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 "msm_gpu.h"
   8 #include "msm_gem.h"
   9 #include "msm_mmu.h"
  10 #include "msm_fence.h"
  11 #include "msm_gpu_trace.h"
  12 #include "adreno/adreno_gpu.h"
  13
  14 #include <generated/utsrelease.h>
  15 #include <linux/string_helpers.h>
  16 #include <linux/devcoredump.h>
  17 #include <linux/sched/task.h>
  18
  19 /*
  20  * Power Management:
  21  */
  22
  23 static int enable_pwrrail(struct msm_gpu *gpu)
  24 {
  25         struct drm_device *dev = gpu->dev;
  26         int ret = 0;
  27
  28         if (gpu->gpu_reg) {
  29                 ret = regulator_enable(gpu->gpu_reg);
  30                 if (ret) {
  31                         DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_reg': %d\n", ret);
  32                         return ret;
  33                 }
  34         }
  35
  36         if (gpu->gpu_cx) {
  37                 ret = regulator_enable(gpu->gpu_cx);
  38                 if (ret) {
  39                         DRM_DEV_ERROR(dev->dev, "failed to enable 'gpu_cx': %d\n", ret);
  40                         return ret;
  41                 }
  42         }
  43
  44         return 0;
  45 }
  46
  47 static int disable_pwrrail(struct msm_gpu *gpu)
  48 {
  49         if (gpu->gpu_cx)
  50                 regulator_disable(gpu->gpu_cx);
  51         if (gpu->gpu_reg)
  52                 regulator_disable(gpu->gpu_reg);
  53         return 0;
  54 }
  55
  56 static int enable_clk(struct msm_gpu *gpu)
  57 {
  58         if (gpu->core_clk && gpu->fast_rate)
  59                 clk_set_rate(gpu->core_clk, gpu->fast_rate);
  60
  61         /* Set the RBBM timer rate to 19.2Mhz */
  62         if (gpu->rbbmtimer_clk)
  63                 clk_set_rate(gpu->rbbmtimer_clk, 19200000);
  64
  65         return clk_bulk_prepare_enable(gpu->nr_clocks, gpu->grp_clks);
  66 }
  67
  68 static int disable_clk(struct msm_gpu *gpu)
  69 {
  70         clk_bulk_disable_unprepare(gpu->nr_clocks, gpu->grp_clks);
  71
  72         /*
  73          * Set the clock to a deliberately low rate. On older targets the clock
  74          * speed had to be non zero to avoid problems. On newer targets this
  75          * will be rounded down to zero anyway so it all works out.
  76          */
  77         if (gpu->core_clk)
  78                 clk_set_rate(gpu->core_clk, 27000000);
  79
  80         if (gpu->rbbmtimer_clk)
  81                 clk_set_rate(gpu->rbbmtimer_clk, 0);
  82
  83         return 0;
  84 }
  85
  86 static int enable_axi(struct msm_gpu *gpu)
  87 {
  88         return clk_prepare_enable(gpu->ebi1_clk);
  89 }
  90
  91 static int disable_axi(struct msm_gpu *gpu)
  92 {
  93         clk_disable_unprepare(gpu->ebi1_clk);
  94         return 0;
  95 }
  96
  97 int msm_gpu_pm_resume(struct msm_gpu *gpu)
  98 {
  99         int ret;
 100
 101         DBG("%s", gpu->name);
 102         trace_msm_gpu_resume(0);
 103
 104         ret = enable_pwrrail(gpu);
 105         if (ret)
 106                 return ret;
 107
 108         ret = enable_clk(gpu);
 109         if (ret)
 110                 return ret;
 111
 112         ret = enable_axi(gpu);
 113         if (ret)
 114                 return ret;
 115
 116         msm_devfreq_resume(gpu);
 117
 118         gpu->needs_hw_init = true;
 119
 120         return 0;
 121 }
 122
 123 int msm_gpu_pm_suspend(struct msm_gpu *gpu)
 124 {
 125         int ret;
 126
 127         DBG("%s", gpu->name);
 128         trace_msm_gpu_suspend(0);
 129
 130         msm_devfreq_suspend(gpu);
 131
 132         ret = disable_axi(gpu);
 133         if (ret)
 134                 return ret;
 135
 136         ret = disable_clk(gpu);
 137         if (ret)
 138                 return ret;
 139
 140         ret = disable_pwrrail(gpu);
 141         if (ret)
 142                 return ret;
 143
 144         gpu->suspend_count++;
 145
 146         return 0;
 147 }
 148
 149 int msm_gpu_hw_init(struct msm_gpu *gpu)
 150 {
 151         int ret;
 152
 153         WARN_ON(!mutex_is_locked(&gpu->dev->struct_mutex));
 154
 155         if (!gpu->needs_hw_init)
 156                 return 0;
 157
 158         disable_irq(gpu->irq);
 159         ret = gpu->funcs->hw_init(gpu);
 160         if (!ret)
 161                 gpu->needs_hw_init = false;
 162         enable_irq(gpu->irq);
 163
 164         return ret;
 165 }
 166
 167 static void update_fences(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
 168                 uint32_t fence)
 169 {
