drivers/gpu/drm/i915/intel_breadcrumbs.c

   1 /*
   2  * Copyright © 2015 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  */
  24
  25 #include <linux/kthread.h>
  26
  27 #include "i915_drv.h"
  28
  29 static void intel_breadcrumbs_fake_irq(unsigned long data)
  30 {
  31         struct intel_engine_cs *engine = (struct intel_engine_cs *)data;
  32
  33         /*
  34          * The timer persists in case we cannot enable interrupts,
  35          * or if we have previously seen seqno/interrupt incoherency
  36          * ("missed interrupt" syndrome). Here the worker will wake up
  37          * every jiffie in order to kick the oldest waiter to do the
  38          * coherent seqno check.
  39          */
  40         rcu_read_lock();
  41         if (intel_engine_wakeup(engine))
  42                 mod_timer(&engine->breadcrumbs.fake_irq, jiffies + 1);
  43         rcu_read_unlock();
  44 }
  45
  46 static void irq_enable(struct intel_engine_cs *engine)
  47 {
  48         /* Enabling the IRQ may miss the generation of the interrupt, but
  49          * we still need to force the barrier before reading the seqno,
  50          * just in case.
  51          */
  52         engine->breadcrumbs.irq_posted = true;
  53
  54         spin_lock_irq(&engine->i915->irq_lock);
  55         engine->irq_enable(engine);
  56         spin_unlock_irq(&engine->i915->irq_lock);
  57 }
  58
  59 static void irq_disable(struct intel_engine_cs *engine)
  60 {
  61         spin_lock_irq(&engine->i915->irq_lock);
  62         engine->irq_disable(engine);
  63         spin_unlock_irq(&engine->i915->irq_lock);
  64
  65         engine->breadcrumbs.irq_posted = false;
  66 }
  67
  68 static void __intel_breadcrumbs_enable_irq(struct intel_breadcrumbs *b)
  69 {
  70         struct intel_engine_cs *engine =
  71                 container_of(b, struct intel_engine_cs, breadcrumbs);
  72         struct drm_i915_private *i915 = engine->i915;
  73
  74         assert_spin_locked(&b->lock);
  75         if (b->rpm_wakelock)
  76                 return;
  77
  78         /* Since we are waiting on a request, the GPU should be busy
  79          * and should have its own rpm reference. For completeness,
  80          * record an rpm reference for ourselves to cover the
  81          * interrupt we unmask.
  82          */
  83         intel_runtime_pm_get_noresume(i915);
  84         b->rpm_wakelock = true;
  85
  86         /* No interrupts? Kick the waiter every jiffie! */
  87         if (intel_irqs_enabled(i915)) {
  88                 if (!test_bit(engine->id, &i915->gpu_error.test_irq_rings))
  89                         irq_enable(engine);
  90                 b->irq_enabled = true;
  91         }
  92
  93         if (!b->irq_enabled ||
  94             test_bit(engine->id, &i915->gpu_error.missed_irq_rings))
  95                 mod_timer(&b->fake_irq, jiffies + 1);
  96
  97         /* Ensure that even if the GPU hangs, we get woken up.
  98          *
  99          * However, note that if no one is waiting, we never notice
 100          * a gpu hang. Eventually, we will have to wait for a resource
 101          * held by the GPU and so trigger a hangcheck. In the most
 102          * pathological case, this will be upon memory starvation!
