drivers/gpu/drm/i915/gt/intel_context.c

   1 // SPDX-License-Identifier: MIT
   2 /*
   3  * Copyright © 2019 Intel Corporation
   4  */
   5
   6 #include "gem/i915_gem_context.h"
   7 #include "gem/i915_gem_pm.h"
   8
   9 #include "i915_drv.h"
  10 #include "i915_trace.h"
  11
  12 #include "intel_context.h"
  13 #include "intel_engine.h"
  14 #include "intel_engine_pm.h"
  15 #include "intel_ring.h"
  16
  17 static struct kmem_cache *slab_ce;
  18
  19 static struct intel_context *intel_context_alloc(void)
  20 {
  21         return kmem_cache_zalloc(slab_ce, GFP_KERNEL);
  22 }
  23
  24 static void rcu_context_free(struct rcu_head *rcu)
  25 {
  26         struct intel_context *ce = container_of(rcu, typeof(*ce), rcu);
  27
  28         trace_intel_context_free(ce);
  29         kmem_cache_free(slab_ce, ce);
  30 }
  31
  32 void intel_context_free(struct intel_context *ce)
  33 {
  34         call_rcu(&ce->rcu, rcu_context_free);
  35 }
  36
  37 struct intel_context *
  38 intel_context_create(struct intel_engine_cs *engine)
  39 {
  40         struct intel_context *ce;
  41
  42         ce = intel_context_alloc();
  43         if (!ce)
  44                 return ERR_PTR(-ENOMEM);
  45
  46         intel_context_init(ce, engine);
  47         trace_intel_context_create(ce);
  48         return ce;
  49 }
  50
  51 int intel_context_alloc_state(struct intel_context *ce)
  52 {
  53         int err = 0;
  54
  55         if (mutex_lock_interruptible(&ce->pin_mutex))
  56                 return -EINTR;
  57
  58         if (!test_bit(CONTEXT_ALLOC_BIT, &ce->flags)) {
  59                 if (intel_context_is_banned(ce)) {
  60                         err = -EIO;
  61                         goto unlock;
  62                 }
  63
  64                 err = ce->ops->alloc(ce);
  65                 if (unlikely(err))
  66                         goto unlock;
  67
  68                 set_bit(CONTEXT_ALLOC_BIT, &ce->flags);
  69         }
  70
  71 unlock:
  72         mutex_unlock(&ce->pin_mutex);
  73         return err;
  74 }
  75
  76 static int intel_context_active_acquire(struct intel_context *ce)
  77 {
  78         int err;
  79
  80         __i915_active_acquire(&ce->active);
  81
  82         if (intel_context_is_barrier(ce) || intel_engine_uses_guc(ce->engine) ||
  83             intel_context_is_parallel(ce))
  84                 return 0;
  85
  86         /* Preallocate tracking nodes */
  87         err = i915_active_acquire_preallocate_barrier(&ce->active,
  88                                                       ce->engine);
  89         if (err)
  90                 i915_active_release(&ce->active);
  91
  92         return err;
  93 }
  94
  95 static void intel_context_active_release(struct intel_context *ce)
  96 {
  97         /* Nodes preallocated in intel_context_active() */
  98         i915_active_acquire_barrier(&ce->active);
  99         i915_active_release(&ce->active);
 100 }
 101
 102 static int __context_pin_state(struct i915_vma *vma, struct i915_gem_ww_ctx *ww)
 103 {
 104         unsigned int bias = i915_ggtt_pin_bias(vma) | PIN_OFFSET_BIAS;
 105         int err;
 106
 107         err = i915_ggtt_pin(vma, ww, 0, bias | PIN_HIGH);
 108         if (err)
 109                 return err;
 110
 111         err = i915_active_acquire(&vma->active);
 112         if (err)
 113                 goto err_unpin;
 114
 115         /*
 116          * And mark it as a globally pinned object to let the shrinker know
 117          * it cannot reclaim the object until we release it.
