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
  30 #include "uapi/drm/vc4_drm.h"
  31 #include "vc4_drv.h"
  32 #include "vc4_regs.h"
  33 #include "vc4_trace.h"
  34
  35 static void
  36 vc4_queue_hangcheck(struct drm_device *dev)
  37 {
  38         struct vc4_dev *vc4 = to_vc4_dev(dev);
  39
  40         mod_timer(&vc4->hangcheck.timer,
  41                   round_jiffies_up(jiffies + msecs_to_jiffies(100)));
  42 }
  43
  44 struct vc4_hang_state {
  45         struct drm_vc4_get_hang_state user_state;
  46
  47         u32 bo_count;
  48         struct drm_gem_object **bo;
  49 };
  50
  51 static void
  52 vc4_free_hang_state(struct drm_device *dev, struct vc4_hang_state *state)
  53 {
  54         unsigned int i;
  55
  56         for (i = 0; i < state->user_state.bo_count; i++)
  57                 drm_gem_object_unreference_unlocked(state->bo[i]);
  58
  59         kfree(state);
  60 }
  61
  62 int
  63 vc4_get_hang_state_ioctl(struct drm_device *dev, void *data,
  64                          struct drm_file *file_priv)
  65 {
  66         struct drm_vc4_get_hang_state *get_state = data;
  67         struct drm_vc4_get_hang_state_bo *bo_state;
  68         struct vc4_hang_state *kernel_state;
  69         struct drm_vc4_get_hang_state *state;
  70         struct vc4_dev *vc4 = to_vc4_dev(dev);
  71         unsigned long irqflags;
  72         u32 i;
  73         int ret = 0;
  74
  75         spin_lock_irqsave(&vc4->job_lock, irqflags);
  76         kernel_state = vc4->hang_state;
  77         if (!kernel_state) {
  78                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
  79                 return -ENOENT;
  80         }
  81         state = &kernel_state->user_state;
  82
  83         /* If the user's array isn't big enough, just return the
  84          * required array size.
  85          */
  86         if (get_state->bo_count < state->bo_count) {
  87                 get_state->bo_count = state->bo_count;
  88                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
  89                 return 0;
  90         }
  91
  92         vc4->hang_state = NULL;
  93         spin_unlock_irqrestore(&vc4->job_lock, irqflags);
  94
  95         /* Save the user's BO pointer, so we don't stomp it with the memcpy. */
  96         state->bo = get_state->bo;
  97         memcpy(get_state, state, sizeof(*state));
  98
  99         bo_state = kcalloc(state->bo_count, sizeof(*bo_state), GFP_KERNEL);
 100         if (!bo_state) {
 101                 ret = -ENOMEM;
 102                 goto err_free;
 103         }
 104
 105         for (i = 0; i < state->bo_count; i++) {
 106                 struct vc4_bo *vc4_bo = to_vc4_bo(kernel_state->bo[i]);
 107                 u32 handle;
 108
 109                 ret = drm_gem_handle_create(file_priv, kernel_state->bo[i],
 110                                             &handle);
 111
 112                 if (ret) {
 113                         state->bo_count = i - 1;
 114                         goto err;
 115                 }
 116                 bo_state[i].handle = handle;
 117                 bo_state[i].paddr = vc4_bo->base.paddr;
 118                 bo_state[i].size = vc4_bo->base.base.size;
 119         }
 120
 121         if (copy_to_user((void __user *)(uintptr_t)get_state->bo,
 122                          bo_state,
 123                          state->bo_count * sizeof(*bo_state)))
 124                 ret = -EFAULT;
 125
 126         kfree(bo_state);
 127
 128 err_free:
 129
 130         vc4_free_hang_state(dev, kernel_state);
 131
 132 err:
 133         return ret;
 134 }
 135
 136 static void
 137 vc4_save_hang_state(struct drm_device *dev)
 138 {
 139         struct vc4_dev *vc4 = to_vc4_dev(dev);
 140         struct drm_vc4_get_hang_state *state;
 141         struct vc4_hang_state *kernel_state;
 142         struct vc4_exec_info *exec[2];
 143         struct vc4_bo *bo;
 144         unsigned long irqflags;
 145         unsigned int i, j, unref_list_count, prev_idx;
 146
 147         kernel_state = kcalloc(1, sizeof(*kernel_state), GFP_KERNEL);
 148         if (!kernel_state)
 149                 return;
 150
 151         state = &kernel_state->user_state;
 152
 153         spin_lock_irqsave(&vc4->job_lock, irqflags);
