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

   1 /*
   2  * SPDX-License-Identifier: MIT
   3  *
   4  * Copyright © 2019 Intel Corporation
   5  */
   6
   7 #include "gem/i915_gem_context.h"
   8 #include "gem/i915_gem_pm.h"
   9
  10 #include "i915_drv.h"
  11 #include "i915_globals.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 i915_global_context {
  19         struct i915_global base;
  20         struct kmem_cache *slab_ce;
  21 } global;
  22
  23 static struct intel_context *intel_context_alloc(void)
  24 {
  25         return kmem_cache_zalloc(global.slab_ce, GFP_KERNEL);
  26 }
  27
  28 void intel_context_free(struct intel_context *ce)
  29 {
  30         kmem_cache_free(global.slab_ce, ce);
  31 }
  32
  33 struct intel_context *
  34 intel_context_create(struct intel_engine_cs *engine)
  35 {
  36         struct intel_context *ce;
  37
  38         ce = intel_context_alloc();
  39         if (!ce)
  40                 return ERR_PTR(-ENOMEM);
  41
  42         intel_context_init(ce, engine);
  43         return ce;
  44 }
  45
  46 int intel_context_alloc_state(struct intel_context *ce)
  47 {
  48         int err = 0;
  49
  50         if (mutex_lock_interruptible(&ce->pin_mutex))
  51                 return -EINTR;
  52
  53         if (!test_bit(CONTEXT_ALLOC_BIT, &ce->flags)) {
  54                 if (intel_context_is_banned(ce)) {
  55                         err = -EIO;
  56                         goto unlock;
  57                 }
  58
  59                 err = ce->ops->alloc(ce);
  60                 if (unlikely(err))
  61                         goto unlock;
  62
  63                 set_bit(CONTEXT_ALLOC_BIT, &ce->flags);
  64         }
  65
  66 unlock:
  67         mutex_unlock(&ce->pin_mutex);
  68         return err;
  69 }
  70
  71 static int intel_context_active_acquire(struct intel_context *ce)
  72 {
  73         int err;
  74
  75         __i915_active_acquire(&ce->active);
  76
  77         if (intel_context_is_barrier(ce))
  78                 return 0;
  79
  80         /* Preallocate tracking nodes */
  81         err = i915_active_acquire_preallocate_barrier(&ce->active,
  82                                                       ce->engine);
  83         if (err)
  84                 i915_active_release(&ce->active);
  85
  86         return err;
  87 }
  88
  89 static void intel_context_active_release(struct intel_context *ce)
  90 {
  91         /* Nodes preallocated in intel_context_active() */
  92         i915_active_acquire_barrier(&ce->active);
  93         i915_active_release(&ce->active);
  94 }
  95
  96 static int __context_pin_state(struct i915_vma *vma, struct i915_gem_ww_ctx *ww)
  97 {
  98         unsigned int bias = i915_ggtt_pin_bias(vma) | PIN_OFFSET_BIAS;
  99         int err;
 100
 101         err = i915_ggtt_pin(vma, ww, 0, bias | PIN_HIGH);
 102         if (err)
 103                 return err;
 104
 105         err = i915_active_acquire(&vma->active);
 106         if (err)
 107                 goto err_unpin;
 108
 109         /*
 110          * And mark it as a globally pinned object to let the shrinker know
 111          * it cannot reclaim the object until we release it.
