drivers/gpu/drm/v3d/v3d_gem.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /* Copyright (C) 2014-2018 Broadcom */
   3
   4 #include <drm/drmP.h>
   5 #include <drm/drm_syncobj.h>
   6 #include <linux/module.h>
   7 #include <linux/platform_device.h>
   8 #include <linux/pm_runtime.h>
   9 #include <linux/reset.h>
  10 #include <linux/device.h>
  11 #include <linux/io.h>
  12 #include <linux/sched/signal.h>
  13
  14 #include "uapi/drm/v3d_drm.h"
  15 #include "v3d_drv.h"
  16 #include "v3d_regs.h"
  17 #include "v3d_trace.h"
  18
  19 static void
  20 v3d_init_core(struct v3d_dev *v3d, int core)
  21 {
  22         /* Set OVRTMUOUT, which means that the texture sampler uniform
  23          * configuration's tmu output type field is used, instead of
  24          * using the hardware default behavior based on the texture
  25          * type.  If you want the default behavior, you can still put
  26          * "2" in the indirect texture state's output_type field.
  27          */
  28         if (v3d->ver < 40)
  29                 V3D_CORE_WRITE(core, V3D_CTL_MISCCFG, V3D_MISCCFG_OVRTMUOUT);
  30
  31         /* Whenever we flush the L2T cache, we always want to flush
  32          * the whole thing.
  33          */
  34         V3D_CORE_WRITE(core, V3D_CTL_L2TFLSTA, 0);
  35         V3D_CORE_WRITE(core, V3D_CTL_L2TFLEND, ~0);
  36 }
  37
  38 /* Sets invariant state for the HW. */
  39 static void
  40 v3d_init_hw_state(struct v3d_dev *v3d)
  41 {
  42         v3d_init_core(v3d, 0);
  43 }
  44
  45 static void
  46 v3d_idle_axi(struct v3d_dev *v3d, int core)
  47 {
  48         V3D_CORE_WRITE(core, V3D_GMP_CFG, V3D_GMP_CFG_STOP_REQ);
  49
  50         if (wait_for((V3D_CORE_READ(core, V3D_GMP_STATUS) &
  51                       (V3D_GMP_STATUS_RD_COUNT_MASK |
  52                        V3D_GMP_STATUS_WR_COUNT_MASK |
  53                        V3D_GMP_STATUS_CFG_BUSY)) == 0, 100)) {
  54                 DRM_ERROR("Failed to wait for safe GMP shutdown\n");
  55         }
  56 }
  57
  58 static void
  59 v3d_idle_gca(struct v3d_dev *v3d)
  60 {
  61         if (v3d->ver >= 41)
  62                 return;
  63
  64         V3D_GCA_WRITE(V3D_GCA_SAFE_SHUTDOWN, V3D_GCA_SAFE_SHUTDOWN_EN);
  65
  66         if (wait_for((V3D_GCA_READ(V3D_GCA_SAFE_SHUTDOWN_ACK) &
  67                       V3D_GCA_SAFE_SHUTDOWN_ACK_ACKED) ==
  68                      V3D_GCA_SAFE_SHUTDOWN_ACK_ACKED, 100)) {
  69                 DRM_ERROR("Failed to wait for safe GCA shutdown\n");
  70         }
  71 }
  72
  73 static void
  74 v3d_reset_by_bridge(struct v3d_dev *v3d)
  75 {
  76         int version = V3D_BRIDGE_READ(V3D_TOP_GR_BRIDGE_REVISION);
  77
  78         if (V3D_GET_FIELD(version, V3D_TOP_GR_BRIDGE_MAJOR) == 2) {
  79                 V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_0,
  80                                  V3D_TOP_GR_BRIDGE_SW_INIT_0_V3D_CLK_108_SW_INIT);
  81                 V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_0, 0);
  82
  83                 /* GFXH-1383: The SW_INIT may cause a stray write to address 0
  84                  * of the unit, so reset it to its power-on value here.
