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/platform_device.h> 8 #include <linux/pm_runtime.h> 9 #include <linux/dma-resv.h> 10 #include <drm/gpu_scheduler.h> 11 #include <drm/panfrost_drm.h> 12 13 #include "panfrost_device.h" 14 #include "panfrost_devfreq.h" 15 #include "panfrost_job.h" 16 #include "panfrost_features.h" 17 #include "panfrost_issues.h" 18 #include "panfrost_gem.h" 19 #include "panfrost_regs.h" 20 #include "panfrost_gpu.h" 21 #include "panfrost_mmu.h" 22 23 #define job_write(dev, reg, data) writel(data, dev->iomem + (reg)) 24 #define job_read(dev, reg) readl(dev->iomem + (reg)) 25 26 struct panfrost_queue_state { 27 struct drm_gpu_scheduler sched; 28 29 u64 fence_context; 30 u64 emit_seqno; 31 }; 32 33 struct panfrost_job_slot { 34 struct panfrost_queue_state queue[NUM_JOB_SLOTS]; 35 spinlock_t job_lock; 36 }; 37 38 static struct panfrost_job * 39 to_panfrost_job(struct drm_sched_job *sched_job) 40 { 41 return container_of(sched_job, struct panfrost_job, base); 42 } 43 44 struct panfrost_fence { 45 struct dma_fence base; 46 struct drm_device *dev; 47 /* panfrost seqno for signaled() test */ 48 u64 seqno; 49 int queue; 50 }; 51 52 static inline struct panfrost_fence * 53 to_panfrost_fence(struct dma_fence *fence) 54 { 55 return (struct panfrost_fence *)fence; 56 } 57 58 static const char *panfrost_fence_get_driver_name(struct dma_fence *fence) 59 { 60 return "panfrost"; 61 } 62 63 static const char *panfrost_fence_get_timeline_name(struct dma_fence *fence) 64 { 65 struct panfrost_fence *f = to_panfrost_fence(fence); 66 67 switch (f->queue) { 68 case 0: 69 return "panfrost-js-0"; 70 case 1: 71 return "panfrost-js-1"; 72 case 2: 73 return "panfrost-js-2"; 74 default: 75 return NULL; 76 } 77 } 78 79 static const struct dma_fence_ops panfrost_fence_ops = { 80 .get_driver_name = panfrost_fence_get_driver_name, 81 .get_timeline_name = panfrost_fence_get_timeline_name, 82 }; 83 84 static struct dma_fence *panfrost_fence_create(struct panfrost_device *pfdev, int js_num) 85 { 86 struct panfrost_fence *fence; 87 struct panfrost_job_slot *js = pfdev->js; 88 89 fence = kzalloc(sizeof(*fence), GFP_KERNEL); 90 if (!fence) 91 return ERR_PTR(-ENOMEM); 92 93 fence->dev = pfdev->ddev; 94 fence->queue = js_num; 95 fence->seqno = ++js->queue[js_num].emit_seqno; 96 dma_fence_init(&fence->base, &panfrost_fence_ops, &js->job_lock, 97 js->queue[js_num].fence_context, fence->seqno); 98 99 return &fence->base; 100 } 101 102 static int panfrost_job_get_slot(struct panfrost_job *job) 103 { 104 /* JS0: fragment jobs. 105 * JS1: vertex/tiler jobs 106 * JS2: compute jobs 107 */ 108 if (job->requirements & PANFROST_JD_REQ_FS) 109 return 0; 110 111 /* Not exposed to userspace yet */ 112 #if 0 113 if (job->requirements & PANFROST_JD_REQ_ONLY_COMPUTE) { 114 if ((job->requirements & PANFROST_JD_REQ_CORE_GRP_MASK) && 115 (job->pfdev->features.nr_core_groups == 2)) 116 return 2; 117 if (panfrost_has_hw_issue(job->pfdev, HW_ISSUE_8987)) 118 return 2; 119 } 120 #endif 121 return 1; 122 } 123 124 static void