drivers/gpu/drm/v3d/v3d_gem.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /* Copyright (C) 2014-2018 Broadcom */
   3
   4 #include <linux/device.h>
   5 #include <linux/dma-mapping.h>
   6 #include <linux/io.h>
   7 #include <linux/module.h>
   8 #include <linux/platform_device.h>
   9 #include <linux/reset.h>
  10 #include <linux/sched/signal.h>
  11 #include <linux/uaccess.h>
  12
  13 #include <drm/drm_managed.h>
  14 #include <drm/drm_syncobj.h>
  15 #include <uapi/drm/v3d_drm.h>
  16
  17 #include "v3d_drv.h"
  18 #include "v3d_regs.h"
  19 #include "v3d_trace.h"
  20
  21 static void
  22 v3d_init_core(struct v3d_dev *v3d, int core)
  23 {
  24         /* Set OVRTMUOUT, which means that the texture sampler uniform
  25          * configuration's tmu output type field is used, instead of
  26          * using the hardware default behavior based on the texture
  27          * type.  If you want the default behavior, you can still put
  28          * "2" in the indirect texture state's output_type field.
  29          */
  30         if (v3d->ver < 40)
  31                 V3D_CORE_WRITE(core, V3D_CTL_MISCCFG, V3D_MISCCFG_OVRTMUOUT);
  32
  33         /* Whenever we flush the L2T cache, we always want to flush
  34          * the whole thing.
  35          */
  36         V3D_CORE_WRITE(core, V3D_CTL_L2TFLSTA, 0);
  37         V3D_CORE_WRITE(core, V3D_CTL_L2TFLEND, ~0);
  38 }
  39
  40 /* Sets invariant state for the HW. */
  41 static void
  42 v3d_init_hw_state(struct v3d_dev *v3d)
  43 {
  44         v3d_init_core(v3d, 0);
  45 }
  46
  47 static void
  48 v3d_idle_axi(struct v3d_dev *v3d, int core)
  49 {
  50         V3D_CORE_WRITE(core, V3D_GMP_CFG, V3D_GMP_CFG_STOP_REQ);
  51
  52         if (wait_for((V3D_CORE_READ(core, V3D_GMP_STATUS) &
  53                       (V3D_GMP_STATUS_RD_COUNT_MASK |
  54                        V3D_GMP_STATUS_WR_COUNT_MASK |
  55                        V3D_GMP_STATUS_CFG_BUSY)) == 0, 100)) {
  56                 DRM_ERROR("Failed to wait for safe GMP shutdown\n");
  57         }
  58 }
  59
  60 static void
  61 v3d_idle_gca(struct v3d_dev *v3d)
  62 {
  63         if (v3d->ver >= 41)
  64                 return;
  65
  66         V3D_GCA_WRITE(V3D_GCA_SAFE_SHUTDOWN, V3D_GCA_SAFE_SHUTDOWN_EN);
  67
  68         if (wait_for((V3D_GCA_READ(V3D_GCA_SAFE_SHUTDOWN_ACK) &
  69                       V3D_GCA_SAFE_SHUTDOWN_ACK_ACKED) ==
  70                      V3D_GCA_SAFE_SHUTDOWN_ACK_ACKED, 100)) {
  71                 DRM_ERROR("Failed to wait for safe GCA shutdown\n");
  72         }
  73 }
  74
  75 static void
  76 v3d_reset_by_bridge(struct v3d_dev *v3d)
  77 {
  78         int version = V3D_BRIDGE_READ(V3D_TOP_GR_BRIDGE_REVISION);
  79
  80         if (V3D_GET_FIELD(version, V3D_TOP_GR_BRIDGE_MAJOR) == 2) {
  81                 V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_0,
  82                                  V3D_TOP_GR_BRIDGE_SW_INIT_0_V3D_CLK_108_SW_INIT);
  83                 V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_0, 0);
  84
  85                 /* GFXH-1383: The SW_INIT may cause a stray write to address 0
  86                  * of the unit, so reset it to its power-on value here.
  87                  */
  88                 V3D_WRITE(V3D_HUB_AXICFG, V3D_HUB_AXICFG_MAX_LEN_MASK);
  89         } else {
  90                 WARN_ON_ONCE(V3D_GET_FIELD(version,
  91                                            V3D_TOP_GR_BRIDGE_MAJOR) != 7);
  92                 V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_1,
  93                                  V3D_TOP_GR_BRIDGE_SW_INIT_1_V3D_CLK_108_SW_INIT);
  94                 V3D_BRIDGE_WRITE(V3D_TOP_GR_BRIDGE_SW_INIT_1, 0);
  95         }
  96 }
  97
  98 static void
  99 v3d_reset_v3d(struct v3d_dev *v3d)
 100 {
 101         if (v3d->reset)
 102                 reset_control_reset(v3d->reset);
 103         else
 104                 v3d_reset_by_bridge(v3d);
 105
 106         v3d_init_hw_state(v3d);
 107 }
 108
 109 void
 110 v3d_reset(struct v3d_dev *v3d)
 111 {
 112         struct drm_device *dev = &v3d->drm;
 113
 114         DRM_DEV_ERROR(dev->dev, "Resetting GPU for hang.\n");
 115         DRM_DEV_ERROR(dev->dev, "V3D_ERR_STAT: 0x%08x\n",
 116                       V3D_CORE_READ(0, V3D_ERR_STAT));
 117         trace_v3d_reset_begin(dev);
 118
 119         /* XXX: only needed for safe powerdown, not reset. */
 120         if (false)
 121                 v3d_idle_axi(v3d, 0);
 122
 123         v3d_idle_gca(v3d);
 124         v3d_reset_v3d(v3d);
 125
 126         v3d_mmu_set_page_table(v3d);
 127         v3d_irq_reset(v3d);
 128
 129         v3d_perfmon_stop(v3d, v3d->active_perfmon, false);
 130
 131         trace_v3d_reset_end(dev);
 132 }
 133
 134 static void
 135 v3d_flush_l3(struct v3d_dev *v3d)
 136 {
 137         if (v3d->ver < 41) {
 138                 u32 gca_ctrl = V3D_GCA_READ(V3D_GCA_CACHE_CTRL);
 139
 140                 V3D_GCA_WRITE(V3D_GCA_CACHE_CTRL,
 141                               gca_ctrl | V3D_GCA_CACHE_CTRL_FLUSH);
 142
 143                 if (v3d->ver < 33) {
 144                         V3D_GCA_WRITE(V3D_GCA_CACHE_CTRL,
 145                                       gca_ctrl & ~V3D_GCA_CACHE_CTRL_FLUSH);
 146                 }
 147         }
 148 }
 149
 150 /* Invalidates the (read-only) L2C cache.  This was the L2 cache for
 151  * uniforms and instructions on V3D 3.2.