  85                  */
  86                 V3D_WRITE(V3D_HUB_AXICFG, V3D_HUB_AXICFG_MAX_LEN_MASK);
  87         } else {
  88                 WARN_ON_ONCE(V3D_GET_FIELD(version,
  89                                            V3D_TOP_GR_BRIDGE_MAJOR) != 7);
  90                 V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_1,
  91                                  V3D_TOP_GR_BRIDGE_SW_INIT_1_V3D_CLK_108_SW_INIT);
  92                 V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_1, 0);
  93         }
  94 }
  95
  96 static void
  97 v3d_reset_v3d(struct v3d_dev *v3d)
  98 {
  99         if (v3d->reset)
 100                 reset_control_reset(v3d->reset);
 101         else
 102                 v3d_reset_by_bridge(v3d);
 103
 104         v3d_init_hw_state(v3d);
 105 }
 106
 107 void
 108 v3d_reset(struct v3d_dev *v3d)
 109 {
 110         struct drm_device *dev = &v3d->drm;
 111
 112         DRM_DEV_ERROR(dev->dev, "Resetting GPU for hang.\n");
 113         DRM_DEV_ERROR(dev->dev, "V3D_ERR_STAT: 0x%08x\n",
 114                       V3D_CORE_READ(0, V3D_ERR_STAT));
 115         trace_v3d_reset_begin(dev);
 116
 117         /* XXX: only needed for safe powerdown, not reset. */
 118         if (false)
 119                 v3d_idle_axi(v3d, 0);
 120
 121         v3d_idle_gca(v3d);
 122         v3d_reset_v3d(v3d);
 123
 124         v3d_mmu_set_page_table(v3d);
 125         v3d_irq_reset(v3d);
 126
 127         trace_v3d_reset_end(dev);
 128 }
 129
 130 static void
 131 v3d_flush_l3(struct v3d_dev *v3d)
 132 {
 133         if (v3d->ver < 41) {
 134                 u32 gca_ctrl = V3D_GCA_READ(V3D_GCA_CACHE_CTRL);
 135
 136                 V3D_GCA_WRITE(V3D_GCA_CACHE_CTRL,
 137                               gca_ctrl | V3D_GCA_CACHE_CTRL_FLUSH);
 138
 139                 if (v3d->ver < 33) {
 140                         V3D_GCA_WRITE(V3D_GCA_CACHE_CTRL,
 141                                       gca_ctrl & ~V3D_GCA_CACHE_CTRL_FLUSH);
 142                 }
 143         }
 144 }
 145
 146 /* Invalidates the (read-only) L2C cache.  This was the L2 cache for
 147  * uniforms and instructions on V3D 3.2.
 148  */
 149 static void
 150 v3d_invalidate_l2c(struct v3d_dev *v3d, int core)
 151 {
 152         if (v3d->ver > 32)
 153                 return;
 154
 155         V3D_CORE_WRITE(core, V3D_CTL_L2CACTL,
 156                        V3D_L2CACTL_L2CCLR |
 157                        V3D_L2CACTL_L2CENA);
 158 }
 159
 160 /* Invalidates texture L2 cachelines */
 161 static void
 162 v3d_flush_l2t(struct v3d_dev *v3d, int core)
 163 {
 164         /* While there is a busy bit (V3D_L2TCACTL_L2TFLS), we don't
 165          * need to wait for completion before dispatching the job --
 166          * L2T accesses will be stalled until the flush has completed.
 167          * However, we do need to make sure we don't try to trigger a
 168          * new flush while the L2_CLEAN queue is trying to
 169          * synchronously clean after a job.
 170          */
 171         mutex_lock(&v3d->cache_clean_lock);
 172         V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL,
 173                        V3D_L2TCACTL_L2TFLS |
 174                        V3D_SET_FIELD(V3D_L2TCACTL_FLM_FLUSH, V3D_L2TCACTL_FLM));
 175         mutex_unlock(&v3d->cache_clean_lock);
 176 }
 177
 178 /* Cleans texture L1 and L2 cachelines (writing back dirty data).
 179  *
 180  * For cleaning, which happens from the CACHE_CLEAN queue after CSD has
 181  * executed, we need to make sure that the clean is done before
 182  * signaling job completion.  So, we synchronously wait before
 183  * returning, and we make sure that L2 invalidates don't happen in the
 184  * meantime to confuse our are-we-done checks.