 154         exec[0] = vc4_first_bin_job(vc4);
 155         exec[1] = vc4_first_render_job(vc4);
 156         if (!exec[0] && !exec[1]) {
 157                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 158                 return;
 159         }
 160
 161         /* Get the bos from both binner and renderer into hang state. */
 162         state->bo_count = 0;
 163         for (i = 0; i < 2; i++) {
 164                 if (!exec[i])
 165                         continue;
 166
 167                 unref_list_count = 0;
 168                 list_for_each_entry(bo, &exec[i]->unref_list, unref_head)
 169                         unref_list_count++;
 170                 state->bo_count += exec[i]->bo_count + unref_list_count;
 171         }
 172
 173         kernel_state->bo = kcalloc(state->bo_count,
 174                                    sizeof(*kernel_state->bo), GFP_ATOMIC);
 175
 176         if (!kernel_state->bo) {
 177                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 178                 return;
 179         }
 180
 181         prev_idx = 0;
 182         for (i = 0; i < 2; i++) {
 183                 if (!exec[i])
 184                         continue;
 185
 186                 for (j = 0; j < exec[i]->bo_count; j++) {
 187                         drm_gem_object_reference(&exec[i]->bo[j]->base);
 188                         kernel_state->bo[j + prev_idx] = &exec[i]->bo[j]->base;
 189                 }
 190
 191                 list_for_each_entry(bo, &exec[i]->unref_list, unref_head) {
 192                         drm_gem_object_reference(&bo->base.base);
 193                         kernel_state->bo[j + prev_idx] = &bo->base.base;
 194                         j++;
 195                 }
 196                 prev_idx = j + 1;
 197         }
 198
 199         if (exec[0])
 200                 state->start_bin = exec[0]->ct0ca;
 201         if (exec[1])
 202                 state->start_render = exec[1]->ct1ca;
 203
 204         spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 205
 206         state->ct0ca = V3D_READ(V3D_CTNCA(0));
 207         state->ct0ea = V3D_READ(V3D_CTNEA(0));
 208
 209         state->ct1ca = V3D_READ(V3D_CTNCA(1));
 210         state->ct1ea = V3D_READ(V3D_CTNEA(1));
 211
 212         state->ct0cs = V3D_READ(V3D_CTNCS(0));
 213         state->ct1cs = V3D_READ(V3D_CTNCS(1));
 214
 215         state->ct0ra0 = V3D_READ(V3D_CT00RA0);
 216         state->ct1ra0 = V3D_READ(V3D_CT01RA0);
 217
 218         state->bpca = V3D_READ(V3D_BPCA);
 219         state->bpcs = V3D_READ(V3D_BPCS);
 220         state->bpoa = V3D_READ(V3D_BPOA);
 221         state->bpos = V3D_READ(V3D_BPOS);
 222
 223         state->vpmbase = V3D_READ(V3D_VPMBASE);
 224
 225         state->dbge = V3D_READ(V3D_DBGE);
 226         state->fdbgo = V3D_READ(V3D_FDBGO);
 227         state->fdbgb = V3D_READ(V3D_FDBGB);
 228         state->fdbgr = V3D_READ(V3D_FDBGR);
 229         state->fdbgs = V3D_READ(V3D_FDBGS);
 230         state->errstat = V3D_READ(V3D_ERRSTAT);
 231
 232         spin_lock_irqsave(&vc4->job_lock, irqflags);
 233         if (vc4->hang_state) {
 234                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 235                 vc4_free_hang_state(dev, kernel_state);
 236         } else {
 237                 vc4->hang_state = kernel_state;
 238                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 239         }
 240 }
 241
 242 static void
 243 vc4_reset(struct drm_device *dev)
 244 {
 245         struct vc4_dev *vc4 = to_vc4_dev(dev);
 246
 247         DRM_INFO("Resetting GPU.\n");
 248
 249         mutex_lock(&vc4->power_lock);
 250         if (vc4->power_refcount) {
 251                 /* Power the device off and back on the by dropping the
 252                  * reference on runtime PM.
 253                  */
 254                 pm_runtime_put_sync_suspend(&vc4->v3d->pdev->dev);
 255                 pm_runtime_get_sync(&vc4->v3d->pdev->dev);
 256         }
 257         mutex_unlock(&vc4->power_lock);
 258
 259         vc4_irq_reset(dev);
 260
 261         /* Rearm the hangcheck -- another job might have been waiting
 262          * for our hung one to get kicked off, and vc4_irq_reset()
 263          * would have started it.
