1 // SPDX-License-Identifier: GPL-2.0 2 /* Copyright 2019 Linaro, Ltd, Rob Herring <robh@kernel.org> */ 3 /* Copyright 2019 Collabora ltd. */ 4 #include <linux/delay.h> 5 #include <linux/interrupt.h> 6 #include <linux/io.h> 7 #include <linux/iopoll.h> 8 #include <linux/platform_device.h> 9 #include <linux/pm_runtime.h> 10 #include <linux/dma-resv.h> 11 #include <drm/gpu_scheduler.h> 12 #include <drm/panfrost_drm.h> 13 14 #include "panfrost_device.h" 15 #include "panfrost_devfreq.h" 16 #include "panfrost_job.h" 17 #include "panfrost_features.h" 18 #include "panfrost_issues.h" 19 #include "panfrost_gem.h" 20 #include "panfrost_regs.h" 21 #include "panfrost_gpu.h" 22 #include "panfrost_mmu.h" 23 24 #define JOB_TIMEOUT_MS 500 25 26 #define job_write(dev, reg, data) writel(data, dev->iomem + (reg)) 27 #define job_read(dev, reg) readl(dev->iomem + (reg)) 28 29 struct panfrost_queue_state { 30 struct drm_gpu_scheduler sched; 31 u64 fence_context; 32 u64 emit_seqno; 33 }; 34 35 struct panfrost_job_slot { 36 struct panfrost_queue_state queue[NUM_JOB_SLOTS]; 37 spinlock_t job_lock; 38 int irq; 39 }; 40 41 static struct panfrost_job * 42 to_panfrost_job(struct drm_sched_job *sched_job) 43 { 44 return container_of(sched_job, struct panfrost_job, base); 45 } 46 47 struct panfrost_fence { 48 struct dma_fence base; 49 struct drm_device *dev; 50 /* panfrost seqno for signaled() test */ 51 u64 seqno; 52 int queue; 53 }; 54 55 static inline struct panfrost_fence * 56 to_panfrost_fence(struct dma_fence *fence) 57 { 58 return (struct panfrost_fence *)fence; 59 } 60 61 static const char *panfrost_fence_get_driver_name(struct dma_fence *fence) 62 { 63 return "panfrost"; 64 } 65 66 static const char *panfrost_fence_get_timeline_name(struct dma_fence *fence) 67 { 68 struct panfrost_fence *f = to_panfrost_fence(fence); 69 70 switch (f->queue) { 71 case 0: 72 return "panfrost-js-0"; 73 case 1: 74 return "panfrost-js-1"; 75 case 2: 76 return "panfrost-js-2"; 77 default: 78 return NULL; 79 } 80 } 81 82 static const struct dma_fence_ops panfrost_fence_ops = { 83 .get_driver_name = panfrost_fence_get_driver_name, 84 .get_timeline_name = panfrost_fence_get_timeline_name, 85 }; 86 87 static struct dma_fence *panfrost_fence_create(struct panfrost_device *pfdev, int js_num) 88 { 89 struct panfrost_fence *fence; 90 struct panfrost_job_slot *js = pfdev->js; 91 92 fence = kzalloc(sizeof(*fence), GFP_KERNEL); 93 if (!fence) 94 return ERR_PTR(-ENOMEM); 95 96 fence->dev = pfdev->ddev; 97 fence->queue = js_num; 98 fence->seqno = ++js->queue[js_num].emit_seqno; 99 dma_fence_init(&fence->base, &panfrost_fence_ops, &js->job_lock, 100 js->queue[js_num].fence_context, fence->seqno); 101 102 return &fence->base; 103 } 104 105 int panfrost_job_get_slot(struct panfrost_job *job) 106 { 107 /* JS0: fragment jobs. 108 * JS1: vertex/tiler jobs 109 * JS2: compute jobs 110 */ 111 if (job->requirements & PANFROST_JD_REQ_FS) 112 return 0; 113 114 /* Not exposed to userspace yet */ 115 #if 0 116 if (job->requirements & PANFROST_JD_REQ_ONLY_COMPUTE) { 117 if ((job->requirements & PANFROST_JD_REQ_CORE_GRP_MASK) && 118 (job->pfdev->features.nr_core_groups == 2)) 119 return 2; 120 if (panfrost_has_hw_issue(job->pfdev, HW_ISSUE_8987)) 121 return 2; 122 } 123 #endif 124 return 1; 125 } 126 127 static void panfrost_job_write_affinity(struct