drivers/gpu/drm/msm/msm_gpu.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2013 Red Hat
   4  * Author: Rob Clark <[email protected]>
   5  */
   6
   7 #include "drm/drm_drv.h"
   8
   9 #include "msm_gpu.h"
  10 #include "msm_gem.h"
  11 #include "msm_mmu.h"
  12 #include "msm_fence.h"
  13 #include "msm_gpu_trace.h"
  14 #include "adreno/adreno_gpu.h"
  15
  16 #include <generated/utsrelease.h>
  17 #include <linux/string_helpers.h>
  18 #include <linux/devcoredump.h>
  19 #include <linux/sched/task.h>
  20
  21 /*
  22  * Power Management:
  23  */
  24
  25 static int enable_pwrrail(struct msm_gpu *gpu)
  26 {
  27         struct drm_device *dev = gpu->dev;
  28         int ret = 0;
  29
  30         if (gpu->gpu_reg) {
  31                 ret = regulator_enable(gpu->gpu_reg);
  32                 if (ret) {
  33                         DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_reg': %d\n", ret);
  34                         return ret;
  35                 }
  36         }
  37
  38         if (gpu->gpu_cx) {
  39                 ret = regulator_enable(gpu->gpu_cx);
  40                 if (ret) {
  41                         DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_cx': %d\n", ret);
  42                         return ret;
  43                 }
  44         }
  45
  46         return 0;
  47 }
  48
  49 static int disable_pwrrail(struct msm_gpu *gpu)
  50 {
  51         if (gpu->gpu_cx)
  52                 regulator_disable(gpu->gpu_cx);
  53         if (gpu->gpu_reg)
  54                 regulator_disable(gpu->gpu_reg);
  55         return 0;
  56 }
  57
  58 static int enable_clk(struct msm_gpu *gpu)
  59 {
  60         if (gpu->core_clk && gpu->fast_rate)
  61                 dev_pm_opp_set_rate(&gpu->pdev->dev, gpu->fast_rate);
  62
  63         /* Set the RBBM timer rate to 19.2Mhz */
  64         if (gpu->rbbmtimer_clk)
  65                 clk_set_rate(gpu->rbbmtimer_clk, 19200000);
  66
  67         return clk_bulk_prepare_enable(gpu->nr_clocks, gpu->grp_clks);
  68 }
  69
  70 static int disable_clk(struct msm_gpu *gpu)
  71 {
  72         clk_bulk_disable_unprepare(gpu->nr_clocks, gpu->grp_clks);
  73
  74         /*
  75          * Set the clock to a deliberately low rate. On older targets the clock
  76          * speed had to be non zero to avoid problems. On newer targets this
  77          * will be rounded down to zero anyway so it all works out.
  78          */
  79         if (gpu->core_clk)
  80                 dev_pm_opp_set_rate(&gpu->pdev->dev, 27000000);
  81
  82         if (gpu->rbbmtimer_clk)
  83                 clk_set_rate(gpu->rbbmtimer_clk, 0);
  84
  85         return 0;
  86 }
  87
  88 static int enable_axi(struct msm_gpu *gpu)
  89 {
  90         return clk_prepare_enable(gpu->ebi1_clk);
  91 }
  92
  93 static int disable_axi(struct msm_gpu *gpu)
  94 {
  95         clk_disable_unprepare(gpu->ebi1_clk);
  96         return 0;
  97 }
  98
  99 int msm_gpu_pm_resume(struct msm_gpu *gpu)
 100 {
 101         int ret;
 102
 103         DBG("%s", gpu->name);
 104         trace_msm_gpu_resume(0);
 105
 106         ret = enable_pwrrail(gpu);
 107         if (ret)
 108                 return ret;
 109
 110         ret = enable_clk(gpu);
 111         if (ret)
 112                 return ret;
 113
 114         ret = enable_axi(gpu);
 115         if (ret)
 116                 return ret;
 117
 118         msm_devfreq_resume(gpu);
 119
 120         gpu->needs_hw_init = true;
 121
 122         return 0;
 123 }
 124
 125 int msm_gpu_pm_suspend(struct msm_gpu *gpu)
 126 {
 127         int ret;
 128
 129         DBG("%s", gpu->name);
 130         trace_msm_gpu_suspend(0);
 131
 132         msm_devfreq_suspend(gpu);
 133
 134         ret = disable_axi(gpu);
 135         if (ret)
 136                 return ret;
 137
 138         ret = disable_clk(gpu);
 139         if (ret)
 140                 return ret;
 141
 142         ret = disable_pwrrail(gpu);
 143         if (ret)
 144                 return ret;
 145
 146         gpu->suspend_count++;
 147
 148         return 0;
 149 }
 150
 151 void msm_gpu_show_fdinfo(struct msm_gpu *gpu, struct msm_file_private *ctx,
 152                          struct drm_printer *p)
 153 {
