drivers/gpu/drm/vc4/vc4_gem.c

   1 /*
   2  * Copyright © 2014 Broadcom
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  */
  23
  24 #include <linux/module.h>
  25 #include <linux/platform_device.h>
  26 #include <linux/pm_runtime.h>
  27 #include <linux/device.h>
  28 #include <linux/io.h>
  29 #include <linux/sched/signal.h>
  30 #include <linux/dma-fence-array.h>
  31
  32 #include "uapi/drm/vc4_drm.h"
  33 #include "vc4_drv.h"
  34 #include "vc4_regs.h"
  35 #include "vc4_trace.h"
  36
  37 static void
  38 vc4_queue_hangcheck(struct drm_device *dev)
  39 {
  40         struct vc4_dev *vc4 = to_vc4_dev(dev);
  41
  42         mod_timer(&vc4->hangcheck.timer,
  43                   round_jiffies_up(jiffies + msecs_to_jiffies(100)));
  44 }
  45
  46 struct vc4_hang_state {
  47         struct drm_vc4_get_hang_state user_state;
  48
  49         u32 bo_count;
  50         struct drm_gem_object **bo;
  51 };
  52
  53 static void
  54 vc4_free_hang_state(struct drm_device *dev, struct vc4_hang_state *state)
  55 {
  56         unsigned int i;
  57
  58         for (i = 0; i < state->user_state.bo_count; i++)
  59                 drm_gem_object_put_unlocked(state->bo[i]);
  60
  61         kfree(state);
  62 }
  63
  64 int
  65 vc4_get_hang_state_ioctl(struct drm_device *dev, void *data,
  66                          struct drm_file *file_priv)
  67 {
  68         struct drm_vc4_get_hang_state *get_state = data;
  69         struct drm_vc4_get_hang_state_bo *bo_state;
  70         struct vc4_hang_state *kernel_state;
  71         struct drm_vc4_get_hang_state *state;
  72         struct vc4_dev *vc4 = to_vc4_dev(dev);
  73         unsigned long irqflags;
  74         u32 i;
  75         int ret = 0;
  76
  77         if (!vc4->v3d) {
  78                 DRM_DEBUG("VC4_GET_HANG_STATE with no VC4 V3D probed\n");
  79                 return -ENODEV;
  80         }
  81
  82         spin_lock_irqsave(&vc4->job_lock, irqflags);
  83         kernel_state = vc4->hang_state;
  84         if (!kernel_state) {
  85                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
  86                 return -ENOENT;
  87         }
  88         state = &kernel_state->user_state;
  89
  90         /* If the user's array isn't big enough, just return the
  91          * required array size.
  92          */
  93         if (get_state->bo_count < state->bo_count) {
  94                 get_state->bo_count = state->bo_count;
  95                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
  96                 return 0;
  97         }
  98
  99         vc4->hang_state = NULL;
 100         spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 101
 102         /* Save the user's BO pointer, so we don't stomp it with the memcpy. */
 103         state->bo = get_state->bo;
 104         memcpy(get_state, state, sizeof(*state));
 105
 106         bo_state = kcalloc(state->bo_count, sizeof(*bo_state), GFP_KERNEL);
 107         if (!bo_state) {
 108                 ret = -ENOMEM;
 109                 goto err_free;
 110         }
 111
 112         for (i = 0; i < state->bo_count; i++) {
 113                 struct vc4_bo *vc4_bo = to_vc4_bo(kernel_state->bo[i]);
 114                 u32 handle;
 115
 116                 ret = drm_gem_handle_create(file_priv, kernel_state->bo[i],
 117                                             &handle);
 118
 119                 if (ret) {
 120                         state->bo_count = i;
 121                         goto err_delete_handle;
 122                 }
 123                 bo_state[i].handle = handle;
 124                 bo_state[i].paddr = vc4_bo->base.paddr;
 125                 bo_state[i].size = vc4_bo->base.base.size;
 126         }
 127
 128         if (copy_to_user(u64_to_user_ptr(get_state->bo),
 129                          bo_state,
 130                          state->bo_count * sizeof(*bo_state)))
 131                 ret = -EFAULT;
 132
 133 err_delete_handle:
 134         if (ret) {
 135                 for (i = 0; i < state->bo_count; i++)
 136                         drm_gem_handle_delete(file_priv, bo_state[i].handle);
 137         }
 138
 139 err_free:
 140         vc4_free_hang_state(dev, kernel_state);
 141         kfree(bo_state);
 142
 143         return ret;
 144 }
 145
 146 static void
 147 vc4_save_hang_state(struct drm_device *dev)
 148 {
 149         struct vc4_dev *vc4 = to_vc4_dev(dev);
 150         struct drm_vc4_get_hang_state *state;
 151         struct vc4_hang_state *kernel_state;
 152         struct vc4_exec_info *exec[2];
 153         struct vc4_bo *bo;
 154         unsigned long irqflags;
 155         unsigned int i, j, k, unref_list_count;
 156
 157         kernel_state = kcalloc(1, sizeof(*kernel_state), GFP_KERNEL);
 158         if (!kernel_state)
 159                 return;
 160
 161         state = &kernel_state->user_state;
 162
 163         spin_lock_irqsave(&vc4->job_lock, irqflags);
 164         exec[0] = vc4_first_bin_job(vc4);
 165         exec[1] = vc4_first_render_job(vc4);
 166         if (!exec[0] && !exec[1]) {
 167                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 168                 return;
 169         }
 170
 171         /* Get the bos from both binner and renderer into hang state. */
 172         state->bo_count = 0;
 173         for (i = 0; i < 2; i++) {
 174                 if (!exec[i])
 175                         continue;
 176
 177                 unref_list_count = 0;
 178                 list_for_each_entry(bo, &exec[i]->unref_list, unref_head)
 179                         unref_list_count++;
 180                 state->bo_count += exec[i]->bo_count + unref_list_count;
 181         }
 182
 183         kernel_state->bo = kcalloc(state->bo_count,
 184                                    sizeof(*kernel_state->bo), GFP_ATOMIC);
 185
 186         if (!kernel_state->bo) {
 187                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 188                 return;
 189         }
 190
 191         k = 0;
 192         for (i = 0; i < 2; i++) {
 193                 if (!exec[i])
 194                         continue;
 195
 196                 for (j = 0; j < exec[i]->bo_count; j++) {
 197                         bo = to_vc4_bo(&exec[i]->bo[j]->base);
 198
 199                         /* Retain BOs just in case they were marked purgeable.