 152  */
 153 static void
 154 v3d_invalidate_l2c(struct v3d_dev *v3d, int core)
 155 {
 156         if (v3d->ver > 32)
 157                 return;
 158
 159         V3D_CORE_WRITE(core, V3D_CTL_L2CACTL,
 160                        V3D_L2CACTL_L2CCLR |
 161                        V3D_L2CACTL_L2CENA);
 162 }
 163
 164 /* Invalidates texture L2 cachelines */
 165 static void
 166 v3d_flush_l2t(struct v3d_dev *v3d, int core)
 167 {
 168         /* While there is a busy bit (V3D_L2TCACTL_L2TFLS), we don't
 169          * need to wait for completion before dispatching the job --
 170          * L2T accesses will be stalled until the flush has completed.
 171          * However, we do need to make sure we don't try to trigger a
 172          * new flush while the L2_CLEAN queue is trying to
 173          * synchronously clean after a job.
 174          */
 175         mutex_lock(&v3d->cache_clean_lock);
 176         V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL,
 177                        V3D_L2TCACTL_L2TFLS |
 178                        V3D_SET_FIELD(V3D_L2TCACTL_FLM_FLUSH, V3D_L2TCACTL_FLM));
 179         mutex_unlock(&v3d->cache_clean_lock);
 180 }
 181
 182 /* Cleans texture L1 and L2 cachelines (writing back dirty data).
 183  *
 184  * For cleaning, which happens from the CACHE_CLEAN queue after CSD has
 185  * executed, we need to make sure that the clean is done before
 186  * signaling job completion.  So, we synchronously wait before
 187  * returning, and we make sure that L2 invalidates don't happen in the
 188  * meantime to confuse our are-we-done checks.
 189  */
 190 void
 191 v3d_clean_caches(struct v3d_dev *v3d)
 192 {
 193         struct drm_device *dev = &v3d->drm;
 194         int core = 0;
 195
 196         trace_v3d_cache_clean_begin(dev);
 197
 198         V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL, V3D_L2TCACTL_TMUWCF);
 199         if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
 200                        V3D_L2TCACTL_TMUWCF), 100)) {
 201                 DRM_ERROR("Timeout waiting for TMU write combiner flush\n");
 202         }
 203
 204         mutex_lock(&v3d->cache_clean_lock);
 205         V3D_CORE_WRITE(core, V3D_CTL_L2TCACTL,
 206                        V3D_L2TCACTL_L2TFLS |
 207                        V3D_SET_FIELD(V3D_L2TCACTL_FLM_CLEAN, V3D_L2TCACTL_FLM));
 208
 209         if (wait_for(!(V3D_CORE_READ(core, V3D_CTL_L2TCACTL) &
 210                        V3D_L2TCACTL_L2TFLS), 100)) {
 211                 DRM_ERROR("Timeout waiting for L2T clean\n");
 212         }
 213
 214         mutex_unlock(&v3d->cache_clean_lock);
 215
 216         trace_v3d_cache_clean_end(dev);
 217 }
 218
 219 /* Invalidates the slice caches.  These are read-only caches. */
 220 static void
 221 v3d_invalidate_slices(struct v3d_dev *v3d, int core)
 222 {
 223         V3D_CORE_WRITE(core, V3D_CTL_SLCACTL,
 224                        V3D_SET_FIELD(0xf, V3D_SLCACTL_TVCCS) |
 225                        V3D_SET_FIELD(0xf, V3D_SLCACTL_TDCCS) |
 226                        V3D_SET_FIELD(0xf, V3D_SLCACTL_UCC) |
 227                        V3D_SET_FIELD(0xf, V3D_SLCACTL_ICC));
 228 }
 229
 230 void
 231 v3d_invalidate_caches(struct v3d_dev *v3d)
 232 {
 233         /* Invalidate the caches from the outside in.  That way if
 234          * another CL's concurrent use of nearby memory were to pull
 235          * an invalidated cacheline back in, we wouldn't leave stale
 236          * data in the inner cache.
 237          */
 238         v3d_flush_l3(v3d);
 239         v3d_invalidate_l2c(v3d, 0);
 240         v3d_flush_l2t(v3d, 0);
 241         v3d_invalidate_slices(v3d, 0);
 242 }
 243
 244 /* Takes the reservation lock on all the BOs being referenced, so that
 245  * at queue submit time we can update the reservations.