 185  */
 186 void
 187 v3d_clean_caches(struct v3d_dev *v3d)
 188 {
 189         struct drm_device *dev = &v3d->drm;
 190         int core = 0;
 191
 192         trace_v3d_cache_clean_begin(dev);
 193
 194         V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL, V3D_L2TCACTL_TMUWCF);
 195         if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
 196                        V3D_L2TCACTL_L2TFLS), 100)) {
 197                 DRM_ERROR("Timeout waiting for L1T write combiner flush\n");
 198         }
 199
 200         mutex_lock(&v3d->cache_clean_lock);
 201         V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL,
 202                        V3D_L2TCACTL_L2TFLS |
 203                        V3D_SET_FIELD(V3D_L2TCACTL_FLM_CLEAN, V3D_L2TCACTL_FLM));
 204
 205         if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
 206                        V3D_L2TCACTL_L2TFLS), 100)) {
 207                 DRM_ERROR("Timeout waiting for L2T clean\n");
 208         }
 209
 210         mutex_unlock(&v3d->cache_clean_lock);
 211
 212         trace_v3d_cache_clean_end(dev);
 213 }
 214
 215 /* Invalidates the slice caches.  These are read-only caches. */
 216 static void
 217 v3d_invalidate_slices(struct v3d_dev *v3d, int core)
 218 {
 219         V3D_CORE_WRITE(core, V3D_CTL_SLCACTL,
 220                        V3D_SET_FIELD(0xf, V3D_SLCACTL_TVCCS) |
 221                        V3D_SET_FIELD(0xf, V3D_SLCACTL_TDCCS) |
 222                        V3D_SET_FIELD(0xf, V3D_SLCACTL_UCC) |
 223                        V3D_SET_FIELD(0xf, V3D_SLCACTL_ICC));
 224 }
 225
 226 void
 227 v3d_invalidate_caches(struct v3d_dev *v3d)
 228 {
 229         /* Invalidate the caches from the outside in.  That way if
 230          * another CL's concurrent use of nearby memory were to pull
 231          * an invalidated cacheline back in, we wouldn't leave stale
 232          * data in the inner cache.
 233          */
 234         v3d_flush_l3(v3d);
 235         v3d_invalidate_l2c(v3d, 0);
 236         v3d_flush_l2t(v3d, 0);
 237         v3d_invalidate_slices(v3d, 0);
 238 }
 239
 240 /* Takes the reservation lock on all the BOs being referenced, so that
 241  * at queue submit time we can update the reservations.
 242  *
 243  * We don't lock the RCL the tile alloc/state BOs, or overflow memory
 244  * (all of which are on exec->unref_list).  They're entirely private
 245  * to v3d, so we don't attach dma-buf fences to them.
 246  */
 247 static int
 248 v3d_lock_bo_reservations(struct v3d_job *job,
 249                          struct ww_acquire_ctx *acquire_ctx)
 250 {
 251         int i, ret;
 252
 253         ret = drm_gem_lock_reservations(job->bo, job->bo_count, acquire_ctx);
 254         if (ret)
 255                 return ret;
 256
 257         for (i = 0; i < job->bo_count; i++) {
 258                 ret = drm_gem_fence_array_add_implicit(&job->deps,
 259                                                        job->bo[i], true);
 260                 if (ret) {
 261                         drm_gem_unlock_reservations(job->bo, job->bo_count,
 262                                                     acquire_ctx);
 263                         return ret;
 264                 }
 265         }
 266
 267         return 0;
 268 }
 269
 270 /**
 271  * v3d_lookup_bos() - Sets up job->bo[] with the GEM objects
 272  * referenced by the job.
 273  * @dev: DRM device
 274  * @file_priv: DRM file for this fd
 275  * @job: V3D job being set up
 276  *
 277  * The command validator needs to reference BOs by their index within
 278  * the submitted job's BO list.  This does the validation of the job's
 279  * BO list and reference counting for the lifetime of the job.
 280  *
 281  * Note that this function doesn't need to unreference the BOs on
 282  * failure, because that will happen at v3d_exec_cleanup() time.
 283  */
 284 static int
 285 v3d_lookup_bos(struct drm_device *dev,
 286                struct drm_file *file_priv,
 287                struct v3d_job *job,
 288                u64 bo_handles,
 289                u32 bo_count)
 290 {
 291         u32 *handles;
 292         int ret = 0;
 293         int i;
 294
 295         job->bo_count = bo_count;
 296
 297         if (!job->bo_count) {
 298                 /* See comment on bo_index for why we have to check
 299                  * this.