 154         drm_printf(p, "drm-driver:\t%s\n", gpu->dev->driver->name);
 155         drm_printf(p, "drm-client-id:\t%u\n", ctx->seqno);
 156         drm_printf(p, "drm-engine-gpu:\t%llu ns\n", ctx->elapsed_ns);
 157         drm_printf(p, "drm-cycles-gpu:\t%llu\n", ctx->cycles);
 158         drm_printf(p, "drm-maxfreq-gpu:\t%u Hz\n", gpu->fast_rate);
 159 }
 160
 161 int msm_gpu_hw_init(struct msm_gpu *gpu)
 162 {
 163         int ret;
 164
 165         WARN_ON(!mutex_is_locked(&gpu->lock));
 166
 167         if (!gpu->needs_hw_init)
 168                 return 0;
 169
 170         disable_irq(gpu->irq);
 171         ret = gpu->funcs->hw_init(gpu);
 172         if (!ret)
 173                 gpu->needs_hw_init = false;
 174         enable_irq(gpu->irq);
 175
 176         return ret;
 177 }
 178
 179 #ifdef CONFIG_DEV_COREDUMP
 180 static ssize_t msm_gpu_devcoredump_read(char *buffer, loff_t offset,
 181                 size_t count, void *data, size_t datalen)
 182 {
 183         struct msm_gpu *gpu = data;
 184         struct drm_print_iterator iter;
 185         struct drm_printer p;
 186         struct msm_gpu_state *state;
 187
 188         state = msm_gpu_crashstate_get(gpu);
 189         if (!state)
 190                 return 0;
 191
 192         iter.data = buffer;
 193         iter.offset = 0;
 194         iter.start = offset;
 195         iter.remain = count;
 196
 197         p = drm_coredump_printer(&iter);
 198
 199         drm_printf(&p, "---\n");
 200         drm_printf(&p, "kernel: " UTS_RELEASE "\n");
 201         drm_printf(&p, "module: " KBUILD_MODNAME "\n");
 202         drm_printf(&p, "time: %lld.%09ld\n",
 203                 state->time.tv_sec, state->time.tv_nsec);
 204         if (state->comm)
 205                 drm_printf(&p, "comm: %s\n", state->comm);
 206         if (state->cmd)
 207                 drm_printf(&p, "cmdline: %s\n", state->cmd);
 208
 209         gpu->funcs->show(gpu, state, &p);
 210
 211         msm_gpu_crashstate_put(gpu);
 212
 213         return count - iter.remain;
 214 }
 215
 216 static void msm_gpu_devcoredump_free(void *data)
 217 {
 218         struct msm_gpu *gpu = data;
 219
 220         msm_gpu_crashstate_put(gpu);
 221 }
 222
 223 static void msm_gpu_crashstate_get_bo(struct msm_gpu_state *state,
 224                 struct msm_gem_object *obj, u64 iova, bool full)
 225 {
 226         struct msm_gpu_state_bo *state_bo = &state->bos[state->nr_bos];
 227
 228         /* Don't record write only objects */
 229         state_bo->size = obj->base.size;
 230         state_bo->iova = iova;
 231
 232         BUILD_BUG_ON(sizeof(state_bo->name) != sizeof(obj->name));
 233
 234         memcpy(state_bo->name, obj->name, sizeof(state_bo->name));
 235
 236         if (full) {
 237                 void *ptr;
 238
 239                 state_bo->data = kvmalloc(obj->base.size, GFP_KERNEL);
 240                 if (!state_bo->data)
 241                         goto out;
 242
 243                 msm_gem_lock(&obj->base);
 244                 ptr = msm_gem_get_vaddr_active(&obj->base);
 245                 msm_gem_unlock(&obj->base);
 246                 if (IS_ERR(ptr)) {
 247                         kvfree(state_bo->data);
 248                         state_bo->data = NULL;
 249                         goto out;
 250                 }
 251
 252                 memcpy(state_bo->data, ptr, obj->base.size);
 253                 msm_gem_put_vaddr(&obj->base);
 254         }
 255 out:
 256         state->nr_bos++;
 257 }
 258
 259 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
 260                 struct msm_gem_submit *submit, char *comm, char *cmd)
 261 {
 262         struct msm_gpu_state *state;
 263
 264         /* Check if the target supports capturing crash state */
 265         if (!gpu->funcs->gpu_state_get)
 266                 return;
 267
 268         /* Only save one crash state at a time */
 269         if (gpu->crashstate)
 270                 return;
 271
 272         state = gpu->funcs->gpu_state_get(gpu);
 273         if (IS_ERR_OR_NULL(state))
 274                 return;
 275
 276         /* Fill in the additional crash state information */
 277         state->comm = kstrdup(comm, GFP_KERNEL);
 278         state->cmd = kstrdup(cmd, GFP_KERNEL);