 103          */
 104         i915_queue_hangcheck(i915);
 105 }
 106
 107 static void __intel_breadcrumbs_disable_irq(struct intel_breadcrumbs *b)
 108 {
 109         struct intel_engine_cs *engine =
 110                 container_of(b, struct intel_engine_cs, breadcrumbs);
 111
 112         assert_spin_locked(&b->lock);
 113         if (!b->rpm_wakelock)
 114                 return;
 115
 116         if (b->irq_enabled) {
 117                 irq_disable(engine);
 118                 b->irq_enabled = false;
 119         }
 120
 121         intel_runtime_pm_put(engine->i915);
 122         b->rpm_wakelock = false;
 123 }
 124
 125 static inline struct intel_wait *to_wait(struct rb_node *node)
 126 {
 127         return container_of(node, struct intel_wait, node);
 128 }
 129
 130 static inline void __intel_breadcrumbs_finish(struct intel_breadcrumbs *b,
 131                                               struct intel_wait *wait)
 132 {
 133         assert_spin_locked(&b->lock);
 134
 135         /* This request is completed, so remove it from the tree, mark it as
 136          * complete, and *then* wake up the associated task.
 137          */
 138         rb_erase(&wait->node, &b->waiters);
 139         RB_CLEAR_NODE(&wait->node);
 140
 141         wake_up_process(wait->tsk); /* implicit smp_wmb() */
 142 }
 143
 144 static bool __intel_engine_add_wait(struct intel_engine_cs *engine,
 145                                     struct intel_wait *wait)
 146 {
 147         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 148         struct rb_node **p, *parent, *completed;
 149         bool first;
 150         u32 seqno;
 151
 152         /* Insert the request into the retirement ordered list
 153          * of waiters by walking the rbtree. If we are the oldest
 154          * seqno in the tree (the first to be retired), then
 155          * set ourselves as the bottom-half.
 156          *
 157          * As we descend the tree, prune completed branches since we hold the
 158          * spinlock we know that the first_waiter must be delayed and can
 159          * reduce some of the sequential wake up latency if we take action
 160          * ourselves and wake up the completed tasks in parallel. Also, by
 161          * removing stale elements in the tree, we may be able to reduce the
 162          * ping-pong between the old bottom-half and ourselves as first-waiter.
 163          */
 164         first = true;
 165         parent = NULL;
 166         completed = NULL;
 167         seqno = intel_engine_get_seqno(engine);
 168
 169          /* If the request completed before we managed to grab the spinlock,
 170           * return now before adding ourselves to the rbtree. We let the
 171           * current bottom-half handle any pending wakeups and instead
 172           * try and get out of the way quickly.
 173           */
 174         if (i915_seqno_passed(seqno, wait->seqno)) {
 175                 RB_CLEAR_NODE(&wait->node);
 176                 return first;
 177         }
 178
 179         p = &b->waiters.rb_node;
 180         while (*p) {
 181                 parent = *p;
 182                 if (wait->seqno == to_wait(parent)->seqno) {
 183                         /* We have multiple waiters on the same seqno, select
 184                          * the highest priority task (that with the smallest
 185                          * task->prio) to serve as the bottom-half for this
 186                          * group.
 187                          */
 188                         if (wait->tsk->prio > to_wait(parent)->tsk->prio) {
 189                                 p = &parent->rb_right;
 190                                 first = false;
 191                         } else {
 192                                 p = &parent->rb_left;
 193                         }
 194                 } else if (i915_seqno_passed(wait->seqno,
 195                                              to_wait(parent)->seqno)) {
 196                         p = &parent->rb_right;
 197                         if (i915_seqno_passed(seqno, to_wait(parent)->seqno))
 198                                 completed = parent;
 199                         else
 200                                 first = false;
 201                 } else {
 202                         p = &parent->rb_left;
 203                 }
 204         }
 205         rb_link_node(&wait->node, parent, p);
 206         rb_insert_color(&wait->node, &b->waiters);
 207         GEM_BUG_ON(!first && !b->irq_seqno_bh);
 208
 209         if (completed) {
 210                 struct rb_node *next = rb_next(completed);
 211
 212                 GEM_BUG_ON(!next && !first);
 213                 if (next && next != &wait->node) {
 214                         GEM_BUG_ON(first);
 215                         b->first_wait = to_wait(next);
 216                         smp_store_mb(b->irq_seqno_bh, b->first_wait->tsk);
 217                         /* As there is a delay between reading the current
 218                          * seqno, processing the completed tasks and selecting
 219                          * the next waiter, we may have missed the interrupt
 220                          * and so need for the next bottom-half to wakeup.