panfrost_job_write_affinity(struct panfrost_device *pfdev, 125 u32 requirements, 126 int js) 127 { 128 u64 affinity; 129 130 /* 131 * Use all cores for now. 132 * Eventually we may need to support tiler only jobs and h/w with 133 * multiple (2) coherent core groups 134 */ 135 affinity = pfdev->features.shader_present; 136 137 job_write(pfdev, JS_AFFINITY_NEXT_LO(js), affinity & 0xFFFFFFFF); 138 job_write(pfdev, JS_AFFINITY_NEXT_HI(js), affinity >> 32); 139 } 140 141 static void panfrost_job_hw_submit(struct panfrost_job *job, int js) 142 { 143 struct panfrost_device *pfdev = job->pfdev; 144 u32 cfg; 145 u64 jc_head = job->jc; 146 int ret; 147 148 ret = pm_runtime_get_sync(pfdev->dev); 149 if (ret < 0) 150 return; 151 152 if (WARN_ON(job_read(pfdev, JS_COMMAND_NEXT(js)))) { 153 pm_runtime_put_sync_autosuspend(pfdev->dev); 154 return; 155 } 156 157 cfg = panfrost_mmu_as_get(pfdev, &job->file_priv->mmu); 158 panfrost_devfreq_record_busy(pfdev); 159 160 job_write(pfdev, JS_HEAD_NEXT_LO(js), jc_head & 0xFFFFFFFF); 161 job_write(pfdev, JS_HEAD_NEXT_HI(js), jc_head >> 32); 162 163 panfrost_job_write_affinity(pfdev, job->requirements, js); 164 165 /* start MMU, medium priority, cache clean/flush on end, clean/flush on 166 * start */ 167 cfg |= JS_CONFIG_THREAD_PRI(8) | 168 JS_CONFIG_START_FLUSH_CLEAN_INVALIDATE | 169 JS_CONFIG_END_FLUSH_CLEAN_INVALIDATE; 170 171 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_FLUSH_REDUCTION)) 172 cfg |= JS_CONFIG_ENABLE_FLUSH_REDUCTION; 173 174 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_10649)) 175 cfg |= JS_CONFIG_START_MMU; 176 177 job_write(pfdev, JS_CONFIG_NEXT(js), cfg); 178 179 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_FLUSH_REDUCTION)) 180 job_write(pfdev, JS_FLUSH_ID_NEXT(js), job->flush_id); 181 182 /* GO ! */ 183 dev_dbg(pfdev->dev, "JS: Submitting atom %p to js[%d] with head=0x%llx", 184 job, js, jc_head); 185 186 job_write(pfdev, JS_COMMAND_NEXT(js), JS_COMMAND_START); 187 } 188 189 static void panfrost_acquire_object_fences(struct drm_gem_object **bos, 190 int bo_count, 191 struct dma_fence **implicit_fences) 192 { 193 int i; 194 195 for (i = 0; i < bo_count; i++) 196 implicit_fences[i] = dma_resv_get_excl_rcu(bos[i]->resv); 197 } 198 199 static void panfrost_attach_object_fences(struct drm_gem_object **bos, 200 int bo_count, 201 struct dma_fence *fence) 202 { 203 int i; 204 205 for (i = 0; i < bo_count; i++) 206 dma_resv_add_excl_fence(bos[i]->resv, fence); 207 } 208 209 int panfrost_job_push(struct panfrost_job *job) 210 { 211 struct panfrost_device *pfdev = job->pfdev; 212 int slot = panfrost_job_get_slot(job); 213 struct drm_sched_entity *entity = &job->file_priv->sched_entity[slot]; 214 struct ww_acquire_ctx acquire_ctx; 215 int ret = 0; 216 217 mutex_lock(&pfdev->sched_lock); 218 219 ret = drm_gem_lock_reservations(job->bos, job->bo_count, 220 &acquire_ctx); 221 if (ret) { 222 mutex_unlock(&pfdev->sched_lock); 223 return ret; 224 } 225 226 ret = drm_sched_job_init(&job->base, entity, NULL); 227 if (ret) { 228 mutex_unlock(&pfdev->sched_lock); 