panfrost_device *pfdev, 128 u32 requirements, 129 int js) 130 { 131 u64 affinity; 132 133 /* 134 * Use all cores for now. 135 * Eventually we may need to support tiler only jobs and h/w with 136 * multiple (2) coherent core groups 137 */ 138 affinity = pfdev->features.shader_present; 139 140 job_write(pfdev, JS_AFFINITY_NEXT_LO(js), lower_32_bits(affinity)); 141 job_write(pfdev, JS_AFFINITY_NEXT_HI(js), upper_32_bits(affinity)); 142 } 143 144 static u32 145 panfrost_get_job_chain_flag(const struct panfrost_job *job) 146 { 147 struct panfrost_fence *f = to_panfrost_fence(job->done_fence); 148 149 if (!panfrost_has_hw_feature(job->pfdev, HW_FEATURE_JOBCHAIN_DISAMBIGUATION)) 150 return 0; 151 152 return (f->seqno & 1) ? JS_CONFIG_JOB_CHAIN_FLAG : 0; 153 } 154 155 static struct panfrost_job * 156 panfrost_dequeue_job(struct panfrost_device *pfdev, int slot) 157 { 158 struct panfrost_job *job = pfdev->jobs[slot][0]; 159 160 WARN_ON(!job); 161 pfdev->jobs[slot][0] = pfdev->jobs[slot][1]; 162 pfdev->jobs[slot][1] = NULL; 163 164 return job; 165 } 166 167 static unsigned int 168 panfrost_enqueue_job(struct panfrost_device *pfdev, int slot, 169 struct panfrost_job *job) 170 { 171 if (WARN_ON(!job)) 172 return 0; 173 174 if (!pfdev->jobs[slot][0]) { 175 pfdev->jobs[slot][0] = job; 176 return 0; 177 } 178 179 WARN_ON(pfdev->jobs[slot][1]); 180 pfdev->jobs[slot][1] = job; 181 WARN_ON(panfrost_get_job_chain_flag(job) == 182 panfrost_get_job_chain_flag(pfdev->jobs[slot][0])); 183 return 1; 184 } 185 186 static void panfrost_job_hw_submit(struct panfrost_job *job, int js) 187 { 188 struct panfrost_device *pfdev = job->pfdev; 189 unsigned int subslot; 190 u32 cfg; 191 u64 jc_head = job->jc; 192 int ret; 193 194 panfrost_devfreq_record_busy(&pfdev->pfdevfreq); 195 196 ret = pm_runtime_get_sync(pfdev->dev); 197 if (ret < 0) 198 return; 199 200 if (WARN_ON(job_read(pfdev, JS_COMMAND_NEXT(js)))) { 201 return; 202 } 203 204 cfg = panfrost_mmu_as_get(pfdev, job->file_priv->mmu); 205 206 job_write(pfdev, JS_HEAD_NEXT_LO(js), lower_32_bits(jc_head)); 207 job_write(pfdev, JS_HEAD_NEXT_HI(js), upper_32_bits(jc_head)); 208 209 panfrost_job_write_affinity(pfdev, job->requirements, js); 210 211 /* start MMU, medium priority, cache clean/flush on end, clean/flush on 212 * start */ 213 cfg |= JS_CONFIG_THREAD_PRI(8) | 214 JS_CONFIG_START_FLUSH_CLEAN_INVALIDATE | 215 JS_CONFIG_END_FLUSH_CLEAN_INVALIDATE | 216 panfrost_get_job_chain_flag(job); 217 218 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_FLUSH_REDUCTION)) 219 cfg |= JS_CONFIG_ENABLE_FLUSH_REDUCTION; 220 221 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_10649)) 222 cfg |= JS_CONFIG_START_MMU; 223 224 job_write(pfdev, JS_CONFIG_NEXT(js), cfg); 225 226 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_FLUSH_REDUCTION)) 227 job_write(pfdev, JS_FLUSH_ID_NEXT(js), job->flush_id); 228 229 /* GO ! */ 230 231 spin_lock(&pfdev->js->job_lock); 232 subslot = panfrost_enqueue_job(pfdev, js, job); 233 /* Don't queue the job if a reset is in progress */ 234 if (!atomic_read(&pfdev->reset.pending)) { 235 job_write(pfdev, JS_COMMAND_NEXT(js), JS_COMMAND_START); 236 dev_dbg(pfdev->dev, 237 "JS: Submitting atom %p to js[%d][%d] with head=0x%llx AS %d", 238 job, js, subslot, jc_head, cfg & 0xf); 239 } 240 spin_unlock(&pfdev->js->job_lock); 241 } 242 