 300                  */
 301                 DRM_DEBUG("Rendering requires BOs\n");
 302                 return -EINVAL;
 303         }
 304
 305         job->bo = kvmalloc_array(job->bo_count,
 306                                  sizeof(struct drm_gem_cma_object *),
 307                                  GFP_KERNEL | __GFP_ZERO);
 308         if (!job->bo) {
 309                 DRM_DEBUG("Failed to allocate validated BO pointers\n");
 310                 return -ENOMEM;
 311         }
 312
 313         handles = kvmalloc_array(job->bo_count, sizeof(u32), GFP_KERNEL);
 314         if (!handles) {
 315                 ret = -ENOMEM;
 316                 DRM_DEBUG("Failed to allocate incoming GEM handles\n");
 317                 goto fail;
 318         }
 319
 320         if (copy_from_user(handles,
 321                            (void __user *)(uintptr_t)bo_handles,
 322                            job->bo_count * sizeof(u32))) {
 323                 ret = -EFAULT;
 324                 DRM_DEBUG("Failed to copy in GEM handles\n");
 325                 goto fail;
 326         }
 327
 328         spin_lock(&file_priv->table_lock);
 329         for (i = 0; i < job->bo_count; i++) {
 330                 struct drm_gem_object *bo = idr_find(&file_priv->object_idr,
 331                                                      handles[i]);
 332                 if (!bo) {
 333                         DRM_DEBUG("Failed to look up GEM BO %d: %d\n",
 334                                   i, handles[i]);
 335                         ret = -ENOENT;
 336                         spin_unlock(&file_priv->table_lock);
 337                         goto fail;
 338                 }
 339                 drm_gem_object_get(bo);
 340                 job->bo[i] = bo;
 341         }
 342         spin_unlock(&file_priv->table_lock);
 343
 344 fail:
 345         kvfree(handles);
 346         return ret;
 347 }
 348
 349 static void
 350 v3d_job_free(struct kref *ref)
 351 {
 352         struct v3d_job *job = container_of(ref, struct v3d_job, refcount);
 353         unsigned long index;
 354         struct dma_fence *fence;
 355         int i;
 356
 357         for (i = 0; i < job->bo_count; i++) {
 358                 if (job->bo[i])
 359                         drm_gem_object_put_unlocked(job->bo[i]);
 360         }
 361         kvfree(job->bo);
 362
 363         xa_for_each(&job->deps, index, fence) {
 364                 dma_fence_put(fence);
 365         }
 366         xa_destroy(&job->deps);
 367
 368         dma_fence_put(job->irq_fence);
 369         dma_fence_put(job->done_fence);
 370
 371         pm_runtime_mark_last_busy(job->v3d->dev);
 372         pm_runtime_put_autosuspend(job->v3d->dev);
 373
 374         kfree(job);
 375 }
 376
 377 static void
 378 v3d_render_job_free(struct kref *ref)
 379 {
 380         struct v3d_render_job *job = container_of(ref, struct v3d_render_job,
 381                                                   base.refcount);
 382         struct v3d_bo *bo, *save;
 383
 384         list_for_each_entry_safe(bo, save, &job->unref_list, unref_head) {
 385                 drm_gem_object_put_unlocked(&bo->base.base);
 386         }
 387
 388         v3d_job_free(ref);
 389 }
 390
 391 void v3d_job_put(struct v3d_job *job)
 392 {
 393         kref_put(&job->refcount, job->free);
 394 }
 395
 396 int
 397 v3d_wait_bo_ioctl(struct drm_device *dev, void *data,
 398                   struct drm_file *file_priv)
 399 {
 400         int ret;
 401         struct drm_v3d_wait_bo *args = data;
 402         ktime_t start = ktime_get();
 403         u64 delta_ns;
 404         unsigned long timeout_jiffies =
 405                 nsecs_to_jiffies_timeout(args->timeout_ns);
 406
 407         if (args->pad != 0)
 408                 return -EINVAL;
 409
 410         ret = drm_gem_reservation_object_wait(file_priv, args->handle,
 411                                               true, timeout_jiffies);
 412
 413         /* Decrement the user's timeout, in case we got interrupted
 414          * such that the ioctl will be restarted.
