1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2013 Red Hat 4 * Author: Rob Clark <robdclark@gmail.com> 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->lock)); 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 (fence_after(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; state->bos && 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(&gpu->lock); 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 submit->queue->faults++; 374 submit->aspace->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 } else { 393 /* 394 * We couldn't attribute this fault to any particular context, 395 * so increment the global fault count instead. 396 */ 397 gpu->global_faults++; 398 } 399 400 /* Record the crash state */ 401 pm_runtime_get_sync(&gpu->pdev->dev); 402 msm_gpu_crashstate_capture(gpu, submit, comm, cmd); 403 pm_runtime_put_sync(&gpu->pdev->dev); 404 405 kfree(cmd); 406 kfree(comm); 407 408 /* 409 * Update all the rings with the latest and greatest fence.. this 410 * needs to happen after msm_rd_dump_submit() to ensure that the 411 * bo's referenced by the offending submit are still around. 412 */ 413 for (i = 0; i < gpu->nr_rings; i++) { 414 struct msm_ringbuffer *ring = gpu->rb[i]; 415 416 uint32_t fence = ring->memptrs->fence; 417 418 /* 419 * For the current (faulting?) ring/submit advance the fence by 420 * one more to clear the faulting submit 421 */ 422 if (ring == cur_ring) 423 fence++; 424 425 update_fences(gpu, ring, fence); 426 } 427 428 if (msm_gpu_active(gpu)) { 429 /* retire completed submits, plus the one that hung: */ 430 retire_submits(gpu); 431 432 pm_runtime_get_sync(&gpu->pdev->dev); 433 gpu->funcs->recover(gpu); 434 pm_runtime_put_sync(&gpu->pdev->dev); 435 436 /* 437 * Replay all remaining submits starting with highest priority 438 * ring 439 */ 440 for (i = 0; i < gpu->nr_rings; i++) { 441 struct msm_ringbuffer *ring = gpu->rb[i]; 442 unsigned long flags; 443 444 spin_lock_irqsave(&ring->submit_lock, flags); 445 list_for_each_entry(submit, &ring->submits, node) 446 gpu->funcs->submit(gpu, submit); 447 spin_unlock_irqrestore(&ring->submit_lock, flags); 448 } 449 } 450 451 mutex_unlock(&gpu->lock); 452 453 msm_gpu_retire(gpu); 454 } 455 456 static void fault_worker(struct kthread_work *work) 457 { 458 struct msm_gpu *gpu = container_of(work, struct msm_gpu, fault_work); 459 struct msm_gem_submit *submit; 460 struct msm_ringbuffer *cur_ring = gpu->funcs->active_ring(gpu); 461 char *comm = NULL, *cmd = NULL; 462 463 mutex_lock(&gpu->lock); 464 465 submit = find_submit(cur_ring, cur_ring->memptrs->fence + 1); 466 if (submit && submit->fault_dumped) 467 goto resume_smmu; 468 469 if (submit) { 470 struct task_struct *task; 471 472 task = get_pid_task(submit->pid, PIDTYPE_PID); 473 if (task) { 474 comm = kstrdup(task->comm, GFP_KERNEL); 475 cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL); 476 put_task_struct(task); 477 } 478 479 /* 480 * When we get GPU iova faults, we can get 1000s of them, 481 * but we really only want to log the first one. 482 */ 483 submit->fault_dumped = true; 484 } 485 486 /* Record the crash state */ 487 pm_runtime_get_sync(&gpu->pdev->dev); 488 msm_gpu_crashstate_capture(gpu, submit, comm, cmd); 489 pm_runtime_put_sync(&gpu->pdev->dev); 490 491 kfree(cmd); 492 kfree(comm); 493 494 resume_smmu: 495 memset(&gpu->fault_info, 0, sizeof(gpu->fault_info)); 496 gpu->aspace->mmu->funcs->resume_translation(gpu->aspace->mmu); 497 498 mutex_unlock(&gpu->lock); 