 279         state->fault_info = gpu->fault_info;
 280
 281         if (submit) {
 282                 int i;
 283
 284                 state->bos = kcalloc(submit->nr_bos,
 285                         sizeof(struct msm_gpu_state_bo), GFP_KERNEL);
 286
 287                 for (i = 0; state->bos && i < submit->nr_bos; i++) {
 288                         msm_gpu_crashstate_get_bo(state, submit->bos[i].obj,
 289                                                   submit->bos[i].iova,
 290                                                   should_dump(submit, i));
 291                 }
 292         }
 293
 294         /* Set the active crash state to be dumped on failure */
 295         gpu->crashstate = state;
 296
 297         /* FIXME: Release the crashstate if this errors out? */
 298         dev_coredumpm(gpu->dev->dev, THIS_MODULE, gpu, 0, GFP_KERNEL,
 299                 msm_gpu_devcoredump_read, msm_gpu_devcoredump_free);
 300 }
 301 #else
 302 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
 303                 struct msm_gem_submit *submit, char *comm, char *cmd)
 304 {
 305 }
 306 #endif
 307
 308 /*
 309  * Hangcheck detection for locked gpu:
 310  */
 311
 312 static struct msm_gem_submit *
 313 find_submit(struct msm_ringbuffer *ring, uint32_t fence)
 314 {
 315         struct msm_gem_submit *submit;
 316         unsigned long flags;
 317
 318         spin_lock_irqsave(&ring->submit_lock, flags);
 319         list_for_each_entry(submit, &ring->submits, node) {
 320                 if (submit->seqno == fence) {
 321                         spin_unlock_irqrestore(&ring->submit_lock, flags);
 322                         return submit;
 323                 }
 324         }
 325         spin_unlock_irqrestore(&ring->submit_lock, flags);
 326
 327         return NULL;
 328 }
 329
 330 static void retire_submits(struct msm_gpu *gpu);
 331
 332 static void get_comm_cmdline(struct msm_gem_submit *submit, char **comm, char **cmd)
 333 {
 334         struct msm_file_private *ctx = submit->queue->ctx;
 335         struct task_struct *task;
 336
 337         WARN_ON(!mutex_is_locked(&submit->gpu->lock));
 338
 339         /* Note that kstrdup will return NULL if argument is NULL: */
 340         *comm = kstrdup(ctx->comm, GFP_KERNEL);
 341         *cmd  = kstrdup(ctx->cmdline, GFP_KERNEL);
 342
 343         task = get_pid_task(submit->pid, PIDTYPE_PID);
 344         if (!task)
 345                 return;
 346
 347         if (!*comm)
 348                 *comm = kstrdup(task->comm, GFP_KERNEL);
 349
 350         if (!*cmd)
 351                 *cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
 352
 353         put_task_struct(task);
 354 }
 355
 356 static void recover_worker(struct kthread_work *work)
 357 {
 358         struct msm_gpu *gpu = container_of(work, struct msm_gpu, recover_work);
 359         struct drm_device *dev = gpu->dev;
 360         struct msm_drm_private *priv = dev->dev_private;
 361         struct msm_gem_submit *submit;
 362         struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
 363         char *comm = NULL, *cmd = NULL;
 364         int i;
 365
 366         mutex_lock(&gpu->lock);
 367
 368         DRM_DEV_ERROR(dev->dev, "%s: hangcheck recover!\n", gpu->name);
 369
 370         submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
 371         if (submit) {
 372                 /* Increment the fault counts */
 373                 submit->queue->faults++;
 374                 if (submit->aspace)
 375                         submit->aspace->faults++;
 376
 377                 get_comm_cmdline(submit, &comm, &cmd);
 378
 379                 if (comm && cmd) {
 380                         DRM_DEV_ERROR(dev->dev, "%s: offending task: %s (%s)\n",
 381                                 gpu->name, comm, cmd);
 382
 383                         msm_rd_dump_submit(priv->hangrd, submit,
 384                                 "offending task: %s (%s)", comm, cmd);
 385                 } else {
 386                         msm_rd_dump_submit(priv->hangrd, submit, NULL);
 387                 }
 388         } else {
 389                 /*
 390                  * We couldn't attribute this fault to any particular context,
 391                  * so increment the global fault count instead.