 264          */
 265         vc4_queue_hangcheck(dev);
 266 }
 267
 268 static void
 269 vc4_reset_work(struct work_struct *work)
 270 {
 271         struct vc4_dev *vc4 =
 272                 container_of(work, struct vc4_dev, hangcheck.reset_work);
 273
 274         vc4_save_hang_state(vc4->dev);
 275
 276         vc4_reset(vc4->dev);
 277 }
 278
 279 static void
 280 vc4_hangcheck_elapsed(unsigned long data)
 281 {
 282         struct drm_device *dev = (struct drm_device *)data;
 283         struct vc4_dev *vc4 = to_vc4_dev(dev);
 284         uint32_t ct0ca, ct1ca;
 285         unsigned long irqflags;
 286         struct vc4_exec_info *bin_exec, *render_exec;
 287
 288         spin_lock_irqsave(&vc4->job_lock, irqflags);
 289
 290         bin_exec = vc4_first_bin_job(vc4);
 291         render_exec = vc4_first_render_job(vc4);
 292
 293         /* If idle, we can stop watching for hangs. */
 294         if (!bin_exec && !render_exec) {
 295                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 296                 return;
 297         }
 298
 299         ct0ca = V3D_READ(V3D_CTNCA(0));
 300         ct1ca = V3D_READ(V3D_CTNCA(1));
 301
 302         /* If we've made any progress in execution, rearm the timer
 303          * and wait.
 304          */
 305         if ((bin_exec && ct0ca != bin_exec->last_ct0ca) ||
 306             (render_exec && ct1ca != render_exec->last_ct1ca)) {
 307                 if (bin_exec)
 308                         bin_exec->last_ct0ca = ct0ca;
 309                 if (render_exec)
 310                         render_exec->last_ct1ca = ct1ca;
 311                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 312                 vc4_queue_hangcheck(dev);
 313                 return;
 314         }
 315
 316         spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 317
 318         /* We've gone too long with no progress, reset.  This has to
 319          * be done from a work struct, since resetting can sleep and
 320          * this timer hook isn't allowed to.
 321          */
 322         schedule_work(&vc4->hangcheck.reset_work);
 323 }
 324
 325 static void
 326 submit_cl(struct drm_device *dev, uint32_t thread, uint32_t start, uint32_t end)
 327 {
 328         struct vc4_dev *vc4 = to_vc4_dev(dev);
 329
 330         /* Set the current and end address of the control list.
 331          * Writing the end register is what starts the job.
 332          */
 333         V3D_WRITE(V3D_CTNCA(thread), start);
 334         V3D_WRITE(V3D_CTNEA(thread), end);
 335 }
 336
 337 int
 338 vc4_wait_for_seqno(struct drm_device *dev, uint64_t seqno, uint64_t timeout_ns,
 339                    bool interruptible)
 340 {
 341         struct vc4_dev *vc4 = to_vc4_dev(dev);
 342         int ret = 0;
 343         unsigned long timeout_expire;
 344         DEFINE_WAIT(wait);
 345
 346         if (vc4->finished_seqno >= seqno)
 347                 return 0;
 348
 349         if (timeout_ns == 0)
 350                 return -ETIME;
 351
 352         timeout_expire = jiffies + nsecs_to_jiffies(timeout_ns);
 353
 354         trace_vc4_wait_for_seqno_begin(dev, seqno, timeout_ns);
 355         for (;;) {
 356                 prepare_to_wait(&vc4->job_wait_queue, &wait,
 357                                 interruptible ? TASK_INTERRUPTIBLE :
 358                                 TASK_UNINTERRUPTIBLE);
 359
 360                 if (interruptible && signal_pending(current)) {
 361                         ret = -ERESTARTSYS;
 362                         break;
 363                 }
 364
 365                 if (vc4->finished_seqno >= seqno)
 366                         break;
 367
 368                 if (timeout_ns != ~0ull) {
 369                         if (time_after_eq(jiffies, timeout_expire)) {
 370                                 ret = -ETIME;
 371                                 break;
 372                         }
 373                         schedule_timeout(timeout_expire - jiffies);
 374                 } else {
 375                         schedule();
 376                 }
 377         }
 378
 379         finish_wait(&vc4->job_wait_queue, &wait);
 380         trace_vc4_wait_for_seqno_end(dev, seqno);
 381
 382         return ret;
 383 }
 384
 385 static void
 386 vc4_flush_caches(struct drm_device *dev)
 387 {
 388         struct vc4_dev *vc4 = to_vc4_dev(dev);
 389
 390         /* Flush the GPU L2 caches.  These caches sit on top of system
 391          * L3 (the 128kb or so shared with the CPU), and are
 392          * non-allocating in the L3.
