a newly kvmalloc'd buffer on the first call
 794  * when the unencoded raw data is encoded
 795  */
 796 void adreno_show_object(struct drm_printer *p, void **ptr, int len,
 797                 bool *encoded)
 798 {
 799         if (!*ptr || !len)
 800                 return;
 801
 802         if (!*encoded) {
 803                 long datalen, i;
 804                 u32 *buf = *ptr;
 805
 806                 /*
 807                  * Only dump the non-zero part of the buffer - rarely will
 808                  * any data completely fill the entire allocated size of
 809                  * the buffer.
 810                  */
 811                 for (datalen = 0, i = 0; i < len >> 2; i++)
 812                         if (buf[i])
 813                                 datalen = ((i + 1) << 2);
 814
 815                 /*
 816                  * If we reach here, then the originally captured binary buffer
 817                  * will be replaced with the ascii85 encoded string
 818                  */
 819                 *ptr = adreno_gpu_ascii85_encode(buf, datalen);
 820
 821                 kvfree(buf);
 822
 823                 *encoded = true;
 824         }
 825
 826         if (!*ptr)
 827                 return;
 828
 829         drm_puts(p, "    data: !!ascii85 |\n");
 830         drm_puts(p, "     ");
 831
 832         drm_puts(p, *ptr);
 833
 834         drm_puts(p, "\n");
 835 }
 836
 837 void adreno_show(struct msm_gpu *gpu, struct msm_gpu_state *state,
 838                 struct drm_printer *p)
 839 {
 840         struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
 841         int i;
 842
 843         if (IS_ERR_OR_NULL(state))
 844                 return;
 845
 846         drm_printf(p, "revision: %d (%d.%d.%d.%d)\n",
 847                         adreno_gpu->info->revn, adreno_gpu->rev.core,
 848                         adreno_gpu->rev.major, adreno_gpu->rev.minor,
 849                         adreno_gpu->rev.patchid);
 850         /*
 851          * If this is state collected due to iova fault, so fault related info
 852          *
 853          * TTBR0 would not be zero, so this is a good way to distinguish
 854          */
 855         if (state->fault_info.ttbr0) {
 856                 const struct msm_gpu_fault_info *info = &state->fault_info;
 857
 858                 drm_puts(p, "fault-info:\n");
 859                 drm_printf(p, "  - ttbr0=%.16llx\n", info->ttbr0);
 860                 drm_printf(p, "  - iova=%.16lx\n", info->iova);
 861                 drm_printf(p, "  - dir=%s\n", info->flags & IOMMU_FAULT_WRITE ? "WRITE" : "READ");
 862                 drm_printf(p, "  - type=%s\n", info->type);
 863                 drm_printf(p, "  - source=%s\n", info->block);
 864         }
 865
 866         drm_printf(p, "rbbm-status: 0x%08x\n", state->rbbm_status);
 867
 868         drm_puts(p, "ringbuffer:\n");
 869
 870         for (i = 0; i < gpu->nr_rings; i++) {
 871                 drm_printf(p, "  - id: %d\n", i);
 872                 drm_printf(p, "    iova: 0x%016llx\n", state->ring[i].iova);
 873                 drm_printf(p, "    last-fence: %u\n", state->ring[i].seqno);
 874                 drm_printf(p, "    retired-fence: %u\n", state->ring[i].fence);
 875                 drm_printf(p, "    rptr: %u\n", state->ring[i].rptr);
 876                 drm_printf(p, "    wptr: %u\n", state->ring[i].wptr);
 877                 drm_printf(p, "    size: %u\n", MSM_GPU_RINGBUFFER_SZ);
 878
 879                 adreno_show_object(p, &state->ring[i].data,
 880                         state->ring[i].data_size, &state->ring[i].encoded);
 881         }
 882
 883         if (state->bos) {
 884                 drm_puts(p, "bos:\n");
 885
 886                 for (i = 0; i < state->nr_bos; i++) {
 887                         drm_printf(p, "  - iova: 0x%016llx\n",
 888                                 state->bos[i].iova);
 889                         drm_printf(p, "    size: %zd\n", state->bos[i].size);
 890                         drm_printf(p, "    name: %-32s\n", state->bos[i].name);
 891
 892                         adreno_show_object(p, &state->bos[i].data,
 893                                 state->bos[i].size, &state->bos[i].encoded);
 894                 }
 895         }
 896
 897         if (state->nr_registers) {
 898                 drm_puts(p, "registers:\n");
 899
 900                 for (i = 0; i < state->nr_registers; i++) {
 901                         drm_printf(p, "  - { offset: 0x%04x, value: 0x%08x }\n",
 902                                 state->registers[i * 2] << 2,
 903                                 state->registers[(i * 2) + 1]);
 904                 }
 905         }
 906 }
 907 #endif
 908
 909 /* Dump common gpu status and scratch registers on any hang, to make
 910  * the hangcheck logs more useful.  The scratch registers seem always
 911  * safe to read when GPU has hung (unlike some other regs, depending
 912  * on how the GPU hung), and they are useful to match up to cmdstream