 392                  */
 393                 gpu->global_faults++;
 394         }
 395
 396         /* Record the crash state */
 397         pm_runtime_get_sync(&gpu->pdev->dev);
 398         msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
 399
 400         kfree(cmd);
 401         kfree(comm);
 402
 403         /*
 404          * Update all the rings with the latest and greatest fence.. this
 405          * needs to happen after msm_rd_dump_submit() to ensure that the
 406          * bo's referenced by the offending submit are still around.
 407          */
 408         for (i = 0; i < gpu->nr_rings; i++) {
 409                 struct msm_ringbuffer *ring = gpu->rb[i];
 410
 411                 uint32_t fence = ring->memptrs->fence;
 412
 413                 /*
 414                  * For the current (faulting?) ring/submit advance the fence by
 415                  * one more to clear the faulting submit
 416                  */
 417                 if (ring == cur_ring)
 418                         ring->memptrs->fence = ++fence;
 419
 420                 msm_update_fence(ring->fctx, fence);
 421         }
 422
 423         if (msm_gpu_active(gpu)) {
 424                 /* retire completed submits, plus the one that hung: */
 425                 retire_submits(gpu);
 426
 427                 gpu->funcs->recover(gpu);
 428
 429                 /*
 430                  * Replay all remaining submits starting with highest priority
 431                  * ring
 432                  */
 433                 for (i = 0; i < gpu->nr_rings; i++) {
 434                         struct msm_ringbuffer *ring = gpu->rb[i];
 435                         unsigned long flags;
 436
 437                         spin_lock_irqsave(&ring->submit_lock, flags);
 438                         list_for_each_entry(submit, &ring->submits, node)
 439                                 gpu->funcs->submit(gpu, submit);
 440                         spin_unlock_irqrestore(&ring->submit_lock, flags);
 441                 }
 442         }
 443
 444         pm_runtime_put(&gpu->pdev->dev);
 445
 446         mutex_unlock(&gpu->lock);
 447
 448         msm_gpu_retire(gpu);
 449 }
 450
 451 static void fault_worker(struct kthread_work *work)
 452 {
 453         struct msm_gpu *gpu = container_of(work, struct msm_gpu, fault_work);
 454         struct msm_gem_submit *submit;
 455         struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
 456         char *comm = NULL, *cmd = NULL;
 457
 458         mutex_lock(&gpu->lock);
 459
 460         submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
 461         if (submit && submit->fault_dumped)
 462                 goto resume_smmu;
 463
 464         if (submit) {
 465                 get_comm_cmdline(submit, &comm, &cmd);
 466
 467                 /*
 468                  * When we get GPU iova faults, we can get 1000s of them,
 469                  * but we really only want to log the first one.
 470                  */
 471                 submit->fault_dumped = true;
 472         }
 473
 474         /* Record the crash state */
 475         pm_runtime_get_sync(&gpu->pdev->dev);
 476         msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
 477         pm_runtime_put_sync(&gpu->pdev->dev);
 478
 479         kfree(cmd);
 480         kfree(comm);
 481
 482 resume_smmu:
 483         memset(&gpu->fault_info, 0, sizeof(gpu->fault_info));
 484         gpu->aspace->mmu->funcs->resume_translation(gpu->aspace->mmu);
 485
 486         mutex_unlock(&gpu->lock);
 487 }
 488
 489 static void hangcheck_timer_reset(struct msm_gpu *gpu)
 490 {
 491         struct msm_drm_private *priv = gpu->dev->dev_private;
 492         mod_timer(&gpu->hangcheck_timer,
 493                         round_jiffies_up(jiffies + msecs_to_jiffies(priv->hangcheck_period)));
 494 }
 495
 496 static bool made_progress(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
 497 {
 498         if (ring->hangcheck_progress_retries >= DRM_MSM_HANGCHECK_PROGRESS_RETRIES)
 499                 return false;
 500
 501         if (!gpu->funcs->progress)
 502                 return false;
 503
 504         if (!gpu->funcs->progress(gpu, ring))
 505                 return false;
 506
 507         ring->hangcheck_progress_retries++;
 508         return true;
 509 }
 510
 511 static void hangcheck_handler(struct timer_list *t)
 512 {
 513         struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer);
 514         struct drm_device *dev = gpu->dev;
 515         struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