 170         struct msm_gem_submit *submit;
 171         unsigned long flags;
 172
 173         spin_lock_irqsave(&ring->submit_lock, flags);
 174         list_for_each_entry(submit, &ring->submits, node) {
 175                 if (submit->seqno > fence)
 176                         break;
 177
 178                 msm_update_fence(submit->ring->fctx,
 179                         submit->hw_fence->seqno);
 180                 dma_fence_signal(submit->hw_fence);
 181         }
 182         spin_unlock_irqrestore(&ring->submit_lock, flags);
 183 }
 184
 185 #ifdef CONFIG_DEV_COREDUMP
 186 static ssize_t msm_gpu_devcoredump_read(char *buffer, loff_t offset,
 187                 size_t count, void *data, size_t datalen)
 188 {
 189         struct msm_gpu *gpu = data;
 190         struct drm_print_iterator iter;
 191         struct drm_printer p;
 192         struct msm_gpu_state *state;
 193
 194         state = msm_gpu_crashstate_get(gpu);
 195         if (!state)
 196                 return 0;
 197
 198         iter.data = buffer;
 199         iter.offset = 0;
 200         iter.start = offset;
 201         iter.remain = count;
 202
 203         p = drm_coredump_printer(&iter);
 204
 205         drm_printf(&p, "---\n");
 206         drm_printf(&p, "kernel: " UTS_RELEASE "\n");
 207         drm_printf(&p, "module: " KBUILD_MODNAME "\n");
 208         drm_printf(&p, "time: %lld.%09ld\n",
 209                 state->time.tv_sec, state->time.tv_nsec);
 210         if (state->comm)
 211                 drm_printf(&p, "comm: %s\n", state->comm);
 212         if (state->cmd)
 213                 drm_printf(&p, "cmdline: %s\n", state->cmd);
 214
 215         gpu->funcs->show(gpu, state, &p);
 216
 217         msm_gpu_crashstate_put(gpu);
 218
 219         return count - iter.remain;
 220 }
 221
 222 static void msm_gpu_devcoredump_free(void *data)
 223 {
 224         struct msm_gpu *gpu = data;
 225
 226         msm_gpu_crashstate_put(gpu);
 227 }
 228
 229 static void msm_gpu_crashstate_get_bo(struct msm_gpu_state *state,
 230                 struct msm_gem_object *obj, u64 iova, u32 flags)
 231 {
 232         struct msm_gpu_state_bo *state_bo = &state->bos[state->nr_bos];
 233
 234         /* Don't record write only objects */
 235         state_bo->size = obj->base.size;
 236         state_bo->iova = iova;
 237
 238         /* Only store data for non imported buffer objects marked for read */
 239         if ((flags & MSM_SUBMIT_BO_READ) && !obj->base.import_attach) {
 240                 void *ptr;
 241
 242                 state_bo->data = kvmalloc(obj->base.size, GFP_KERNEL);
 243                 if (!state_bo->data)
 244                         goto out;
 245
 246                 msm_gem_lock(&obj->base);
 247                 ptr = msm_gem_get_vaddr_active(&obj->base);
 248                 msm_gem_unlock(&obj->base);
 249                 if (IS_ERR(ptr)) {
 250                         kvfree(state_bo->data);
 251                         state_bo->data = NULL;
 252                         goto out;
 253                 }
 254
 255                 memcpy(state_bo->data, ptr, obj->base.size);
 256                 msm_gem_put_vaddr(&obj->base);
 257         }
 258 out:
 259         state->nr_bos++;
 260 }
 261
 262 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
 263                 struct msm_gem_submit *submit, char *comm, char *cmd)
 264 {
 265         struct msm_gpu_state *state;
 266
 267         /* Check if the target supports capturing crash state */
 268         if (!gpu->funcs->gpu_state_get)
 269                 return;
 270
 271         /* Only save one crash state at a time */
 272         if (gpu->crashstate)
 273                 return;
 274
 275         state = gpu->funcs->gpu_state_get(gpu);
 276         if (IS_ERR_OR_NULL(state))
 277                 return;
 278
 279         /* Fill in the additional crash state information */
 280         state->comm = kstrdup(comm, GFP_KERNEL);
 281         state->cmd = kstrdup(cmd, GFP_KERNEL);
 282         state->fault_info = gpu->fault_info;
 283
 284         if (submit) {
 285                 int i, nr = 0;
 286
 287                 /* count # of buffers to dump: */
 288                 for (i = 0; i < submit->nr_bos; i++)
 289                         if (should_dump(submit, i))
 290                                 nr++;
 291                 /* always dump cmd bo's, but don't double count them: */
 292                 for (i = 0; i < submit->nr_cmds; i++)
 293                         if (!should_dump(submit, submit->cmd[i].idx))