 415          */
 416         delta_ns = ktime_to_ns(ktime_sub(ktime_get(), start));
 417         if (delta_ns < args->timeout_ns)
 418                 args->timeout_ns -= delta_ns;
 419         else
 420                 args->timeout_ns = 0;
 421
 422         /* Asked to wait beyond the jiffie/scheduler precision? */
 423         if (ret == -ETIME && args->timeout_ns)
 424                 ret = -EAGAIN;
 425
 426         return ret;
 427 }
 428
 429 static int
 430 v3d_job_init(struct v3d_dev *v3d, struct drm_file *file_priv,
 431              struct v3d_job *job, void (*free)(struct kref *ref),
 432              u32 in_sync)
 433 {
 434         struct dma_fence *in_fence = NULL;
 435         int ret;
 436
 437         job->v3d = v3d;
 438         job->free = free;
 439
 440         ret = pm_runtime_get_sync(v3d->dev);
 441         if (ret < 0)
 442                 return ret;
 443
 444         xa_init_flags(&job->deps, XA_FLAGS_ALLOC);
 445
 446         ret = drm_syncobj_find_fence(file_priv, in_sync, 0, 0, &in_fence);
 447         if (ret == -EINVAL)
 448                 goto fail;
 449
 450         ret = drm_gem_fence_array_add(&job->deps, in_fence);
 451         if (ret)
 452                 goto fail;
 453
 454         kref_init(&job->refcount);
 455
 456         return 0;
 457 fail:
 458         xa_destroy(&job->deps);
 459         pm_runtime_put_autosuspend(v3d->dev);
 460         return ret;
 461 }
 462
 463 static int
 464 v3d_push_job(struct v3d_file_priv *v3d_priv,
 465              struct v3d_job *job, enum v3d_queue queue)
 466 {
 467         int ret;
 468
 469         ret = drm_sched_job_init(&job->base, &v3d_priv->sched_entity[queue],
 470                                  v3d_priv);
 471         if (ret)
 472                 return ret;
 473
 474         job->done_fence = dma_fence_get(&job->base.s_fence->finished);
 475
 476         /* put by scheduler job completion */
 477         kref_get(&job->refcount);
 478
 479         drm_sched_entity_push_job(&job->base, &v3d_priv->sched_entity[queue]);
 480
 481         return 0;
 482 }
 483
 484 static void
 485 v3d_attach_fences_and_unlock_reservation(struct drm_file *file_priv,
 486                                          struct v3d_job *job,
 487                                          struct ww_acquire_ctx *acquire_ctx,
 488                                          u32 out_sync,
 489                                          struct dma_fence *done_fence)
 490 {
 491         struct drm_syncobj *sync_out;
 492         int i;
 493
 494         for (i = 0; i < job->bo_count; i++) {
 495                 /* XXX: Use shared fences for read-only objects. */
 496                 reservation_object_add_excl_fence(job->bo[i]->resv,
 497                                                   job->done_fence);
 498         }
 499
 500         drm_gem_unlock_reservations(job->bo, job->bo_count, acquire_ctx);
 501
 502         /* Update the return sync object for the job */
 503         sync_out = drm_syncobj_find(file_priv, out_sync);
 504         if (sync_out) {
 505                 drm_syncobj_replace_fence(sync_out, done_fence);
 506                 drm_syncobj_put(sync_out);
 507         }
 508 }
 509
 510 /**
 511  * v3d_submit_cl_ioctl() - Submits a job (frame) to the V3D.
 512  * @dev: DRM device
 513  * @data: ioctl argument
 514  * @file_priv: DRM file for this fd
 515  *
 516  * This is the main entrypoint for userspace to submit a 3D frame to
 517  * the GPU.  Userspace provides the binner command list (if
 518  * applicable), and the kernel sets up the render command list to draw
 519  * to the framebuffer described in the ioctl, using the command lists
 520  * that the 3D engine's binner will produce.