 393          */
 394         V3D_WRITE(V3D_L2CACTL,
 395                   V3D_L2CACTL_L2CCLR);
 396
 397         V3D_WRITE(V3D_SLCACTL,
 398                   VC4_SET_FIELD(0xf, V3D_SLCACTL_T1CC) |
 399                   VC4_SET_FIELD(0xf, V3D_SLCACTL_T0CC) |
 400                   VC4_SET_FIELD(0xf, V3D_SLCACTL_UCC) |
 401                   VC4_SET_FIELD(0xf, V3D_SLCACTL_ICC));
 402 }
 403
 404 /* Sets the registers for the next job to be actually be executed in
 405  * the hardware.
 406  *
 407  * The job_lock should be held during this.
 408  */
 409 void
 410 vc4_submit_next_bin_job(struct drm_device *dev)
 411 {
 412         struct vc4_dev *vc4 = to_vc4_dev(dev);
 413         struct vc4_exec_info *exec;
 414
 415 again:
 416         exec = vc4_first_bin_job(vc4);
 417         if (!exec)
 418                 return;
 419
 420         vc4_flush_caches(dev);
 421
 422         /* Either put the job in the binner if it uses the binner, or
 423          * immediately move it to the to-be-rendered queue.
 424          */
 425         if (exec->ct0ca != exec->ct0ea) {
 426                 submit_cl(dev, 0, exec->ct0ca, exec->ct0ea);
 427         } else {
 428                 vc4_move_job_to_render(dev, exec);
 429                 goto again;
 430         }
 431 }
 432
 433 void
 434 vc4_submit_next_render_job(struct drm_device *dev)
 435 {
 436         struct vc4_dev *vc4 = to_vc4_dev(dev);
 437         struct vc4_exec_info *exec = vc4_first_render_job(vc4);
 438
 439         if (!exec)
 440                 return;
 441
 442         submit_cl(dev, 1, exec->ct1ca, exec->ct1ea);
 443 }
 444
 445 void
 446 vc4_move_job_to_render(struct drm_device *dev, struct vc4_exec_info *exec)
 447 {
 448         struct vc4_dev *vc4 = to_vc4_dev(dev);
 449         bool was_empty = list_empty(&vc4->render_job_list);
 450
 451         list_move_tail(&exec->head, &vc4->render_job_list);
 452         if (was_empty)
 453                 vc4_submit_next_render_job(dev);
 454 }
 455
 456 static void
 457 vc4_update_bo_seqnos(struct vc4_exec_info *exec, uint64_t seqno)
 458 {
 459         struct vc4_bo *bo;
 460         unsigned i;
 461
 462         for (i = 0; i < exec->bo_count; i++) {
 463                 bo = to_vc4_bo(&exec->bo[i]->base);
 464                 bo->seqno = seqno;
 465         }
 466
 467         list_for_each_entry(bo, &exec->unref_list, unref_head) {
 468                 bo->seqno = seqno;
 469         }
 470
 471         for (i = 0; i < exec->rcl_write_bo_count; i++) {
 472                 bo = to_vc4_bo(&exec->rcl_write_bo[i]->base);
 473                 bo->write_seqno = seqno;
 474         }
 475 }
 476
 477 /* Queues a struct vc4_exec_info for execution.  If no job is
 478  * currently executing, then submits it.
 479  *
 480  * Unlike most GPUs, our hardware only handles one command list at a
 481  * time.  To queue multiple jobs at once, we'd need to edit the
 482  * previous command list to have a jump to the new one at the end, and
 483  * then bump the end address.  That's a change for a later date,
 484  * though.
 485  */
 486 static void
 487 vc4_queue_submit(struct drm_device *dev, struct vc4_exec_info *exec)
 488 {
 489         struct vc4_dev *vc4 = to_vc4_dev(dev);
 490         uint64_t seqno;
 491         unsigned long irqflags;
 492
 493         spin_lock_irqsave(&vc4->job_lock, irqflags);
 494
 495         seqno = ++vc4->emit_seqno;
 496         exec->seqno = seqno;
 497         vc4_update_bo_seqnos(exec, seqno);
 498
 499         list_add_tail(&exec->head, &vc4->bin_job_list);
 500
 501         /* If no job was executing, kick ours off.  Otherwise, it'll
 502          * get started when the previous job's flush done interrupt
 503          * occurs.