243 static int panfrost_acquire_object_fences(struct drm_gem_object **bos, 244 int bo_count, 245 struct drm_sched_job *job) 246 { 247 int i, ret; 248 249 for (i = 0; i < bo_count; i++) { 250 /* panfrost always uses write mode in its current uapi */ 251 ret = drm_sched_job_add_implicit_dependencies(job, bos[i], 252 true); 253 if (ret) 254 return ret; 255 } 256 257 return 0; 258 } 259 260 static void panfrost_attach_object_fences(struct drm_gem_object **bos, 261 int bo_count, 262 struct dma_fence *fence) 263 { 264 int i; 265 266 for (i = 0; i < bo_count; i++) 267 dma_resv_add_excl_fence(bos[i]->resv, fence); 268 } 269 270 int panfrost_job_push(struct panfrost_job *job) 271 { 272 struct panfrost_device *pfdev = job->pfdev; 273 struct ww_acquire_ctx acquire_ctx; 274 int ret = 0; 275 276 ret = drm_gem_lock_reservations(job->bos, job->bo_count, 277 &acquire_ctx); 278 if (ret) 279 return ret; 280 281 mutex_lock(&pfdev->sched_lock); 282 drm_sched_job_arm(&job->base); 283 284 job->render_done_fence = dma_fence_get(&job->base.s_fence->finished); 285 286 ret = panfrost_acquire_object_fences(job->bos, job->bo_count, 287 &job->base); 288 if (ret) { 289 mutex_unlock(&pfdev->sched_lock); 290 goto unlock; 291 } 292 293 kref_get(&job->refcount); /* put by scheduler job completion */ 294 295 drm_sched_entity_push_job(&job->base); 296 297 mutex_unlock(&pfdev->sched_lock); 298 299 panfrost_attach_object_fences(job->bos, job->bo_count, 300 job->render_done_fence); 301 302 unlock: 303 drm_gem_unlock_reservations(job->bos, job->bo_count, &acquire_ctx); 304 305 return ret; 306 } 307 308 static void panfrost_job_cleanup(struct kref *ref) 309 { 310 struct panfrost_job *job = container_of(ref, struct panfrost_job, 311 refcount); 312 unsigned int i; 313 314 dma_fence_put(job->done_fence); 315 dma_fence_put(job->render_done_fence); 316 317 if (job->mappings) { 318 for (i = 0; i < job->bo_count; i++) { 319 if (!job->mappings[i]) 320 break; 321 322 atomic_dec(&job->mappings[i]->obj->gpu_usecount); 323 panfrost_gem_mapping_put(job->mappings[i]); 324 } 325 kvfree(job->mappings); 326 } 327 328 if (job->bos) { 329 for (i = 0; i < job->bo_count; i++) 330 drm_gem_object_put(job->bos[i]); 331 332 kvfree(job->bos); 333 } 334 335 kfree(job); 336 } 337 338 void panfrost_job_put(struct panfrost_job *job) 339 { 340 kref_put(&job->refcount, panfrost_job_cleanup); 341 } 342 343 static void panfrost_job_free(struct drm_sched_job *sched_job) 344 { 345 struct panfrost_job *job = to_panfrost_job(sched_job); 346 347 drm_sched_job_cleanup(sched_job); 348 349 panfrost_job_put(job); 350 } 351 352 static struct dma_fence *panfrost_job_run(struct drm_sched_job *sched_job) 353 { 354 struct panfrost_job *job = to_panfrost_job(sched_job); 355 struct panfrost_device *pfdev = job->pfdev; 356 int slot = panfrost_job_get_slot(job); 357 struct dma_fence *fence = NULL; 358 359 if (unlikely(job->base.s_fence->finished.error)) 360 return NULL; 361 362 /* Nothing to execute: can happen if the job has finished while 363 * we were resetting the GPU. 364 */ 365 if (!job->jc) 366 return NULL; 367 368 fence = panfrost_fence_create(pfdev, slot); 369 if (IS_ERR(fence)) 370 return fence; 371 372 if (job->done_fence) 373 dma_fence_put(job->done_fence); 374 job->done_fence = dma_fence_get(fence); 375 376 panfrost_job_hw_submit(job, slot); 377 378 return