 200                          * This prevents the BO from being purged before
 201                          * someone had a chance to dump the hang state.
 202                          */
 203                         WARN_ON(!refcount_read(&bo->usecnt));
 204                         refcount_inc(&bo->usecnt);
 205                         drm_gem_object_get(&exec[i]->bo[j]->base);
 206                         kernel_state->bo[k++] = &exec[i]->bo[j]->base;
 207                 }
 208
 209                 list_for_each_entry(bo, &exec[i]->unref_list, unref_head) {
 210                         /* No need to retain BOs coming from the ->unref_list
 211                          * because they are naturally unpurgeable.
 212                          */
 213                         drm_gem_object_get(&bo->base.base);
 214                         kernel_state->bo[k++] = &bo->base.base;
 215                 }
 216         }
 217
 218         WARN_ON_ONCE(k != state->bo_count);
 219
 220         if (exec[0])
 221                 state->start_bin = exec[0]->ct0ca;
 222         if (exec[1])
 223                 state->start_render = exec[1]->ct1ca;
 224
 225         spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 226
 227         state->ct0ca = V3D_READ(V3D_CTNCA(0));
 228         state->ct0ea = V3D_READ(V3D_CTNEA(0));
 229
 230         state->ct1ca = V3D_READ(V3D_CTNCA(1));
 231         state->ct1ea = V3D_READ(V3D_CTNEA(1));
 232
 233         state->ct0cs = V3D_READ(V3D_CTNCS(0));
 234         state->ct1cs = V3D_READ(V3D_CTNCS(1));
 235
 236         state->ct0ra0 = V3D_READ(V3D_CT00RA0);
 237         state->ct1ra0 = V3D_READ(V3D_CT01RA0);
 238
 239         state->bpca = V3D_READ(V3D_BPCA);
 240         state->bpcs = V3D_READ(V3D_BPCS);
 241         state->bpoa = V3D_READ(V3D_BPOA);
 242         state->bpos = V3D_READ(V3D_BPOS);
 243
 244         state->vpmbase = V3D_READ(V3D_VPMBASE);
 245
 246         state->dbge = V3D_READ(V3D_DBGE);
 247         state->fdbgo = V3D_READ(V3D_FDBGO);
 248         state->fdbgb = V3D_READ(V3D_FDBGB);
 249         state->fdbgr = V3D_READ(V3D_FDBGR);
 250         state->fdbgs = V3D_READ(V3D_FDBGS);
 251         state->errstat = V3D_READ(V3D_ERRSTAT);
 252
 253         /* We need to turn purgeable BOs into unpurgeable ones so that
 254          * userspace has a chance to dump the hang state before the kernel
 255          * decides to purge those BOs.
 256          * Note that BO consistency at dump time cannot be guaranteed. For
 257          * example, if the owner of these BOs decides to re-use them or mark
 258          * them purgeable again there's nothing we can do to prevent it.
 259          */
 260         for (i = 0; i < kernel_state->user_state.bo_count; i++) {
 261                 struct vc4_bo *bo = to_vc4_bo(kernel_state->bo[i]);
 262
 263                 if (bo->madv == __VC4_MADV_NOTSUPP)
 264                         continue;
 265
 266                 mutex_lock(&bo->madv_lock);
 267                 if (!WARN_ON(bo->madv == __VC4_MADV_PURGED))
 268                         bo->madv = VC4_MADV_WILLNEED;
 269                 refcount_dec(&bo->usecnt);
 270                 mutex_unlock(&bo->madv_lock);
 271         }
 272
 273         spin_lock_irqsave(&vc4->job_lock, irqflags);
 274         if (vc4->hang_state) {
 275                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 276                 vc4_free_hang_state(dev, kernel_state);
 277         } else {
 278                 vc4->hang_state = kernel_state;
 279                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 280         }
 281 }
 282
 283 static void
 284 vc4_reset(struct drm_device *dev)
 285 {
 286         struct vc4_dev *vc4 = to_vc4_dev(dev);
 287
 288         DRM_INFO("Resetting GPU.\n");
 289
 290         mutex_lock(&vc4->power_lock);
 291         if (vc4->power_refcount) {
 292                 /* Power the device off and back on the by dropping the
 293                  * reference on runtime PM.
 294                  */
 295                 pm_runtime_put_sync_suspend(&vc4->v3d->pdev->dev);
 296                 pm_runtime_get_sync(&vc4->v3d->pdev->dev);
 297         }
 298         mutex_unlock(&vc4->power_lock);
 299
 300         vc4_irq_reset(dev);
 301
 302         /* Rearm the hangcheck -- another job might have been waiting
 303          * for our hung one to get kicked off, and vc4_irq_reset()
 304          * would have started it.
 305          */
 306         vc4_queue_hangcheck(dev);
 307 }
 308
 309 static void
 310 vc4_reset_work(struct work_struct *work)
 311 {
 312         struct vc4_dev *vc4 =
 313                 container_of(work, struct vc4_dev, hangcheck.reset_work);
 314
 315         vc4_save_hang_state(vc4->dev);
 316
 317         vc4_reset(vc4->dev);
 318 }
 319
 320 static void
 321 vc4_hangcheck_elapsed(struct timer_list *t)
 322 {
 323         struct vc4_dev *vc4 = from_timer(vc4, t, hangcheck.timer);
 324         struct drm_device *dev = vc4->dev;
 325         uint32_t ct0ca, ct1ca;
 326         unsigned long irqflags;
 327         struct vc4_exec_info *bin_exec, *render_exec;
 328
 329         spin_lock_irqsave(&vc4->job_lock, irqflags);
 330
 331         bin_exec = vc4_first_bin_job(vc4);
 332         render_exec = vc4_first_render_job(vc4);
 333
 334         /* If idle, we can stop watching for hangs. */
 335         if (!bin_exec && !render_exec) {
 336                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 337                 return;
 338         }
 339
 340         ct0ca = V3D_READ(V3D_CTNCA(0));
 341         ct1ca = V3D_READ(V3D_CTNCA(1));
 342
 343         /* If we've made any progress in execution, rearm the timer
 344          * and wait.
 345          */
 346         if ((bin_exec && ct0ca != bin_exec->last_ct0ca) ||
 347             (render_exec && ct1ca != render_exec->last_ct1ca)) {
 348                 if (bin_exec)
 349                         bin_exec->last_ct0ca = ct0ca;
 350                 if (render_exec)
 351                         render_exec->last_ct1ca = ct1ca;
 352                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 353                 vc4_queue_hangcheck(dev);
 354                 return;
 355         }
 356
 357         spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 358
 359         /* We've gone too long with no progress, reset.  This has to
 360          * be done from a work struct, since resetting can sleep and
 361          * this timer hook isn't allowed to.