 294                                 nr++;
 295
 296                 state->bos = kcalloc(nr,
 297                         sizeof(struct msm_gpu_state_bo), GFP_KERNEL);
 298
 299                 for (i = 0; i < submit->nr_bos; i++) {
 300                         if (should_dump(submit, i)) {
 301                                 msm_gpu_crashstate_get_bo(state, submit->bos[i].obj,
 302                                         submit->bos[i].iova, submit->bos[i].flags);
 303                         }
 304                 }
 305
 306                 for (i = 0; state->bos && i < submit->nr_cmds; i++) {
 307                         int idx = submit->cmd[i].idx;
 308
 309                         if (!should_dump(submit, submit->cmd[i].idx)) {
 310                                 msm_gpu_crashstate_get_bo(state, submit->bos[idx].obj,
 311                                         submit->bos[idx].iova, submit->bos[idx].flags);
 312                         }
 313                 }
 314         }
 315
 316         /* Set the active crash state to be dumped on failure */
 317         gpu->crashstate = state;
 318
 319         /* FIXME: Release the crashstate if this errors out? */
 320         dev_coredumpm(gpu->dev->dev, THIS_MODULE, gpu, 0, GFP_KERNEL,
 321                 msm_gpu_devcoredump_read, msm_gpu_devcoredump_free);
 322 }
 323 #else
 324 static void msm_gpu_crashstate_capture(struct msm_gpu *gpu,
 325                 struct msm_gem_submit *submit, char *comm, char *cmd)
 326 {
 327 }
 328 #endif
 329
 330 /*
 331  * Hangcheck detection for locked gpu:
 332  */
 333
 334 static struct msm_gem_submit *
 335 find_submit(struct msm_ringbuffer *ring, uint32_t fence)
 336 {
 337         struct msm_gem_submit *submit;
 338         unsigned long flags;
 339
 340         spin_lock_irqsave(&ring->submit_lock, flags);
 341         list_for_each_entry(submit, &ring->submits, node) {
 342                 if (submit->seqno == fence) {
 343                         spin_unlock_irqrestore(&ring->submit_lock, flags);
 344                         return submit;
 345                 }
 346         }
 347         spin_unlock_irqrestore(&ring->submit_lock, flags);
 348
 349         return NULL;
 350 }
 351
 352 static void retire_submits(struct msm_gpu *gpu);
 353
 354 static void recover_worker(struct kthread_work *work)
 355 {
 356         struct msm_gpu *gpu = container_of(work, struct msm_gpu, recover_work);
 357         struct drm_device *dev = gpu->dev;
 358         struct msm_drm_private *priv = dev->dev_private;
 359         struct msm_gem_submit *submit;
 360         struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu);
 361         char *comm = NULL, *cmd = NULL;
 362         int i;
 363
 364         mutex_lock(&dev->struct_mutex);
 365
 366         DRM_DEV_ERROR(dev->dev, "%s: hangcheck recover!\n", gpu->name);
 367
 368         submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1);
 369         if (submit) {
 370                 struct task_struct *task;
 371
 372                 /* Increment the fault counts */
 373                 gpu->global_faults++;
 374                 submit->queue->faults++;
 375
 376                 task = get_pid_task(submit->pid, PIDTYPE_PID);
 377                 if (task) {
 378                         comm = kstrdup(task->comm, GFP_KERNEL);
 379                         cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
 380                         put_task_struct(task);
 381                 }
 382
 383                 if (comm && cmd) {
 384                         DRM_DEV_ERROR(dev->dev, "%s: offending task: %s (%s)\n",
 385                                 gpu->name, comm, cmd);
 386
 387                         msm_rd_dump_submit(priv->hangrd, submit,
 388                                 "offending task: %s (%s)", comm, cmd);
 389                 } else {
 390                         msm_rd_dump_submit(priv->hangrd, submit, NULL);
 391                 }
 392         }
 393
 394         /* Record the crash state */
 395         pm_runtime_get_sync(&gpu->pdev->dev);
 396         msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
 397         pm_runtime_put_sync(&gpu->pdev->dev);
 398
 399         kfree(cmd);
 400         kfree(comm);
 401
 402         /*
 403          * Update all the rings with the latest and greatest fence.. this
 404          * needs to happen after msm_rd_dump_submit() to ensure that the
 405          * bo's referenced by the offending submit are still around.
