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_drm_client.h"
  10 #include "i915_drv.h"
  11 #include "i915_trace.h"
  12
  13 #include "intel_context.h"
  14 #include "intel_engine.h"
  15 #include "intel_engine_pm.h"
  16 #include "intel_ring.h"
  17
  18 static struct kmem_cache *slab_ce;
  19
  20 static struct intel_context *intel_context_alloc(void)
  21 {
  22         return kmem_cache_zalloc(slab_ce, GFP_KERNEL);
  23 }
  24
  25 static void rcu_context_free(struct rcu_head *rcu)
  26 {
  27         struct intel_context *ce = container_of(rcu, typeof(*ce), rcu);
  28
  29         trace_intel_context_free(ce);
  30         kmem_cache_free(slab_ce, ce);
  31 }
  32
  33 void intel_context_free(struct intel_context *ce)
  34 {
  35         call_rcu(&ce->rcu, rcu_context_free);
  36 }
  37
  38 struct intel_context *
  39 intel_context_create(struct intel_engine_cs *engine)
  40 {
  41         struct intel_context *ce;
  42
  43         ce = intel_context_alloc();
  44         if (!ce)
  45                 return ERR_PTR(-ENOMEM);
  46
  47         intel_context_init(ce, engine);
  48         trace_intel_context_create(ce);
  49         return ce;
  50 }
  51
  52 int intel_context_alloc_state(struct intel_context *ce)
  53 {
  54         struct i915_gem_context *ctx;
  55         int err = 0;
  56
  57         if (mutex_lock_interruptible(&ce->pin_mutex))
  58                 return -EINTR;
  59
  60         if (!test_bit(CONTEXT_ALLOC_BIT, &ce->flags)) {
  61                 if (intel_context_is_banned(ce)) {
  62                         err = -EIO;
  63                         goto unlock;
  64                 }
  65
  66                 err = ce->ops->alloc(ce);
  67                 if (unlikely(err))
  68                         goto unlock;
  69
  70                 set_bit(CONTEXT_ALLOC_BIT, &ce->flags);
  71
  72                 rcu_read_lock();
  73                 ctx = rcu_dereference(ce->gem_context);
  74                 if (ctx && !kref_get_unless_zero(&ctx->ref))
  75                         ctx = NULL;
  76                 rcu_read_unlock();
  77                 if (ctx) {
  78                         if (ctx->client)
  79                                 i915_drm_client_add_context_objects(ctx->client,
  80                                                                     ce);
  81                         i915_gem_context_put(ctx);
  82                 }
  83         }
  84
  85 unlock:
  86         mutex_unlock(&ce->pin_mutex);
  87         return err;
  88 }
  89
  90 static int intel_context_active_acquire(struct intel_context *ce)
  91 {
  92         int err;
  93
  94         __i915_active_acquire(&ce->active);
  95
  96         if (intel_context_is_barrier(ce) || intel_engine_uses_guc(ce->engine) ||
  97             intel_context_is_parallel(ce))
  98                 return 0;
  99
 100         /* Preallocate tracking nodes */
 101         err = i915_active_acquire_preallocate_barrier(&ce->active,
 102                                                       ce->engine);
 103         if (err)
 104                 i915_active_release(&ce->active);
 105
 106         return err;
 107 }
 108
 109 static void intel_context_active_release(struct intel_context *ce)
 110 {
 111         /* Nodes preallocated in intel_context_active() */
 112         i915_active_acquire_barrier(&ce->active);
 113         i915_active_release(&ce->active);
 114 }
 115
 116 static int __context_pin_state(struct i915_vma *vma, struct i915_gem_ww_ctx *ww)
 117 {
 118         unsigned int bias = i915_ggtt_pin_bias(vma) | PIN_OFFSET_BIAS;
 119         int err;
 120
 121         err = i915_ggtt_pin(vma, ww, 0, bias | PIN_HIGH);
 122         if (err)
 123                 return err;
 124
 125         err = i915_active_acquire(&vma->active);
 126         if (err)
 127                 goto err_unpin;
 128
 129         /*
 130          * And mark it as a globally pinned object to let the shrinker know
 131          * it cannot reclaim the object until we release it.