499 } 500 501 static void hangcheck_timer_reset(struct msm_gpu *gpu) 502 { 503 struct msm_drm_private *priv = gpu->dev->dev_private; 504 mod_timer(&gpu->hangcheck_timer, 505 round_jiffies_up(jiffies + msecs_to_jiffies(priv->hangcheck_period))); 506 } 507 508 static void hangcheck_handler(struct timer_list *t) 509 { 510 struct msm_gpu *gpu = from_timer(gpu, t, hangcheck_timer); 511 struct drm_device *dev = gpu->dev; 512 struct msm_ringbuffer *ring = gpu->funcs->active_ring(gpu); 513 uint32_t fence = ring->memptrs->fence; 514 515 if (fence != ring->hangcheck_fence) { 516 /* some progress has been made.. ya! */ 517 ring->hangcheck_fence = fence; 518 } else if (fence_before(fence, ring->seqno)) { 519 /* no progress and not done.. hung! */ 520 ring->hangcheck_fence = fence; 521 DRM_DEV_ERROR(dev->dev, "%s: hangcheck detected gpu lockup rb %d!\n", 522 gpu->name, ring->id); 523 DRM_DEV_ERROR(dev->dev, "%s: completed fence: %u\n", 524 gpu->name, fence); 525 DRM_DEV_ERROR(dev->dev, "%s: submitted fence: %u\n", 526 gpu->name, ring->seqno); 527 528 kthread_queue_work(gpu->worker, &gpu->recover_work); 529 } 530 531 /* if still more pending work, reset the hangcheck timer: */ 532 if (fence_after(ring->seqno, ring->hangcheck_fence)) 533 hangcheck_timer_reset(gpu); 534 535 /* workaround for missing irq: */ 536 msm_gpu_retire(gpu); 537 } 538 539 /* 540 * Performance Counters: 541 */ 542 543 /* called under perf_lock */ 544 static int update_hw_cntrs(struct msm_gpu *gpu, uint32_t ncntrs, uint32_t *cntrs) 545 { 546 uint32_t current_cntrs[ARRAY_SIZE(gpu->last_cntrs)]; 547 int i, n = min(ncntrs, gpu->num_perfcntrs); 548 549 /* read current values: */ 550 for (i = 0; i < gpu->num_perfcntrs; i++) 551 current_cntrs[i] = gpu_read(gpu, gpu->perfcntrs[i].sample_reg); 552 553 /* update cntrs: */ 554 for (i = 0; i < n; i++) 555 cntrs[i] = current_cntrs[i] - gpu->last_cntrs[i]; 556 557 /* save current values: */ 558 for (i = 0; i < gpu->num_perfcntrs; i++) 559 gpu->last_cntrs[i] = current_cntrs[i]; 560 561 return n; 562 } 563 564 static void update_sw_cntrs(struct msm_gpu *gpu) 565 { 566 ktime_t time; 567 uint32_t elapsed; 568 unsigned long flags; 569 570 spin_lock_irqsave(&gpu->perf_lock, flags); 571 if (!gpu->perfcntr_active) 572 goto out; 573 574 time = ktime_get(); 575 elapsed = ktime_to_us(ktime_sub(time, gpu->last_sample.time)); 576 577 gpu->totaltime += elapsed; 578 if (gpu->last_sample.active) 579 gpu->activetime += elapsed; 580 581 gpu->last_sample.active = msm_gpu_active(gpu); 582 gpu->last_sample.time = time; 583 584 out: 585 spin_unlock_irqrestore(&gpu->perf_lock, flags); 586 } 587 588 void msm_gpu_perfcntr_start(struct msm_gpu *gpu) 589 { 590 unsigned long flags; 591 592 pm_runtime_get_sync(&gpu->pdev->dev); 593 594 spin_lock_irqsave(&gpu->perf_lock, flags); 595 /* we could dynamically enable/disable perfcntr registers too.. */ 596 gpu->last_sample.active = msm_gpu_active(gpu); 597 gpu->last_sample.time = ktime_get(); 598 gpu->activetime = gpu->totaltime = 0; 599 gpu->perfcntr_active = true; 600 update_hw_cntrs(gpu, 0, NULL); 601 spin_unlock_irqrestore(&gpu->perf_lock, flags); 602 } 603 604 void msm_gpu_perfcntr_stop(struct msm_gpu *gpu) 605 { 606 gpu->perfcntr_active = false; 607 pm_runtime_put_sync(&gpu->pdev->dev); 608 } 609 610 /* returns -errno or # of cntrs sampled */ 611 int msm_gpu_perfcntr_sample(struct msm_gpu *gpu, uint32_t *activetime, 612 uint32_t *totaltime, uint32_t ncntrs, uint32_t *cntrs) 613 { 