 504          */
 505         if (vc4_first_bin_job(vc4) == exec) {
 506                 vc4_submit_next_bin_job(dev);
 507                 vc4_queue_hangcheck(dev);
 508         }
 509
 510         spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 511 }
 512
 513 /**
 514  * Looks up a bunch of GEM handles for BOs and stores the array for
 515  * use in the command validator that actually writes relocated
 516  * addresses pointing to them.
 517  */
 518 static int
 519 vc4_cl_lookup_bos(struct drm_device *dev,
 520                   struct drm_file *file_priv,
 521                   struct vc4_exec_info *exec)
 522 {
 523         struct drm_vc4_submit_cl *args = exec->args;
 524         uint32_t *handles;
 525         int ret = 0;
 526         int i;
 527
 528         exec->bo_count = args->bo_handle_count;
 529
 530         if (!exec->bo_count) {
 531                 /* See comment on bo_index for why we have to check
 532                  * this.
 533                  */
 534                 DRM_ERROR("Rendering requires BOs to validate\n");
 535                 return -EINVAL;
 536         }
 537
 538         exec->bo = drm_calloc_large(exec->bo_count,
 539                                     sizeof(struct drm_gem_cma_object *));
 540         if (!exec->bo) {
 541                 DRM_ERROR("Failed to allocate validated BO pointers\n");
 542                 return -ENOMEM;
 543         }
 544
 545         handles = drm_malloc_ab(exec->bo_count, sizeof(uint32_t));
 546         if (!handles) {
 547                 ret = -ENOMEM;
 548                 DRM_ERROR("Failed to allocate incoming GEM handles\n");
 549                 goto fail;
 550         }
 551
 552         if (copy_from_user(handles,
 553                            (void __user *)(uintptr_t)args->bo_handles,
 554                            exec->bo_count * sizeof(uint32_t))) {
 555                 ret = -EFAULT;
 556                 DRM_ERROR("Failed to copy in GEM handles\n");
 557                 goto fail;
 558         }
 559
 560         spin_lock(&file_priv->table_lock);
 561         for (i = 0; i < exec->bo_count; i++) {
 562                 struct drm_gem_object *bo = idr_find(&file_priv->object_idr,
 563                                                      handles[i]);
 564                 if (!bo) {
 565                         DRM_ERROR("Failed to look up GEM BO %d: %d\n",
 566                                   i, handles[i]);
 567                         ret = -EINVAL;
 568                         spin_unlock(&file_priv->table_lock);
 569                         goto fail;
 570                 }
 571                 drm_gem_object_reference(bo);
 572                 exec->bo[i] = (struct drm_gem_cma_object *)bo;
 573         }
 574         spin_unlock(&file_priv->table_lock);
 575
 576 fail:
 577         drm_free_large(handles);
 578         return ret;
 579 }
 580
 581 static int
 582 vc4_get_bcl(struct drm_device *dev, struct vc4_exec_info *exec)
 583 {
 584         struct drm_vc4_submit_cl *args = exec->args;
 585         void *temp = NULL;
 586         void *bin;
 587         int ret = 0;
 588         uint32_t bin_offset = 0;
 589         uint32_t shader_rec_offset = roundup(bin_offset + args->bin_cl_size,
 590                                              16);
 591         uint32_t uniforms_offset = shader_rec_offset + args->shader_rec_size;
 592         uint32_t exec_size = uniforms_offset + args->uniforms_size;
 593         uint32_t temp_size = exec_size + (sizeof(struct vc4_shader_state) *
 594                                           args->shader_rec_count);
 595         struct vc4_bo *bo;
 596
 597         if (shader_rec_offset < args->bin_cl_size ||
 598             uniforms_offset < shader_rec_offset ||
 599             exec_size < uniforms_offset ||
 600             args->shader_rec_count >= (UINT_MAX /
 601                                           sizeof(struct vc4_shader_state)) ||
 602             temp_size < exec_size) {
 603                 DRM_ERROR("overflow in exec arguments\n");
 604                 ret = -EINVAL;
 605                 goto fail;
 606         }
 607
 608         /* Allocate space where we'll store the copied in user command lists
 609          * and shader records.