 246  *
 247  * We don't lock the RCL the tile alloc/state BOs, or overflow memory
 248  * (all of which are on exec->unref_list).  They're entirely private
 249  * to v3d, so we don't attach dma-buf fences to them.
 250  */
 251 static int
 252 v3d_lock_bo_reservations(struct v3d_job *job,
 253                          struct ww_acquire_ctx *acquire_ctx)
 254 {
 255         int i, ret;
 256
 257         ret = drm_gem_lock_reservations(job->bo, job->bo_count, acquire_ctx);
 258         if (ret)
 259                 return ret;
 260
 261         for (i = 0; i < job->bo_count; i++) {
 262                 ret = dma_resv_reserve_fences(job->bo[i]->resv, 1);
 263                 if (ret)
 264                         goto fail;
 265
 266                 ret = drm_sched_job_add_implicit_dependencies(&job->base,
 267                                                               job->bo[i], true);
 268                 if (ret)
 269                         goto fail;
 270         }
 271
 272         return 0;
 273
 274 fail:
 275         drm_gem_unlock_reservations(job->bo, job->bo_count, acquire_ctx);
 276         return ret;
 277 }
 278
 279 /**
 280  * v3d_lookup_bos() - Sets up job->bo[] with the GEM objects
 281  * referenced by the job.
 282  * @dev: DRM device
 283  * @file_priv: DRM file for this fd
 284  * @job: V3D job being set up
 285  * @bo_handles: GEM handles
 286  * @bo_count: Number of GEM handles passed in
 287  *
 288  * The command validator needs to reference BOs by their index within
 289  * the submitted job's BO list.  This does the validation of the job's
 290  * BO list and reference counting for the lifetime of the job.
 291  *
 292  * Note that this function doesn't need to unreference the BOs on
 293  * failure, because that will happen at v3d_exec_cleanup() time.
 294  */
 295 static int
 296 v3d_lookup_bos(struct drm_device *dev,
 297                struct drm_file *file_priv,
 298                struct v3d_job *job,
 299                u64 bo_handles,
 300                u32 bo_count)
 301 {
 302         u32 *handles;
 303         int ret = 0;
 304         int i;
 305
 306         job->bo_count = bo_count;
 307
 308         if (!job->bo_count) {
 309                 /* See comment on bo_index for why we have to check
 310                  * this.
 311                  */
 312                 DRM_DEBUG("Rendering requires BOs\n");
 313                 return -EINVAL;
 314         }
 315
 316         job->bo = kvmalloc_array(job->bo_count,
 317                                  sizeof(struct drm_gem_dma_object *),
 318                                  GFP_KERNEL | __GFP_ZERO);
 319         if (!job->bo) {
 320                 DRM_DEBUG("Failed to allocate validated BO pointers\n");
 321                 return -ENOMEM;
 322         }
 323
 324         handles = kvmalloc_array(job->bo_count, sizeof(u32), GFP_KERNEL);
 325         if (!handles) {
 326                 ret = -ENOMEM;
 327                 DRM_DEBUG("Failed to allocate incoming GEM handles\n");
 328                 goto fail;
 329         }
 330
 331         if (copy_from_user(handles,
 332                            (void __user *)(uintptr_t)bo_handles,
 333                            job->bo_count * sizeof(u32))) {
 334                 ret = -EFAULT;
 335                 DRM_DEBUG("Failed to copy in GEM handles\n");
 336                 goto fail;
 337         }
 338
 339         spin_lock(&file_priv->table_lock);
 340         for (i = 0; i < job->bo_count; i++) {
 341                 struct drm_gem_object *bo = idr_find(&file_priv->object_idr,
 342                                                      handles[i]);
 343                 if (!bo) {
 344                         DRM_DEBUG("Failed to look up GEM BO %d: %d\n",
 345                                   i, handles[i]);
 346                         ret = -ENOENT;
 347                         spin_unlock(&file_priv->table_lock);
 348                         goto fail;
 349                 }
 350                 drm_gem_object_get(bo);
 351                 job->bo[i] = bo;
 352         }
 353         spin_unlock(&file_priv->table_lock);
 354
 355 fail:
 356         kvfree(handles);
 357         return ret;
 358 }
 359
 360 static void
 361 v3d_job_free(struct kref *ref)
 362 {
 363         struct v3d_job *job = container_of(ref, struct v3d_job, refcount);
 364         int i;
 365
 366         for (i = 0; i < job->bo_count; i++) {
 367                 if (job->bo[i])
 368                         drm_gem_object_put(job->bo[i]);
 369         }
 370         kvfree(job->bo);
 371
 372         dma_fence_put(job->irq_fence);
 373         dma_fence_put(job->done_fence);
 374
 375         if (job->perfmon)
 376                 v3d_perfmon_put(job->perfmon);
 377
 378         kfree(job);
 379 }
 380
 381 static void
 382 v3d_render_job_free(struct kref *ref)
 383 {
 384         struct v3d_render_job *job = container_of(ref, struct v3d_render_job,
 385                                                   base.refcount);
 386         struct v3d_bo *bo, *save;
 387
 388         list_for_each_entry_safe(bo, save, &job->unref_list, unref_head) {
 389                 drm_gem_object_put(&bo->base.base);
 390         }
 391
 392         v3d_job_free(ref);
 393 }
 394
 395 void v3d_job_cleanup(struct v3d_job *job)
 396 {
 397         if (!job)
 398                 return;
 399
 400         drm_sched_job_cleanup(&job->base);
 401         v3d_job_put(job);
 402 }
 403
 404 void v3d_job_put(struct v3d_job *job)
 405 {
 406         kref_put(&job->refcount, job->free);
 407 }
 408
 409 int
 410 v3d_wait_bo_ioctl(struct drm_device *dev, void *data,
 411                   struct drm_file *file_priv)
 412 {
 413         int ret;
 414         struct drm_v3d_wait_bo *args = data;
 415         ktime_t start = ktime_get();
 416         u64 delta_ns;
 417         unsigned long timeout_jiffies =
 418                 nsecs_to_jiffies_timeout(args->timeout_ns);
 419
 420         if (args->pad != 0)
 421                 return -EINVAL;
 422
 423         ret = drm_gem_dma_resv_wait(file_priv, args->handle,
 424                                     true, timeout_jiffies);
 425
 426         /* Decrement the user's timeout, in case we got interrupted
 427          * such that the ioctl will be restarted.