 112          */
 113         i915_vma_make_unshrinkable(vma);
 114         vma->obj->mm.dirty = true;
 115
 116         return 0;
 117
 118 err_unpin:
 119         i915_vma_unpin(vma);
 120         return err;
 121 }
 122
 123 static void __context_unpin_state(struct i915_vma *vma)
 124 {
 125         i915_vma_make_shrinkable(vma);
 126         i915_active_release(&vma->active);
 127         __i915_vma_unpin(vma);
 128 }
 129
 130 static int __ring_active(struct intel_ring *ring,
 131                          struct i915_gem_ww_ctx *ww)
 132 {
 133         int err;
 134
 135         err = intel_ring_pin(ring, ww);
 136         if (err)
 137                 return err;
 138
 139         err = i915_active_acquire(&ring->vma->active);
 140         if (err)
 141                 goto err_pin;
 142
 143         return 0;
 144
 145 err_pin:
 146         intel_ring_unpin(ring);
 147         return err;
 148 }
 149
 150 static void __ring_retire(struct intel_ring *ring)
 151 {
 152         i915_active_release(&ring->vma->active);
 153         intel_ring_unpin(ring);
 154 }
 155
 156 static int intel_context_pre_pin(struct intel_context *ce,
 157                                  struct i915_gem_ww_ctx *ww)
 158 {
 159         int err;
 160
 161         CE_TRACE(ce, "active\n");
 162
 163         err = __ring_active(ce->ring, ww);
 164         if (err)
 165                 return err;
 166
 167         err = intel_timeline_pin(ce->timeline, ww);
 168         if (err)
 169                 goto err_ring;
 170
 171         if (!ce->state)
 172                 return 0;
 173
 174         err = __context_pin_state(ce->state, ww);
 175         if (err)
 176                 goto err_timeline;
 177
 178
 179         return 0;
 180
 181 err_timeline:
 182         intel_timeline_unpin(ce->timeline);
 183 err_ring:
 184         __ring_retire(ce->ring);
 185         return err;
 186 }
 187
 188 static void intel_context_post_unpin(struct intel_context *ce)
 189 {
 190         if (ce->state)
 191                 __context_unpin_state(ce->state);
 192
 193         intel_timeline_unpin(ce->timeline);
 194         __ring_retire(ce->ring);
 195 }
 196
 197 int __intel_context_do_pin_ww(struct intel_context *ce,
 198                               struct i915_gem_ww_ctx *ww)
 199 {
 200         bool handoff = false;
 201         void *vaddr;
 202         int err = 0;
 203
 204         if (unlikely(!test_bit(CONTEXT_ALLOC_BIT, &ce->flags))) {
 205                 err = intel_context_alloc_state(ce);
 206                 if (err)
 207                         return err;
 208         }
 209
 210         /*
 211          * We always pin the context/ring/timeline here, to ensure a pin
 212          * refcount for __intel_context_active(), which prevent a lock
 213          * inversion of ce->pin_mutex vs dma_resv_lock().
 214          */
 215
 216         err = i915_gem_object_lock(ce->timeline->hwsp_ggtt->obj, ww);
 217         if (!err && ce->ring->vma->obj)
 218                 err = i915_gem_object_lock(ce->ring->vma->obj, ww);
 219         if (!err && ce->state)
 220                 err = i915_gem_object_lock(ce->state->obj, ww);
 221         if (!err)
 222                 err = intel_context_pre_pin(ce, ww);
 223         if (err)
 224                 return err;
 225
 226         err = i915_active_acquire(&ce->active);
 227         if (err)
 228                 goto err_ctx_unpin;
 229
 230         err = ce->ops->pre_pin(ce, ww, &vaddr);
 231         if (err)
 232                 goto err_release;
 233
 234         err = mutex_lock_interruptible(&ce->pin_mutex);
 235         if (err)
 236                 goto err_post_unpin;
 237
 238         if (unlikely(intel_context_is_closed(ce))) {
 239                 err = -ENOENT;
 240                 goto err_unlock;
 241         }
 242
 243         if (likely(!atomic_add_unless(&ce->pin_count, 1, 0))) {
 244                 err = intel_context_active_acquire(ce);
 245                 if (unlikely(err))
 246                         goto err_unlock;
 247
 248                 err = ce->ops->pin(ce, vaddr);
 249                 if (err) {
 250                         intel_context_active_release(ce);
 251                         goto err_unlock;
 252                 }
 253
 254                 CE_TRACE(ce, "pin ring:{start:%08x, head:%04x, tail:%04x}\n",
 255                          i915_ggtt_offset(ce->ring->vma),
 256                          ce->ring->head, ce->ring->tail);
 257
 258                 handoff = true;
 259                 smp_mb__before_atomic(); /* flush pin before it is visible */
 260                 atomic_inc(&ce->pin_count);
 261         }
 262
 263         GEM_BUG_ON(!intel_context_is_pinned(ce)); /* no overflow! */
 264
 265 err_unlock:
 266         mutex_unlock(&ce->pin_mutex);
 267 err_post_unpin:
 268         if (!handoff)
 269                 ce->ops->post_unpin(ce);
 270 err_release:
 271         i915_active_release(&ce->active);
 272 err_ctx_unpin:
 273         intel_context_post_unpin(ce);
 274
 275         /*
 276          * Unlock the hwsp_ggtt object since it's shared.