 610          *
 611          * We don't just copy directly into the BOs because we need to
 612          * read the contents back for validation, and I think the
 613          * bo->vaddr is uncached access.
 614          */
 615         temp = drm_malloc_ab(temp_size, 1);
 616         if (!temp) {
 617                 DRM_ERROR("Failed to allocate storage for copying "
 618                           "in bin/render CLs.\n");
 619                 ret = -ENOMEM;
 620                 goto fail;
 621         }
 622         bin = temp + bin_offset;
 623         exec->shader_rec_u = temp + shader_rec_offset;
 624         exec->uniforms_u = temp + uniforms_offset;
 625         exec->shader_state = temp + exec_size;
 626         exec->shader_state_size = args->shader_rec_count;
 627
 628         if (copy_from_user(bin,
 629                            (void __user *)(uintptr_t)args->bin_cl,
 630                            args->bin_cl_size)) {
 631                 ret = -EFAULT;
 632                 goto fail;
 633         }
 634
 635         if (copy_from_user(exec->shader_rec_u,
 636                            (void __user *)(uintptr_t)args->shader_rec,
 637                            args->shader_rec_size)) {
 638                 ret = -EFAULT;
 639                 goto fail;
 640         }
 641
 642         if (copy_from_user(exec->uniforms_u,
 643                            (void __user *)(uintptr_t)args->uniforms,
 644                            args->uniforms_size)) {
 645                 ret = -EFAULT;
 646                 goto fail;
 647         }
 648
 649         bo = vc4_bo_create(dev, exec_size, true);
 650         if (IS_ERR(bo)) {
 651                 DRM_ERROR("Couldn't allocate BO for binning\n");
 652                 ret = PTR_ERR(bo);
 653                 goto fail;
 654         }
 655         exec->exec_bo = &bo->base;
 656
 657         list_add_tail(&to_vc4_bo(&exec->exec_bo->base)->unref_head,
 658                       &exec->unref_list);
 659
 660         exec->ct0ca = exec->exec_bo->paddr + bin_offset;
 661
 662         exec->bin_u = bin;
 663
 664         exec->shader_rec_v = exec->exec_bo->vaddr + shader_rec_offset;
 665         exec->shader_rec_p = exec->exec_bo->paddr + shader_rec_offset;
 666         exec->shader_rec_size = args->shader_rec_size;
 667
 668         exec->uniforms_v = exec->exec_bo->vaddr + uniforms_offset;
 669         exec->uniforms_p = exec->exec_bo->paddr + uniforms_offset;
 670         exec->uniforms_size = args->uniforms_size;
 671
 672         ret = vc4_validate_bin_cl(dev,
 673                                   exec->exec_bo->vaddr + bin_offset,
 674                                   bin,
 675                                   exec);
 676         if (ret)
 677                 goto fail;
 678
 679         ret = vc4_validate_shader_recs(dev, exec);
 680         if (ret)
 681                 goto fail;
 682
 683         /* Block waiting on any previous rendering into the CS's VBO,
 684          * IB, or textures, so that pixels are actually written by the
 685          * time we try to read them.
 686          */
 687         ret = vc4_wait_for_seqno(dev, exec->bin_dep_seqno, ~0ull, true);
 688
 689 fail:
 690         drm_free_large(temp);
 691         return ret;
 692 }
 693
 694 static void
 695 vc4_complete_exec(struct drm_device *dev, struct vc4_exec_info *exec)
 696 {
 697         struct vc4_dev *vc4 = to_vc4_dev(dev);
 698         unsigned i;
 699
 700         if (exec->bo) {
 701                 for (i = 0; i < exec->bo_count; i++)
 702                         drm_gem_object_unreference_unlocked(&exec->bo[i]->base);
 703                 drm_free_large(exec->bo);
 704         }
 705
 706         while (!list_empty(&exec->unref_list)) {
 707                 struct vc4_bo *bo = list_first_entry(&exec->unref_list,
 708                                                      struct vc4_bo, unref_head);
 709                 list_del(&bo->unref_head);
 710                 drm_gem_object_unreference_unlocked(&bo->base.base);
 711         }
 712
 713         mutex_lock(&vc4->power_lock);
 714         if (--vc4->power_refcount == 0) {
 715                 pm_runtime_mark_last_busy(&vc4->v3d->pdev->dev);
 716                 pm_runtime_put_autosuspend(&vc4->v3d->pdev->dev);
 717         }
 718         mutex_unlock(&vc4->power_lock);
 719
 720         kfree(exec);
 721 }
 722
 723 void