 428          */
 429         delta_ns = ktime_to_ns(ktime_sub(ktime_get(), start));
 430         if (delta_ns < args->timeout_ns)
 431                 args->timeout_ns -= delta_ns;
 432         else
 433                 args->timeout_ns = 0;
 434
 435         /* Asked to wait beyond the jiffie/scheduler precision? */
 436         if (ret == -ETIME && args->timeout_ns)
 437                 ret = -EAGAIN;
 438
 439         return ret;
 440 }
 441
 442 static int
 443 v3d_job_add_deps(struct drm_file *file_priv, struct v3d_job *job,
 444                  u32 in_sync, u32 point)
 445 {
 446         struct dma_fence *in_fence = NULL;
 447         int ret;
 448
 449         ret = drm_syncobj_find_fence(file_priv, in_sync, point, 0, &in_fence);
 450         if (ret == -EINVAL)
 451                 return ret;
 452
 453         return drm_sched_job_add_dependency(&job->base, in_fence);
 454 }
 455
 456 static int
 457 v3d_job_init(struct v3d_dev *v3d, struct drm_file *file_priv,
 458              void **container, size_t size, void (*free)(struct kref *ref),
 459              u32 in_sync, struct v3d_submit_ext *se, enum v3d_queue queue)
 460 {
 461         struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
 462         struct v3d_job *job;
 463         bool has_multisync = se && (se->flags & DRM_V3D_EXT_ID_MULTI_SYNC);
 464         int ret, i;
 465
 466         *container = kcalloc(1, size, GFP_KERNEL);
 467         if (!*container) {
 468                 DRM_ERROR("Cannot allocate memory for v3d job.");
 469                 return -ENOMEM;
 470         }
 471
 472         job = *container;
 473         job->v3d = v3d;
 474         job->free = free;
 475
 476         ret = drm_sched_job_init(&job->base, &v3d_priv->sched_entity[queue],
 477                                  v3d_priv);
 478         if (ret)
 479                 goto fail;
 480
 481         if (has_multisync) {
 482                 if (se->in_sync_count && se->wait_stage == queue) {
 483                         struct drm_v3d_sem __user *handle = u64_to_user_ptr(se->in_syncs);
 484
 485                         for (i = 0; i < se->in_sync_count; i++) {
 486                                 struct drm_v3d_sem in;
 487
 488                                 if (copy_from_user(&in, handle++, sizeof(in))) {
 489                                         ret = -EFAULT;
 490                                         DRM_DEBUG("Failed to copy wait dep handle.\n");
 491                                         goto fail_deps;
 492                                 }
 493                                 ret = v3d_job_add_deps(file_priv, job, in.handle, 0);
 494                                 if (ret)
 495                                         goto fail_deps;
 496                         }
 497                 }
 498         } else {
 499                 ret = v3d_job_add_deps(file_priv, job, in_sync, 0);
 500                 if (ret)
 501                         goto fail_deps;
 502         }
 503
 504         kref_init(&job->refcount);
 505
 506         return 0;
 507
 508 fail_deps:
 509         drm_sched_job_cleanup(&job->base);
 510 fail:
 511         kfree(*container);
 512         *container = NULL;
 513
 514         return ret;
 515 }
 516
 517 static void
 518 v3d_push_job(struct v3d_job *job)
 519 {
 520         drm_sched_job_arm(&job->base);
 521
 522         job->done_fence = dma_fence_get(&job->base.s_fence->finished);
 523
 524         /* put by scheduler job completion */
 525         kref_get(&job->refcount);
 526
 527         drm_sched_entity_push_job(&job->base);
 528 }
 529
 530 static void
 531 v3d_attach_fences_and_unlock_reservation(struct drm_file *file_priv,
 532                                          struct v3d_job *job,
 533                                          struct ww_acquire_ctx *acquire_ctx,
 534                                          u32 out_sync,
 535                                          struct v3d_submit_ext *se,
 536                                          struct dma_fence *done_fence)
 537 {
 538         struct drm_syncobj *sync_out;
 539         bool has_multisync = se && (se->flags & DRM_V3D_EXT_ID_MULTI_SYNC);
 540         int i;
 541
 542         for (i = 0; i < job->bo_count; i++) {
 543                 /* XXX: Use shared fences for read-only objects. */
 544                 dma_resv_add_fence(job->bo[i]->resv, job->done_fence,
 545                                    DMA_RESV_USAGE_WRITE);
 546         }
 547
 548         drm_gem_unlock_reservations(job->bo, job->bo_count, acquire_ctx);
 549
 550         /* Update the return sync object for the job */
 551         /* If it only supports a single signal semaphore*/
 552         if (!has_multisync) {
 553                 sync_out = drm_syncobj_find(file_priv, out_sync);
 554                 if (sync_out) {
 555                         drm_syncobj_replace_fence(sync_out, done_fence);
 556                         drm_syncobj_put(sync_out);
 557                 }
 558                 return;
 559         }
 560
 561         /* If multiple semaphores extension is supported */
 562         if (se->out_sync_count) {
 563                 for (i = 0; i < se->out_sync_count; i++) {
 564                         drm_syncobj_replace_fence(se->out_syncs[i].syncobj,