 516         uint32_t fence = ring->memptrs->fence;
 517
 518         if (fence != ring->hangcheck_fence) {
 519                 /* some progress has been made.. ya! */
 520                 ring->hangcheck_fence = fence;
 521                 ring->hangcheck_progress_retries = 0;
 522         } else if (fence_before(fence, ring->fctx->last_fence) &&
 523                         !made_progress(gpu, ring)) {
 524                 /* no progress and not done.. hung! */
 525                 ring->hangcheck_fence = fence;
 526                 ring->hangcheck_progress_retries = 0;
 527                 DRM_DEV_ERROR(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n",
 528                                 gpu->name, ring->id);
 529                 DRM_DEV_ERROR(dev->dev, "%s:     completed fence: %u\n",
 530                                 gpu->name, fence);
 531                 DRM_DEV_ERROR(dev->dev, "%s:     submitted fence: %u\n",
 532                                 gpu->name, ring->fctx->last_fence);
 533
 534                 kthread_queue_work(gpu->worker, &gpu->recover_work);
 535         }
 536
 537         /* if still more pending work, reset the hangcheck timer: */
 538         if (fence_after(ring->fctx->last_fence, ring->hangcheck_fence))
 539                 hangcheck_timer_reset(gpu);
 540
 541         /* workaround for missing irq: */
 542         msm_gpu_retire(gpu);
 543 }
 544
 545 /*
 546  * Performance Counters:
 547  */
 548
 549 /* called under perf_lock */
 550 static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
 551 {
 552         uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
 553         int i, n = min(ncntrs, gpu->num_perfcntrs);
 554
 555         /* read current values: */
 556         for (i = 0; i < gpu->num_perfcntrs; i++)
 557                 current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
 558
 559         /* update cntrs: */
 560         for (i = 0; i < n; i++)
 561                 cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
 562
 563         /* save current values: */
 564         for (i = 0; i < gpu->num_perfcntrs; i++)
 565                 gpu->last_cntrs[i] = current_cntrs[i];
 566
 567         return n;
 568 }
 569
 570 static void update_sw_cntrs(struct msm_gpu *gpu)
 571 {
 572         ktime_t time;
 573         uint32_t elapsed;
 574         unsigned long flags;
 575
 576         spin_lock_irqsave(&gpu->perf_lock, flags);
 577         if (!gpu->perfcntr_active)
 578                 goto out;
 579
 580         time = ktime_get();
 581         elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
 582
 583         gpu->totaltime += elapsed;
 584         if (gpu->last_sample.active)
 585                 gpu->activetime += elapsed;
 586
 587         gpu->last_sample.active = msm_gpu_active(gpu);
 588         gpu->last_sample.time = time;
 589
 590 out:
 591         spin_unlock_irqrestore(&gpu->perf_lock, flags);
 592 }
 593
 594 void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
 595 {
 596         unsigned long flags;
 597
 598         pm_runtime_get_sync(&gpu->pdev->dev);
 599
 600         spin_lock_irqsave(&gpu->perf_lock, flags);
 601         /* we could dynamically enable/disable perfcntr registers too.. */
 602         gpu->last_sample.active = msm_gpu_active(gpu);
 603         gpu->last_sample.time = ktime_get();
 604         gpu->activetime = gpu->totaltime = 0;
 605         gpu->perfcntr_active = true;
 606         update_hw_cntrs(gpu, 0, NULL);
 607         spin_unlock_irqrestore(&gpu->perf_lock, flags);
 608 }
 609
 610 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
 611 {
 612         gpu->perfcntr_active = false;
 613         pm_runtime_put_sync(&gpu->pdev->dev);
 614 }
 615
 616 /* returns -errno or # of cntrs sampled */
 617 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
 618                 uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
 619 {
 620         unsigned long flags;
 621         int ret;
 622
 623         spin_lock_irqsave(&gpu->perf_lock, flags);
 624
 625         if (!gpu->perfcntr_active) {
 626                 ret = -EINVAL;
 627                 goto out;
 628         }
 629
 630         *activetime = gpu->activetime;
 631         *totaltime = gpu->totaltime;
 632
 633         gpu->activetime = gpu->totaltime = 0;
 634
 635         ret = update_hw_cntrs(gpu, ncntrs, cntrs);
 636
 637 out:
 638         spin_unlock_irqrestore(&gpu->perf_lock, flags);
 639
 640         return ret;
 641 }
 642
 643 /*
 644  * Cmdstream submission/retirement:
 645  */
 646
 647 static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