 406          */
 407         for (i = 0; i < gpu->nr_rings; i++) {
 408                 struct msm_ringbuffer *ring = gpu->rb[i];
 409
 410                 uint32_t fence = ring->memptrs->fence;
 411
 412                 /*
 413                  * For the current (faulting?) ring/submit advance the fence by
 414                  * one more to clear the faulting submit
 415                  */
 416                 if (ring == cur_ring)
 417                         fence++;
 418
 419                 update_fences(gpu, ring, fence);
 420         }
 421
 422         if (msm_gpu_active(gpu)) {
 423                 /* retire completed submits, plus the one that hung: */
 424                 retire_submits(gpu);
 425
 426                 pm_runtime_get_sync(&gpu->pdev->dev);
 427                 gpu->funcs->recover(gpu);
 428                 pm_runtime_put_sync(&gpu->pdev->dev);
 429
 430                 /*
 431                  * Replay all remaining submits starting with highest priority
 432                  * ring
 433                  */
 434                 for (i = 0; i < gpu->nr_rings; i++) {
 435                         struct msm_ringbuffer *ring = gpu->rb[i];
 436                         unsigned long flags;
 437
 438                         spin_lock_irqsave(&ring->submit_lock, flags);
 439                         list_for_each_entry(submit, &ring->submits, node)
 440                                 gpu->funcs->submit(gpu, submit);
 441                         spin_unlock_irqrestore(&ring->submit_lock, flags);
 442                 }
 443         }
 444
 445         mutex_unlock(&dev->struct_mutex);
 446
 447         msm_gpu_retire(gpu);
 448 }
 449
 450 static void fault_worker(struct kthread_work *work)
 451 {
 452         struct msm_gpu *gpu = container_of(work, struct msm_gpu, fault_work);
 453         struct drm_device *dev = gpu->dev;
 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(&dev->struct_mutex);
 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                 struct task_struct *task;
 466
 467                 task = get_pid_task(submit->pid, PIDTYPE_PID);
 468                 if (task) {
 469                         comm = kstrdup(task->comm, GFP_KERNEL);
 470                         cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
 471                         put_task_struct(task);
 472                 }
 473
 474                 /*
 475                  * When we get GPU iova faults, we can get 1000s of them,
 476                  * but we really only want to log the first one.