 362          */
 363         schedule_work(&vc4->hangcheck.reset_work);
 364 }
 365
 366 static void
 367 submit_cl(struct drm_device *dev, uint32_t thread, uint32_t start, uint32_t end)
 368 {
 369         struct vc4_dev *vc4 = to_vc4_dev(dev);
 370
 371         /* Set the current and end address of the control list.
 372          * Writing the end register is what starts the job.
 373          */
 374         V3D_WRITE(V3D_CTNCA(thread), start);
 375         V3D_WRITE(V3D_CTNEA(thread), end);
 376 }
 377
 378 int
 379 vc4_wait_for_seqno(struct drm_device *dev, uint64_t seqno, uint64_t timeout_ns,
 380                    bool interruptible)
 381 {
 382         struct vc4_dev *vc4 = to_vc4_dev(dev);
 383         int ret = 0;
 384         unsigned long timeout_expire;
 385         DEFINE_WAIT(wait);
 386
 387         if (vc4->finished_seqno >= seqno)
 388                 return 0;
 389
 390         if (timeout_ns == 0)
 391                 return -ETIME;
 392
 393         timeout_expire = jiffies + nsecs_to_jiffies(timeout_ns);
 394
 395         trace_vc4_wait_for_seqno_begin(dev, seqno, timeout_ns);
 396         for (;;) {
 397                 prepare_to_wait(&vc4->job_wait_queue, &wait,
 398                                 interruptible ? TASK_INTERRUPTIBLE :
 399                                 TASK_UNINTERRUPTIBLE);
 400
 401                 if (interruptible && signal_pending(current)) {
 402                         ret = -ERESTARTSYS;
 403                         break;
 404                 }
 405
 406                 if (vc4->finished_seqno >= seqno)
 407                         break;
 408
 409                 if (timeout_ns != ~0ull) {
 410                         if (time_after_eq(jiffies, timeout_expire)) {
 411                                 ret = -ETIME;
 412                                 break;
 413                         }
 414                         schedule_timeout(timeout_expire - jiffies);
 415                 } else {
 416                         schedule();
 417                 }
 418         }
 419
 420         finish_wait(&vc4->job_wait_queue, &wait);
 421         trace_vc4_wait_for_seqno_end(dev, seqno);
 422
 423         return ret;
 424 }
 425
 426 static void
 427 vc4_flush_caches(struct drm_device *dev)
 428 {
 429         struct vc4_dev *vc4 = to_vc4_dev(dev);
 430
 431         /* Flush the GPU L2 caches.  These caches sit on top of system
 432          * L3 (the 128kb or so shared with the CPU), and are
 433          * non-allocating in the L3.
 434          */
 435         V3D_WRITE(V3D_L2CACTL,
 436                   V3D_L2CACTL_L2CCLR);
 437
 438         V3D_WRITE(V3D_SLCACTL,
 439                   VC4_SET_FIELD(0xf, V3D_SLCACTL_T1CC) |
 440                   VC4_SET_FIELD(0xf, V3D_SLCACTL_T0CC) |
 441                   VC4_SET_FIELD(0xf, V3D_SLCACTL_UCC) |
 442                   VC4_SET_FIELD(0xf, V3D_SLCACTL_ICC));
 443 }
 444
 445 static void
 446 vc4_flush_texture_caches(struct drm_device *dev)
 447 {
 448         struct vc4_dev *vc4 = to_vc4_dev(dev);
 449
 450         V3D_WRITE(V3D_L2CACTL,
 451                   V3D_L2CACTL_L2CCLR);
 452
 453         V3D_WRITE(V3D_SLCACTL,
 454                   VC4_SET_FIELD(0xf, V3D_SLCACTL_T1CC) |
 455                   VC4_SET_FIELD(0xf, V3D_SLCACTL_T0CC));
 456 }
 457
 458 /* Sets the registers for the next job to be actually be executed in
 459  * the hardware.
 460  *
 461  * The job_lock should be held during this.
 462  */
 463 void
 464 vc4_submit_next_bin_job(struct drm_device *dev)
 465 {
 466         struct vc4_dev *vc4 = to_vc4_dev(dev);
 467         struct vc4_exec_info *exec;
 468
 469 again:
 470         exec = vc4_first_bin_job(vc4);
 471         if (!exec)
 472                 return;
 473
 474         vc4_flush_caches(dev);
 475
 476         /* Only start the perfmon if it was not already started by a previous
 477          * job.
 478          */
 479         if (exec->perfmon && vc4->active_perfmon != exec->perfmon)
 480                 vc4_perfmon_start(vc4, exec->perfmon);
 481
 482         /* Either put the job in the binner if it uses the binner, or
 483          * immediately move it to the to-be-rendered queue.
 484          */
 485         if (exec->ct0ca != exec->ct0ea) {
 486                 submit_cl(dev, 0, exec->ct0ca, exec->ct0ea);
 487         } else {
 488                 struct vc4_exec_info *next;
 489
 490                 vc4_move_job_to_render(dev, exec);
 491                 next = vc4_first_bin_job(vc4);
 492
 493                 /* We can't start the next bin job if the previous job had a
 494                  * different perfmon instance attached to it. The same goes
 495                  * if one of them had a perfmon attached to it and the other
 496                  * one doesn't.
 497                  */
 498                 if (next && next->perfmon == exec->perfmon)
 499                         goto again;
 500         }
 501 }
 502
 503 void
 504 vc4_submit_next_render_job(struct drm_device *dev)
 505 {
 506         struct vc4_dev *vc4 = to_vc4_dev(dev);
 507         struct vc4_exec_info *exec = vc4_first_render_job(vc4);
 508
 509         if (!exec)
 510                 return;
 511
 512         /* A previous RCL may have written to one of our textures, and
 513          * our full cache flush at bin time may have occurred before
 514          * that RCL completed.  Flush the texture cache now, but not
 515          * the instructions or uniforms (since we don't write those
 516          * from an RCL).