fence; 379 } 380 381 void panfrost_job_enable_interrupts(struct panfrost_device *pfdev) 382 { 383 int j; 384 u32 irq_mask = 0; 385 386 for (j = 0; j < NUM_JOB_SLOTS; j++) { 387 irq_mask |= MK_JS_MASK(j); 388 } 389 390 job_write(pfdev, JOB_INT_CLEAR, irq_mask); 391 job_write(pfdev, JOB_INT_MASK, irq_mask); 392 } 393 394 static void panfrost_job_handle_err(struct panfrost_device *pfdev, 395 struct panfrost_job *job, 396 unsigned int js) 397 { 398 u32 js_status = job_read(pfdev, JS_STATUS(js)); 399 const char *exception_name = panfrost_exception_name(js_status); 400 bool signal_fence = true; 401 402 if (!panfrost_exception_is_fault(js_status)) { 403 dev_dbg(pfdev->dev, "js event, js=%d, status=%s, head=0x%x, tail=0x%x", 404 js, exception_name, 405 job_read(pfdev, JS_HEAD_LO(js)), 406 job_read(pfdev, JS_TAIL_LO(js))); 407 } else { 408 dev_err(pfdev->dev, "js fault, js=%d, status=%s, head=0x%x, tail=0x%x", 409 js, exception_name, 410 job_read(pfdev, JS_HEAD_LO(js)), 411 job_read(pfdev, JS_TAIL_LO(js))); 412 } 413 414 if (js_status == DRM_PANFROST_EXCEPTION_STOPPED) { 415 /* Update the job head so we can resume */ 416 job->jc = job_read(pfdev, JS_TAIL_LO(js)) | 417 ((u64)job_read(pfdev, JS_TAIL_HI(js)) << 32); 418 419 /* The job will be resumed, don't signal the fence */ 420 signal_fence = false; 421 } else if (js_status == DRM_PANFROST_EXCEPTION_TERMINATED) { 422 /* Job has been hard-stopped, flag it as canceled */ 423 dma_fence_set_error(job->done_fence, -ECANCELED); 424 job->jc = 0; 425 } else if (panfrost_exception_is_fault(js_status)) { 426 /* We might want to provide finer-grained error code based on 427 * the exception type, but unconditionally setting to EINVAL 428 * is good enough for now. 429 */ 430 dma_fence_set_error(job->done_fence, -EINVAL); 431 job->jc = 0; 432 } 433 434 panfrost_mmu_as_put(pfdev, job->file_priv->mmu); 435 panfrost_devfreq_record_idle(&pfdev->pfdevfreq); 436 437 if (signal_fence) 438 dma_fence_signal_locked(job->done_fence); 439 440 pm_runtime_put_autosuspend(pfdev->dev); 441 442 if (panfrost_exception_needs_reset(pfdev, js_status)) { 443 atomic_set(&pfdev->reset.pending, 1); 444 drm_sched_fault(&pfdev->js->queue[js].sched); 445 } 446 } 447 448 static void panfrost_job_handle_done(struct panfrost_device *pfdev, 449 struct panfrost_job *job) 450 { 451 /* Set ->jc to 0 to avoid re-submitting an already finished job (can 452 * happen when we receive the DONE interrupt while doing a GPU reset). 453 */ 454 job->jc = 0; 455 panfrost_mmu_as_put(pfdev, job->file_priv->mmu); 456 panfrost_devfreq_record_idle(&pfdev->pfdevfreq); 457 458 dma_fence_signal_locked(job->done_fence); 459 pm_runtime_put_autosuspend(pfdev->dev); 460 } 461 462 static void panfrost_job_handle_irq(struct panfrost_device *pfdev, u32 status) 463 { 464 struct panfrost_job *done[NUM_JOB_SLOTS][2] = {}; 465 struct panfrost_job *failed[NUM_JOB_SLOTS] = {}; 466 u32 js_state = 0, js_events = 0; 467 unsigned int i, j; 468 469 /* First we collect all failed/done jobs. */ 470 while (status) { 471 u32 js_state_mask = 0; 472 473 for (j = 0; j < NUM_JOB_SLOTS; j++) { 474 if (status & MK_JS_MASK(j)) 475 js_state_mask |= MK_JS_MASK(j); 476 477 if (status & JOB_INT_MASK_DONE(j)) { 478 if (done[j][0]) 479 done[j][1] = panfrost_dequeue_job(pfdev, j); 480 else 481 done[j][0] = panfrost_dequeue_job(pfdev, j); 482 } 483 484 if (status & JOB_INT_MASK_ERR(j)) { 485 /* Cancel the next submission. Will be submitted 486 * after we're done handling this failure if 487 * there's no reset pending. 