 477                  */
 478                 submit->fault_dumped = true;
 479         }
 480
 481         /* Record the crash state */
 482         pm_runtime_get_sync(&gpu->pdev->dev);
 483         msm_gpu_crashstate_capture(gpu, submit, comm, cmd);
 484         pm_runtime_put_sync(&gpu->pdev->dev);
 485
 486         kfree(cmd);
 487         kfree(comm);
 488
 489 resume_smmu:
 490         memset(&gpu->fault_info, 0, sizeof(gpu->fault_info));
 491         gpu->aspace->mmu->funcs->resume_translation(gpu->aspace->mmu);
 492
 493         mutex_unlock(&dev->struct_mutex);
 494 }
 495
 496 static void hangcheck_timer_reset(struct msm_gpu *gpu)
 497 {
 498         struct msm_drm_private *priv = gpu->dev->dev_private;
 499         mod_timer(&gpu->hangcheck_timer,
 500                         round_jiffies_up(jiffies + msecs_to_jiffies(priv->hangcheck_period)));
 501 }
 502
 503 static void hangcheck_handler(struct timer_list *t)
 504 {
 505         struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer);
 506         struct drm_device *dev = gpu->dev;
 507         struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu);
 508         uint32_t fence = ring->memptrs->fence;
 509
 510         if (fence != ring->hangcheck_fence) {
 511                 /* some progress has been made.. ya! */
 512                 ring->hangcheck_fence = fence;
 513         } else if (fence < ring->seqno) {
 514                 /* no progress and not done.. hung! */
 515                 ring->hangcheck_fence = fence;
 516                 DRM_DEV_ERROR(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n",
 517                                 gpu->name, ring->id);
 518                 DRM_DEV_ERROR(dev->dev, "%s:     completed fence: %u\n",
 519                                 gpu->name, fence);
 520                 DRM_DEV_ERROR(dev->dev, "%s:     submitted fence: %u\n",
 521                                 gpu->name, ring->seqno);
 522
 523                 kthread_queue_work(gpu->worker, &gpu->recover_work);
 524         }
 525
 526         /* if still more pending work, reset the hangcheck timer: */
 527         if (ring->seqno > ring->hangcheck_fence)
 528                 hangcheck_timer_reset(gpu);
 529
 530         /* workaround for missing irq: */
 531         msm_gpu_retire(gpu);
 532 }
 533
 534 /*
 535  * Performance Counters:
 536  */
 537
 538 /* called under perf_lock */
 539 static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs)
 540 {
 541         uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)];
 542         int i, n = min(ncntrs, gpu->num_perfcntrs);
 543
 544         /* read current values: */
 545         for (i = 0; i < gpu->num_perfcntrs; i++)
 546                 current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg);
 547
 548         /* update cntrs: */
 549         for (i = 0; i < n; i++)
 550                 cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i];
 551
 552         /* save current values: */
 553         for (i = 0; i < gpu->num_perfcntrs; i++)
 554                 gpu->last_cntrs[i] = current_cntrs[i];
 555
 556         return n;
 557 }
 558
 559 static void update_sw_cntrs(struct msm_gpu *gpu)
 560 {
 561         ktime_t time;
 562         uint32_t elapsed;
 563         unsigned long flags;
 564
 565         spin_lock_irqsave(&gpu->perf_lock, flags);
 566         if (!gpu->perfcntr_active)
 567                 goto out;
 568
 569         time = ktime_get();
 570         elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time));
 571
 572         gpu->totaltime += elapsed;
 573         if (gpu->last_sample.active)
 574                 gpu->activetime += elapsed;
 575
 576         gpu->last_sample.active = msm_gpu_active(gpu);
 577         gpu->last_sample.time = time;
 578
 579 out:
 580         spin_unlock_irqrestore(&gpu->perf_lock, flags);
 581 }
 582
 583 void msm_gpu_perfcntr_start(struct msm_gpu *gpu)
 584 {
 585         unsigned long flags;
 586
 587         pm_runtime_get_sync(&gpu->pdev->dev);
 588
 589         spin_lock_irqsave(&gpu->perf_lock, flags);
 590         /* we could dynamically enable/disable perfcntr registers too.. */
 591         gpu->last_sample.active = msm_gpu_active(gpu);
 592         gpu->last_sample.time = ktime_get();
 593         gpu->activetime = gpu->totaltime = 0;
 594         gpu->perfcntr_active = true;
 595         update_hw_cntrs(gpu, 0, NULL);
 596         spin_unlock_irqrestore(&gpu->perf_lock, flags);
 597 }
 598
 599 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu)
 600 {
 601         gpu->perfcntr_active = false;
 602         pm_runtime_put_sync(&gpu->pdev->dev);
 603 }
 604
 605 /* returns -errno or # of cntrs sampled */
 606 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime,
 607                 uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs)
 608 {
 609         unsigned long flags;
 610         int ret;
 611
 612         spin_lock_irqsave(&gpu->perf_lock, flags);