 118          */
 119         i915_vma_make_unshrinkable(vma);
 120         vma->obj->mm.dirty = true;
 121
 122         return 0;
 123
 124 err_unpin:
 125         i915_vma_unpin(vma);
 126         return err;
 127 }
 128
 129 static void __context_unpin_state(struct i915_vma *vma)
 130 {
 131         i915_vma_make_shrinkable(vma);
 132         i915_active_release(&vma->active);
 133         __i915_vma_unpin(vma);
 134 }
 135
 136 static int __ring_active(struct intel_ring *ring,
 137                          struct i915_gem_ww_ctx *ww)
 138 {
 139         int err;
 140
 141         err = intel_ring_pin(ring, ww);
 142         if (err)
 143                 return err;
 144
 145         err = i915_active_acquire(&ring->vma->active);
 146         if (err)
 147                 goto err_pin;
 148
 149         return 0;
 150
 151 err_pin:
 152         intel_ring_unpin(ring);
 153         return err;
 154 }
 155
 156 static void __ring_retire(struct intel_ring *ring)
 157 {
 158         i915_active_release(&ring->vma->active);
 159         intel_ring_unpin(ring);
 160 }
 161
 162 static int intel_context_pre_pin(struct intel_context *ce,
 163                                  struct i915_gem_ww_ctx *ww)
 164 {
 165         int err;
 166
 167         CE_TRACE(ce, "active\n");
 168
 169         err = __ring_active(ce->ring, ww);
 170         if (err)
 171                 return err;
 172
 173         err = intel_timeline_pin(ce->timeline, ww);
 174         if (err)
 175                 goto err_ring;
 176
 177         if (!ce->state)
 178                 return 0;
 179
 180         err = __context_pin_state(ce->state, ww);
 181         if (err)
 182                 goto err_timeline;
 183
 184
 185         return 0;
 186
 187 err_timeline:
 188         intel_timeline_unpin(ce->timeline);
 189 err_ring:
 190         __ring_retire(ce->ring);
 191         return err;
 192 }
 193
 194 static void intel_context_post_unpin(struct intel_context *ce)
 195 {
 196         if (ce->state)
 197                 __context_unpin_state(ce->state);
 198
 199         intel_timeline_unpin(ce->timeline);
 200         __ring_retire(ce->ring);
 201 }
 202
 203 int __intel_context_do_pin_ww(struct intel_context *ce,
 204                               struct i915_gem_ww_ctx *ww)
 205 {
 206         bool handoff = false;
 207         void *vaddr;
 208         int err = 0;
 209
 210         if (unlikely(!test_bit(CONTEXT_ALLOC_BIT, &ce->flags))) {
 211                 err = intel_context_alloc_state(ce);
 212                 if (err)
 213                         return err;
 214         }
 215
 216         /*
 217          * We always pin the context/ring/timeline here, to ensure a pin
 218          * refcount for __intel_context_active(), which prevent a lock
 219          * inversion of ce->pin_mutex vs dma_resv_lock().
 220          */
 221
 222         err = i915_gem_object_lock(ce->timeline->hwsp_ggtt->obj, ww);
 223         if (!err)
 224                 err = i915_gem_object_lock(ce->ring->vma->obj, ww);
 225         if (!err && ce->state)
 226                 err = i915_gem_object_lock(ce->state->obj, ww);
 227         if (!err)
 228                 err = intel_context_pre_pin(ce, ww);
 229         if (err)
 230                 return err;
 231
 232         err = ce->ops->pre_pin(ce, ww, &vaddr);
 233         if (err)
 234                 goto err_ctx_unpin;
 235
 236         err = i915_active_acquire(&ce->active);
 237         if (err)
 238                 goto err_post_unpin;
 239
 240         err = mutex_lock_interruptible(&ce->pin_mutex);
 241         if (err)
 242                 goto err_release;
 243
 244         intel_engine_pm_might_get(ce->engine);
 245
 246         if (unlikely(intel_context_is_closed(ce))) {
 247                 err = -ENOENT;
 248                 goto err_unlock;
 249         }
 250
 251         if (likely(!atomic_add_unless(&ce->pin_count, 1, 0))) {
 252                 err = intel_context_active_acquire(ce);
 253                 if (unlikely(err))
 254                         goto err_unlock;
 255
 256                 err = ce->ops->pin(ce, vaddr);
 257                 if (err) {
 258                         intel_context_active_release(ce);
 259                         goto err_unlock;
 260                 }
 261
 262                 CE_TRACE(ce, "pin ring:{start:%08x, head:%04x, tail:%04x}\n",
 263                          i915_ggtt_offset(ce->ring->vma),
 264                          ce->ring->head, ce->ring->tail);
 265
 266                 handoff = true;
 267                 smp_mb__before_atomic(); /* flush pin before it is visible */
 268                 atomic_inc(&ce->pin_count);
 269         }
 270
 271         GEM_BUG_ON(!intel_context_is_pinned(ce)); /* no overflow! */
 272
 273         trace_intel_context_do_pin(ce);
 274
 275 err_unlock:
 276         mutex_unlock(&ce->pin_mutex);
 277 err_release:
 278         i915_active_release(&ce->active);
 279 err_post_unpin:
 280         if (!handoff)
 281                 ce->ops->post_unpin(ce);
 282 err_ctx_unpin:
 283         intel_context_post_unpin(ce);
 284
 285         /*
 286          * Unlock the hwsp_ggtt object since it's shared.