 277          * In principle we can unlock all the global state locked above
 278          * since it's pinned and doesn't need fencing, and will
 279          * thus remain resident until it is explicitly unpinned.
 280          */
 281         i915_gem_ww_unlock_single(ce->timeline->hwsp_ggtt->obj);
 282
 283         return err;
 284 }
 285
 286 int __intel_context_do_pin(struct intel_context *ce)
 287 {
 288         struct i915_gem_ww_ctx ww;
 289         int err;
 290
 291         i915_gem_ww_ctx_init(&ww, true);
 292 retry:
 293         err = __intel_context_do_pin_ww(ce, &ww);
 294         if (err == -EDEADLK) {
 295                 err = i915_gem_ww_ctx_backoff(&ww);
 296                 if (!err)
 297                         goto retry;
 298         }
 299         i915_gem_ww_ctx_fini(&ww);
 300         return err;
 301 }
 302
 303 void intel_context_unpin(struct intel_context *ce)
 304 {
 305         if (!atomic_dec_and_test(&ce->pin_count))
 306                 return;
 307
 308         CE_TRACE(ce, "unpin\n");
 309         ce->ops->unpin(ce);
 310         ce->ops->post_unpin(ce);
 311
 312         /*
 313          * Once released, we may asynchronously drop the active reference.
 314          * As that may be the only reference keeping the context alive,
 315          * take an extra now so that it is not freed before we finish
 316          * dereferencing it.
 317          */
 318         intel_context_get(ce);
 319         intel_context_active_release(ce);
 320         intel_context_put(ce);
 321 }
 322
 323 __i915_active_call
 324 static void __intel_context_retire(struct i915_active *active)
 325 {
 326         struct intel_context *ce = container_of(active, typeof(*ce), active);
 327
 328         CE_TRACE(ce, "retire runtime: { total:%lluns, avg:%lluns }\n",
 329                  intel_context_get_total_runtime_ns(ce),
 330                  intel_context_get_avg_runtime_ns(ce));
 331
 332         set_bit(CONTEXT_VALID_BIT, &ce->flags);
 333         intel_context_post_unpin(ce);
 334         intel_context_put(ce);
 335 }
 336
 337 static int __intel_context_active(struct i915_active *active)
 338 {
 339         struct intel_context *ce = container_of(active, typeof(*ce), active);
 340
 341         intel_context_get(ce);
 342
 343         /* everything should already be activated by intel_context_pre_pin() */
 344         GEM_WARN_ON(!i915_active_acquire_if_busy(&ce->ring->vma->active));
 345         __intel_ring_pin(ce->ring);
 346
 347         __intel_timeline_pin(ce->timeline);
 348
 349         if (ce->state) {
 350                 GEM_WARN_ON(!i915_active_acquire_if_busy(&ce->state->active));
 351                 __i915_vma_pin(ce->state);
 352                 i915_vma_make_unshrinkable(ce->state);
 353         }
 354
 355         return 0;
 356 }
 357
 358 void
 359 intel_context_init(struct intel_context *ce,
 360                    struct intel_engine_cs *engine)
 361 {
 362         GEM_BUG_ON(!engine->cops);
 363         GEM_BUG_ON(!engine->gt->vm);
 364
 365         kref_init(&ce->ref);
 366
 367         ce->engine = engine;
 368         ce->ops = engine->cops;
 369         ce->sseu = engine->sseu;
 370         ce->ring = __intel_context_ring_size(SZ_4K);
 371
 372         ewma_runtime_init(&ce->runtime.avg);
 373
 374         ce->vm = i915_vm_get(engine->gt->vm);
 375
 376         INIT_LIST_HEAD(&ce->signal_link);
 377         INIT_LIST_HEAD(&ce->signals);
 378
 379         mutex_init(&ce->pin_mutex);
 380
 381         i915_active_init(&ce->active,
 382                          __intel_context_active, __intel_context_retire);
 383 }
 384
 385 void intel_context_fini(struct intel_context *ce)
 386 {
 387         if (ce->timeline)
 388                 intel_timeline_put(ce->timeline);
 389         i915_vm_put(ce->vm);