 132          */
 133         i915_vma_make_unshrinkable(vma);
 134         vma->obj->mm.dirty = true;
 135
 136         return 0;
 137
 138 err_unpin:
 139         i915_vma_unpin(vma);
 140         return err;
 141 }
 142
 143 static void __context_unpin_state(struct i915_vma *vma)
 144 {
 145         i915_vma_make_shrinkable(vma);
 146         i915_active_release(&vma->active);
 147         __i915_vma_unpin(vma);
 148 }
 149
 150 static int __ring_active(struct intel_ring *ring,
 151                          struct i915_gem_ww_ctx *ww)
 152 {
 153         int err;
 154
 155         err = intel_ring_pin(ring, ww);
 156         if (err)
 157                 return err;
 158
 159         err = i915_active_acquire(&ring->vma->active);
 160         if (err)
 161                 goto err_pin;
 162
 163         return 0;
 164
 165 err_pin:
 166         intel_ring_unpin(ring);
 167         return err;
 168 }
 169
 170 static void __ring_retire(struct intel_ring *ring)
 171 {
 172         i915_active_release(&ring->vma->active);
 173         intel_ring_unpin(ring);
 174 }
 175
 176 static int intel_context_pre_pin(struct intel_context *ce,
 177                                  struct i915_gem_ww_ctx *ww)
 178 {
 179         int err;
 180
 181         CE_TRACE(ce, "active\n");
 182
 183         err = __ring_active(ce->ring, ww);
 184         if (err)
 185                 return err;
 186
 187         err = intel_timeline_pin(ce->timeline, ww);
 188         if (err)
 189                 goto err_ring;
 190
 191         if (!ce->state)
 192                 return 0;
 193
 194         err = __context_pin_state(ce->state, ww);
 195         if (err)
 196                 goto err_timeline;
 197
 198
 199         return 0;
 200
 201 err_timeline:
 202         intel_timeline_unpin(ce->timeline);
 203 err_ring:
 204         __ring_retire(ce->ring);
 205         return err;
 206 }
 207
 208 static void intel_context_post_unpin(struct intel_context *ce)
 209 {
 210         if (ce->state)
 211                 __context_unpin_state(ce->state);
 212
 213         intel_timeline_unpin(ce->timeline);
 214         __ring_retire(ce->ring);
 215 }
 216
 217 int __intel_context_do_pin_ww(struct intel_context *ce,
 218                               struct i915_gem_ww_ctx *ww)
 219 {
 220         bool handoff = false;
 221         void *vaddr;
 222         int err = 0;
 223
 224         if (unlikely(!test_bit(CONTEXT_ALLOC_BIT, &ce->flags))) {
 225                 err = intel_context_alloc_state(ce);
 226                 if (err)
 227                         return err;
 228         }
 229
 230         /*
 231          * We always pin the context/ring/timeline here, to ensure a pin
 232          * refcount for __intel_context_active(), which prevent a lock
 233          * inversion of ce->pin_mutex vs dma_resv_lock().
 234          */
 235
 236         err = i915_gem_object_lock(ce->timeline->hwsp_ggtt->obj, ww);
 237         if (!err)
 238                 err = i915_gem_object_lock(ce->ring->vma->obj, ww);
 239         if (!err && ce->state)
 240                 err = i915_gem_object_lock(ce->state->obj, ww);
 241         if (!err)
 242                 err = intel_context_pre_pin(ce, ww);
 243         if (err)
 244                 return err;
 245
 246         err = ce->ops->pre_pin(ce, ww, &vaddr);
 247         if (err)
 248                 goto err_ctx_unpin;
 249
 250         err = i915_active_acquire(&ce->active);
 251         if (err)
 252                 goto err_post_unpin;
 253
 254         err = mutex_lock_interruptible(&ce->pin_mutex);
 255         if (err)
 256                 goto err_release;
 257
 258         intel_engine_pm_might_get(ce->engine);
 259
 260         if (unlikely(intel_context_is_closed(ce))) {
 261                 err = -ENOENT;
 262                 goto err_unlock;
 263         }
 264
 265         if (likely(!atomic_add_unless(&ce->pin_count, 1, 0))) {
 266                 err = intel_context_active_acquire(ce);
 267                 if (unlikely(err))
 268                         goto err_unlock;
 269
 270                 err = ce->ops->pin(ce, vaddr);
 271                 if (err) {
 272                         intel_context_active_release(ce);
 273                         goto err_unlock;
 274                 }
 275
 276                 CE_TRACE(ce, "pin ring:{start:%08x, head:%04x, tail:%04x}\n",
 277                          i915_ggtt_offset(ce->ring->vma),
 278                          ce->ring->head, ce->ring->tail);
 279
 280                 handoff = true;
 281                 smp_mb__before_atomic(); /* flush pin before it is visible */
 282                 atomic_inc(&ce->pin_count);
 283         }
 284
 285         GEM_BUG_ON(!intel_context_is_pinned(ce)); /* no overflow! */
 286
 287         trace_intel_context_do_pin(ce);
 288
 289 err_unlock:
 290         mutex_unlock(&ce->pin_mutex);
 291 err_release:
 292         i915_active_release(&ce->active);
 293 err_post_unpin:
 294         if (!handoff)
 295                 ce->ops->post_unpin(ce);
 296 err_ctx_unpin:
 297         intel_context_post_unpin(ce);
 298
 299         /*
 300          * Unlock the hwsp_ggtt object since it's shared.