 287          * In principle we can unlock all the global state locked above
 288          * since it's pinned and doesn't need fencing, and will
 289          * thus remain resident until it is explicitly unpinned.
 290          */
 291         i915_gem_ww_unlock_single(ce->timeline->hwsp_ggtt->obj);
 292
 293         return err;
 294 }
 295
 296 int __intel_context_do_pin(struct intel_context *ce)
 297 {
 298         struct i915_gem_ww_ctx ww;
 299         int err;
 300
 301         i915_gem_ww_ctx_init(&ww, true);
 302 retry:
 303         err = __intel_context_do_pin_ww(ce, &ww);
 304         if (err == -EDEADLK) {
 305                 err = i915_gem_ww_ctx_backoff(&ww);
 306                 if (!err)
 307                         goto retry;
 308         }
 309         i915_gem_ww_ctx_fini(&ww);
 310         return err;
 311 }
 312
 313 void __intel_context_do_unpin(struct intel_context *ce, int sub)
 314 {
 315         if (!atomic_sub_and_test(sub, &ce->pin_count))
 316                 return;
 317
 318         CE_TRACE(ce, "unpin\n");
 319         ce->ops->unpin(ce);
 320         ce->ops->post_unpin(ce);
 321
 322         /*
 323          * Once released, we may asynchronously drop the active reference.
 324          * As that may be the only reference keeping the context alive,
 325          * take an extra now so that it is not freed before we finish
 326          * dereferencing it.
 327          */
 328         intel_context_get(ce);
 329         intel_context_active_release(ce);
 330         trace_intel_context_do_unpin(ce);
 331         intel_context_put(ce);
 332 }
 333
 334 static void __intel_context_retire(struct i915_active *active)
 335 {
 336         struct intel_context *ce = container_of(active, typeof(*ce), active);
 337
 338         CE_TRACE(ce, "retire runtime: { total:%lluns, avg:%lluns }\n",
 339                  intel_context_get_total_runtime_ns(ce),
 340                  intel_context_get_avg_runtime_ns(ce));
 341
 342         set_bit(CONTEXT_VALID_BIT, &ce->flags);
 343         intel_context_post_unpin(ce);
 344         intel_context_put(ce);
 345 }
 346
 347 static int __intel_context_active(struct i915_active *active)
 348 {
 349         struct intel_context *ce = container_of(active, typeof(*ce), active);
 350
 351         intel_context_get(ce);
 352
 353         /* everything should already be activated by intel_context_pre_pin() */
 354         GEM_WARN_ON(!i915_active_acquire_if_busy(&ce->ring->vma->active));
 355         __intel_ring_pin(ce->ring);
 356
 357         __intel_timeline_pin(ce->timeline);
 358
 359         if (ce->state) {
 360                 GEM_WARN_ON(!i915_active_acquire_if_busy(&ce->state->active));
 361                 __i915_vma_pin(ce->state);
 362                 i915_vma_make_unshrinkable(ce->state);
 363         }
 364
 365         return 0;
 366 }
 367
 368 static int
 369 sw_fence_dummy_notify(struct i915_sw_fence *sf,
 370                       enum i915_sw_fence_notify state)
 371 {
 372         return NOTIFY_DONE;
 373 }
 374
 375 void
 376 intel_context_init(struct intel_context *ce, struct intel_engine_cs *engine)
 377 {
 378         GEM_BUG_ON(!engine->cops);
 379         GEM_BUG_ON(!engine->gt->vm);
 380
 381         kref_init(&ce->ref);
 382
 383         ce->engine = engine;
 384         ce->ops = engine->cops;
 385         ce->sseu = engine->sseu;
 386         ce->ring = NULL;
 387         ce->ring_size = SZ_4K;
 388
 389         ewma_runtime_init(&ce->stats.runtime.avg);
 390
 391         ce->vm = i915_vm_get(engine->gt->vm);
 392
 393         /* NB ce->signal_link/lock is used under RCU */
 394         spin_lock_init(&ce->signal_lock);
 395         INIT_LIST_HEAD(&ce->signals);
 396
 397         mutex_init(&ce->pin_mutex);
 398
 399         spin_lock_init(&ce->guc_state.lock);
 400         INIT_LIST_HEAD(&ce->guc_state.fences);
 401         INIT_LIST_HEAD(&ce->guc_state.requests);
 402
 403         ce->guc_id.id = GUC_INVALID_CONTEXT_ID;
 404         INIT_LIST_HEAD(&ce->guc_id.link);
 405
 406         INIT_LIST_HEAD(&ce->destroyed_link);
 407
 408         INIT_LIST_HEAD(&ce->parallel.child_list);
 409
 410         /*
 411          * Initialize fence to be complete as this is expected to be complete
 412          * unless there is a pending schedule disable outstanding.