488 */ 489 job_write(pfdev, JS_COMMAND_NEXT(j), JS_COMMAND_NOP); 490 failed[j] = panfrost_dequeue_job(pfdev, j); 491 } 492 } 493 494 /* JS_STATE is sampled when JOB_INT_CLEAR is written. 495 * For each BIT(slot) or BIT(slot + 16) bit written to 496 * JOB_INT_CLEAR, the corresponding bits in JS_STATE 497 * (BIT(slot) and BIT(slot + 16)) are updated, but this 498 * is racy. If we only have one job done at the time we 499 * read JOB_INT_RAWSTAT but the second job fails before we 500 * clear the status, we end up with a status containing 501 * only the DONE bit and consider both jobs as DONE since 502 * JS_STATE reports both NEXT and CURRENT as inactive. 503 * To prevent that, let's repeat this clear+read steps 504 * until status is 0. 505 */ 506 job_write(pfdev, JOB_INT_CLEAR, status); 507 js_state &= ~js_state_mask; 508 js_state |= job_read(pfdev, JOB_INT_JS_STATE) & js_state_mask; 509 js_events |= status; 510 status = job_read(pfdev, JOB_INT_RAWSTAT); 511 } 512 513 /* Then we handle the dequeued jobs. */ 514 for (j = 0; j < NUM_JOB_SLOTS; j++) { 515 if (!(js_events & MK_JS_MASK(j))) 516 continue; 517 518 if (failed[j]) { 519 panfrost_job_handle_err(pfdev, failed[j], j); 520 } else if (pfdev->jobs[j][0] && !(js_state & MK_JS_MASK(j))) { 521 /* When the current job doesn't fail, the JM dequeues 522 * the next job without waiting for an ACK, this means 523 * we can have 2 jobs dequeued and only catch the 524 * interrupt when the second one is done. If both slots 525 * are inactive, but one job remains in pfdev->jobs[j], 526 * consider it done. Of course that doesn't apply if a 527 * failure happened since we cancelled execution of the 528 * job in _NEXT (see above). 529 */ 530 if (WARN_ON(!done[j][0])) 531 done[j][0] = panfrost_dequeue_job(pfdev, j); 532 else 533 done[j][1] = panfrost_dequeue_job(pfdev, j); 534 } 535 536 for (i = 0; i < ARRAY_SIZE(done[0]) && done[j][i]; i++) 537 panfrost_job_handle_done(pfdev, done[j][i]); 538 } 539 540 /* And finally we requeue jobs that were waiting in the second slot 541 * and have been stopped if we detected a failure on the first slot. 542 */ 543 for (j = 0; j < NUM_JOB_SLOTS; j++) { 544 if (!(js_events & MK_JS_MASK(j))) 545 continue; 546 547 if (!failed[j] || !pfdev->jobs[j][0]) 548 continue; 549 550 if (pfdev->jobs[j][0]->jc == 0) { 551 /* The job was cancelled, signal the fence now */ 552 struct panfrost_job *canceled = panfrost_dequeue_job(pfdev, j); 553 554 dma_fence_set_error(canceled->done_fence, -ECANCELED); 555 panfrost_job_handle_done(pfdev, canceled); 556 } else if (!atomic_read(&pfdev->reset.pending)) { 557 /* Requeue the job we removed if no reset is pending */ 558 job_write(pfdev, JS_COMMAND_NEXT(j), JS_COMMAND_START); 559 } 560 } 561 } 562 563 static void panfrost_job_handle_irqs(struct panfrost_device *pfdev) 564 { 565 u32 status = job_read(pfdev, JOB_INT_RAWSTAT); 566 567 while (status) { 568 pm_runtime_mark_last_busy(pfdev->dev); 569 570 spin_lock(&pfdev->js->job_lock); 571 panfrost_job_handle_irq(pfdev, status); 572 spin_unlock(&pfdev->js->job_lock); 573 status = job_read(pfdev, JOB_INT_RAWSTAT); 574 } 575 } 576 577 static u32 panfrost_active_slots(struct panfrost_device *pfdev, 578 u32 *js_state_mask, u32 js_state) 579 { 580 u32 rawstat; 581 582 if (!