614 unsigned long flags; 615 int ret; 616 617 spin_lock_irqsave(&gpu->perf_lock, flags); 618 619 if (!gpu->perfcntr_active) { 620 ret = -EINVAL; 621 goto out; 622 } 623 624 *activetime = gpu->activetime; 625 *totaltime = gpu->totaltime; 626 627 gpu->activetime = gpu->totaltime = 0; 628 629 ret = update_hw_cntrs(gpu, ncntrs, cntrs); 630 631 out: 632 spin_unlock_irqrestore(&gpu->perf_lock, flags); 633 634 return ret; 635 } 636 637 /* 638 * Cmdstream submission/retirement: 639 */ 640 641 static void retire_submit(struct msm_gpu *gpu, struct msm_ringbuffer *ring, 642 struct msm_gem_submit *submit) 643 { 644 int index = submit->seqno % MSM_GPU_SUBMIT_STATS_COUNT; 645 volatile struct msm_gpu_submit_stats *stats; 646 u64 elapsed, clock = 0; 647 unsigned long flags; 648 649 stats = &ring->memptrs->stats[index]; 650 /* Convert 19.2Mhz alwayson ticks to nanoseconds for elapsed time */ 651 elapsed = (stats->alwayson_end - stats->alwayson_start) * 10000; 652 do_div(elapsed, 192); 653 654 /* Calculate the clock frequency from the number of CP cycles */ 655 if (elapsed) { 656 clock = (stats->cpcycles_end - stats->cpcycles_start) * 1000; 657 do_div(clock, elapsed); 658 } 659 660 trace_msm_gpu_submit_retired(submit, elapsed, clock, 661 stats->alwayson_start, stats->alwayson_end); 662 663 msm_submit_retire(submit); 664 665 pm_runtime_mark_last_busy(&gpu->pdev->dev); 666 pm_runtime_put_autosuspend(&gpu->pdev->dev); 667 668 spin_lock_irqsave(&ring->submit_lock, flags); 669 list_del(&submit->node); 670 spin_unlock_irqrestore(&ring->submit_lock, flags); 671 672 /* Update devfreq on transition from active->idle: */ 673 mutex_lock(&gpu->active_lock); 674 gpu->active_submits--; 675 WARN_ON(gpu->active_submits < 0); 676 if (!gpu->active_submits) 677 msm_devfreq_idle(gpu); 678 mutex_unlock(&gpu->active_lock); 679 680 msm_gem_submit_put(submit); 681 } 682 683 static void retire_submits(struct msm_gpu *gpu) 684 { 685 int i; 686 687 /* Retire the commits starting with highest priority */ 688 for (i = 0; i < gpu->nr_rings; i++) { 689 struct msm_ringbuffer *ring = gpu->rb[i]; 690 691 while (true) { 692 struct msm_gem_submit *submit = NULL; 693 unsigned long flags; 694 695 spin_lock_irqsave(&ring->submit_lock, flags); 696 submit = list_first_entry_or_null(&ring->submits, 697 struct msm_gem_submit, node); 698 spin_unlock_irqrestore(&ring->submit_lock, flags); 699 700 /* 701 * If no submit, we are done. If submit->fence hasn't 702 * been signalled, then later submits are not signalled 703 * either, so we are also done. 704 */ 705 if (submit && dma_fence_is_signaled(submit->hw_fence)) { 706 retire_submit(gpu, ring, submit); 707 } else { 708 break; 709 } 710 } 711 } 712 713 wake_up_all(&gpu->retire_event); 714 } 715 716 static void retire_worker(struct kthread_work *work) 717 { 718 struct msm_gpu *gpu = container_of(work, struct msm_gpu, retire_work); 719 720 retire_submits(gpu); 721 } 722 723 /* call from irq handler to schedule work to retire bo's */ 724 void msm_gpu_retire(struct msm_gpu *gpu) 725 { 726 int i; 727 728 for (i = 0; i < gpu->nr_rings; i++) 729 update_fences(gpu, gpu->rb[i], gpu->rb[i]->memptrs->fence); 730 731 kthread_queue_work(gpu->worker, &gpu->retire_work); 732 update_sw_cntrs(gpu); 733 } 734 735 /* add bo's to gpu's ring, and kick gpu: */ 736 void msm_gpu_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit) 737 { 738 struct drm_device *dev = gpu->dev; 739 struct msm_drm_private *priv = dev->dev_private; 740 struct msm_ringbuffer *ring = submit->ring; 741 unsigned long flags; 742 743 WARN_ON(!mutex_is_locked(&gpu->lock)); 744 745 pm_runtime_get_sync(&gpu->pdev->dev); 746 747 msm_gpu_hw_init(gpu); 748 749 submit->seqno = ++ring->seqno; 750 751 msm_rd_dump_submit(priv->rd, submit, NULL); 752 753 update_sw_cntrs(gpu); 754 755 /* 756 * ring->submits holds a ref to the submit, to deal with the case 757 * that a submit completes before msm_ioctl_gem_submit() returns. 