 517          */
 518         vc4_flush_texture_caches(dev);
 519
 520         submit_cl(dev, 1, exec->ct1ca, exec->ct1ea);
 521 }
 522
 523 void
 524 vc4_move_job_to_render(struct drm_device *dev, struct vc4_exec_info *exec)
 525 {
 526         struct vc4_dev *vc4 = to_vc4_dev(dev);
 527         bool was_empty = list_empty(&vc4->render_job_list);
 528
 529         list_move_tail(&exec->head, &vc4->render_job_list);
 530         if (was_empty)
 531                 vc4_submit_next_render_job(dev);
 532 }
 533
 534 static void
 535 vc4_update_bo_seqnos(struct vc4_exec_info *exec, uint64_t seqno)
 536 {
 537         struct vc4_bo *bo;
 538         unsigned i;
 539
 540         for (i = 0; i < exec->bo_count; i++) {
 541                 bo = to_vc4_bo(&exec->bo[i]->base);
 542                 bo->seqno = seqno;
 543
 544                 reservation_object_add_shared_fence(bo->base.base.resv, exec->fence);
 545         }
 546
 547         list_for_each_entry(bo, &exec->unref_list, unref_head) {
 548                 bo->seqno = seqno;
 549         }
 550
 551         for (i = 0; i < exec->rcl_write_bo_count; i++) {
 552                 bo = to_vc4_bo(&exec->rcl_write_bo[i]->base);
 553                 bo->write_seqno = seqno;
 554
 555                 reservation_object_add_excl_fence(bo->base.base.resv, exec->fence);
 556         }
 557 }
 558
 559 static void
 560 vc4_unlock_bo_reservations(struct drm_device *dev,
 561                            struct vc4_exec_info *exec,
 562                            struct ww_acquire_ctx *acquire_ctx)
 563 {
 564         int i;
 565
 566         for (i = 0; i < exec->bo_count; i++) {
 567                 struct drm_gem_object *bo = &exec->bo[i]->base;
 568
 569                 ww_mutex_unlock(&bo->resv->lock);
 570         }
 571
 572         ww_acquire_fini(acquire_ctx);
 573 }
 574
 575 /* Takes the reservation lock on all the BOs being referenced, so that
 576  * at queue submit time we can update the reservations.
 577  *
 578  * We don't lock the RCL the tile alloc/state BOs, or overflow memory
 579  * (all of which are on exec->unref_list).  They're entirely private
 580  * to vc4, so we don't attach dma-buf fences to them.
 581  */
 582 static int
 583 vc4_lock_bo_reservations(struct drm_device *dev,
 584                          struct vc4_exec_info *exec,
 585                          struct ww_acquire_ctx *acquire_ctx)
 586 {
 587         int contended_lock = -1;
 588         int i, ret;
 589         struct drm_gem_object *bo;
 590
 591         ww_acquire_init(acquire_ctx, &reservation_ww_class);
 592
 593 retry:
 594         if (contended_lock != -1) {
 595                 bo = &exec->bo[contended_lock]->base;
 596                 ret = ww_mutex_lock_slow_interruptible(&bo->resv->lock,
 597                                                        acquire_ctx);
 598                 if (ret) {
 599                         ww_acquire_done(acquire_ctx);
 600                         return ret;
 601                 }
 602         }
 603
 604         for (i = 0; i < exec->bo_count; i++) {
 605                 if (i == contended_lock)
 606                         continue;
 607
 608                 bo = &exec->bo[i]->base;
 609
 610                 ret = ww_mutex_lock_interruptible(&bo->resv->lock, acquire_ctx);
 611                 if (ret) {
 612                         int j;
 613
 614                         for (j = 0; j < i; j++) {
 615                                 bo = &exec->bo[j]->base;
 616                                 ww_mutex_unlock(&bo->resv->lock);
 617                         }
 618
 619                         if (contended_lock != -1 && contended_lock >= i) {
 620                                 bo = &exec->bo[contended_lock]->base;
 621
 622                                 ww_mutex_unlock(&bo->resv->lock);
 623                         }
 624
 625                         if (ret == -EDEADLK) {
 626                                 contended_lock = i;
 627                                 goto retry;
 628                         }
 629
 630                         ww_acquire_done(acquire_ctx);
 631                         return ret;
 632                 }
 633         }
 634
 635         ww_acquire_done(acquire_ctx);
 636
 637         /* Reserve space for our shared (read-only) fence references,
 638          * before we commit the CL to the hardware.
 639          */
 640         for (i = 0; i < exec->bo_count; i++) {
 641                 bo = &exec->bo[i]->base;
 642
 643                 ret = reservation_object_reserve_shared(bo->resv, 1);
 644                 if (ret) {
 645                         vc4_unlock_bo_reservations(dev, exec, acquire_ctx);
 646                         return ret;
 647                 }
 648         }
 649
 650         return 0;
 651 }
 652
 653 /* Queues a struct vc4_exec_info for execution.  If no job is
 654  * currently executing, then submits it.
 655  *
 656  * Unlike most GPUs, our hardware only handles one command list at a
 657  * time.  To queue multiple jobs at once, we'd need to edit the
 658  * previous command list to have a jump to the new one at the end, and
 659  * then bump the end address.  That's a change for a later date,
 660  * though.
 661  */
 662 static int
 663 vc4_queue_submit(struct drm_device *dev, struct vc4_exec_info *exec,
 664                  struct ww_acquire_ctx *acquire_ctx,
 665                  struct drm_syncobj *out_sync)
 666 {
 667         struct vc4_dev *vc4 = to_vc4_dev(dev);
 668         struct vc4_exec_info *renderjob;
 669         uint64_t seqno;
 670         unsigned long irqflags;
 671         struct vc4_fence *fence;
 672
 673         fence = kzalloc(sizeof(*fence), GFP_KERNEL);
 674         if (!fence)
 675                 return -ENOMEM;
 676         fence->dev = dev;
 677
 678         spin_lock_irqsave(&vc4->job_lock, irqflags);
 679
 680         seqno = ++vc4->emit_seqno;
 681         exec->seqno = seqno;
 682
 683         dma_fence_init(&fence->base, &vc4_fence_ops, &vc4->job_lock,
 684                        vc4->dma_fence_context, exec->seqno);
 685         fence->seqno = exec->seqno;
 686         exec->fence = &fence->base;
 687
 688         if (out_sync)
 689                 drm_syncobj_replace_fence(out_sync, exec->fence);
 690
 691         vc4_update_bo_seqnos(exec, seqno);
 692
 693         vc4_unlock_bo_reservations(dev, exec, acquire_ctx);
 694
 695         list_add_tail(&exec->head, &vc4->bin_job_list);
 696
 697         /* If no bin job was executing and if the render job (if any) has the
 698          * same perfmon as our job attached to it (or if both jobs don't have
 699          * perfmon activated), then kick ours off.  Otherwise, it'll get
 700          * started when the previous job's flush/render done interrupt occurs.