(js_state & *js_state_mask)) 583 return 0; 584 585 rawstat = job_read(pfdev, JOB_INT_RAWSTAT); 586 if (rawstat) { 587 unsigned int i; 588 589 for (i = 0; i < NUM_JOB_SLOTS; i++) { 590 if (rawstat & MK_JS_MASK(i)) 591 *js_state_mask &= ~MK_JS_MASK(i); 592 } 593 } 594 595 return js_state & *js_state_mask; 596 } 597 598 static void 599 panfrost_reset(struct panfrost_device *pfdev, 600 struct drm_sched_job *bad) 601 { 602 u32 js_state, js_state_mask = 0xffffffff; 603 unsigned int i, j; 604 bool cookie; 605 int ret; 606 607 if (!atomic_read(&pfdev->reset.pending)) 608 return; 609 610 /* Stop the schedulers. 611 * 612 * FIXME: We temporarily get out of the dma_fence_signalling section 613 * because the cleanup path generate lockdep splats when taking locks 614 * to release job resources. We should rework the code to follow this 615 * pattern: 616 * 617 * try_lock 618 * if (locked) 619 * release 620 * else 621 * schedule_work_to_release_later 622 */ 623 for (i = 0; i < NUM_JOB_SLOTS; i++) 624 drm_sched_stop(&pfdev->js->queue[i].sched, bad); 625 626 cookie = dma_fence_begin_signalling(); 627 628 if (bad) 629 drm_sched_increase_karma(bad); 630 631 /* Mask job interrupts and synchronize to make sure we won't be 632 * interrupted during our reset. 633 */ 634 job_write(pfdev, JOB_INT_MASK, 0); 635 synchronize_irq(pfdev->js->irq); 636 637 for (i = 0; i < NUM_JOB_SLOTS; i++) { 638 /* Cancel the next job and soft-stop the running job. */ 639 job_write(pfdev, JS_COMMAND_NEXT(i), JS_COMMAND_NOP); 640 job_write(pfdev, JS_COMMAND(i), JS_COMMAND_SOFT_STOP); 641 } 642 643 /* Wait at most 10ms for soft-stops to complete */ 644 ret = readl_poll_timeout(pfdev->iomem + JOB_INT_JS_STATE, js_state, 645 !panfrost_active_slots(pfdev, &js_state_mask, js_state), 646 10, 10000); 647 648 if (ret) 649 dev_err(pfdev->dev, "Soft-stop failed\n"); 650 651 /* Handle the remaining interrupts before we reset. */ 652 panfrost_job_handle_irqs(pfdev); 653 654 /* Remaining interrupts have been handled, but we might still have 655 * stuck jobs. Let's make sure the PM counters stay balanced by 656 * manually calling pm_runtime_put_noidle() and 657 * panfrost_devfreq_record_idle() for each stuck job. 658 */ 659 spin_lock(&pfdev->js->job_lock); 660 for (i = 0; i < NUM_JOB_SLOTS; i++) { 661 for (j = 0; j < ARRAY_SIZE(pfdev->jobs[0]) && pfdev->jobs[i][j]; j++) { 662 pm_runtime_put_noidle(pfdev->dev); 663 panfrost_devfreq_record_idle(&pfdev->pfdevfreq); 664 } 665 } 666 memset(pfdev->jobs, 0, sizeof(pfdev->jobs)); 667 spin_unlock(&pfdev->js->job_lock); 668 669 /* Proceed with reset now. */ 670 panfrost_device_reset(pfdev); 671 672 /* panfrost_device_reset() unmasks job interrupts, but we want to 673 * keep them masked a bit longer. 674 */ 675 job_write(pfdev, JOB_INT_MASK, 0); 676 677 /* GPU has been reset, we can clear the reset pending bit. */ 678 atomic_set(&pfdev->reset.pending, 0); 679 680 /* Now resubmit jobs that were previously queued but didn't have a 681 * chance to finish. 682 * FIXME: We temporarily get out of the DMA fence signalling section 683 * while resubmitting jobs because the job submission logic will 684 * allocate memory with the GFP_KERNEL flag which can trigger memory 685 * reclaim and exposes a lock ordering issue. 