 301          * In principle we can unlock all the global state locked above
 302          * since it's pinned and doesn't need fencing, and will
 303          * thus remain resident until it is explicitly unpinned.
 304          */
 305         i915_gem_ww_unlock_single(ce->timeline->hwsp_ggtt->obj);
 306
 307         return err;
 308 }
 309
 310 int __intel_context_do_pin(struct intel_context *ce)
 311 {
 312         struct i915_gem_ww_ctx ww;
 313         int err;
 314
 315         i915_gem_ww_ctx_init(&ww, true);
 316 retry:
 317         err = __intel_context_do_pin_ww(ce, &ww);
 318         if (err == -EDEADLK) {
 319                 err = i915_gem_ww_ctx_backoff(&ww);
 320                 if (!err)
 321                         goto retry;
 322         }
 323         i915_gem_ww_ctx_fini(&ww);
 324         return err;
 325 }
 326
 327 void __intel_context_do_unpin(struct intel_context *ce, int sub)
 328 {
 329         if (!atomic_sub_and_test(sub, &ce->pin_count))
 330                 return;
 331
 332         CE_TRACE(ce, "unpin\n");
 333         ce->ops->unpin(ce);
 334         ce->ops->post_unpin(ce);
 335
 336         /*
 337          * Once released, we may asynchronously drop the active reference.
 338          * As that may be the only reference keeping the context alive,
 339          * take an extra now so that it is not freed before we finish
 340          * dereferencing it.
 341          */
 342         intel_context_get(ce);
 343         intel_context_active_release(ce);
 344         trace_intel_context_do_unpin(ce);
 345         intel_context_put(ce);
 346 }
 347
 348 static void __intel_context_retire(struct i915_active *active)
 349 {
 350         struct intel_context *ce = container_of(active, typeof(*ce), active);
 351
 352         CE_TRACE(ce, "retire runtime: { total:%lluns, avg:%lluns }\n",
 353                  intel_context_get_total_runtime_ns(ce),
 354                  intel_context_get_avg_runtime_ns(ce));
 355
 356         set_bit(CONTEXT_VALID_BIT, &ce->flags);
 357         intel_context_post_unpin(ce);
 358         intel_context_put(ce);
 359 }
 360
 361 static int __intel_context_active(struct i915_active *active)
 362 {
 363         struct intel_context *ce = container_of(active, typeof(*ce), active);
 364
 365         intel_context_get(ce);
 366
 367         /* everything should already be activated by intel_context_pre_pin() */
 368         GEM_WARN_ON(!i915_active_acquire_if_busy(&ce->ring->vma->active));
 369         __intel_ring_pin(ce->ring);
 370
 371         __intel_timeline_pin(ce->timeline);
 372
 373         if (ce->state) {
 374                 GEM_WARN_ON(!i915_active_acquire_if_busy(&ce->state->active));
 375                 __i915_vma_pin(ce->state);
 376                 i915_vma_make_unshrinkable(ce->state);
 377         }
 378
 379         return 0;
 380 }
 381
 382 static int
 383 sw_fence_dummy_notify(struct i915_sw_fence *sf,
 384                       enum i915_sw_fence_notify state)
 385 {
 386         return NOTIFY_DONE;
 387 }
 388
 389 void
 390 intel_context_init(struct intel_context *ce, struct intel_engine_cs *engine)
 391 {
 392         GEM_BUG_ON(!engine->cops);
 393         GEM_BUG_ON(!engine->gt->vm);
 394
 395         kref_init(&ce->ref);
 396
 397         ce->engine = engine;
 398         ce->ops = engine->cops;
 399         ce->sseu = engine->sseu;
 400         ce->ring = NULL;
 401         ce->ring_size = SZ_4K;
 402
 403         ewma_runtime_init(&ce->stats.runtime.avg);
 404
 405         ce->vm = i915_vm_get(engine->gt->vm);
 406
 407         /* NB ce->signal_link/lock is used under RCU */
 408         spin_lock_init(&ce->signal_lock);
 409         INIT_LIST_HEAD(&ce->signals);
 410
 411         mutex_init(&ce->pin_mutex);
 412
 413         spin_lock_init(&ce->guc_state.lock);
 414         INIT_LIST_HEAD(&ce->guc_state.fences);
 415         INIT_LIST_HEAD(&ce->guc_state.requests);
 416
 417         ce->guc_id.id = GUC_INVALID_CONTEXT_ID;
 418         INIT_LIST_HEAD(&ce->guc_id.link);
 419
 420         INIT_LIST_HEAD(&ce->destroyed_link);
 421
 422         INIT_LIST_HEAD(&ce->parallel.child_list);
 423
 424         /*
 425          * Initialize fence to be complete as this is expected to be complete
 426          * unless there is a pending schedule disable outstanding.