 613
 614         if (!gpu->perfcntr_active) {
 615                 ret = -EINVAL;
 616                 goto out;
 617         }
 618
 619         *activetime = gpu->activetime;
 620         *totaltime = gpu->totaltime;
 621
 622         gpu->activetime = gpu->totaltime = 0;
 623
 624         ret = update_hw_cntrs(gpu, ncntrs, cntrs);
 625
 626 out:
 627         spin_unlock_irqrestore(&gpu->perf_lock, flags);
 628
 629         return ret;
 630 }
 631
 632 /*
 633  * Cmdstream submission/retirement:
 634  */
 635
 636 static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring,
 637                 struct msm_gem_submit *submit)
 638 {
 639         int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT;
 640         volatile struct msm_gpu_submit_stats *stats;
 641         u64 elapsed, clock = 0;
 642         unsigned long flags;
 643
 644         stats = &ring->memptrs->stats[index];
 645         /* Convert 19.2Mhz alwayson ticks to nanoseconds for elapsed time */
 646         elapsed = (stats->alwayson_end - stats->alwayson_start) * 10000;
 647         do_div(elapsed, 192);
 648
 649         /* Calculate the clock frequency from the number of CP cycles */
 650         if (elapsed) {
 651                 clock = (stats->cpcycles_end - stats->cpcycles_start) * 1000;
 652                 do_div(clock, elapsed);
 653         }
 654
 655         trace_msm_gpu_submit_retired(submit, elapsed, clock,
 656                 stats->alwayson_start, stats->alwayson_end);
 657
 658         msm_submit_retire(submit);
 659
 660         pm_runtime_mark_last_busy(&gpu->pdev->dev);
 661         pm_runtime_put_autosuspend(&gpu->pdev->dev);
 662
 663         spin_lock_irqsave(&ring->submit_lock, flags);
 664         list_del(&submit->node);
 665         spin_unlock_irqrestore(&ring->submit_lock, flags);
 666
 667         /* Update devfreq on transition from active->idle: */
 668         mutex_lock(&gpu->active_lock);
 669         gpu->active_submits--;
 670         WARN_ON(gpu->active_submits < 0);
 671         if (!gpu->active_submits)
 672                 msm_devfreq_idle(gpu);
 673         mutex_unlock(&gpu->active_lock);
 674
 675         msm_gem_submit_put(submit);
 676 }
 677
 678 static void retire_submits(struct msm_gpu *gpu)
 679 {
 680         int i;
 681
 682         /* Retire the commits starting with highest priority */
 683         for (i = 0; i < gpu->nr_rings; i++) {
 684                 struct msm_ringbuffer *ring = gpu->rb[i];
 685
 686                 while (true) {
 687                         struct msm_gem_submit *submit = NULL;
 688                         unsigned long flags;
 689
 690                         spin_lock_irqsave(&ring->submit_lock, flags);
 691                         submit = list_first_entry_or_null(&ring->submits,
 692                                         struct msm_gem_submit, node);
 693                         spin_unlock_irqrestore(&ring->submit_lock, flags);
 694
 695                         /*
 696                          * If no submit, we are done.  If submit->fence hasn't
 697                          * been signalled, then later submits are not signalled
 698                          * either, so we are also done.
 699                          */
 700                         if (submit && dma_fence_is_signaled(submit->hw_fence)) {
 701                                 retire_submit(gpu, ring, submit);
 702                         } else {
 703                                 break;
 704                         }
 705                 }
 706         }
 707 }
 708
 709 static void retire_worker(struct kthread_work *work)
 710 {
 711         struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work);
 712
 713         retire_submits(gpu);
 714 }
 715
 716 /* call from irq handler to schedule work to retire bo's */
 717 void msm_gpu_retire(struct msm_gpu *gpu)
 718 {
 719         int i;
 720
 721         for (i = 0; i < gpu->nr_rings; i++)
 722                 update_fences(gpu, gpu->rb[i], gpu->rb[i]->memptrs->fence);
 723
 724         kthread_queue_work(gpu->worker, &gpu->retire_work);
 725         update_sw_cntrs(gpu);
 726 }
 727
 728 /* add bo's to gpu's ring, and kick gpu: */
 729 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit)
 730 {
 731         struct drm_device *dev = gpu->dev;
 732         struct msm_drm_private *priv = dev->dev_private;
 733         struct msm_ringbuffer *ring = submit->ring;
 734         unsigned long flags;
 735
 736         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
 737
 738         pm_runtime_get_sync(&gpu->pdev->dev);
 739
 740         msm_gpu_hw_init(gpu);
 741
 742         submit->seqno = ++ring->seqno;
 743
 744         msm_rd_dump_submit(priv->rd, submit, NULL);
 745
 746         update_sw_cntrs(gpu);
 747
 748         /*
 749          * ring->submits holds a ref to the submit, to deal with the case
 750          * that a submit completes before msm_ioctl_gem_submit() returns.