 521  */
 522 int
 523 v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
 524                     struct drm_file *file_priv)
 525 {
 526         struct v3d_dev *v3d = to_v3d_dev(dev);
 527         struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
 528         struct drm_v3d_submit_cl *args = data;
 529         struct v3d_bin_job *bin = NULL;
 530         struct v3d_render_job *render;
 531         struct ww_acquire_ctx acquire_ctx;
 532         int ret = 0;
 533
 534         trace_v3d_submit_cl_ioctl(&v3d->drm, args->rcl_start, args->rcl_end);
 535
 536         if (args->pad != 0) {
 537                 DRM_INFO("pad must be zero: %d\n", args->pad);
 538                 return -EINVAL;
 539         }
 540
 541         render = kcalloc(1, sizeof(*render), GFP_KERNEL);
 542         if (!render)
 543                 return -ENOMEM;
 544
 545         render->start = args->rcl_start;
 546         render->end = args->rcl_end;
 547         INIT_LIST_HEAD(&render->unref_list);
 548
 549         ret = v3d_job_init(v3d, file_priv, &render->base,
 550                            v3d_render_job_free, args->in_sync_rcl);
 551         if (ret) {
 552                 kfree(render);
 553                 return ret;
 554         }
 555
 556         if (args->bcl_start != args->bcl_end) {
 557                 bin = kcalloc(1, sizeof(*bin), GFP_KERNEL);
 558                 if (!bin) {
 559                         v3d_job_put(&render->base);
 560                         return -ENOMEM;
 561                 }
 562
 563                 ret = v3d_job_init(v3d, file_priv, &bin->base,
 564                                    v3d_job_free, args->in_sync_bcl);
 565                 if (ret) {
 566                         v3d_job_put(&render->base);
 567                         kfree(bin);
 568                         return ret;
 569                 }
 570
 571                 bin->start = args->bcl_start;
 572                 bin->end = args->bcl_end;
 573                 bin->qma = args->qma;
 574                 bin->qms = args->qms;
 575                 bin->qts = args->qts;
 576                 bin->render = render;
 577         }
 578
 579         ret = v3d_lookup_bos(dev, file_priv, &render->base,
 580                              args->bo_handles, args->bo_handle_count);
 581         if (ret)
 582                 goto fail;
 583
 584         ret = v3d_lock_bo_reservations(&render->base, &acquire_ctx);
 585         if (ret)
 586                 goto fail;
 587
 588         mutex_lock(&v3d->sched_lock);
 589         if (bin) {
 590                 ret = v3d_push_job(v3d_priv, &bin->base, V3D_BIN);
 591                 if (ret)
 592                         goto fail_unreserve;
 593
 594                 ret = drm_gem_fence_array_add(&render->base.deps,
 595                                               dma_fence_get(bin->base.done_fence));
 596                 if (ret)
 597                         goto fail_unreserve;
 598         }
 599
 600         ret = v3d_push_job(v3d_priv, &render->base, V3D_RENDER);
 601         if (ret)
 602                 goto fail_unreserve;
 603         mutex_unlock(&v3d->sched_lock);
 604
 605         v3d_attach_fences_and_unlock_reservation(file_priv,
 606                                                  &render->base,
 607                                                  &acquire_ctx,
 608                                                  args->out_sync,
 609                                                  render->base.done_fence);
 610
 611         if (bin)
 612                 v3d_job_put(&bin->base);
 613         v3d_job_put(&render->base);
 614
 615         return 0;
 616
 617 fail_unreserve:
 618         mutex_unlock(&v3d->sched_lock);
 619         drm_gem_unlock_reservations(render->base.bo,
 620                                     render->base.bo_count, &acquire_ctx);
 621 fail:
 622         if (bin)
 623                 v3d_job_put(&bin->base);
 624         v3d_job_put(&render->base);
 625
 626         return ret;
 627 }
 628
 629 /**
 630  * v3d_submit_tfu_ioctl() - Submits a TFU (texture formatting) job to the V3D.
 631  * @dev: DRM device
 632  * @data: ioctl argument
 633  * @file_priv: DRM file for this fd
 634  *
 635  * Userspace provides the register setup for the TFU, which we don't
 636  * need to validate since the TFU is behind the MMU.