 413          */
 414         i915_sw_fence_init(&ce->guc_state.blocked,
 415                            sw_fence_dummy_notify);
 416         i915_sw_fence_commit(&ce->guc_state.blocked);
 417
 418         i915_active_init(&ce->active,
 419                          __intel_context_active, __intel_context_retire, 0);
 420 }
 421
 422 void intel_context_fini(struct intel_context *ce)
 423 {
 424         struct intel_context *child, *next;
 425
 426         if (ce->timeline)
 427                 intel_timeline_put(ce->timeline);
 428         i915_vm_put(ce->vm);
 429
 430         /* Need to put the creation ref for the children */
 431         if (intel_context_is_parent(ce))
 432                 for_each_child_safe(ce, child, next)
 433                         intel_context_put(child);
 434
 435         mutex_destroy(&ce->pin_mutex);
 436         i915_active_fini(&ce->active);
 437         i915_sw_fence_fini(&ce->guc_state.blocked);
 438 }
 439
 440 void i915_context_module_exit(void)
 441 {
 442         kmem_cache_destroy(slab_ce);
 443 }
 444
 445 int __init i915_context_module_init(void)
 446 {
 447         slab_ce = KMEM_CACHE(intel_context, SLAB_HWCACHE_ALIGN);
 448         if (!slab_ce)
 449                 return -ENOMEM;
 450
 451         return 0;
 452 }
 453
 454 void intel_context_enter_engine(struct intel_context *ce)
 455 {
 456         intel_engine_pm_get(ce->engine);
 457         intel_timeline_enter(ce->timeline);
 458 }
 459
 460 void intel_context_exit_engine(struct intel_context *ce)
 461 {
 462         intel_timeline_exit(ce->timeline);
 463         intel_engine_pm_put(ce->engine);
 464 }
 465
 466 int intel_context_prepare_remote_request(struct intel_context *ce,
 467                                          struct i915_request *rq)
 468 {
 469         struct intel_timeline *tl = ce->timeline;
 470         int err;
 471
 472         /* Only suitable for use in remotely modifying this context */
 473         GEM_BUG_ON(rq->context == ce);
 474
 475         if (rcu_access_pointer(rq->timeline) != tl) { /* timeline sharing! */
 476                 /* Queue this switch after current activity by this context. */
 477                 err = i915_active_fence_set(&tl->last_request, rq);
 478                 if (err)
 479                         return err;
 480         }
 481
 482         /*
 483          * Guarantee context image and the timeline remains pinned until the
 484          * modifying request is retired by setting the ce activity tracker.
 485          *
 486          * But we only need to take one pin on the account of it. Or in other
 487          * words transfer the pinned ce object to tracked active request.
 488          */
 489         GEM_BUG_ON(i915_active_is_idle(&ce->active));
 490         return i915_active_add_request(&ce->active, rq);
 491 }
 492
 493 struct i915_request *intel_context_create_request(struct intel_context *ce)
 494 {
 495         struct i915_gem_ww_ctx ww;
 496         struct i915_request *rq;
 497         int err;
 498
 499         i915_gem_ww_ctx_init(&ww, true);
 500 retry:
 501         err = intel_context_pin_ww(ce, &ww);
 502         if (!err) {
 503                 rq = i915_request_create(ce);
 504                 intel_context_unpin(ce);
 505         } else if (err == -EDEADLK) {
 506                 err = i915_gem_ww_ctx_backoff(&ww);
 507                 if (!err)
 508                         goto retry;
 509                 rq = ERR_PTR(err);
 510         } else {
 511                 rq = ERR_PTR(err);
 512         }
 513
 514         i915_gem_ww_ctx_fini(&ww);
 515
 516         if (IS_ERR(rq))
 517                 return rq;
 518
 519         /*
 520          * timeline->mutex should be the inner lock, but is used as outer lock.