 565                                                   done_fence);
 566                         drm_syncobj_put(se->out_syncs[i].syncobj);
 567                 }
 568                 kvfree(se->out_syncs);
 569         }
 570 }
 571
 572 static void
 573 v3d_put_multisync_post_deps(struct v3d_submit_ext *se)
 574 {
 575         unsigned int i;
 576
 577         if (!(se && se->out_sync_count))
 578                 return;
 579
 580         for (i = 0; i < se->out_sync_count; i++)
 581                 drm_syncobj_put(se->out_syncs[i].syncobj);
 582         kvfree(se->out_syncs);
 583 }
 584
 585 static int
 586 v3d_get_multisync_post_deps(struct drm_file *file_priv,
 587                             struct v3d_submit_ext *se,
 588                             u32 count, u64 handles)
 589 {
 590         struct drm_v3d_sem __user *post_deps;
 591         int i, ret;
 592
 593         if (!count)
 594                 return 0;
 595
 596         se->out_syncs = (struct v3d_submit_outsync *)
 597                         kvmalloc_array(count,
 598                                        sizeof(struct v3d_submit_outsync),
 599                                        GFP_KERNEL);
 600         if (!se->out_syncs)
 601                 return -ENOMEM;
 602
 603         post_deps = u64_to_user_ptr(handles);
 604
 605         for (i = 0; i < count; i++) {
 606                 struct drm_v3d_sem out;
 607
 608                 if (copy_from_user(&out, post_deps++, sizeof(out))) {
 609                         ret = -EFAULT;
 610                         DRM_DEBUG("Failed to copy post dep handles\n");
 611                         goto fail;
 612                 }
 613
 614                 se->out_syncs[i].syncobj = drm_syncobj_find(file_priv,
 615                                                             out.handle);
 616                 if (!se->out_syncs[i].syncobj) {
 617                         ret = -EINVAL;
 618                         goto fail;
 619                 }
 620         }
 621         se->out_sync_count = count;
 622
 623         return 0;
 624
 625 fail:
 626         for (i--; i >= 0; i--)
 627                 drm_syncobj_put(se->out_syncs[i].syncobj);
 628         kvfree(se->out_syncs);
 629
 630         return ret;
 631 }
 632
 633 /* Get data for multiple binary semaphores synchronization. Parse syncobj
 634  * to be signaled when job completes (out_sync).
 635  */
 636 static int
 637 v3d_get_multisync_submit_deps(struct drm_file *file_priv,
 638                               struct drm_v3d_extension __user *ext,
 639                               void *data)
 640 {
 641         struct drm_v3d_multi_sync multisync;
 642         struct v3d_submit_ext *se = data;
 643         int ret;
 644
 645         if (copy_from_user(&multisync, ext, sizeof(multisync)))
 646                 return -EFAULT;
 647
 648         if (multisync.pad)
 649                 return -EINVAL;
 650
 651         ret = v3d_get_multisync_post_deps(file_priv, data, multisync.out_sync_count,
 652                                           multisync.out_syncs);
 653         if (ret)
 654                 return ret;
 655
 656         se->in_sync_count = multisync.in_sync_count;
 657         se->in_syncs = multisync.in_syncs;
 658         se->flags |= DRM_V3D_EXT_ID_MULTI_SYNC;
 659         se->wait_stage = multisync.wait_stage;
 660
 661         return 0;
 662 }
 663
 664 /* Whenever userspace sets ioctl extensions, v3d_get_extensions parses data
 665  * according to the extension id (name).
 666  */
 667 static int
 668 v3d_get_extensions(struct drm_file *file_priv,
 669                    u64 ext_handles,
 670                    void *data)
 671 {
 672         struct drm_v3d_extension __user *user_ext;
 673         int ret;
 674
 675         user_ext = u64_to_user_ptr(ext_handles);
 676         while (user_ext) {
 677                 struct drm_v3d_extension ext;
 678
 679                 if (copy_from_user(&ext, user_ext, sizeof(ext))) {
 680                         DRM_DEBUG("Failed to copy submit extension\n");
 681                         return -EFAULT;
 682                 }
 683
 684                 switch (ext.id) {
 685                 case DRM_V3D_EXT_ID_MULTI_SYNC:
 686                         ret = v3d_get_multisync_submit_deps(file_priv, user_ext, data);
 687                         if (ret)
 688                                 return ret;
 689                         break;
 690                 default:
 691                         DRM_DEBUG_DRIVER("Unknown extension id: %d\n", ext.id);
 692                         return -EINVAL;
 693                 }
 694
 695                 user_ext = u64_to_user_ptr(ext.next);
 696         }
 697
 698         return 0;
 699 }
 700
 701 /**
 702  * v3d_submit_cl_ioctl() - Submits a job (frame) to the V3D.
 703  * @dev: DRM device
 704  * @data: ioctl argument
 705  * @file_priv: DRM file for this fd
 706  *
 707  * This is the main entrypoint for userspace to submit a 3D frame to
 708  * the GPU.  Userspace provides the binner command list (if
 709  * applicable), and the kernel sets up the render command list to draw
 710  * to the framebuffer described in the ioctl, using the command lists
 711  * that the 3D engine's binner will produce.