 637  */
 638 int
 639 v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
 640                      struct drm_file *file_priv)
 641 {
 642         struct v3d_dev *v3d = to_v3d_dev(dev);
 643         struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
 644         struct drm_v3d_submit_tfu *args = data;
 645         struct v3d_tfu_job *job;
 646         struct ww_acquire_ctx acquire_ctx;
 647         int ret = 0;
 648
 649         trace_v3d_submit_tfu_ioctl(&v3d->drm, args->iia);
 650
 651         job = kcalloc(1, sizeof(*job), GFP_KERNEL);
 652         if (!job)
 653                 return -ENOMEM;
 654
 655         ret = v3d_job_init(v3d, file_priv, &job->base,
 656                            v3d_job_free, args->in_sync);
 657         if (ret) {
 658                 kfree(job);
 659                 return ret;
 660         }
 661
 662         job->base.bo = kcalloc(ARRAY_SIZE(args->bo_handles),
 663                                sizeof(*job->base.bo), GFP_KERNEL);
 664         if (!job->base.bo) {
 665                 v3d_job_put(&job->base);
 666                 return -ENOMEM;
 667         }
 668
 669         job->args = *args;
 670
 671         spin_lock(&file_priv->table_lock);
 672         for (job->base.bo_count = 0;
 673              job->base.bo_count < ARRAY_SIZE(args->bo_handles);
 674              job->base.bo_count++) {
 675                 struct drm_gem_object *bo;
 676
 677                 if (!args->bo_handles[job->base.bo_count])
 678                         break;
 679
 680                 bo = idr_find(&file_priv->object_idr,
 681                               args->bo_handles[job->base.bo_count]);
 682                 if (!bo) {
 683                         DRM_DEBUG("Failed to look up GEM BO %d: %d\n",
 684                                   job->base.bo_count,
 685                                   args->bo_handles[job->base.bo_count]);
 686                         ret = -ENOENT;
 687                         spin_unlock(&file_priv->table_lock);
 688                         goto fail;
 689                 }
 690                 drm_gem_object_get(bo);
 691                 job->base.bo[job->base.bo_count] = bo;
 692         }
 693         spin_unlock(&file_priv->table_lock);
 694
 695         ret = v3d_lock_bo_reservations(&job->base, &acquire_ctx);
 696         if (ret)
 697                 goto fail;
 698
 699         mutex_lock(&v3d->sched_lock);
 700         ret = v3d_push_job(v3d_priv, &job->base, V3D_TFU);
 701         if (ret)
 702                 goto fail_unreserve;
 703         mutex_unlock(&v3d->sched_lock);
 704
 705         v3d_attach_fences_and_unlock_reservation(file_priv,
 706                                                  &job->base, &acquire_ctx,
 707                                                  args->out_sync,
 708                                                  job->base.done_fence);
 709
 710         v3d_job_put(&job->base);
 711
 712         return 0;
 713
 714 fail_unreserve:
 715         mutex_unlock(&v3d->sched_lock);
 716         drm_gem_unlock_reservations(job->base.bo, job->base.bo_count,
 717                                     &acquire_ctx);
 718 fail:
 719         v3d_job_put(&job->base);
 720
 721         return ret;
 722 }
 723
 724 /**
 725  * v3d_submit_csd_ioctl() - Submits a CSD (texture formatting) job to the V3D.
 726  * @dev: DRM device
 727  * @data: ioctl argument
 728  * @file_priv: DRM file for this fd
 729  *
 730  * Userspace provides the register setup for the CSD, which we don't
 731  * need to validate since the CSD is behind the MMU.
 732  */
 733 int
 734 v3d_submit_csd_ioctl(struct drm_device *dev, void *data,
 735                      struct drm_file *file_priv)
 736 {
 737         struct v3d_dev *v3d = to_v3d_dev(dev);
 738         struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
 739         struct drm_v3d_submit_csd *args = data;
 740         struct v3d_csd_job *job;
 741         struct v3d_job *clean_job;
 742         struct ww_acquire_ctx acquire_ctx;
 743         int ret;
 744
 745         trace_v3d_submit_csd_ioctl(&v3d->drm, args->cfg[5], args->cfg[6]);
 746
 747         if (!v3d_has_csd(v3d)) {
 748                 DRM_DEBUG("Attempting CSD submit on non-CSD hardware\n");
 749                 return -EINVAL;
 750         }
 751
 752         job = kcalloc(1, sizeof(*job), GFP_KERNEL);
 753         if (!job)
 754                 return -ENOMEM;
 755
 756         ret = v3d_job_init(v3d, file_priv, &job->base,
 757                            v3d_job_free, args->in_sync);