 648                 struct msm_gem_submit *submit)
 649 {
 650         int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
 651         volatile struct msm_gpu_submit_stats *stats;
 652         u64 elapsed, clock = 0, cycles;
 653         unsigned long flags;
 654
 655         stats = &ring->memptrs->stats[index];
 656         /* Convert 19.2Mhz alwayson ticks to nanoseconds for elapsed time */
 657         elapsed = (stats->alwayson_end - stats->alwayson_start) * 10000;
 658         do_div(elapsed, 192);
 659
 660         cycles = stats->cpcycles_end - stats->cpcycles_start;
 661
 662         /* Calculate the clock frequency from the number of CP cycles */
 663         if (elapsed) {
 664                 clock = cycles * 1000;
 665                 do_div(clock, elapsed);
 666         }
 667
 668         submit->queue->ctx->elapsed_ns += elapsed;
 669         submit->queue->ctx->cycles     += cycles;
 670
 671         trace_msm_gpu_submit_retired(submit, elapsed, clock,
 672                 stats->alwayson_start, stats->alwayson_end);
 673
 674         msm_submit_retire(submit);
 675
 676         pm_runtime_mark_last_busy(&gpu->pdev->dev);
 677
 678         spin_lock_irqsave(&ring->submit_lock, flags);
 679         list_del(&submit->node);
 680         spin_unlock_irqrestore(&ring->submit_lock, flags);
 681
 682         /* Update devfreq on transition from active->idle: */
 683         mutex_lock(&gpu->active_lock);
 684         gpu->active_submits--;
 685         WARN_ON(gpu->active_submits < 0);
 686         if (!gpu->active_submits) {
 687                 msm_devfreq_idle(gpu);
 688                 pm_runtime_put_autosuspend(&gpu->pdev->dev);
 689         }
 690
 691         mutex_unlock(&gpu->active_lock);
 692
 693         msm_gem_submit_put(submit);
 694 }
 695
 696 static void retire_submits(struct msm_gpu *gpu)
 697 {
 698         int i;
 699
 700         /* Retire the commits starting with highest priority */
 701         for (i = 0; i < gpu->nr_rings; i++) {
 702                 struct msm_ringbuffer *ring = gpu->rb[i];
 703
 704                 while (true) {
 705                         struct msm_gem_submit *submit = NULL;
 706                         unsigned long flags;
 707
 708                         spin_lock_irqsave(&ring->submit_lock, flags);
 709                         submit = list_first_entry_or_null(&ring->submits,
 710                                         struct msm_gem_submit, node);
 711                         spin_unlock_irqrestore(&ring->submit_lock, flags);
 712
 713                         /*
 714                          * If no submit, we are done.  If submit->fence hasn't
 715                          * been signalled, then later submits are not signalled
 716                          * either, so we are also done.
 717                          */
 718                         if (submit && dma_fence_is_signaled(submit->hw_fence)) {
 719                                 retire_submit(gpu, ring, submit);
 720                         } else {
 721                                 break;
 722                         }
 723                 }
 724         }
 725
 726         wake_up_all(&gpu->retire_event);
 727 }
 728
 729 static void retire_worker(struct kthread_work *work)
 730 {
 731         struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
 732
 733         retire_submits(gpu);
 734 }
 735
 736 /* call from irq handler to schedule work to retire bo's */
 737 void msm_gpu_retire(struct msm_gpu *gpu)
 738 {
 739         int i;
 740
 741         for (i = 0; i < gpu->nr_rings; i++)
 742                 msm_update_fence(gpu->rb[i]->fctx, gpu->rb[i]->memptrs->fence);
 743
 744         kthread_queue_work(gpu->worker, &gpu->retire_work);
 745         update_sw_cntrs(gpu);
 746 }
 747
 748 /* add bo's to gpu's ring, and kick gpu: */
 749 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
 750 {
 751         struct drm_device *dev = gpu->dev;
 752         struct msm_drm_private *priv = dev->dev_private;
 753         struct msm_ringbuffer *ring = submit->ring;
 754         unsigned long flags;
 755
 756         WARN_ON(!mutex_is_locked(&gpu->lock));
 757
 758         pm_runtime_get_sync(&gpu->pdev->dev);
 759
 760         msm_gpu_hw_init(gpu);
 761
 762         submit->seqno = submit->hw_fence->seqno;
 763
 764         msm_rd_dump_submit(priv->rd, submit, NULL);
 765
 766         update_sw_cntrs(gpu);
 767
 768         /*
 769          * ring->submits holds a ref to the submit, to deal with the case
 770          * that a submit completes before msm_ioctl_gem_submit() returns.