 913  * dumps when debugging hangs:
 914  */
 915 void adreno_dump_info(struct msm_gpu *gpu)
 916 {
 917         struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
 918         int i;
 919
 920         printk("revision: %d (%d.%d.%d.%d)\n",
 921                         adreno_gpu->info->revn, adreno_gpu->rev.core,
 922                         adreno_gpu->rev.major, adreno_gpu->rev.minor,
 923                         adreno_gpu->rev.patchid);
 924
 925         for (i = 0; i < gpu->nr_rings; i++) {
 926                 struct msm_ringbuffer *ring = gpu->rb[i];
 927
 928                 printk("rb %d: fence:    %d/%d\n", i,
 929                         ring->memptrs->fence,
 930                         ring->fctx->last_fence);
 931
 932                 printk("rptr:     %d\n", get_rptr(adreno_gpu, ring));
 933                 printk("rb wptr:  %d\n", get_wptr(ring));
 934         }
 935 }
 936
 937 /* would be nice to not have to duplicate the _show() stuff with printk(): */
 938 void adreno_dump(struct msm_gpu *gpu)
 939 {
 940         struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
 941         int i;
 942
 943         if (!adreno_gpu->registers)
 944                 return;
 945
 946         /* dump these out in a form that can be parsed by demsm: */
 947         printk("IO:region %s 00000000 00020000\n", gpu->name);
 948         for (i = 0; adreno_gpu->registers[i] != ~0; i += 2) {
 949                 uint32_t start = adreno_gpu->registers[i];
 950                 uint32_t end   = adreno_gpu->registers[i+1];
 951                 uint32_t addr;
 952
 953                 for (addr = start; addr <= end; addr++) {
 954                         uint32_t val = gpu_read(gpu, addr);
 955                         printk("IO:R %08x %08x\n", addr<<2, val);
 956                 }
 957         }
 958 }
 959
 960 static uint32_t ring_freewords(struct msm_ringbuffer *ring)
 961 {
 962         struct adreno_gpu *adreno_gpu = to_adreno_gpu(ring->gpu);
 963         uint32_t size = MSM_GPU_RINGBUFFER_SZ >> 2;
 964         /* Use ring->next to calculate free size */
 965         uint32_t wptr = ring->next - ring->start;
 966         uint32_t rptr = get_rptr(adreno_gpu, ring);
 967         return (rptr + (size - 1) - wptr) % size;
 968 }
 969
 970 void adreno_wait_ring(struct msm_ringbuffer *ring, uint32_t ndwords)
 971 {
 972         if (spin_until(ring_freewords(ring) >= ndwords))
 973                 DRM_DEV_ERROR(ring->gpu->dev->dev,
 974                         "timeout waiting for space in ringbuffer %d\n",
 975                         ring->id);
 976 }
 977
 978 static int adreno_get_pwrlevels(struct device *dev,
 979                 struct msm_gpu *gpu)
 980 {
 981         struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
 982         unsigned long freq = ULONG_MAX;
 983         struct dev_pm_opp *opp;
 984         int ret;
 985
 986         gpu->fast_rate = 0;
 987
 988         /* devm_pm_opp_of_add_table may error out but will still create an OPP table */
 989         ret = devm_pm_opp_of_add_table(dev);
 990         if (ret == -ENODEV) {
 991                 /* Special cases for ancient hw with ancient DT bindings */
 992                 if (adreno_is_a2xx(adreno_gpu)) {
 993                         dev_warn(dev, "Unable to find the OPP table. Falling back to 200 MHz.\n");
 994                         dev_pm_opp_add(dev, 200000000, 0);
 995                 } else if (adreno_is_a320(adreno_gpu)) {
 996                         dev_warn(dev, "Unable to find the OPP table. Falling back to 450 MHz.\n");
 997                         dev_pm_opp_add(dev, 450000000, 0);
 998                 } else {
 999                         DRM_DEV_ERROR(dev, "Unable to find the OPP table\n");
1000                         return -ENODEV;
1001                 }
1002         } else if (ret) {
1003                 DRM_DEV_ERROR(dev, "Unable to set the OPP table\n");
1004                 return ret;
1005         }
1006
1007         /* Find the fastest defined rate */
1008         opp = dev_pm_opp_find_freq_floor(dev, &freq);
1009         if (IS_ERR(opp))
1010                 return PTR_ERR(opp);
1011
1012         gpu->fast_rate = freq;
1013         dev_pm_opp_put(opp);
1014
1015         DBG("fast_rate=%u, slow_rate=27000000", gpu->fast_rate);
1016
1017         return 0;
1018 }
1019
1020 int adreno_gpu_ocmem_init(struct device *dev, struct adreno_gpu *adreno_gpu,
1021                           struct adreno_ocmem *adreno_ocmem)
1022 {
1023         struct ocmem_buf *ocmem_hdl;
1024         struct ocmem *ocmem;
1025
1026         ocmem = of_get_ocmem(dev);
1027         if (IS_ERR(ocmem)) {
1028                 if (PTR_ERR(ocmem) == -ENODEV) {
1029                         /*
1030                          * Return success since either the ocmem property was
1031                          * not specified in device tree, or ocmem support is
1032                          * not compiled into the kernel.