 751          */
 752         msm_gem_submit_get(submit);
 753
 754         spin_lock_irqsave(&ring->submit_lock, flags);
 755         list_add_tail(&submit->node, &ring->submits);
 756         spin_unlock_irqrestore(&ring->submit_lock, flags);
 757
 758         /* Update devfreq on transition from idle->active: */
 759         mutex_lock(&gpu->active_lock);
 760         if (!gpu->active_submits)
 761                 msm_devfreq_active(gpu);
 762         gpu->active_submits++;
 763         mutex_unlock(&gpu->active_lock);
 764
 765         gpu->funcs->submit(gpu, submit);
 766         priv->lastctx = submit->queue->ctx;
 767
 768         hangcheck_timer_reset(gpu);
 769 }
 770
 771 /*
 772  * Init/Cleanup:
 773  */
 774
 775 static irqreturn_t irq_handler(int irq, void *data)
 776 {
 777         struct msm_gpu *gpu = data;
 778         return gpu->funcs->irq(gpu);
 779 }
 780
 781 static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu)
 782 {
 783         int ret = devm_clk_bulk_get_all(&pdev->dev, &gpu->grp_clks);
 784
 785         if (ret < 1) {
 786                 gpu->nr_clocks = 0;
 787                 return ret;
 788         }
 789
 790         gpu->nr_clocks = ret;
 791
 792         gpu->core_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
 793                 gpu->nr_clocks, "core");
 794
 795         gpu->rbbmtimer_clk = msm_clk_bulk_get_clock(gpu->grp_clks,
 796                 gpu->nr_clocks, "rbbmtimer");
 797
 798         return 0;
 799 }
 800
 801 /* Return a new address space for a msm_drm_private instance */
 802 struct msm_gem_address_space *
 803 msm_gpu_create_private_address_space(struct msm_gpu *gpu, struct task_struct *task)
 804 {
 805         struct msm_gem_address_space *aspace = NULL;
 806         if (!gpu)
 807                 return NULL;
 808
 809         /*
 810          * If the target doesn't support private address spaces then return
 811          * the global one
 812          */
 813         if (gpu->funcs->create_private_address_space) {
 814                 aspace = gpu->funcs->create_private_address_space(gpu);
 815                 if (!IS_ERR(aspace))
 816                         aspace->pid = get_pid(task_pid(task));
 817         }
 818
 819         if (IS_ERR_OR_NULL(aspace))
 820                 aspace = msm_gem_address_space_get(gpu->aspace);
 821
 822         return aspace;
 823 }
 824
 825 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev,
 826                 struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs,
 827                 const char *name, struct msm_gpu_config *config)
 828 {
 829         int i, ret, nr_rings = config->nr_rings;
 830         void *memptrs;
 831         uint64_t memptrs_iova;
 832
 833         if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs)))
 834                 gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs);
 835
 836         gpu->dev = drm;
 837         gpu->funcs = funcs;
 838         gpu->name = name;
 839
 840         gpu->worker = kthread_create_worker(0, "%s-worker", gpu->name);
 841         if (IS_ERR(gpu->worker)) {
 842                 ret = PTR_ERR(gpu->worker);
 843                 gpu->worker = NULL;
 844                 goto fail;
 845         }
 846
 847         sched_set_fifo_low(gpu->worker->task);
 848
 849         INIT_LIST_HEAD(&gpu->active_list);
 850         mutex_init(&gpu->active_lock);
 851         kthread_init_work(&gpu->retire_work, retire_worker);
 852         kthread_init_work(&gpu->recover_work, recover_worker);
 853         kthread_init_work(&gpu->fault_work, fault_worker);
 854
 855         timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0);
 856
 857         spin_lock_init(&gpu->perf_lock);
 858
 859
 860         /* Map registers: */
 861         gpu->mmio = msm_ioremap(pdev, config->ioname, name);
 862         if (IS_ERR(gpu->mmio)) {
 863                 ret = PTR_ERR(gpu->mmio);
 864                 goto fail;
 865         }
 866
 867         /* Get Interrupt: */
 868         gpu->irq = platform_get_irq(pdev, 0);
 869         if (gpu->irq < 0) {
 870                 ret = gpu->irq;
 871                 DRM_DEV_ERROR(drm->dev, "failed to get irq: %d\n", ret);
 872                 goto fail;