 771          */
 772         msm_gem_submit_get(submit);
 773
 774         spin_lock_irqsave(&ring->submit_lock, flags);
 775         list_add_tail(&submit->node, &ring->submits);
 776         spin_unlock_irqrestore(&ring->submit_lock, flags);
 777
 778         /* Update devfreq on transition from idle->active: */
 779         mutex_lock(&gpu->active_lock);
 780         if (!gpu->active_submits) {
 781                 pm_runtime_get(&gpu->pdev->dev);
 782                 msm_devfreq_active(gpu);
 783         }
 784         gpu->active_submits++;
 785         mutex_unlock(&gpu->active_lock);
 786
 787         gpu->funcs->submit(gpu, submit);
 788         gpu->cur_ctx_seqno = submit->queue->ctx->seqno;
 789
 790         pm_runtime_put(&gpu->pdev->dev);
 791         hangcheck_timer_reset(gpu);
 792 }
 793
 794 /*
 795  * Init/Cleanup:
 796  */
 797
 798 static irqreturn_t irq_handler(int irq, void *data)
 799 {
 800         struct msm_gpu *gpu = data;
 801         return gpu->funcs->irq(gpu);
 802 }
 803
 804 static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)
 805 {
 806         int ret = devm_clk_bulk_get_all(&pdev->dev, &gpu->grp_clks);
 807
 808         if (ret < 1) {
 809                 gpu->nr_clocks = 0;
 810                 return ret;
 811         }
 812
 813         gpu->nr_clocks = ret;
 814
 815         gpu->core_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
 816                 gpu->nr_clocks, "core");
 817
 818         gpu->rbbmtimer_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
 819                 gpu->nr_clocks, "rbbmtimer");
 820
 821         return 0;
 822 }
 823
 824 /* Return a new address space for a msm_drm_private instance */
 825 struct msm_gem_address_space *
 826 msm_gpu_create_private_address_space(struct msm_gpu *gpu, struct task_struct *task)
 827 {
 828         struct msm_gem_address_space *aspace = NULL;
 829         if (!gpu)
 830                 return NULL;
 831
 832         /*
 833          * If the target doesn't support private address spaces then return
 834          * the global one
 835          */
 836         if (gpu->funcs->create_private_address_space) {
 837                 aspace = gpu->funcs->create_private_address_space(gpu);
 838                 if (!IS_ERR(aspace))
 839                         aspace->pid = get_pid(task_pid(task));
 840         }
 841
 842         if (IS_ERR_OR_NULL(aspace))
 843                 aspace = msm_gem_address_space_get(gpu->aspace);
 844
 845         return aspace;
 846 }
 847
 848 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
 849                 struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
 850                 const char *name, struct msm_gpu_config *config)
 851 {
 852         struct msm_drm_private *priv = drm->dev_private;
 853         int i, ret, nr_rings = config->nr_rings;
 854         void *memptrs;
 855         uint64_t memptrs_iova;
 856
 857         if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
 858                 gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
 859
 860         gpu->dev = drm;
 861         gpu->funcs = funcs;
 862         gpu->name = name;
 863
 864         gpu->worker = kthread_create_worker(0, "gpu-worker");
 865         if (IS_ERR(gpu->worker)) {
 866                 ret = PTR_ERR(gpu->worker);
 867                 gpu->worker = NULL;
 868                 goto fail;
 869         }
 870
 871         sched_set_fifo_low(gpu->worker->task);
 872
 873         mutex_init(&gpu->active_lock);
 874         mutex_init(&gpu->lock);
 875         init_waitqueue_head(&gpu->retire_event);
 876         kthread_init_work(&gpu->retire_work, retire_worker);
 877         kthread_init_work(&gpu->recover_work, recover_worker);
 878         kthread_init_work(&gpu->fault_work, fault_worker);
 879
 880         priv->hangcheck_period = DRM_MSM_HANGCHECK_DEFAULT_PERIOD;
 881
 882         /*
 883          * If progress detection is supported, halve the hangcheck timer
 884          * duration, as it takes two iterations of the hangcheck handler
 885          * to detect a hang.