686 */ 687 dma_fence_end_signalling(cookie); 688 for (i = 0; i < NUM_JOB_SLOTS; i++) 689 drm_sched_resubmit_jobs(&pfdev->js->queue[i].sched); 690 cookie = dma_fence_begin_signalling(); 691 692 /* Restart the schedulers */ 693 for (i = 0; i < NUM_JOB_SLOTS; i++) 694 drm_sched_start(&pfdev->js->queue[i].sched, true); 695 696 /* Re-enable job interrupts now that everything has been restarted. */ 697 job_write(pfdev, JOB_INT_MASK, 698 GENMASK(16 + NUM_JOB_SLOTS - 1, 16) | 699 GENMASK(NUM_JOB_SLOTS - 1, 0)); 700 701 dma_fence_end_signalling(cookie); 702 } 703 704 static enum drm_gpu_sched_stat panfrost_job_timedout(struct drm_sched_job 705 *sched_job) 706 { 707 struct panfrost_job *job = to_panfrost_job(sched_job); 708 struct panfrost_device *pfdev = job->pfdev; 709 int js = panfrost_job_get_slot(job); 710 711 /* 712 * If the GPU managed to complete this jobs fence, the timeout is 713 * spurious. Bail out. 714 */ 715 if (dma_fence_is_signaled(job->done_fence)) 716 return DRM_GPU_SCHED_STAT_NOMINAL; 717 718 dev_err(pfdev->dev, "gpu sched timeout, js=%d, config=0x%x, status=0x%x, head=0x%x, tail=0x%x, sched_job=%p", 719 js, 720 job_read(pfdev, JS_CONFIG(js)), 721 job_read(pfdev, JS_STATUS(js)), 722 job_read(pfdev, JS_HEAD_LO(js)), 723 job_read(pfdev, JS_TAIL_LO(js)), 724 sched_job); 725 726 atomic_set(&pfdev->reset.pending, 1); 727 panfrost_reset(pfdev, sched_job); 728 729 return DRM_GPU_SCHED_STAT_NOMINAL; 730 } 731 732 static void panfrost_reset_work(struct work_struct *work) 733 { 734 struct panfrost_device *pfdev; 735 736 pfdev = container_of(work, struct panfrost_device, reset.work); 737 panfrost_reset(pfdev, NULL); 738 } 739 740 static const struct drm_sched_backend_ops panfrost_sched_ops = { 741 .run_job = panfrost_job_run, 742 .timedout_job = panfrost_job_timedout, 743 .free_job = panfrost_job_free 744 }; 745 746 static irqreturn_t panfrost_job_irq_handler_thread(int irq, void *data) 747 { 748 struct panfrost_device *pfdev = data; 749 750 panfrost_job_handle_irqs(pfdev); 751 job_write(pfdev, JOB_INT_MASK, 752 GENMASK(16 + NUM_JOB_SLOTS - 1, 16) | 753 GENMASK(NUM_JOB_SLOTS - 1, 0)); 754 return IRQ_HANDLED; 755 } 756 757 static irqreturn_t panfrost_job_irq_handler(int irq, void *data) 758 { 759 struct panfrost_device *pfdev = data; 760 u32 status = job_read(pfdev, JOB_INT_STAT); 761 762 if (!status) 763 return IRQ_NONE; 764 765 job_write(pfdev, JOB_INT_MASK, 0); 766 return IRQ_WAKE_THREAD; 767 } 768 769 int panfrost_job_init(struct panfrost_device *pfdev) 770 { 771 struct panfrost_job_slot *js; 772 unsigned int nentries = 2; 773 int ret, j; 774 775 /* All GPUs have two entries per queue, but without jobchain 776 * disambiguation stopping the right job in the close path is tricky, 777 * so let's just advertise one entry in that case. 778 */ 779 if (!panfrost_has_hw_feature(pfdev, HW_FEATURE_JOBCHAIN_DISAMBIGUATION)) 780 nentries = 1; 781 782 pfdev->js = js = devm_kzalloc(pfdev->dev, sizeof(*js), GFP_KERNEL); 783 if (!js) 784 return -ENOMEM; 785 786 INIT_WORK(&pfdev->reset.work, panfrost_reset_work); 787 spin_lock_init(&js->job_lock); 788 789 js->irq = platform_get_irq_byname(to_platform_device(pfdev->dev), "job"); 790 if (js->irq <= 0) 791 return -ENODEV; 792 793 ret = devm_request_threaded_irq(pfdev->dev, js->irq, 794 panfrost_job_irq_handler, 795 