 701          */
 702         renderjob = vc4_first_render_job(vc4);
 703         if (vc4_first_bin_job(vc4) == exec &&
 704             (!renderjob || renderjob->perfmon == exec->perfmon)) {
 705                 vc4_submit_next_bin_job(dev);
 706                 vc4_queue_hangcheck(dev);
 707         }
 708
 709         spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 710
 711         return 0;
 712 }
 713
 714 /**
 715  * vc4_cl_lookup_bos() - Sets up exec->bo[] with the GEM objects
 716  * referenced by the job.
 717  * @dev: DRM device
 718  * @file_priv: DRM file for this fd
 719  * @exec: V3D job being set up
 720  *
 721  * The command validator needs to reference BOs by their index within
 722  * the submitted job's BO list.  This does the validation of the job's
 723  * BO list and reference counting for the lifetime of the job.
 724  */
 725 static int
 726 vc4_cl_lookup_bos(struct drm_device *dev,
 727                   struct drm_file *file_priv,
 728                   struct vc4_exec_info *exec)
 729 {
 730         struct drm_vc4_submit_cl *args = exec->args;
 731         uint32_t *handles;
 732         int ret = 0;
 733         int i;
 734
 735         exec->bo_count = args->bo_handle_count;
 736
 737         if (!exec->bo_count) {
 738                 /* See comment on bo_index for why we have to check
 739                  * this.
 740                  */
 741                 DRM_DEBUG("Rendering requires BOs to validate\n");
 742                 return -EINVAL;
 743         }
 744
 745         exec->bo = kvmalloc_array(exec->bo_count,
 746                                     sizeof(struct drm_gem_cma_object *),
 747                                     GFP_KERNEL | __GFP_ZERO);
 748         if (!exec->bo) {
 749                 DRM_ERROR("Failed to allocate validated BO pointers\n");
 750                 return -ENOMEM;
 751         }
 752
 753         handles = kvmalloc_array(exec->bo_count, sizeof(uint32_t), GFP_KERNEL);
 754         if (!handles) {
 755                 ret = -ENOMEM;
 756                 DRM_ERROR("Failed to allocate incoming GEM handles\n");
 757                 goto fail;
 758         }
 759
 760         if (copy_from_user(handles, u64_to_user_ptr(args->bo_handles),
 761                            exec->bo_count * sizeof(uint32_t))) {
 762                 ret = -EFAULT;
 763                 DRM_ERROR("Failed to copy in GEM handles\n");
 764                 goto fail;
 765         }
 766
 767         spin_lock(&file_priv->table_lock);
 768         for (i = 0; i < exec->bo_count; i++) {
 769                 struct drm_gem_object *bo = idr_find(&file_priv->object_idr,
 770                                                      handles[i]);
 771                 if (!bo) {
 772                         DRM_DEBUG("Failed to look up GEM BO %d: %d\n",
 773                                   i, handles[i]);
 774                         ret = -EINVAL;
 775                         break;
 776                 }
 777
 778                 drm_gem_object_get(bo);
 779                 exec->bo[i] = (struct drm_gem_cma_object *)bo;
 780         }
 781         spin_unlock(&file_priv->table_lock);
 782
 783         if (ret)
 784                 goto fail_put_bo;
 785
 786         for (i = 0; i < exec->bo_count; i++) {
 787                 ret = vc4_bo_inc_usecnt(to_vc4_bo(&exec->bo[i]->base));
 788                 if (ret)
 789                         goto fail_dec_usecnt;
 790         }
 791
 792         kvfree(handles);
 793         return 0;
 794
 795 fail_dec_usecnt:
 796         /* Decrease usecnt on acquired objects.
 797          * We cannot rely on  vc4_complete_exec() to release resources here,
 798          * because vc4_complete_exec() has no information about which BO has
 799          * had its ->usecnt incremented.
 800          * To make things easier we just free everything explicitly and set
 801          * exec->bo to NULL so that vc4_complete_exec() skips the 'BO release'
 802          * step.
 803          */
 804         for (i-- ; i >= 0; i--)
 805                 vc4_bo_dec_usecnt(to_vc4_bo(&exec->bo[i]->base));
 806
 807 fail_put_bo:
 808         /* Release any reference to acquired objects. */
 809         for (i = 0; i < exec->bo_count && exec->bo[i]; i++)
 810                 drm_gem_object_put_unlocked(&exec->bo[i]->base);
 811
 812 fail:
 813         kvfree(handles);
 814         kvfree(exec->bo);
 815         exec->bo = NULL;
 816         return ret;
 817 }
 818
 819 static int
 820 vc4_get_bcl(struct drm_device *dev, struct vc4_exec_info *exec)
 821 {
 822         struct drm_vc4_submit_cl *args = exec->args;
 823         void *temp = NULL;
 824         void *bin;
 825         int ret = 0;
 826         uint32_t bin_offset = 0;
 827         uint32_t shader_rec_offset = roundup(bin_offset + args->bin_cl_size,
 828                                              16);
 829         uint32_t uniforms_offset = shader_rec_offset + args->shader_rec_size;
 830         uint32_t exec_size = uniforms_offset + args->uniforms_size;
 831         uint32_t temp_size = exec_size + (sizeof(struct vc4_shader_state) *
 832                                           args->shader_rec_count);
 833         struct vc4_bo *bo;
 834
 835         if (shader_rec_offset < args->bin_cl_size ||
 836             uniforms_offset < shader_rec_offset ||
 837             exec_size < uniforms_offset ||
 838             args->shader_rec_count >= (UINT_MAX /
 839                                           sizeof(struct vc4_shader_state)) ||
 840             temp_size < exec_size) {
 841                 DRM_DEBUG("overflow in exec arguments\n");
 842                 ret = -EINVAL;
 843                 goto fail;
 844         }
 845
 846         /* Allocate space where we'll store the copied in user command lists
 847          * and shader records.
 848          *
 849          * We don't just copy directly into the BOs because we need to
 850          * read the contents back for validation, and I think the
 851          * bo->vaddr is uncached access.