 427          */
 428         i915_sw_fence_init(&ce->guc_state.blocked,
 429                            sw_fence_dummy_notify);
 430         i915_sw_fence_commit(&ce->guc_state.blocked);
 431
 432         i915_active_init(&ce->active,
 433                          __intel_context_active, __intel_context_retire, 0);
 434 }
 435
 436 void intel_context_fini(struct intel_context *ce)
 437 {
 438         struct intel_context *child, *next;
 439
 440         if (ce->timeline)
 441                 intel_timeline_put(ce->timeline);
 442         i915_vm_put(ce->vm);
 443
 444         /* Need to put the creation ref for the children */
 445         if (intel_context_is_parent(ce))
 446                 for_each_child_safe(ce, child, next)
 447                         intel_context_put(child);
 448
 449         mutex_destroy(&ce->pin_mutex);
 450         i915_active_fini(&ce->active);
 451         i915_sw_fence_fini(&ce->guc_state.blocked);
 452 }
 453
 454 void i915_context_module_exit(void)
 455 {
 456         kmem_cache_destroy(slab_ce);
 457 }
 458
 459 int __init i915_context_module_init(void)
 460 {
 461         slab_ce = KMEM_CACHE(intel_context, SLAB_HWCACHE_ALIGN);
 462         if (!slab_ce)
 463                 return -ENOMEM;
 464
 465         return 0;
 466 }
 467
 468 void intel_context_enter_engine(struct intel_context *ce)
 469 {
 470         intel_engine_pm_get(ce->engine);
 471         intel_timeline_enter(ce->timeline);
 472 }
 473
 474 void intel_context_exit_engine(struct intel_context *ce)
 475 {
 476         intel_timeline_exit(ce->timeline);
 477         intel_engine_pm_put(ce->engine);
 478 }
 479
 480 int intel_context_prepare_remote_request(struct intel_context *ce,
 481                                          struct i915_request *rq)
 482 {
 483         struct intel_timeline *tl = ce->timeline;
 484         int err;
 485
 486         /* Only suitable for use in remotely modifying this context */
 487         GEM_BUG_ON(rq->context == ce);
 488
 489         if (rcu_access_pointer(rq->timeline) != tl) { /* timeline sharing! */
 490                 /* Queue this switch after current activity by this context. */
 491                 err = i915_active_fence_set(&tl->last_request, rq);
 492                 if (err)
 493                         return err;
 494         }
 495
 496         /*
 497          * Guarantee context image and the timeline remains pinned until the
 498          * modifying request is retired by setting the ce activity tracker.
 499          *
 500          * But we only need to take one pin on the account of it. Or in other
 501          * words transfer the pinned ce object to tracked active request.
 502          */
 503         GEM_BUG_ON(i915_active_is_idle(&ce->active));
 504         return i915_active_add_request(&ce->active, rq);
 505 }
 506
 507 struct i915_request *intel_context_create_request(struct intel_context *ce)
 508 {
 509         struct i915_gem_ww_ctx ww;
 510         struct i915_request *rq;
 511         int err;
 512
 513         i915_gem_ww_ctx_init(&ww, true);
 514 retry:
 515         err = intel_context_pin_ww(ce, &ww);
 516         if (!err) {
 517                 rq = i915_request_create(ce);
 518                 intel_context_unpin(ce);
 519         } else if (err == -EDEADLK) {
 520                 err = i915_gem_ww_ctx_backoff(&ww);
 521                 if (!err)
 522                         goto retry;
 523                 rq = ERR_PTR(err);
 524         } else {
 525                 rq = ERR_PTR(err);
 526         }
 527
 528         i915_gem_ww_ctx_fini(&ww);
 529
 530         if (IS_ERR(rq))
 531                 return rq;
 532
 533         /*
 534          * timeline->mutex should be the inner lock, but is used as outer lock.