229 goto unlock; 230 } 231 232 job->render_done_fence = dma_fence_get(&job->base.s_fence->finished); 233 234 kref_get(&job->refcount); /* put by scheduler job completion */ 235 236 panfrost_acquire_object_fences(job->bos, job->bo_count, 237 job->implicit_fences); 238 239 drm_sched_entity_push_job(&job->base, entity); 240 241 mutex_unlock(&pfdev->sched_lock); 242 243 panfrost_attach_object_fences(job->bos, job->bo_count, 244 job->render_done_fence); 245 246 unlock: 247 drm_gem_unlock_reservations(job->bos, job->bo_count, &acquire_ctx); 248 249 return ret; 250 } 251 252 static void panfrost_job_cleanup(struct kref *ref) 253 { 254 struct panfrost_job *job = container_of(ref, struct panfrost_job, 255 refcount); 256 unsigned int i; 257 258 if (job->in_fences) { 259 for (i = 0; i < job->in_fence_count; i++) 260 dma_fence_put(job->in_fences[i]); 261 kvfree(job->in_fences); 262 } 263 if (job->implicit_fences) { 264 for (i = 0; i < job->bo_count; i++) 265 dma_fence_put(job->implicit_fences[i]); 266 kvfree(job->implicit_fences); 267 } 268 dma_fence_put(job->done_fence); 269 dma_fence_put(job->render_done_fence); 270 271 if (job->bos) { 272 for (i = 0; i < job->bo_count; i++) 273 drm_gem_object_put_unlocked(job->bos[i]); 274 kvfree(job->bos); 275 } 276 277 kfree(job); 278 } 279 280 void panfrost_job_put(struct panfrost_job *job) 281 { 282 kref_put(&job->refcount, panfrost_job_cleanup); 283 } 284 285 static void panfrost_job_free(struct drm_sched_job *sched_job) 286 { 287 struct panfrost_job *job = to_panfrost_job(sched_job); 288 289 drm_sched_job_cleanup(sched_job); 290 291 panfrost_job_put(job); 292 } 293 294 static struct dma_fence *panfrost_job_dependency(struct drm_sched_job *sched_job, 295 struct drm_sched_entity *s_entity) 296 { 297 struct panfrost_job *job = to_panfrost_job(sched_job); 298 struct dma_fence *fence; 299 unsigned int i; 300 301 /* Explicit fences */ 302 for (i = 0; i < job->in_fence_count; i++) { 303 if (job->in_fences[i]) { 304 fence = job->in_fences[i]; 305 job->in_fences[i] = NULL; 306 return fence; 307 } 308 } 309 310 /* Implicit fences, max. one per BO */ 311 for (i = 0; i < job->bo_count; i++) { 312 if (job->implicit_fences[i]) { 313 fence = job->implicit_fences[i]; 314 job->implicit_fences[i] = NULL; 315 return fence; 316 } 317 } 318 319 return NULL; 320 } 321 322 static struct dma_fence *panfrost_job_run(struct drm_sched_job *sched_job) 323 { 324 struct panfrost_job *job = to_panfrost_job(sched_job); 325 struct panfrost_device *pfdev = job->pfdev; 326 int slot = panfrost_job_get_slot(job); 327 struct dma_fence *fence = NULL; 328 329 if (unlikely(job->base.s_fence->finished.error)) 330 return NULL; 331 332 pfdev->jobs[slot] = job; 333 334 fence = panfrost_fence_create(pfdev, slot); 335 if (IS_ERR(fence)) 336 return NULL; 337 338 if (job->done_fence) 339 dma_fence_put(job->done_fence); 340 job->done_fence = dma_fence_get(fence); 341 342 panfrost_job_hw_submit(job, slot); 343 344 return fence; 345 } 346 347 void panfrost_job_enable_interrupts(struct panfrost_device *pfdev) 348 { 349 int j; 350 