 758         if (ret) {
 759                 kfree(job);
 760                 return ret;
 761         }
 762
 763         clean_job = kcalloc(1, sizeof(*clean_job), GFP_KERNEL);
 764         if (!clean_job) {
 765                 v3d_job_put(&job->base);
 766                 kfree(job);
 767                 return -ENOMEM;
 768         }
 769
 770         ret = v3d_job_init(v3d, file_priv, clean_job, v3d_job_free, 0);
 771         if (ret) {
 772                 v3d_job_put(&job->base);
 773                 kfree(clean_job);
 774                 return ret;
 775         }
 776
 777         job->args = *args;
 778
 779         ret = v3d_lookup_bos(dev, file_priv, clean_job,
 780                              args->bo_handles, args->bo_handle_count);
 781         if (ret)
 782                 goto fail;
 783
 784         ret = v3d_lock_bo_reservations(clean_job, &acquire_ctx);
 785         if (ret)
 786                 goto fail;
 787
 788         mutex_lock(&v3d->sched_lock);
 789         ret = v3d_push_job(v3d_priv, &job->base, V3D_CSD);
 790         if (ret)
 791                 goto fail_unreserve;
 792
 793         ret = drm_gem_fence_array_add(&clean_job->deps,
 794                                       dma_fence_get(job->base.done_fence));
 795         if (ret)
 796                 goto fail_unreserve;
 797
 798         ret = v3d_push_job(v3d_priv, clean_job, V3D_CACHE_CLEAN);
 799         if (ret)
 800                 goto fail_unreserve;
 801         mutex_unlock(&v3d->sched_lock);
 802
 803         v3d_attach_fences_and_unlock_reservation(file_priv,
 804                                                  clean_job,
 805                                                  &acquire_ctx,
 806                                                  args->out_sync,
 807                                                  clean_job->done_fence);
 808
 809         v3d_job_put(&job->base);
 810         v3d_job_put(clean_job);
 811
 812         return 0;
 813
 814 fail_unreserve:
 815         mutex_unlock(&v3d->sched_lock);
 816         drm_gem_unlock_reservations(clean_job->bo, clean_job->bo_count,
 817                                     &acquire_ctx);
 818 fail:
 819         v3d_job_put(&job->base);
 820         v3d_job_put(clean_job);
 821
 822         return ret;
 823 }
 824
 825 int
 826 v3d_gem_init(struct drm_device *dev)
 827 {
 828         struct v3d_dev *v3d = to_v3d_dev(dev);
 829         u32 pt_size = 4096 * 1024;
 830         int ret, i;
 831
 832         for (i = 0; i < V3D_MAX_QUEUES; i++)
 833                 v3d->queue[i].fence_context = dma_fence_context_alloc(1);
 834
 835         spin_lock_init(&v3d->mm_lock);
 836         spin_lock_init(&v3d->job_lock);
 837         mutex_init(&v3d->bo_lock);
 838         mutex_init(&v3d->reset_lock);
 839         mutex_init(&v3d->sched_lock);
 840         mutex_init(&v3d->cache_clean_lock);
 841
 842         /* Note: We don't allocate address 0.  Various bits of HW
 843          * treat 0 as special, such as the occlusion query counters
 844          * where 0 means "disabled".
 845          */
 846         drm_mm_init(&v3d->mm, 1, pt_size / sizeof(u32) - 1);
 847
 848         v3d->pt = dma_alloc_wc(v3d->dev, pt_size,
 849                                &v3d->pt_paddr,
 850                                GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO);
 851         if (!v3d->pt) {
 852                 drm_mm_takedown(&v3d->mm);
 853                 dev_err(v3d->dev,
 854                         "Failed to allocate page tables. "
 855                         "Please ensure you have CMA enabled.\n");
 856                 return -ENOMEM;
 857         }
 858
 859         v3d_init_hw_state(v3d);
 860         v3d_mmu_set_page_table(v3d);
 861
 862         ret = v3d_sched_init(v3d);
 863         if (ret) {
 864                 drm_mm_takedown(&v3d->mm);
 865                 dma_free_coherent(v3d->dev, 4096 * 1024, (void *)v3d->pt,
 866                                   v3d->pt_paddr);
 867         }
 868
 869         return 0;
 870 }
 871
 872 void
 873 v3d_gem_destroy(struct drm_device *dev)
 874 {
 875         struct v3d_dev *v3d = to_v3d_dev(dev);
 876
 877         v3d_sched_fini(v3d);
 878
 879         /* Waiting for jobs to finish would need to be done before
 880          * unregistering V3D.
 881          */
 882         WARN_ON(v3d->bin_job);
 883         WARN_ON(v3d->render_job);
 884
 885         drm_mm_takedown(&v3d->mm);
 886
 887         dma_free_coherent(v3d->dev, 4096 * 1024, (void *)v3d->pt, v3d->pt_paddr);
 888 }