panfrost_job_irq_handler_thread, 796 IRQF_SHARED, KBUILD_MODNAME "-job", 797 pfdev); 798 if (ret) { 799 dev_err(pfdev->dev, "failed to request job irq"); 800 return ret; 801 } 802 803 pfdev->reset.wq = alloc_ordered_workqueue("panfrost-reset", 0); 804 if (!pfdev->reset.wq) 805 return -ENOMEM; 806 807 for (j = 0; j < NUM_JOB_SLOTS; j++) { 808 js->queue[j].fence_context = dma_fence_context_alloc(1); 809 810 ret = drm_sched_init(&js->queue[j].sched, 811 &panfrost_sched_ops, 812 nentries, 0, 813 msecs_to_jiffies(JOB_TIMEOUT_MS), 814 pfdev->reset.wq, 815 NULL, "pan_js", pfdev->dev); 816 if (ret) { 817 dev_err(pfdev->dev, "Failed to create scheduler: %d.", ret); 818 goto err_sched; 819 } 820 } 821 822 panfrost_job_enable_interrupts(pfdev); 823 824 return 0; 825 826 err_sched: 827 for (j--; j >= 0; j--) 828 drm_sched_fini(&js->queue[j].sched); 829 830 destroy_workqueue(pfdev->reset.wq); 831 return ret; 832 } 833 834 void panfrost_job_fini(struct panfrost_device *pfdev) 835 { 836 struct panfrost_job_slot *js = pfdev->js; 837 int j; 838 839 job_write(pfdev, JOB_INT_MASK, 0); 840 841 for (j = 0; j < NUM_JOB_SLOTS; j++) { 842 drm_sched_fini(&js->queue[j].sched); 843 } 844 845 cancel_work_sync(&pfdev->reset.work); 846 destroy_workqueue(pfdev->reset.wq); 847 } 848 849 int panfrost_job_open(struct panfrost_file_priv *panfrost_priv) 850 { 851 struct panfrost_device *pfdev = panfrost_priv->pfdev; 852 struct panfrost_job_slot *js = pfdev->js; 853 struct drm_gpu_scheduler *sched; 854 int ret, i; 855 856 for (i = 0; i < NUM_JOB_SLOTS; i++) { 857 sched = &js->queue[i].sched; 858 ret = drm_sched_entity_init(&panfrost_priv->sched_entity[i], 859 DRM_SCHED_PRIORITY_NORMAL, &sched, 860 1, NULL); 861 if (WARN_ON(ret)) 862 return ret; 863 } 864 return 0; 865 } 866 867 void panfrost_job_close(struct panfrost_file_priv *panfrost_priv) 868 { 869 struct panfrost_device *pfdev = panfrost_priv->pfdev; 870 int i; 871 872 for (i = 0; i < NUM_JOB_SLOTS; i++) 873 drm_sched_entity_destroy(&panfrost_priv->sched_entity[i]); 874 875 /* Kill in-flight jobs */ 876 spin_lock(&pfdev->js->job_lock); 877 for (i = 0; i < NUM_JOB_SLOTS; i++) { 878 struct drm_sched_entity *entity = &panfrost_priv->sched_entity[i]; 879 int j; 880 881 for (j = ARRAY_SIZE(pfdev->jobs[0]) - 1; j >= 0; j--) { 882 struct panfrost_job *job = pfdev->jobs[i][j]; 883 u32 cmd; 884 885 if (!job || job->base.entity != entity) 886 continue; 887 888 if (j == 1) { 889 /* Try to cancel the job before it starts */ 890 job_write(pfdev, JS_COMMAND_NEXT(i), JS_COMMAND_NOP); 891 /* Reset the job head so it doesn't get restarted if 892 * the job in the first slot failed. 893 */ 894 job->jc = 0; 895 } 896 897 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_JOBCHAIN_DISAMBIGUATION)) { 898 cmd = panfrost_get_job_chain_flag(job) ? 899 JS_COMMAND_HARD_STOP_1 : 900 JS_COMMAND_HARD_STOP_0; 901 } else { 902 cmd = JS_COMMAND_HARD_STOP; 903 } 904 905 job_write(pfdev, JS_COMMAND(i), cmd); 906 } 907 } 908 spin_unlock(&pfdev->js->job_lock); 909 } 910 911 int panfrost_job_is_idle(struct panfrost_device *pfdev) 912 { 913 struct panfrost_job_slot *js = pfdev->js; 914 int i; 915 916 for (i = 0; i < NUM_JOB_SLOTS; i++) { 917 /* If there are any jobs in the HW queue, we're not idle */ 918 if (atomic_read(&js->queue[i].sched.hw_rq_count)) 919 return false; 920 } 921 922 return true; 923 } 924