u32 irq_mask = 0; 351 352 for (j = 0; j < NUM_JOB_SLOTS; j++) { 353 irq_mask |= MK_JS_MASK(j); 354 } 355 356 job_write(pfdev, JOB_INT_CLEAR, irq_mask); 357 job_write(pfdev, JOB_INT_MASK, irq_mask); 358 } 359 360 static void panfrost_job_timedout(struct drm_sched_job *sched_job) 361 { 362 struct panfrost_job *job = to_panfrost_job(sched_job); 363 struct panfrost_device *pfdev = job->pfdev; 364 int js = panfrost_job_get_slot(job); 365 unsigned long flags; 366 int i; 367 368 /* 369 * If the GPU managed to complete this jobs fence, the timeout is 370 * spurious. Bail out. 371 */ 372 if (dma_fence_is_signaled(job->done_fence)) 373 return; 374 375 dev_err(pfdev->dev, "gpu sched timeout, js=%d, config=0x%x, status=0x%x, head=0x%x, tail=0x%x, sched_job=%p", 376 js, 377 job_read(pfdev, JS_CONFIG(js)), 378 job_read(pfdev, JS_STATUS(js)), 379 job_read(pfdev, JS_HEAD_LO(js)), 380 job_read(pfdev, JS_TAIL_LO(js)), 381 sched_job); 382 383 if (!mutex_trylock(&pfdev->reset_lock)) 384 return; 385 386 for (i = 0; i < NUM_JOB_SLOTS; i++) { 387 struct drm_gpu_scheduler *sched = &pfdev->js->queue[i].sched; 388 389 drm_sched_stop(sched, sched_job); 390 if (js != i) 391 /* Ensure any timeouts on other slots have finished */ 392 cancel_delayed_work_sync(&sched->work_tdr); 393 } 394 395 drm_sched_increase_karma(sched_job); 396 397 spin_lock_irqsave(&pfdev->js->job_lock, flags); 398 for (i = 0; i < NUM_JOB_SLOTS; i++) { 399 if (pfdev->jobs[i]) { 400 pm_runtime_put_noidle(pfdev->dev); 401 pfdev->jobs[i] = NULL; 402 } 403 } 404 spin_unlock_irqrestore(&pfdev->js->job_lock, flags); 405 406 panfrost_devfreq_record_idle(pfdev); 407 panfrost_device_reset(pfdev); 408 409 for (i = 0; i < NUM_JOB_SLOTS; i++) 410 drm_sched_resubmit_jobs(&pfdev->js->queue[i].sched); 411 412 /* restart scheduler after GPU is usable again */ 413 for (i = 0; i < NUM_JOB_SLOTS; i++) 414 drm_sched_start(&pfdev->js->queue[i].sched, true); 415 416 mutex_unlock(&pfdev->reset_lock); 417 } 418 419 static const struct drm_sched_backend_ops panfrost_sched_ops = { 420 .dependency = panfrost_job_dependency, 421 .run_job = panfrost_job_run, 422 .timedout_job = panfrost_job_timedout, 423 .free_job = panfrost_job_free 424 }; 425 426 static irqreturn_t panfrost_job_irq_handler(int irq, void *data) 427 { 428 struct panfrost_device *pfdev = data; 429 u32 status = job_read(pfdev, JOB_INT_STAT); 430 int j; 431 432 dev_dbg(pfdev->dev, "jobslot irq status=%x\n", status); 433 434 if (!status) 435 return IRQ_NONE; 436 437 pm_runtime_mark_last_busy(pfdev->dev); 438 439 for (j = 0; status; j++) { 440 u32 mask = MK_JS_MASK(j); 441 442 if (!