758 */ 759 msm_gem_submit_get(submit); 760 761 spin_lock_irqsave(&ring->submit_lock, flags); 762 list_add_tail(&submit->node, &ring->submits); 763 spin_unlock_irqrestore(&ring->submit_lock, flags); 764 765 /* Update devfreq on transition from idle->active: */ 766 mutex_lock(&gpu->active_lock); 767 if (!gpu->active_submits) 768 msm_devfreq_active(gpu); 769 gpu->active_submits++; 770 mutex_unlock(&gpu->active_lock); 771 772 gpu->funcs->submit(gpu, submit); 773 gpu->cur_ctx_seqno = submit->queue->ctx->seqno; 774 775 hangcheck_timer_reset(gpu); 776 } 777 778 /* 779 * Init/Cleanup: 780 */ 781 782 static irqreturn_t irq_handler(int irq, void *data) 783 { 784 struct msm_gpu *gpu = data; 785 return gpu->funcs->irq(gpu); 786 } 787 788 static int get_clocks(struct platform_device *pdev, struct msm_gpu *gpu) 789 { 790 int ret = devm_clk_bulk_get_all(&pdev->dev, &gpu->grp_clks); 791 792 if (ret < 1) { 793 gpu->nr_clocks = 0; 794 return ret; 795 } 796 797 gpu->nr_clocks = ret; 798 799 gpu->core_clk = msm_clk_bulk_get_clock(gpu->grp_clks, 800 gpu->nr_clocks, "core"); 801 802 gpu->rbbmtimer_clk = msm_clk_bulk_get_clock(gpu->grp_clks, 803 gpu->nr_clocks, "rbbmtimer"); 804 805 return 0; 806 } 807 808 /* Return a new address space for a msm_drm_private instance */ 809 struct msm_gem_address_space * 810 msm_gpu_create_private_address_space(struct msm_gpu *gpu, struct task_struct *task) 811 { 812 struct msm_gem_address_space *aspace = NULL; 813 if (!gpu) 814 return NULL; 815 816 /* 817 * If the target doesn't support private address spaces then return 818 * the global one 819 */ 820 if (gpu->funcs->create_private_address_space) { 821 aspace = gpu->funcs->create_private_address_space(gpu); 822 if (!IS_ERR(aspace)) 823 aspace->pid = get_pid(task_pid(task)); 824 } 825 826 if (IS_ERR_OR_NULL(aspace)) 827 aspace = msm_gem_address_space_get(gpu->aspace); 828 829 return aspace; 830 } 831 832 int msm_gpu_init(struct drm_device *drm, struct platform_device *pdev, 833 struct msm_gpu *gpu, const struct msm_gpu_funcs *funcs, 834 const char *name, struct msm_gpu_config *config) 835 { 836 int i, ret, nr_rings = config->nr_rings; 837 void *memptrs; 838 uint64_t memptrs_iova; 839 840 if (WARN_ON(gpu->num_perfcntrs > ARRAY_SIZE(gpu->last_cntrs))) 841 gpu->num_perfcntrs = ARRAY_SIZE(gpu->last_cntrs); 842 843 gpu->dev = drm; 844 gpu->funcs = funcs; 845 gpu->name = name; 846 847 gpu->worker = kthread_create_worker(0, "gpu-worker"); 848 if (IS_ERR(gpu->worker)) { 849 ret = PTR_ERR(gpu->worker); 850 gpu->worker = NULL; 851 goto fail; 852 } 853 854 sched_set_fifo_low(gpu->worker->task); 855 856 INIT_LIST_HEAD(&gpu->active_list); 857 mutex_init(&gpu->active_lock); 858 mutex_init(&gpu->lock); 859 init_waitqueue_head(&gpu->retire_event); 860 kthread_init_work(&gpu->retire_work, retire_worker); 861 kthread_init_work(&gpu->recover_work, recover_worker); 862 kthread_init_work(&gpu->fault_work, fault_worker); 863 864 timer_setup(&gpu->hangcheck_timer, hangcheck_handler, 0); 865 866 spin_lock_init(&gpu->perf_lock); 