 521          * Hack around this to shut up lockdep in selftests..
 522          */
 523         lockdep_unpin_lock(&ce->timeline->mutex, rq->cookie);
 524         mutex_release(&ce->timeline->mutex.dep_map, _RET_IP_);
 525         mutex_acquire(&ce->timeline->mutex.dep_map, SINGLE_DEPTH_NESTING, 0, _RET_IP_);
 526         rq->cookie = lockdep_pin_lock(&ce->timeline->mutex);
 527
 528         return rq;
 529 }
 530
 531 struct i915_request *intel_context_find_active_request(struct intel_context *ce)
 532 {
 533         struct intel_context *parent = intel_context_to_parent(ce);
 534         struct i915_request *rq, *active = NULL;
 535         unsigned long flags;
 536
 537         GEM_BUG_ON(!intel_engine_uses_guc(ce->engine));
 538
 539         /*
 540          * We search the parent list to find an active request on the submitted
 541          * context. The parent list contains the requests for all the contexts
 542          * in the relationship so we have to do a compare of each request's
 543          * context.
 544          */
 545         spin_lock_irqsave(&parent->guc_state.lock, flags);
 546         list_for_each_entry_reverse(rq, &parent->guc_state.requests,
 547                                     sched.link) {
 548                 if (rq->context != ce)
 549                         continue;
 550                 if (i915_request_completed(rq))
 551                         break;
 552
 553                 active = rq;
 554         }
 555         spin_unlock_irqrestore(&parent->guc_state.lock, flags);
 556
 557         return active;
 558 }
 559
 560 void intel_context_bind_parent_child(struct intel_context *parent,
 561                                      struct intel_context *child)
 562 {
 563         /*
 564          * Callers responsibility to validate that this function is used
 565          * correctly but we use GEM_BUG_ON here ensure that they do.
 566          */
 567         GEM_BUG_ON(intel_context_is_pinned(parent));
 568         GEM_BUG_ON(intel_context_is_child(parent));
 569         GEM_BUG_ON(intel_context_is_pinned(child));
 570         GEM_BUG_ON(intel_context_is_child(child));
 571         GEM_BUG_ON(intel_context_is_parent(child));
 572
 573         parent->parallel.child_index = parent->parallel.number_children++;
 574         list_add_tail(&child->parallel.child_link,
 575                       &parent->parallel.child_list);
 576         child->parallel.parent = parent;
 577 }
 578
 579 u64 intel_context_get_total_runtime_ns(const struct intel_context *ce)
 580 {
 581         u64 total, active;
 582
 583         total = ce->stats.runtime.total;
 584         if (ce->ops->flags & COPS_RUNTIME_CYCLES)
 585                 total *= ce->engine->gt->clock_period_ns;
 586
 587         active = READ_ONCE(ce->stats.active);
 588         if (active)
 589                 active = intel_context_clock() - active;
 590
 591         return total + active;
 592 }
 593
 594 u64 intel_context_get_avg_runtime_ns(struct intel_context *ce)
 595 {
 596         u64 avg = ewma_runtime_read(&ce->stats.runtime.avg);
 597
 598         if (ce->ops->flags & COPS_RUNTIME_CYCLES)
 599                 avg *= ce->engine->gt->clock_period_ns;
 600
 601         return avg;
 602 }
 603
 604 bool intel_context_ban(struct intel_context *ce, struct i915_request *rq)
 605 {
 606         bool ret = intel_context_set_banned(ce);
 607
 608         trace_intel_context_ban(ce);
 609
 610         if (ce->ops->revoke)
 611                 ce->ops->revoke(ce, rq,
 612                                 INTEL_CONTEXT_BANNED_PREEMPT_TIMEOUT_MS);
 613
 614         return ret;
 615 }
 616
 617 bool intel_context_exit_nonpersistent(struct intel_context *ce,
 618                                       struct i915_request *rq)
 619 {
 620         bool ret = intel_context_set_exiting(ce);
 621
 622         if (ce->ops->revoke)
 623                 ce->ops->revoke(ce, rq, ce->engine->props.preempt_timeout_ms);
 624
 625         return ret;
 626 }
 627
 628 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
 629 #include "selftest_context.c"
 630 #endif