 886          */
 887         if (funcs->progress)
 888                 priv->hangcheck_period /= 2;
 889
 890         timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
 891
 892         spin_lock_init(&gpu->perf_lock);
 893
 894
 895         /* Map registers: */
 896         gpu->mmio = msm_ioremap(pdev, config->ioname);
 897         if (IS_ERR(gpu->mmio)) {
 898                 ret = PTR_ERR(gpu->mmio);
 899                 goto fail;
 900         }
 901
 902         /* Get Interrupt: */
 903         gpu->irq = platform_get_irq(pdev, 0);
 904         if (gpu->irq < 0) {
 905                 ret = gpu->irq;
 906                 DRM_DEV_ERROR(drm->dev, "failed to get irq: %d\n", ret);
 907                 goto fail;
 908         }
 909
 910         ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
 911                         IRQF_TRIGGER_HIGH, "gpu-irq", gpu);
 912         if (ret) {
 913                 DRM_DEV_ERROR(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
 914                 goto fail;
 915         }
 916
 917         ret = get_clocks(pdev, gpu);
 918         if (ret)
 919                 goto fail;
 920
 921         gpu->ebi1_clk = msm_clk_get(pdev, "bus");
 922         DBG("ebi1_clk: %p", gpu->ebi1_clk);
 923         if (IS_ERR(gpu->ebi1_clk))
 924                 gpu->ebi1_clk = NULL;
 925
 926         /* Acquire regulators: */
 927         gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
 928         DBG("gpu_reg: %p", gpu->gpu_reg);
 929         if (IS_ERR(gpu->gpu_reg))
 930                 gpu->gpu_reg = NULL;
 931
 932         gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
 933         DBG("gpu_cx: %p", gpu->gpu_cx);
 934         if (IS_ERR(gpu->gpu_cx))
 935                 gpu->gpu_cx = NULL;
 936
 937         gpu->pdev = pdev;
 938         platform_set_drvdata(pdev, &gpu->adreno_smmu);
 939
 940         msm_devfreq_init(gpu);
 941
 942
 943         gpu->aspace = gpu->funcs->create_address_space(gpu, pdev);
 944
 945         if (gpu->aspace == NULL)
 946                 DRM_DEV_INFO(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
 947         else if (IS_ERR(gpu->aspace)) {
 948                 ret = PTR_ERR(gpu->aspace);
 949                 goto fail;
 950         }
 951
 952         memptrs = msm_gem_kernel_new(drm,
 953                 sizeof(struct msm_rbmemptrs) * nr_rings,
 954                 check_apriv(gpu, MSM_BO_WC), gpu->aspace, &gpu->memptrs_bo,
 955                 &memptrs_iova);
 956
 957         if (IS_ERR(memptrs)) {
 958                 ret = PTR_ERR(memptrs);
 959                 DRM_DEV_ERROR(drm->dev, "could not allocate memptrs: %d\n", ret);
 960                 goto fail;
 961         }
 962
 963         msm_gem_object_set_name(gpu->memptrs_bo, "memptrs");
 964
 965         if (nr_rings > ARRAY_SIZE(gpu->rb)) {
 966                 DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n",
 967                         ARRAY_SIZE(gpu->rb));
 968                 nr_rings = ARRAY_SIZE(gpu->rb);
 969         }
 970
 971         /* Create ringbuffer(s): */
 972         for (i = 0; i < nr_rings; i++) {
 973                 gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova);
 974
 975                 if (IS_ERR(gpu->rb[i])) {
 976                         ret = PTR_ERR(gpu->rb[i]);
 977                         DRM_DEV_ERROR(drm->dev,
 978                                 "could not create ringbuffer %d: %d\n", i, ret);
 979                         goto fail;
 980                 }
 981
 982                 memptrs += sizeof(struct msm_rbmemptrs);
 983                 memptrs_iova += sizeof(struct msm_rbmemptrs);
 984         }
 985
 986         gpu->nr_rings = nr_rings;
 987
 988         refcount_set(&gpu->sysprof_active, 1);
 989
 990         return 0;
 991
 992 fail:
 993         for (i = 0; i < ARRAY_SIZE(gpu->rb); i++)  {
 994                 msm_ringbuffer_destroy(gpu->rb[i]);
 995                 gpu->rb[i] = NULL;
 996         }
 997
 998         msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
 999
1000         platform_set_drvdata(pdev, NULL);
1001         return ret;
1002 }
1003
1004 void msm_gpu_cleanup(struct msm_gpu *gpu)
1005 {
1006         int i;
1007
1008         DBG("%s", gpu->name);
1009
1010         for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
1011                 msm_ringbuffer_destroy(gpu->rb[i]);
1012                 gpu->rb[i] = NULL;
1013         }
1014
1015         msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
1016
1017         if (!IS_ERR_OR_NULL(gpu->aspace)) {
1018                 gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu);
1019                 msm_gem_address_space_put(gpu->aspace);
1020         }
1021
1022         if (gpu->worker) {
1023                 kthread_destroy_worker(gpu->worker);
1024         }
1025
1026         msm_devfreq_cleanup(gpu);
1027
1028         platform_set_drvdata(gpu->pdev, NULL);
1029 }