867 868 869 /* Map registers: */ 870 gpu->mmio = msm_ioremap(pdev, config->ioname); 871 if (IS_ERR(gpu->mmio)) { 872 ret = PTR_ERR(gpu->mmio); 873 goto fail; 874 } 875 876 /* Get Interrupt: */ 877 gpu->irq = platform_get_irq(pdev, 0); 878 if (gpu->irq < 0) { 879 ret = gpu->irq; 880 DRM_DEV_ERROR(drm->dev, "failed to get irq: %d\n", ret); 881 goto fail; 882 } 883 884 ret = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 885 IRQF_TRIGGER_HIGH, "gpu-irq", gpu); 886 if (ret) { 887 DRM_DEV_ERROR(drm->dev, "failed to request IRQ%u: %d\n", gpu->irq, ret); 888 goto fail; 889 } 890 891 ret = get_clocks(pdev, gpu); 892 if (ret) 893 goto fail; 894 895 gpu->ebi1_clk = msm_clk_get(pdev, "bus"); 896 DBG("ebi1_clk: %p", gpu->ebi1_clk); 897 if (IS_ERR(gpu->ebi1_clk)) 898 gpu->ebi1_clk = NULL; 899 900 /* Acquire regulators: */ 901 gpu->gpu_reg = devm_regulator_get(&pdev->dev, "vdd"); 902 DBG("gpu_reg: %p", gpu->gpu_reg); 903 if (IS_ERR(gpu->gpu_reg)) 904 gpu->gpu_reg = NULL; 905 906 gpu->gpu_cx = devm_regulator_get(&pdev->dev, "vddcx"); 907 DBG("gpu_cx: %p", gpu->gpu_cx); 908 if (IS_ERR(gpu->gpu_cx)) 909 gpu->gpu_cx = NULL; 910 911 gpu->pdev = pdev; 912 platform_set_drvdata(pdev, &gpu->adreno_smmu); 913 914 msm_devfreq_init(gpu); 915 916 917 gpu->aspace = gpu->funcs->create_address_space(gpu, pdev); 918 919 if (gpu->aspace == NULL) 920 DRM_DEV_INFO(drm->dev, "%s: no IOMMU, fallback to VRAM carveout!\n", name); 921 else if (IS_ERR(gpu->aspace)) { 922 ret = PTR_ERR(gpu->aspace); 923 goto fail; 924 } 925 926 memptrs = msm_gem_kernel_new(drm, 927 sizeof(struct msm_rbmemptrs) * nr_rings, 928 check_apriv(gpu, MSM_BO_UNCACHED), gpu->aspace, &gpu->memptrs_bo, 929 &memptrs_iova); 930 931 if (IS_ERR(memptrs)) { 932 ret = PTR_ERR(memptrs); 933 DRM_DEV_ERROR(drm->dev, "could not allocate memptrs: %d\n", ret); 934 goto fail; 935 } 936 937 msm_gem_object_set_name(gpu->memptrs_bo, "memptrs"); 938 939 if (nr_rings > ARRAY_SIZE(gpu->rb)) { 940 DRM_DEV_INFO_ONCE(drm->dev, "Only creating %zu ringbuffers\n", 941 ARRAY_SIZE(gpu->rb)); 942 nr_rings = ARRAY_SIZE(gpu->rb); 943 } 944 945 /* Create ringbuffer(s): */ 946 for (i = 0; i < nr_rings; i++) { 947 gpu->rb[i] = msm_ringbuffer_new(gpu, i, memptrs, memptrs_iova); 948 949 if (IS_ERR(gpu->rb[i])) { 950 ret = PTR_ERR(gpu->rb[i]); 951 DRM_DEV_ERROR(drm->dev, 952 "could not create ringbuffer %d: %d\n", i, ret); 953 goto fail; 954 } 955 956 memptrs += sizeof(struct msm_rbmemptrs); 957 memptrs_iova += sizeof(struct msm_rbmemptrs); 958 } 959 960 gpu->nr_rings = nr_rings; 961 962 refcount_set(&gpu->sysprof_active, 1); 963 964 return 0; 965 966 fail: 967 for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) { 968 msm_ringbuffer_destroy(gpu->rb[i]); 969 gpu->rb[i] = NULL; 970 } 971 972 msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace); 973 974 platform_set_drvdata(pdev, NULL); 975 return ret; 976 } 977 978 void msm_gpu_cleanup(struct msm_gpu *gpu) 979 { 980 int i; 981 982 DBG("%s", gpu->name); 983 984 WARN_ON(!list_empty(&gpu->active_list)); 985 986 for (i = 0; i < ARRAY_SIZE(gpu->rb); i++) { 987 msm_ringbuffer_destroy(gpu->rb[i]); 988 gpu->rb[i] = NULL; 989 } 990 991 msm_gem_kernel_put(gpu->memptrs_bo, gpu->aspace); 992 993 if (!IS_ERR_OR_NULL(gpu->aspace)) { 994 gpu->aspace->mmu->funcs->detach(gpu->aspace->mmu); 995 msm_gem_address_space_put(gpu->aspace); 996 } 997 998 if (gpu->worker) { 999 kthread_destroy_worker(gpu->worker); 1000 } 1001 1002 msm_devfreq_cleanup(gpu); 1003 } 1004