 390
 391         mutex_destroy(&ce->pin_mutex);
 392         i915_active_fini(&ce->active);
 393 }
 394
 395 static void i915_global_context_shrink(void)
 396 {
 397         kmem_cache_shrink(global.slab_ce);
 398 }
 399
 400 static void i915_global_context_exit(void)
 401 {
 402         kmem_cache_destroy(global.slab_ce);
 403 }
 404
 405 static struct i915_global_context global = { {
 406         .shrink = i915_global_context_shrink,
 407         .exit = i915_global_context_exit,
 408 } };
 409
 410 int __init i915_global_context_init(void)
 411 {
 412         global.slab_ce = KMEM_CACHE(intel_context, SLAB_HWCACHE_ALIGN);
 413         if (!global.slab_ce)
 414                 return -ENOMEM;
 415
 416         i915_global_register(&global.base);
 417         return 0;
 418 }
 419
 420 void intel_context_enter_engine(struct intel_context *ce)
 421 {
 422         intel_engine_pm_get(ce->engine);
 423         intel_timeline_enter(ce->timeline);
 424 }
 425
 426 void intel_context_exit_engine(struct intel_context *ce)
 427 {
 428         intel_timeline_exit(ce->timeline);
 429         intel_engine_pm_put(ce->engine);
 430 }
 431
 432 int intel_context_prepare_remote_request(struct intel_context *ce,
 433                                          struct i915_request *rq)
 434 {
 435         struct intel_timeline *tl = ce->timeline;
 436         int err;
 437
 438         /* Only suitable for use in remotely modifying this context */
 439         GEM_BUG_ON(rq->context == ce);
 440
 441         if (rcu_access_pointer(rq->timeline) != tl) { /* timeline sharing! */
 442                 /* Queue this switch after current activity by this context. */
 443                 err = i915_active_fence_set(&tl->last_request, rq);
 444                 if (err)
 445                         return err;
 446         }
 447
 448         /*
 449          * Guarantee context image and the timeline remains pinned until the
 450          * modifying request is retired by setting the ce activity tracker.
 451          *
 452          * But we only need to take one pin on the account of it. Or in other
 453          * words transfer the pinned ce object to tracked active request.
 454          */
 455         GEM_BUG_ON(i915_active_is_idle(&ce->active));
 456         return i915_active_add_request(&ce->active, rq);
 457 }
 458
 459 struct i915_request *intel_context_create_request(struct intel_context *ce)
 460 {
 461         struct i915_gem_ww_ctx ww;
 462         struct i915_request *rq;
 463         int err;
 464
 465         i915_gem_ww_ctx_init(&ww, true);
 466 retry:
 467         err = intel_context_pin_ww(ce, &ww);
 468         if (!err) {
 469                 rq = i915_request_create(ce);
 470                 intel_context_unpin(ce);
 471         } else if (err == -EDEADLK) {
 472                 err = i915_gem_ww_ctx_backoff(&ww);
 473                 if (!err)
 474                         goto retry;
 475                 rq = ERR_PTR(err);
 476         } else {
 477                 rq = ERR_PTR(err);
 478         }
 479
 480         i915_gem_ww_ctx_fini(&ww);
 481
 482         if (IS_ERR(rq))
 483                 return rq;
 484
 485         /*
 486          * timeline->mutex should be the inner lock, but is used as outer lock.
 487          * Hack around this to shut up lockdep in selftests..
 488          */
 489         lockdep_unpin_lock(&ce->timeline->mutex, rq->cookie);
 490         mutex_release(&ce->timeline->mutex.dep_map, _RET_IP_);
 491         mutex_acquire(&ce->timeline->mutex.dep_map, SINGLE_DEPTH_NESTING, 0, _RET_IP_);
 492         rq->cookie = lockdep_pin_lock(&ce->timeline->mutex);
 493
 494         return rq;
 495 }
 496
 497 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
 498 #include "selftest_context.c"
 499 #endif