 852          */
 853         temp = kvmalloc_array(temp_size, 1, GFP_KERNEL);
 854         if (!temp) {
 855                 DRM_ERROR("Failed to allocate storage for copying "
 856                           "in bin/render CLs.\n");
 857                 ret = -ENOMEM;
 858                 goto fail;
 859         }
 860         bin = temp + bin_offset;
 861         exec->shader_rec_u = temp + shader_rec_offset;
 862         exec->uniforms_u = temp + uniforms_offset;
 863         exec->shader_state = temp + exec_size;
 864         exec->shader_state_size = args->shader_rec_count;
 865
 866         if (copy_from_user(bin,
 867                            u64_to_user_ptr(args->bin_cl),
 868                            args->bin_cl_size)) {
 869                 ret = -EFAULT;
 870                 goto fail;
 871         }
 872
 873         if (copy_from_user(exec->shader_rec_u,
 874                            u64_to_user_ptr(args->shader_rec),
 875                            args->shader_rec_size)) {
 876                 ret = -EFAULT;
 877                 goto fail;
 878         }
 879
 880         if (copy_from_user(exec->uniforms_u,
 881                            u64_to_user_ptr(args->uniforms),
 882                            args->uniforms_size)) {
 883                 ret = -EFAULT;
 884                 goto fail;
 885         }
 886
 887         bo = vc4_bo_create(dev, exec_size, true, VC4_BO_TYPE_BCL);
 888         if (IS_ERR(bo)) {
 889                 DRM_ERROR("Couldn't allocate BO for binning\n");
 890                 ret = PTR_ERR(bo);
 891                 goto fail;
 892         }
 893         exec->exec_bo = &bo->base;
 894
 895         list_add_tail(&to_vc4_bo(&exec->exec_bo->base)->unref_head,
 896                       &exec->unref_list);
 897
 898         exec->ct0ca = exec->exec_bo->paddr + bin_offset;
 899
 900         exec->bin_u = bin;
 901
 902         exec->shader_rec_v = exec->exec_bo->vaddr + shader_rec_offset;
 903         exec->shader_rec_p = exec->exec_bo->paddr + shader_rec_offset;
 904         exec->shader_rec_size = args->shader_rec_size;
 905
 906         exec->uniforms_v = exec->exec_bo->vaddr + uniforms_offset;
 907         exec->uniforms_p = exec->exec_bo->paddr + uniforms_offset;
 908         exec->uniforms_size = args->uniforms_size;
 909
 910         ret = vc4_validate_bin_cl(dev,
 911                                   exec->exec_bo->vaddr + bin_offset,
 912                                   bin,
 913                                   exec);
 914         if (ret)
 915                 goto fail;
 916
 917         ret = vc4_validate_shader_recs(dev, exec);
 918         if (ret)
 919                 goto fail;
 920
 921         /* Block waiting on any previous rendering into the CS's VBO,
 922          * IB, or textures, so that pixels are actually written by the
 923          * time we try to read them.
 924          */
 925         ret = vc4_wait_for_seqno(dev, exec->bin_dep_seqno, ~0ull, true);
 926
 927 fail:
 928         kvfree(temp);
 929         return ret;
 930 }
 931
 932 static void
 933 vc4_complete_exec(struct drm_device *dev, struct vc4_exec_info *exec)
 934 {
 935         struct vc4_dev *vc4 = to_vc4_dev(dev);
 936         unsigned long irqflags;
 937         unsigned i;
 938
 939         /* If we got force-completed because of GPU reset rather than
 940          * through our IRQ handler, signal the fence now.
 941          */
 942         if (exec->fence) {
 943                 dma_fence_signal(exec->fence);
 944                 dma_fence_put(exec->fence);
 945         }
 946
 947         if (exec->bo) {
 948                 for (i = 0; i < exec->bo_count; i++) {
 949                         struct vc4_bo *bo = to_vc4_bo(&exec->bo[i]->base);
 950
 951                         vc4_bo_dec_usecnt(bo);
 952                         drm_gem_object_put_unlocked(&exec->bo[i]->base);
 953                 }
 954                 kvfree(exec->bo);
 955         }
 956
 957         while (!list_empty(&exec->unref_list)) {
 958                 struct vc4_bo *bo = list_first_entry(&exec->unref_list,
 959                                                      struct vc4_bo, unref_head);
 960                 list_del(&bo->unref_head);
 961                 drm_gem_object_put_unlocked(&bo->base.base);
 962         }
 963
 964         /* Free up the allocation of any bin slots we used. */
 965         spin_lock_irqsave(&vc4->job_lock, irqflags);
 966         vc4->bin_alloc_used &= ~exec->bin_slots;
 967         spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 968
 969         /* Release the reference we had on the perf monitor. */
 970         vc4_perfmon_put(exec->perfmon);
 971
 972         vc4_v3d_pm_put(vc4);
 973
 974         kfree(exec);
 975 }
 976
 977 void
 978 vc4_job_handle_completed(struct vc4_dev *vc4)
 979 {
 980         unsigned long irqflags;
 981         struct vc4_seqno_cb *cb, *cb_temp;
 982
 983         spin_lock_irqsave(&vc4->job_lock, irqflags);
 984         while (!list_empty(&vc4->job_done_list)) {
 985                 struct vc4_exec_info *exec =
 986                         list_first_entry(&vc4->job_done_list,
 987                                          struct vc4_exec_info, head);
 988                 list_del(&exec->head);
 989
 990                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 991                 vc4_complete_exec(vc4->dev, exec);
 992                 spin_lock_irqsave(&vc4->job_lock, irqflags);
 993         }
 994
 995         list_for_each_entry_safe(cb, cb_temp, &vc4->seqno_cb_list, work.entry) {
 996                 if (cb->seqno <= vc4->finished_seqno) {
 997                         list_del_init(&cb->work.entry);
 998                         schedule_work(&cb->work);
 999                 }
1000         }
1001
1002         spin_unlock_irqrestore(&vc4->job_lock, irqflags);
1003 }
1004
1005 static void vc4_seqno_cb_work(struct work_struct *work)
1006 {
1007         struct vc4_seqno_cb *cb = container_of(work, struct vc4_seqno_cb, work);
1008
1009         cb->func(cb);
1010 }
1011
1012 int vc4_queue_seqno_cb(struct drm_device *dev,
1013                        struct vc4_seqno_cb *cb, uint64_t seqno,
1014                        void (*func)(struct vc4_seqno_cb *cb))
1015 {
1016         struct vc4_dev *vc4 = to_vc4_dev(dev);
1017         int ret = 0;
1018         unsigned long irqflags;
1019
1020         cb->func = func;
1021         INIT_WORK(&cb->work, vc4_seqno_cb_work);
1022
1023         spin_lock_irqsave(&vc4->job_lock, irqflags);
1024         if (seqno > vc4->finished_seqno) {
1025                 cb->seqno = seqno;
1026                 list_add_tail(&cb->work.entry, &vc4->seqno_cb_list);
1027         } else {
1028                 schedule_work(&cb->work);
1029         }
1030         spin_unlock_irqrestore(&vc4->job_lock, irqflags);
1031
1032         return ret;
1033 }
1034
1035 /* Scheduled when any job has been completed, this walks the list of
1036  * jobs that had completed and unrefs their BOs and frees their exec
1037  * structs.