 712  */
 713 int
 714 v3d_submit_cl_ioctl(struct drm_device *dev, void *data,
 715                     struct drm_file *file_priv)
 716 {
 717         struct v3d_dev *v3d = to_v3d_dev(dev);
 718         struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
 719         struct drm_v3d_submit_cl *args = data;
 720         struct v3d_submit_ext se = {0};
 721         struct v3d_bin_job *bin = NULL;
 722         struct v3d_render_job *render = NULL;
 723         struct v3d_job *clean_job = NULL;
 724         struct v3d_job *last_job;
 725         struct ww_acquire_ctx acquire_ctx;
 726         int ret = 0;
 727
 728         trace_v3d_submit_cl_ioctl(&v3d->drm, args->rcl_start, args->rcl_end);
 729
 730         if (args->pad)
 731                 return -EINVAL;
 732
 733         if (args->flags &&
 734             args->flags & ~(DRM_V3D_SUBMIT_CL_FLUSH_CACHE |
 735                             DRM_V3D_SUBMIT_EXTENSION)) {
 736                 DRM_INFO("invalid flags: %d\n", args->flags);
 737                 return -EINVAL;
 738         }
 739
 740         if (args->flags & DRM_V3D_SUBMIT_EXTENSION) {
 741                 ret = v3d_get_extensions(file_priv, args->extensions, &se);
 742                 if (ret) {
 743                         DRM_DEBUG("Failed to get extensions.\n");
 744                         return ret;
 745                 }
 746         }
 747
 748         ret = v3d_job_init(v3d, file_priv, (void *)&render, sizeof(*render),
 749                            v3d_render_job_free, args->in_sync_rcl, &se, V3D_RENDER);
 750         if (ret)
 751                 goto fail;
 752
 753         render->start = args->rcl_start;
 754         render->end = args->rcl_end;
 755         INIT_LIST_HEAD(&render->unref_list);
 756
 757         if (args->bcl_start != args->bcl_end) {
 758                 ret = v3d_job_init(v3d, file_priv, (void *)&bin, sizeof(*bin),
 759                                    v3d_job_free, args->in_sync_bcl, &se, V3D_BIN);
 760                 if (ret)
 761                         goto fail;
 762
 763                 bin->start = args->bcl_start;
 764                 bin->end = args->bcl_end;
 765                 bin->qma = args->qma;
 766                 bin->qms = args->qms;
 767                 bin->qts = args->qts;
 768                 bin->render = render;
 769         }
 770
 771         if (args->flags & DRM_V3D_SUBMIT_CL_FLUSH_CACHE) {
 772                 ret = v3d_job_init(v3d, file_priv, (void *)&clean_job, sizeof(*clean_job),
 773                                    v3d_job_free, 0, NULL, V3D_CACHE_CLEAN);
 774                 if (ret)
 775                         goto fail;
 776
 777                 last_job = clean_job;
 778         } else {
 779                 last_job = &render->base;
 780         }
 781
 782         ret = v3d_lookup_bos(dev, file_priv, last_job,
 783                              args->bo_handles, args->bo_handle_count);
 784         if (ret)
 785                 goto fail;
 786
 787         ret = v3d_lock_bo_reservations(last_job, &acquire_ctx);
 788         if (ret)
 789                 goto fail;
 790
 791         if (args->perfmon_id) {
 792                 render->base.perfmon = v3d_perfmon_find(v3d_priv,
 793                                                         args->perfmon_id);
 794
 795                 if (!render->base.perfmon) {
 796                         ret = -ENOENT;
 797                         goto fail_perfmon;
 798                 }
 799         }
 800
 801         mutex_lock(&v3d->sched_lock);
 802         if (bin) {
 803                 bin->base.perfmon = render->base.perfmon;
 804                 v3d_perfmon_get(bin->base.perfmon);
 805                 v3d_push_job(&bin->base);
 806
 807                 ret = drm_sched_job_add_dependency(&render->base.base,
 808                                                    dma_fence_get(bin->base.done_fence));
 809                 if (ret)
 810                         goto fail_unreserve;
 811         }
 812
 813         v3d_push_job(&render->base);
 814
 815         if (clean_job) {
 816                 struct dma_fence *render_fence =
 817                         dma_fence_get(render->base.done_fence);
 818                 ret = drm_sched_job_add_dependency(&clean_job->base,
 819                                                    render_fence);
 820                 if (ret)
 821                         goto fail_unreserve;
 822                 clean_job->perfmon = render->base.perfmon;
 823                 v3d_perfmon_get(clean_job->perfmon);
 824                 v3d_push_job(clean_job);
 825         }
 826
 827         mutex_unlock(&v3d->sched_lock);
 828
 829         v3d_attach_fences_and_unlock_reservation(file_priv,
 830                                                  last_job,
 831                                                  &acquire_ctx,
 832                                                  args->out_sync,
 833                                                  &se,
 834                                                  last_job->done_fence);
 835
 836         if (bin)
 837                 v3d_job_put(&bin->base);
 838         v3d_job_put(&render->base);
 839         if (clean_job)
 840                 v3d_job_put(clean_job);
 841
 842         return 0;
 843
 844 fail_unreserve:
 845         mutex_unlock(&v3d->sched_lock);
 846 fail_perfmon:
 847         drm_gem_unlock_reservations(last_job->bo,
 848                                     last_job->bo_count, &acquire_ctx);
 849 fail:
 850         v3d_job_cleanup((void *)bin);
 851         v3d_job_cleanup((void *)render);
 852         v3d_job_cleanup(clean_job);
 853         v3d_put_multisync_post_deps(&se);
 854
 855         return ret;
 856 }
 857
 858 /**
 859  * v3d_submit_tfu_ioctl() - Submits a TFU (texture formatting) job to the V3D.