 535          * Hack around this to shut up lockdep in selftests..
 536          */
 537         lockdep_unpin_lock(&ce->timeline->mutex, rq->cookie);
 538         mutex_release(&ce->timeline->mutex.dep_map, _RET_IP_);
 539         mutex_acquire(&ce->timeline->mutex.dep_map, SINGLE_DEPTH_NESTING, 0, _RET_IP_);
 540         rq->cookie = lockdep_pin_lock(&ce->timeline->mutex);
 541
 542         return rq;
 543 }
 544
 545 struct i915_request *intel_context_get_active_request(struct intel_context *ce)
 546 {
 547         struct intel_context *parent = intel_context_to_parent(ce);
 548         struct i915_request *rq, *active = NULL;
 549         unsigned long flags;
 550
 551         GEM_BUG_ON(!intel_engine_uses_guc(ce->engine));
 552
 553         /*
 554          * We search the parent list to find an active request on the submitted
 555          * context. The parent list contains the requests for all the contexts
 556          * in the relationship so we have to do a compare of each request's
 557          * context.
 558          */
 559         spin_lock_irqsave(&parent->guc_state.lock, flags);
 560         list_for_each_entry_reverse(rq, &parent->guc_state.requests,
 561                                     sched.link) {
 562                 if (rq->context != ce)
 563                         continue;
 564                 if (i915_request_completed(rq))
 565                         break;
 566
 567                 active = rq;
 568         }
 569         if (active)
 570                 active = i915_request_get_rcu(active);
 571         spin_unlock_irqrestore(&parent->guc_state.lock, flags);
 572
 573         return active;
 574 }
 575
 576 void intel_context_bind_parent_child(struct intel_context *parent,
 577                                      struct intel_context *child)
 578 {
 579         /*
 580          * Callers responsibility to validate that this function is used
 581          * correctly but we use GEM_BUG_ON here ensure that they do.
 582          */
 583         GEM_BUG_ON(intel_context_is_pinned(parent));
 584         GEM_BUG_ON(intel_context_is_child(parent));
 585         GEM_BUG_ON(intel_context_is_pinned(child));
 586         GEM_BUG_ON(intel_context_is_child(child));
 587         GEM_BUG_ON(intel_context_is_parent(child));
 588
 589         parent->parallel.child_index = parent->parallel.number_children++;
 590         list_add_tail(&child->parallel.child_link,
 591                       &parent->parallel.child_list);
 592         child->parallel.parent = parent;
 593 }
 594
 595 u64 intel_context_get_total_runtime_ns(struct intel_context *ce)
 596 {
 597         u64 total, active;
 598
 599         if (ce->ops->update_stats)
 600                 ce->ops->update_stats(ce);
 601
 602         total = ce->stats.runtime.total;
 603         if (ce->ops->flags & COPS_RUNTIME_CYCLES)
 604                 total *= ce->engine->gt->clock_period_ns;
 605
 606         active = READ_ONCE(ce->stats.active);
 607         if (active)
 608                 active = intel_context_clock() - active;
 609
 610         return total + active;
 611 }
 612
 613 u64 intel_context_get_avg_runtime_ns(struct intel_context *ce)
 614 {
 615         u64 avg = ewma_runtime_read(&ce->stats.runtime.avg);
 616
 617         if (ce->ops->flags & COPS_RUNTIME_CYCLES)
 618                 avg *= ce->engine->gt->clock_period_ns;
 619
 620         return avg;
 621 }
 622
 623 bool intel_context_ban(struct intel_context *ce, struct i915_request *rq)
 624 {
 625         bool ret = intel_context_set_banned(ce);
 626
 627         trace_intel_context_ban(ce);
 628
 629         if (ce->ops->revoke)
 630                 ce->ops->revoke(ce, rq,
 631                                 INTEL_CONTEXT_BANNED_PREEMPT_TIMEOUT_MS);
 632
 633         return ret;
 634 }
 635
 636 bool intel_context_revoke(struct intel_context *ce)
 637 {
 638         bool ret = intel_context_set_exiting(ce);
 639
 640         if (ce->ops->revoke)
 641                 ce->ops->revoke(ce, NULL, ce->engine->props.preempt_timeout_ms);
 642
 643         return ret;
 644 }
 645
 646 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
 647 #include "selftest_context.c"
 648 #endif