1038  */
1039 static void
1040 vc4_job_done_work(struct work_struct *work)
1041 {
1042         struct vc4_dev *vc4 =
1043                 container_of(work, struct vc4_dev, job_done_work);
1044
1045         vc4_job_handle_completed(vc4);
1046 }
1047
1048 static int
1049 vc4_wait_for_seqno_ioctl_helper(struct drm_device *dev,
1050                                 uint64_t seqno,
1051                                 uint64_t *timeout_ns)
1052 {
1053         unsigned long start = jiffies;
1054         int ret = vc4_wait_for_seqno(dev, seqno, *timeout_ns, true);
1055
1056         if ((ret == -EINTR || ret == -ERESTARTSYS) && *timeout_ns != ~0ull) {
1057                 uint64_t delta = jiffies_to_nsecs(jiffies - start);
1058
1059                 if (*timeout_ns >= delta)
1060                         *timeout_ns -= delta;
1061         }
1062
1063         return ret;
1064 }
1065
1066 int
1067 vc4_wait_seqno_ioctl(struct drm_device *dev, void *data,
1068                      struct drm_file *file_priv)
1069 {
1070         struct drm_vc4_wait_seqno *args = data;
1071
1072         return vc4_wait_for_seqno_ioctl_helper(dev, args->seqno,
1073                                                &args->timeout_ns);
1074 }
1075
1076 int
1077 vc4_wait_bo_ioctl(struct drm_device *dev, void *data,
1078                   struct drm_file *file_priv)
1079 {
1080         int ret;
1081         struct drm_vc4_wait_bo *args = data;
1082         struct drm_gem_object *gem_obj;
1083         struct vc4_bo *bo;
1084
1085         if (args->pad != 0)
1086                 return -EINVAL;
1087
1088         gem_obj = drm_gem_object_lookup(file_priv, args->handle);
1089         if (!gem_obj) {
1090                 DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
1091                 return -EINVAL;
1092         }
1093         bo = to_vc4_bo(gem_obj);
1094
1095         ret = vc4_wait_for_seqno_ioctl_helper(dev, bo->seqno,
1096                                               &args->timeout_ns);
1097
1098         drm_gem_object_put_unlocked(gem_obj);
1099         return ret;
1100 }
1101
1102 /**
1103  * vc4_submit_cl_ioctl() - Submits a job (frame) to the VC4.
1104  * @dev: DRM device
1105  * @data: ioctl argument
1106  * @file_priv: DRM file for this fd
1107  *
1108  * This is the main entrypoint for userspace to submit a 3D frame to
1109  * the GPU.  Userspace provides the binner command list (if
1110  * applicable), and the kernel sets up the render command list to draw
1111  * to the framebuffer described in the ioctl, using the command lists
1112  * that the 3D engine's binner will produce.
1113  */
1114 int
1115 vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
1116                     struct drm_file *file_priv)
1117 {
1118         struct vc4_dev *vc4 = to_vc4_dev(dev);
1119         struct vc4_file *vc4file = file_priv->driver_priv;
1120         struct drm_vc4_submit_cl *args = data;
1121         struct drm_syncobj *out_sync = NULL;
1122         struct vc4_exec_info *exec;
1123         struct ww_acquire_ctx acquire_ctx;
1124         struct dma_fence *in_fence;
1125         int ret = 0;
1126
1127         if (!vc4->v3d) {
1128                 DRM_DEBUG("VC4_SUBMIT_CL with no VC4 V3D probed\n");
1129                 return -ENODEV;
1130         }
1131
1132         if ((args->flags & ~(VC4_SUBMIT_CL_USE_CLEAR_COLOR |
1133                              VC4_SUBMIT_CL_FIXED_RCL_ORDER |
1134                              VC4_SUBMIT_CL_RCL_ORDER_INCREASING_X |
1135                              VC4_SUBMIT_CL_RCL_ORDER_INCREASING_Y)) != 0) {
1136                 DRM_DEBUG("Unknown flags: 0x%02x\n", args->flags);
1137                 return -EINVAL;
1138         }
1139
1140         if (args->pad2 != 0) {
1141                 DRM_DEBUG("Invalid pad: 0x%08x\n", args->pad2);
1142                 return -EINVAL;
1143         }
1144
1145         exec = kcalloc(1, sizeof(*exec), GFP_KERNEL);
1146         if (!exec) {
1147                 DRM_ERROR("malloc failure on exec struct\n");
1148                 return -ENOMEM;
1149         }
1150
1151         ret = vc4_v3d_pm_get(vc4);
1152         if (ret) {
1153                 kfree(exec);
1154                 return ret;
1155         }
1156
1157         exec->args = args;
1158         INIT_LIST_HEAD(&exec->unref_list);
1159
1160         ret = vc4_cl_lookup_bos(dev, file_priv, exec);
1161         if (ret)
1162                 goto fail;
1163
1164         if (args->perfmonid) {
1165                 exec->perfmon = vc4_perfmon_find(vc4file,
1166                                                  args->perfmonid);
1167                 if (!exec->perfmon) {
1168                         ret = -ENOENT;
1169                         goto fail;
1170                 }
1171         }
1172
1173         if (args->in_sync) {
1174                 ret = drm_syncobj_find_fence(file_priv, args->in_sync,
1175                                              0, 0, &in_fence);
1176                 if (ret)
1177                         goto fail;
1178
1179                 /* When the fence (or fence array) is exclusively from our
1180                  * context we can skip the wait since jobs are executed in
1181                  * order of their submission through this ioctl and this can
1182                  * only have fences from a prior job.