 724 vc4_job_handle_completed(struct vc4_dev *vc4)
 725 {
 726         unsigned long irqflags;
 727         struct vc4_seqno_cb *cb, *cb_temp;
 728
 729         spin_lock_irqsave(&vc4->job_lock, irqflags);
 730         while (!list_empty(&vc4->job_done_list)) {
 731                 struct vc4_exec_info *exec =
 732                         list_first_entry(&vc4->job_done_list,
 733                                          struct vc4_exec_info, head);
 734                 list_del(&exec->head);
 735
 736                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 737                 vc4_complete_exec(vc4->dev, exec);
 738                 spin_lock_irqsave(&vc4->job_lock, irqflags);
 739         }
 740
 741         list_for_each_entry_safe(cb, cb_temp, &vc4->seqno_cb_list, work.entry) {
 742                 if (cb->seqno <= vc4->finished_seqno) {
 743                         list_del_init(&cb->work.entry);
 744                         schedule_work(&cb->work);
 745                 }
 746         }
 747
 748         spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 749 }
 750
 751 static void vc4_seqno_cb_work(struct work_struct *work)
 752 {
 753         struct vc4_seqno_cb *cb = container_of(work, struct vc4_seqno_cb, work);
 754
 755         cb->func(cb);
 756 }
 757
 758 int vc4_queue_seqno_cb(struct drm_device *dev,
 759                        struct vc4_seqno_cb *cb, uint64_t seqno,
 760                        void (*func)(struct vc4_seqno_cb *cb))
 761 {
 762         struct vc4_dev *vc4 = to_vc4_dev(dev);
 763         int ret = 0;
 764         unsigned long irqflags;
 765
 766         cb->func = func;
 767         INIT_WORK(&cb->work, vc4_seqno_cb_work);
 768
 769         spin_lock_irqsave(&vc4->job_lock, irqflags);
 770         if (seqno > vc4->finished_seqno) {
 771                 cb->seqno = seqno;
 772                 list_add_tail(&cb->work.entry, &vc4->seqno_cb_list);
 773         } else {
 774                 schedule_work(&cb->work);
 775         }
 776         spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 777
 778         return ret;
 779 }
 780
 781 /* Scheduled when any job has been completed, this walks the list of
 782  * jobs that had completed and unrefs their BOs and frees their exec
 783  * structs.
 784  */
 785 static void
 786 vc4_job_done_work(struct work_struct *work)
 787 {
 788         struct vc4_dev *vc4 =
 789                 container_of(work, struct vc4_dev, job_done_work);
 790
 791         vc4_job_handle_completed(vc4);
 792 }
 793
 794 static int
 795 vc4_wait_for_seqno_ioctl_helper(struct drm_device *dev,
 796                                 uint64_t seqno,
 797                                 uint64_t *timeout_ns)
 798 {
 799         unsigned long start = jiffies;
 800         int ret = vc4_wait_for_seqno(dev, seqno, *timeout_ns, true);
 801
 802         if ((ret == -EINTR || ret == -ERESTARTSYS) && *timeout_ns != ~0ull) {
 803                 uint64_t delta = jiffies_to_nsecs(jiffies - start);
 804
 805                 if (*timeout_ns >= delta)
 806                         *timeout_ns -= delta;
 807         }
 808
 809         return ret;
 810 }
 811
 812 int
 813 vc4_wait_seqno_ioctl(struct drm_device *dev, void *data,
 814                      struct drm_file *file_priv)
 815 {
 816         struct drm_vc4_wait_seqno *args = data;
 817
 818         return vc4_wait_for_seqno_ioctl_helper(dev, args->seqno,
 819                                                &args->timeout_ns);
 820 }
 821
 822 int
 823 vc4_wait_bo_ioctl(struct drm_device *dev, void *data,
 824                   struct drm_file *file_priv)
 825 {
 826         int ret;
 827         struct drm_vc4_wait_bo *args = data;
 828         struct drm_gem_object *gem_obj;
 829         struct vc4_bo *bo;
 830
 831         if (args->pad != 0)
 832                 return -EINVAL;
 833
 834         gem_obj = drm_gem_object_lookup(file_priv, args->handle);
 835         if (!gem_obj) {
 836                 DRM_ERROR("Failed to look up GEM BO %d\n", args->handle);
 837                 return -EINVAL;
 838         }
 839         bo = to_vc4_bo(gem_obj);
 840
 841         ret = vc4_wait_for_seqno_ioctl_helper(dev, bo->seqno,
 842                                               &args->timeout_ns);
 843
 844         drm_gem_object_unreference_unlocked(gem_obj);
 845         return ret;
 846 }
 847
 848 /**
 849  * Submits a command list to the VC4.