(status & mask)) 443 continue; 444 445 job_write(pfdev, JOB_INT_CLEAR, mask); 446 447 if (status & JOB_INT_MASK_ERR(j)) { 448 job_write(pfdev, JS_COMMAND_NEXT(j), JS_COMMAND_NOP); 449 450 dev_err(pfdev->dev, "js fault, js=%d, status=%s, head=0x%x, tail=0x%x", 451 j, 452 panfrost_exception_name(pfdev, job_read(pfdev, JS_STATUS(j))), 453 job_read(pfdev, JS_HEAD_LO(j)), 454 job_read(pfdev, JS_TAIL_LO(j))); 455 456 drm_sched_fault(&pfdev->js->queue[j].sched); 457 } 458 459 if (status & JOB_INT_MASK_DONE(j)) { 460 struct panfrost_job *job; 461 462 spin_lock(&pfdev->js->job_lock); 463 job = pfdev->jobs[j]; 464 /* Only NULL if job timeout occurred */ 465 if (job) { 466 pfdev->jobs[j] = NULL; 467 468 panfrost_mmu_as_put(pfdev, &job->file_priv->mmu); 469 panfrost_devfreq_record_idle(pfdev); 470 471 dma_fence_signal_locked(job->done_fence); 472 pm_runtime_put_autosuspend(pfdev->dev); 473 } 474 spin_unlock(&pfdev->js->job_lock); 475 } 476 477 status &= ~mask; 478 } 479 480 return IRQ_HANDLED; 481 } 482 483 int panfrost_job_init(struct panfrost_device *pfdev) 484 { 485 struct panfrost_job_slot *js; 486 int ret, j, irq; 487 488 pfdev->js = js = devm_kzalloc(pfdev->dev, sizeof(*js), GFP_KERNEL); 489 if (!js) 490 return -ENOMEM; 491 492 spin_lock_init(&js->job_lock); 493 494 irq = platform_get_irq_byname(to_platform_device(pfdev->dev), "job"); 495 if (irq <= 0) 496 return -ENODEV; 497 498 ret = devm_request_irq(pfdev->dev, irq, panfrost_job_irq_handler, 499 IRQF_SHARED, "job", pfdev); 500 if (ret) { 501 dev_err(pfdev->dev, "failed to request job irq"); 502 return ret; 503 } 504 505 for (j = 0; j < NUM_JOB_SLOTS; j++) { 506 js->queue[j].fence_context = dma_fence_context_alloc(1); 507 508 ret = drm_sched_init(&js->queue[j].sched, 509 &panfrost_sched_ops, 510 1, 0, msecs_to_jiffies(500), 511 "pan_js"); 512 if (ret) { 513 dev_err(pfdev->dev, "Failed to create scheduler: %d.", ret); 514 goto err_sched; 515 } 516 } 517 518 panfrost_job_enable_interrupts(pfdev); 519 520 return 0; 521 522 err_sched: 523 for (j--; j >= 0; j--) 524 drm_sched_fini(&js->queue[j].sched); 525 526 return ret; 527 } 528 529 void panfrost_job_fini(struct panfrost_device *pfdev) 530 { 531 struct panfrost_job_slot *js = pfdev->js; 532 int j; 533 534 job_write(pfdev, JOB_INT_MASK, 0); 535 536 for (j = 0; j < NUM_JOB_SLOTS; j++) 537 drm_sched_fini(&js->queue[j].sched); 538 539 } 540 541 int panfrost_job_open(struct panfrost_file_priv *panfrost_priv) 542 { 543 struct panfrost_device *pfdev = panfrost_priv->pfdev; 544 struct panfrost_job_slot *js = pfdev->js; 545 struct drm_sched_rq *rq; 546 int ret, i; 547 548 for (i = 0; i < NUM_JOB_SLOTS; i++) { 549 rq = &js->queue[i].sched.sched_rq[DRM_SCHED_PRIORITY_NORMAL]; 550 ret = drm_sched_entity_init(&panfrost_priv->sched_entity[i], &rq, 1, NULL); 551 if (WARN_ON(ret)) 552 return ret; 553 } 554 return 0; 555 } 556 557 void panfrost_job_close(struct panfrost_file_priv *panfrost_priv) 558 { 559 int i; 560 561 for (i = 0; i < NUM_JOB_SLOTS; i++) 562 drm_sched_entity_destroy(&panfrost_priv->sched_entity[i]); 563 } 564 565 int panfrost_job_is_idle(struct panfrost_device *pfdev) 566 { 567 struct panfrost_job_slot *js = pfdev->js; 568 int i; 569 570 /* Check whether the hardware is idle */ 571 if (atomic_read(&pfdev->devfreq.busy_count)) 572 return false; 573 574 for (i = 0; i < NUM_JOB_SLOTS; i++) { 575 /* If there are any jobs in the HW queue, we're not idle */ 576 if (atomic_read(&js->queue[i].sched.hw_rq_count)) 577 return false; 578 } 579 580 return true; 581 } 582