1183                  */
1184                 if (!dma_fence_match_context(in_fence,
1185                                              vc4->dma_fence_context)) {
1186                         ret = dma_fence_wait(in_fence, true);
1187                         if (ret) {
1188                                 dma_fence_put(in_fence);
1189                                 goto fail;
1190                         }
1191                 }
1192
1193                 dma_fence_put(in_fence);
1194         }
1195
1196         if (exec->args->bin_cl_size != 0) {
1197                 ret = vc4_get_bcl(dev, exec);
1198                 if (ret)
1199                         goto fail;
1200         } else {
1201                 exec->ct0ca = 0;
1202                 exec->ct0ea = 0;
1203         }
1204
1205         ret = vc4_get_rcl(dev, exec);
1206         if (ret)
1207                 goto fail;
1208
1209         ret = vc4_lock_bo_reservations(dev, exec, &acquire_ctx);
1210         if (ret)
1211                 goto fail;
1212
1213         if (args->out_sync) {
1214                 out_sync = drm_syncobj_find(file_priv, args->out_sync);
1215                 if (!out_sync) {
1216                         ret = -EINVAL;
1217                         goto fail;
1218                 }
1219
1220                 /* We replace the fence in out_sync in vc4_queue_submit since
1221                  * the render job could execute immediately after that call.
1222                  * If it finishes before our ioctl processing resumes the
1223                  * render job fence could already have been freed.
1224                  */
1225         }
1226
1227         /* Clear this out of the struct we'll be putting in the queue,
1228          * since it's part of our stack.
1229          */
1230         exec->args = NULL;
1231
1232         ret = vc4_queue_submit(dev, exec, &acquire_ctx, out_sync);
1233
1234         /* The syncobj isn't part of the exec data and we need to free our
1235          * reference even if job submission failed.
1236          */
1237         if (out_sync)
1238                 drm_syncobj_put(out_sync);
1239
1240         if (ret)
1241                 goto fail;
1242
1243         /* Return the seqno for our job. */
1244         args->seqno = vc4->emit_seqno;
1245
1246         return 0;
1247
1248 fail:
1249         vc4_complete_exec(vc4->dev, exec);
1250
1251         return ret;
1252 }
1253
1254 void
1255 vc4_gem_init(struct drm_device *dev)
1256 {
1257         struct vc4_dev *vc4 = to_vc4_dev(dev);
1258
1259         vc4->dma_fence_context = dma_fence_context_alloc(1);
1260
1261         INIT_LIST_HEAD(&vc4->bin_job_list);
1262         INIT_LIST_HEAD(&vc4->render_job_list);
1263         INIT_LIST_HEAD(&vc4->job_done_list);
1264         INIT_LIST_HEAD(&vc4->seqno_cb_list);
1265         spin_lock_init(&vc4->job_lock);
1266
1267         INIT_WORK(&vc4->hangcheck.reset_work, vc4_reset_work);
1268         timer_setup(&vc4->hangcheck.timer, vc4_hangcheck_elapsed, 0);
1269
1270         INIT_WORK(&vc4->job_done_work, vc4_job_done_work);
1271
1272         mutex_init(&vc4->power_lock);
1273
1274         INIT_LIST_HEAD(&vc4->purgeable.list);
1275         mutex_init(&vc4->purgeable.lock);
1276 }
1277
1278 void
1279 vc4_gem_destroy(struct drm_device *dev)
1280 {
1281         struct vc4_dev *vc4 = to_vc4_dev(dev);
1282
1283         /* Waiting for exec to finish would need to be done before
1284          * unregistering V3D.
1285          */
1286         WARN_ON(vc4->emit_seqno != vc4->finished_seqno);
1287
1288         /* V3D should already have disabled its interrupt and cleared
1289          * the overflow allocation registers.  Now free the object.
1290          */
1291         if (vc4->bin_bo) {
1292                 drm_gem_object_put_unlocked(&vc4->bin_bo->base.base);
1293                 vc4->bin_bo = NULL;
1294         }
1295
1296         if (vc4->hang_state)
1297                 vc4_free_hang_state(dev, vc4->hang_state);
1298 }
1299
1300 int vc4_gem_madvise_ioctl(struct drm_device *dev, void *data,
1301                           struct drm_file *file_priv)
1302 {
1303         struct drm_vc4_gem_madvise *args = data;
1304         struct drm_gem_object *gem_obj;
1305         struct vc4_bo *bo;
1306         int ret;
1307
1308         switch (args->madv) {
1309         case VC4_MADV_DONTNEED:
1310         case VC4_MADV_WILLNEED:
1311                 break;
1312         default:
1313                 return -EINVAL;
1314         }
1315
1316         if (args->pad != 0)
1317                 return -EINVAL;
1318
1319         gem_obj = drm_gem_object_lookup(file_priv, args->handle);
1320         if (!gem_obj) {
1321                 DRM_DEBUG("Failed to look up GEM BO %d\n", args->handle);
1322                 return -ENOENT;
1323         }
1324
1325         bo = to_vc4_bo(gem_obj);
1326
1327         /* Only BOs exposed to userspace can be purged. */
1328         if (bo->madv == __VC4_MADV_NOTSUPP) {
1329                 DRM_DEBUG("madvise not supported on this BO\n");
1330                 ret = -EINVAL;
1331                 goto out_put_gem;
1332         }
1333
1334         /* Not sure it's safe to purge imported BOs. Let's just assume it's
1335          * not until proven otherwise.
1336          */
1337         if (gem_obj->import_attach) {
1338                 DRM_DEBUG("madvise not supported on imported BOs\n");
1339                 ret = -EINVAL;
1340                 goto out_put_gem;
1341         }
1342
1343         mutex_lock(&bo->madv_lock);
1344
1345         if (args->madv == VC4_MADV_DONTNEED && bo->madv == VC4_MADV_WILLNEED &&
1346             !refcount_read(&bo->usecnt)) {
1347                 /* If the BO is about to be marked as purgeable, is not used
1348                  * and is not already purgeable or purged, add it to the
1349                  * purgeable list.
1350                  */
1351                 vc4_bo_add_to_purgeable_pool(bo);
1352         } else if (args->madv == VC4_MADV_WILLNEED &&
1353                    bo->madv == VC4_MADV_DONTNEED &&
1354                    !refcount_read(&bo->usecnt)) {
1355                 /* The BO has not been purged yet, just remove it from
1356                  * the purgeable list.
1357                  */
1358                 vc4_bo_remove_from_purgeable_pool(bo);
1359         }
1360
1361         /* Save the purged state. */
1362         args->retained = bo->madv != __VC4_MADV_PURGED;
1363
1364         /* Update internal madv state only if the bo was not purged. */
1365         if (bo->madv != __VC4_MADV_PURGED)
1366                 bo->madv = args->madv;
1367
1368         mutex_unlock(&bo->madv_lock);
1369
1370         ret = 0;
1371
1372 out_put_gem:
1373         drm_gem_object_put_unlocked(gem_obj);
1374
1375         return ret;
1376 }