 850  *
 851  * This is what is called batchbuffer emitting on other hardware.
 852  */
 853 int
 854 vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
 855                     struct drm_file *file_priv)
 856 {
 857         struct vc4_dev *vc4 = to_vc4_dev(dev);
 858         struct drm_vc4_submit_cl *args = data;
 859         struct vc4_exec_info *exec;
 860         int ret = 0;
 861
 862         if ((args->flags & ~VC4_SUBMIT_CL_USE_CLEAR_COLOR) != 0) {
 863                 DRM_ERROR("Unknown flags: 0x%02x\n", args->flags);
 864                 return -EINVAL;
 865         }
 866
 867         exec = kcalloc(1, sizeof(*exec), GFP_KERNEL);
 868         if (!exec) {
 869                 DRM_ERROR("malloc failure on exec struct\n");
 870                 return -ENOMEM;
 871         }
 872
 873         mutex_lock(&vc4->power_lock);
 874         if (vc4->power_refcount++ == 0)
 875                 ret = pm_runtime_get_sync(&vc4->v3d->pdev->dev);
 876         mutex_unlock(&vc4->power_lock);
 877         if (ret < 0) {
 878                 kfree(exec);
 879                 return ret;
 880         }
 881
 882         exec->args = args;
 883         INIT_LIST_HEAD(&exec->unref_list);
 884
 885         ret = vc4_cl_lookup_bos(dev, file_priv, exec);
 886         if (ret)
 887                 goto fail;
 888
 889         if (exec->args->bin_cl_size != 0) {
 890                 ret = vc4_get_bcl(dev, exec);
 891                 if (ret)
 892                         goto fail;
 893         } else {
 894                 exec->ct0ca = 0;
 895                 exec->ct0ea = 0;
 896         }
 897
 898         ret = vc4_get_rcl(dev, exec);
 899         if (ret)
 900                 goto fail;
 901
 902         /* Clear this out of the struct we'll be putting in the queue,
 903          * since it's part of our stack.
 904          */
 905         exec->args = NULL;
 906
 907         vc4_queue_submit(dev, exec);
 908
 909         /* Return the seqno for our job. */
 910         args->seqno = vc4->emit_seqno;
 911
 912         return 0;
 913
 914 fail:
 915         vc4_complete_exec(vc4->dev, exec);
 916
 917         return ret;
 918 }
 919
 920 void
 921 vc4_gem_init(struct drm_device *dev)
 922 {
 923         struct vc4_dev *vc4 = to_vc4_dev(dev);
 924
 925         INIT_LIST_HEAD(&vc4->bin_job_list);
 926         INIT_LIST_HEAD(&vc4->render_job_list);
 927         INIT_LIST_HEAD(&vc4->job_done_list);
 928         INIT_LIST_HEAD(&vc4->seqno_cb_list);
 929         spin_lock_init(&vc4->job_lock);
 930
 931         INIT_WORK(&vc4->hangcheck.reset_work, vc4_reset_work);
 932         setup_timer(&vc4->hangcheck.timer,
 933                     vc4_hangcheck_elapsed,
 934                     (unsigned long)dev);
 935
 936         INIT_WORK(&vc4->job_done_work, vc4_job_done_work);
 937
 938         mutex_init(&vc4->power_lock);
 939 }
 940
 941 void
 942 vc4_gem_destroy(struct drm_device *dev)
 943 {
 944         struct vc4_dev *vc4 = to_vc4_dev(dev);
 945
 946         /* Waiting for exec to finish would need to be done before
 947          * unregistering V3D.
 948          */
 949         WARN_ON(vc4->emit_seqno != vc4->finished_seqno);
 950
 951         /* V3D should already have disabled its interrupt and cleared
 952          * the overflow allocation registers.  Now free the object.
 953          */
 954         if (vc4->overflow_mem) {
 955                 drm_gem_object_unreference_unlocked(&vc4->overflow_mem->base.base);
 956                 vc4->overflow_mem = NULL;
 957         }
 958
 959         if (vc4->hang_state)
 960                 vc4_free_hang_state(dev, vc4->hang_state);
 961
 962         vc4_bo_cache_destroy(dev);
 963 }