 860  * @dev: DRM device
 861  * @data: ioctl argument
 862  * @file_priv: DRM file for this fd
 863  *
 864  * Userspace provides the register setup for the TFU, which we don't
 865  * need to validate since the TFU is behind the MMU.
 866  */
 867 int
 868 v3d_submit_tfu_ioctl(struct drm_device *dev, void *data,
 869                      struct drm_file *file_priv)
 870 {
 871         struct v3d_dev *v3d = to_v3d_dev(dev);
 872         struct drm_v3d_submit_tfu *args = data;
 873         struct v3d_submit_ext se = {0};
 874         struct v3d_tfu_job *job = NULL;
 875         struct ww_acquire_ctx acquire_ctx;
 876         int ret = 0;
 877
 878         trace_v3d_submit_tfu_ioctl(&v3d->drm, args->iia);
 879
 880         if (args->flags && !(args->flags & DRM_V3D_SUBMIT_EXTENSION)) {
 881                 DRM_DEBUG("invalid flags: %d\n", args->flags);
 882                 return -EINVAL;
 883         }
 884
 885         if (args->flags & DRM_V3D_SUBMIT_EXTENSION) {
 886                 ret = v3d_get_extensions(file_priv, args->extensions, &se);
 887                 if (ret) {
 888                         DRM_DEBUG("Failed to get extensions.\n");
 889                         return ret;
 890                 }
 891         }
 892
 893         ret = v3d_job_init(v3d, file_priv, (void *)&job, sizeof(*job),
 894                            v3d_job_free, args->in_sync, &se, V3D_TFU);
 895         if (ret)
 896                 goto fail;
 897
 898         job->base.bo = kcalloc(ARRAY_SIZE(args->bo_handles),
 899                                sizeof(*job->base.bo), GFP_KERNEL);
 900         if (!job->base.bo) {
 901                 ret = -ENOMEM;
 902                 goto fail;
 903         }
 904
 905         job->args = *args;
 906
 907         spin_lock(&file_priv->table_lock);
 908         for (job->base.bo_count = 0;
 909              job->base.bo_count < ARRAY_SIZE(args->bo_handles);
 910              job->base.bo_count++) {
 911                 struct drm_gem_object *bo;
 912
 913                 if (!args->bo_handles[job->base.bo_count])
 914                         break;
 915
 916                 bo = idr_find(&file_priv->object_idr,
 917                               args->bo_handles[job->base.bo_count]);
 918                 if (!bo) {
 919                         DRM_DEBUG("Failed to look up GEM BO %d: %d\n",
 920                                   job->base.bo_count,
 921                                   args->bo_handles[job->base.bo_count]);
 922                         ret = -ENOENT;
 923                         spin_unlock(&file_priv->table_lock);
 924                         goto fail;
 925                 }
 926                 drm_gem_object_get(bo);
 927                 job->base.bo[job->base.bo_count] = bo;
 928         }
 929         spin_unlock(&file_priv->table_lock);
 930
 931         ret = v3d_lock_bo_reservations(&job->base, &acquire_ctx);
 932         if (ret)
 933                 goto fail;
 934
 935         mutex_lock(&v3d->sched_lock);
 936         v3d_push_job(&job->base);
 937         mutex_unlock(&v3d->sched_lock);
 938
 939         v3d_attach_fences_and_unlock_reservation(file_priv,
 940                                                  &job->base, &acquire_ctx,
 941                                                  args->out_sync,
 942                                                  &se,
 943                                                  job->base.done_fence);
 944
 945         v3d_job_put(&job->base);
 946
 947         return 0;
 948
 949 fail:
 950         v3d_job_cleanup((void *)job);
 951         v3d_put_multisync_post_deps(&se);
 952
 953         return ret;
 954 }
 955
 956 /**
 957  * v3d_submit_csd_ioctl() - Submits a CSD (texture formatting) job to the V3D.
 958  * @dev: DRM device
 959  * @data: ioctl argument
 960  * @file_priv: DRM file for this fd
 961  *
 962  * Userspace provides the register setup for the CSD, which we don't
 963  * need to validate since the CSD is behind the MMU.