1033                          */
1034                         return 0;
1035                 }
1036
1037                 return PTR_ERR(ocmem);
1038         }
1039
1040         ocmem_hdl = ocmem_allocate(ocmem, OCMEM_GRAPHICS, adreno_gpu->gmem);
1041         if (IS_ERR(ocmem_hdl))
1042                 return PTR_ERR(ocmem_hdl);
1043
1044         adreno_ocmem->ocmem = ocmem;
1045         adreno_ocmem->base = ocmem_hdl->addr;
1046         adreno_ocmem->hdl = ocmem_hdl;
1047         adreno_gpu->gmem = ocmem_hdl->len;
1048
1049         return 0;
1050 }
1051
1052 void adreno_gpu_ocmem_cleanup(struct adreno_ocmem *adreno_ocmem)
1053 {
1054         if (adreno_ocmem && adreno_ocmem->base)
1055                 ocmem_free(adreno_ocmem->ocmem, OCMEM_GRAPHICS,
1056                            adreno_ocmem->hdl);
1057 }
1058
1059 int adreno_read_speedbin(struct device *dev, u32 *speedbin)
1060 {
1061         return nvmem_cell_read_variable_le_u32(dev, "speed_bin", speedbin);
1062 }
1063
1064 int adreno_gpu_init(struct drm_device *drm, struct platform_device *pdev,
1065                 struct adreno_gpu *adreno_gpu,
1066                 const struct adreno_gpu_funcs *funcs, int nr_rings)
1067 {
1068         struct device *dev = &pdev->dev;
1069         struct adreno_platform_config *config = dev->platform_data;
1070         struct msm_gpu_config adreno_gpu_config  = { 0 };
1071         struct msm_gpu *gpu = &adreno_gpu->base;
1072         struct adreno_rev *rev = &config->rev;
1073         const char *gpu_name;
1074         u32 speedbin;
1075         int ret;
1076
1077         /* Only handle the core clock when GMU is not in use */
1078         if (config->rev.core < 6) {
1079                 /*
1080                  * This can only be done before devm_pm_opp_of_add_table(), or
1081                  * dev_pm_opp_set_config() will WARN_ON()
1082                  */
1083                 if (IS_ERR(devm_clk_get(dev, "core"))) {
1084                         /*
1085                          * If "core" is absent, go for the legacy clock name.
1086                          * If we got this far in probing, it's a given one of
1087                          * them exists.
1088                          */
1089                         devm_pm_opp_set_clkname(dev, "core_clk");
1090                 } else
1091                         devm_pm_opp_set_clkname(dev, "core");
1092         }
1093
1094         adreno_gpu->funcs = funcs;
1095         adreno_gpu->info = adreno_info(config->rev);
1096         adreno_gpu->gmem = adreno_gpu->info->gmem;
1097         adreno_gpu->revn = adreno_gpu->info->revn;
1098         adreno_gpu->rev = *rev;
1099
1100         if (adreno_read_speedbin(dev, &speedbin) || !speedbin)
1101                 speedbin = 0xffff;
1102         adreno_gpu->speedbin = (uint16_t) (0xffff & speedbin);
1103
1104         gpu_name = adreno_gpu->info->name;
1105         if (!gpu_name) {
1106                 gpu_name = devm_kasprintf(dev, GFP_KERNEL, "%d.%d.%d.%d",
1107                                 rev->core, rev->major, rev->minor,
1108                                 rev->patchid);
1109                 if (!gpu_name)
1110                         return -ENOMEM;
1111         }
1112
1113         adreno_gpu_config.ioname = "kgsl_3d0_reg_memory";
1114
1115         adreno_gpu_config.nr_rings = nr_rings;
1116
1117         ret = adreno_get_pwrlevels(dev, gpu);
1118         if (ret)
1119                 return ret;
1120
1121         pm_runtime_set_autosuspend_delay(dev,
1122                 adreno_gpu->info->inactive_period);
1123         pm_runtime_use_autosuspend(dev);
1124
1125         return msm_gpu_init(drm, pdev, &adreno_gpu->base, &funcs->base,
1126                         gpu_name, &adreno_gpu_config);
1127 }
1128
1129 void adreno_gpu_cleanup(struct adreno_gpu *adreno_gpu)
1130 {
1131         struct msm_gpu *gpu = &adreno_gpu->base;
1132         struct msm_drm_private *priv = gpu->dev ? gpu->dev->dev_private : NULL;
1133         unsigned int i;
1134
1135         for (i = 0; i < ARRAY_SIZE(adreno_gpu->info->fw); i++)
1136                 release_firmware(adreno_gpu->fw[i]);
1137
1138         if (priv && pm_runtime_enabled(&priv->gpu_pdev->dev))
1139                 pm_runtime_disable(&priv->gpu_pdev->dev);
1140
1141         msm_gpu_cleanup(&adreno_gpu->base);
1142 }