 873         }
 874
 875         ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler,
 876                         IRQF_TRIGGER_HIGH, gpu->name, gpu);
 877         if (ret) {
 878                 DRM_DEV_ERROR(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret);
 879                 goto fail;
 880         }
 881
 882         ret = get_clocks(pdev, gpu);
 883         if (ret)
 884                 goto fail;
 885
 886         gpu->ebi1_clk = msm_clk_get(pdev, "bus");
 887         DBG("ebi1_clk: %p", gpu->ebi1_clk);
 888         if (IS_ERR(gpu->ebi1_clk))
 889                 gpu->ebi1_clk = NULL;
 890
 891         /* Acquire regulators: */
 892         gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd");
 893         DBG("gpu_reg: %p", gpu->gpu_reg);
 894         if (IS_ERR(gpu->gpu_reg))
 895                 gpu->gpu_reg = NULL;
 896
 897         gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx");
 898         DBG("gpu_cx: %p", gpu->gpu_cx);
 899         if (IS_ERR(gpu->gpu_cx))
 900                 gpu->gpu_cx = NULL;
 901
 902         gpu->pdev = pdev;
 903         platform_set_drvdata(pdev, &gpu->adreno_smmu);
 904
 905         msm_devfreq_init(gpu);
 906
 907
 908         gpu->aspace = gpu->funcs->create_address_space(gpu, pdev);
 909
 910         if (gpu->aspace == NULL)
 911                 DRM_DEV_INFO(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name);
 912         else if (IS_ERR(gpu->aspace)) {
 913                 ret = PTR_ERR(gpu->aspace);
 914                 goto fail;
 915         }
 916
 917         memptrs = msm_gem_kernel_new(drm,
 918                 sizeof(struct msm_rbmemptrs) * nr_rings,
 919                 check_apriv(gpu, MSM_BO_UNCACHED), gpu->aspace, &gpu->memptrs_bo,
 920                 &memptrs_iova);
 921
 922         if (IS_ERR(memptrs)) {
 923                 ret = PTR_ERR(memptrs);
 924                 DRM_DEV_ERROR(drm->dev, "could not allocate memptrs: %d\n", ret);
 925                 goto fail;
 926         }
 927
 928         msm_gem_object_set_name(gpu->memptrs_bo, "memptrs");
 929
 930         if (nr_rings > ARRAY_SIZE(gpu->rb)) {
 931                 DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n",
 932                         ARRAY_SIZE(gpu->rb));
 933                 nr_rings = ARRAY_SIZE(gpu->rb);
 934         }
 935
 936         /* Create ringbuffer(s): */
 937         for (i = 0; i < nr_rings; i++) {
 938                 gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova);
 939
 940                 if (IS_ERR(gpu->rb[i])) {
 941                         ret = PTR_ERR(gpu->rb[i]);
 942                         DRM_DEV_ERROR(drm->dev,
 943                                 "could not create ringbuffer %d: %d\n", i, ret);
 944                         goto fail;
 945                 }
 946
 947                 memptrs += sizeof(struct msm_rbmemptrs);
 948                 memptrs_iova += sizeof(struct msm_rbmemptrs);
 949         }
 950
 951         gpu->nr_rings = nr_rings;
 952
 953         return 0;
 954
 955 fail:
 956         for (i = 0; i < ARRAY_SIZE(gpu->rb); i++)  {
 957                 msm_ringbuffer_destroy(gpu->rb[i]);
 958                 gpu->rb[i] = NULL;
 959         }
 960
 961         msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
 962
 963         platform_set_drvdata(pdev, NULL);
 964         return ret;
 965 }
 966
 967 void msm_gpu_cleanup(struct msm_gpu *gpu)
 968 {
 969         int i;
 970
 971         DBG("%s", gpu->name);
 972
 973         WARN_ON(!list_empty(&gpu->active_list));
 974
 975         for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) {
 976                 msm_ringbuffer_destroy(gpu->rb[i]);
 977                 gpu->rb[i] = NULL;
 978         }
 979
 980         msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace);
 981
 982         if (!IS_ERR_OR_NULL(gpu->aspace)) {
 983                 gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu);
 984                 msm_gem_address_space_put(gpu->aspace);
 985         }
 986
 987         if (gpu->worker) {
 988                 kthread_destroy_worker(gpu->worker);
 989         }
 990
 991         msm_devfreq_cleanup(gpu);
 992 }