 964  */
 965 int
 966 v3d_submit_csd_ioctl(struct drm_device *dev, void *data,
 967                      struct drm_file *file_priv)
 968 {
 969         struct v3d_dev *v3d = to_v3d_dev(dev);
 970         struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
 971         struct drm_v3d_submit_csd *args = data;
 972         struct v3d_submit_ext se = {0};
 973         struct v3d_csd_job *job = NULL;
 974         struct v3d_job *clean_job = NULL;
 975         struct ww_acquire_ctx acquire_ctx;
 976         int ret;
 977
 978         trace_v3d_submit_csd_ioctl(&v3d->drm, args->cfg[5], args->cfg[6]);
 979
 980         if (args->pad)
 981                 return -EINVAL;
 982
 983         if (!v3d_has_csd(v3d)) {
 984                 DRM_DEBUG("Attempting CSD submit on non-CSD hardware\n");
 985                 return -EINVAL;
 986         }
 987
 988         if (args->flags && !(args->flags & DRM_V3D_SUBMIT_EXTENSION)) {
 989                 DRM_INFO("invalid flags: %d\n", args->flags);
 990                 return -EINVAL;
 991         }
 992
 993         if (args->flags & DRM_V3D_SUBMIT_EXTENSION) {
 994                 ret = v3d_get_extensions(file_priv, args->extensions, &se);
 995                 if (ret) {
 996                         DRM_DEBUG("Failed to get extensions.\n");
 997                         return ret;
 998                 }
 999         }
1000
1001         ret = v3d_job_init(v3d, file_priv, (void *)&job, sizeof(*job),
1002                            v3d_job_free, args->in_sync, &se, V3D_CSD);
1003         if (ret)
1004                 goto fail;
1005
1006         ret = v3d_job_init(v3d, file_priv, (void *)&clean_job, sizeof(*clean_job),
1007                            v3d_job_free, 0, NULL, V3D_CACHE_CLEAN);
1008         if (ret)
1009                 goto fail;
1010
1011         job->args = *args;
1012
1013         ret = v3d_lookup_bos(dev, file_priv, clean_job,
1014                              args->bo_handles, args->bo_handle_count);
1015         if (ret)
1016                 goto fail;
1017
1018         ret = v3d_lock_bo_reservations(clean_job, &acquire_ctx);
1019         if (ret)
1020                 goto fail;
1021
1022         if (args->perfmon_id) {
1023                 job->base.perfmon = v3d_perfmon_find(v3d_priv,
1024                                                      args->perfmon_id);
1025                 if (!job->base.perfmon) {
1026                         ret = -ENOENT;
1027                         goto fail_perfmon;
1028                 }
1029         }
1030
1031         mutex_lock(&v3d->sched_lock);
1032         v3d_push_job(&job->base);
1033
1034         ret = drm_sched_job_add_dependency(&clean_job->base,
1035                                            dma_fence_get(job->base.done_fence));
1036         if (ret)
1037                 goto fail_unreserve;
1038
1039         v3d_push_job(clean_job);
1040         mutex_unlock(&v3d->sched_lock);
1041
1042         v3d_attach_fences_and_unlock_reservation(file_priv,
1043                                                  clean_job,
1044                                                  &acquire_ctx,
1045                                                  args->out_sync,
1046                                                  &se,
1047                                                  clean_job->done_fence);
1048
1049         v3d_job_put(&job->base);
1050         v3d_job_put(clean_job);
1051
1052         return 0;
1053
1054 fail_unreserve:
1055         mutex_unlock(&v3d->sched_lock);
1056 fail_perfmon:
1057         drm_gem_unlock_reservations(clean_job->bo, clean_job->bo_count,
1058                                     &acquire_ctx);
1059 fail:
1060         v3d_job_cleanup((void *)job);
1061         v3d_job_cleanup(clean_job);
1062         v3d_put_multisync_post_deps(&se);
1063
1064         return ret;
1065 }
1066
1067 int
1068 v3d_gem_init(struct drm_device *dev)
1069 {
1070         struct v3d_dev *v3d = to_v3d_dev(dev);
1071         u32 pt_size = 4096 * 1024;
1072         int ret, i;
1073
1074         for (i = 0; i < V3D_MAX_QUEUES; i++)
1075                 v3d->queue[i].fence_context = dma_fence_context_alloc(1);
1076
1077         spin_lock_init(&v3d->mm_lock);
1078         spin_lock_init(&v3d->job_lock);
1079         ret = drmm_mutex_init(dev, &v3d->bo_lock);
1080         if (ret)
1081                 return ret;
1082         ret = drmm_mutex_init(dev, &v3d->reset_lock);
1083         if (ret)
1084                 return ret;
1085         ret = drmm_mutex_init(dev, &v3d->sched_lock);
1086         if (ret)
1087                 return ret;
1088         ret = drmm_mutex_init(dev, &v3d->cache_clean_lock);
1089         if (ret)
1090                 return ret;
1091
1092         /* Note: We don't allocate address 0.  Various bits of HW
1093          * treat 0 as special, such as the occlusion query counters
1094          * where 0 means "disabled".
1095          */
1096         drm_mm_init(&v3d->mm, 1, pt_size / sizeof(u32) - 1);
1097
1098         v3d->pt = dma_alloc_wc(v3d->drm.dev, pt_size,
1099                                &v3d->pt_paddr,
1100                                GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO);
1101         if (!v3d->pt) {
1102                 drm_mm_takedown(&v3d->mm);
1103                 dev_err(v3d->drm.dev,
1104                         "Failed to allocate page tables. Please ensure you have DMA enabled.\n");
1105                 return -ENOMEM;
1106         }
1107
1108         v3d_init_hw_state(v3d);
1109         v3d_mmu_set_page_table(v3d);
1110
1111         ret = v3d_sched_init(v3d);
1112         if (ret) {
1113                 drm_mm_takedown(&v3d->mm);
1114                 dma_free_coherent(v3d->drm.dev, 4096 * 1024, (void *)v3d->pt,
1115                                   v3d->pt_paddr);
1116         }
1117
1118         return 0;
1119 }
1120
1121 void
1122 v3d_gem_destroy(struct drm_device *dev)
1123 {
1124         struct v3d_dev *v3d = to_v3d_dev(dev);
1125
1126         v3d_sched_fini(v3d);
1127
1128         /* Waiting for jobs to finish would need to be done before
1129          * unregistering V3D.
1130          */
1131         WARN_ON(v3d->bin_job);
1132         WARN_ON(v3d->render_job);
1133
1134         drm_mm_takedown(&v3d->mm);
1135
1136         dma_free_coherent(v3d->drm.dev, 4096 * 1024, (void *)v3d->pt,
1137                           v3d->pt_paddr);
1138 }