 221                          *
 222                          * Also as we enable the IRQ, we may miss the
 223                          * interrupt for that seqno, so we have to wake up
 224                          * the next bottom-half in order to do a coherent check
 225                          * in case the seqno passed.
 226                          */
 227                         __intel_breadcrumbs_enable_irq(b);
 228                         if (READ_ONCE(b->irq_posted))
 229                                 wake_up_process(to_wait(next)->tsk);
 230                 }
 231
 232                 do {
 233                         struct intel_wait *crumb = to_wait(completed);
 234                         completed = rb_prev(completed);
 235                         __intel_breadcrumbs_finish(b, crumb);
 236                 } while (completed);
 237         }
 238
 239         if (first) {
 240                 GEM_BUG_ON(rb_first(&b->waiters) != &wait->node);
 241                 b->first_wait = wait;
 242                 smp_store_mb(b->irq_seqno_bh, wait->tsk);
 243                 /* After assigning ourselves as the new bottom-half, we must
 244                  * perform a cursory check to prevent a missed interrupt.
 245                  * Either we miss the interrupt whilst programming the hardware,
 246                  * or if there was a previous waiter (for a later seqno) they
 247                  * may be woken instead of us (due to the inherent race
 248                  * in the unlocked read of b->irq_seqno_bh in the irq handler)
 249                  * and so we miss the wake up.
 250                  */
 251                 __intel_breadcrumbs_enable_irq(b);
 252         }
 253         GEM_BUG_ON(!b->irq_seqno_bh);
 254         GEM_BUG_ON(!b->first_wait);
 255         GEM_BUG_ON(rb_first(&b->waiters) != &b->first_wait->node);
 256
 257         return first;
 258 }
 259
 260 bool intel_engine_add_wait(struct intel_engine_cs *engine,
 261                            struct intel_wait *wait)
 262 {
 263         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 264         bool first;
 265
 266         spin_lock(&b->lock);
 267         first = __intel_engine_add_wait(engine, wait);
 268         spin_unlock(&b->lock);
 269
 270         return first;
 271 }
 272
 273 void intel_engine_enable_fake_irq(struct intel_engine_cs *engine)
 274 {
 275         mod_timer(&engine->breadcrumbs.fake_irq, jiffies + 1);
 276 }
 277
 278 static inline bool chain_wakeup(struct rb_node *rb, int priority)
 279 {
 280         return rb && to_wait(rb)->tsk->prio <= priority;
 281 }
 282
 283 static inline int wakeup_priority(struct intel_breadcrumbs *b,
 284                                   struct task_struct *tsk)
 285 {
 286         if (tsk == b->signaler)
 287                 return INT_MIN;
 288         else
 289                 return tsk->prio;
 290 }
 291
 292 void intel_engine_remove_wait(struct intel_engine_cs *engine,
 293                               struct intel_wait *wait)
 294 {
 295         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 296
 297         /* Quick check to see if this waiter was already decoupled from
 298          * the tree by the bottom-half to avoid contention on the spinlock
 299          * by the herd.
 300          */
 301         if (RB_EMPTY_NODE(&wait->node))
 302                 return;
 303
 304         spin_lock(&b->lock);
 305
 306         if (RB_EMPTY_NODE(&wait->node))
 307                 goto out_unlock;
 308
 309         if (b->first_wait == wait) {
 310                 const int priority = wakeup_priority(b, wait->tsk);
 311                 struct rb_node *next;
 312
 313                 GEM_BUG_ON(b->irq_seqno_bh != wait->tsk);
 314
 315                 /* We are the current bottom-half. Find the next candidate,
 316                  * the first waiter in the queue on the remaining oldest
 317                  * request. As multiple seqnos may complete in the time it
 318                  * takes us to wake up and find the next waiter, we have to
 319                  * wake up that waiter for it to perform its own coherent
 320                  * completion check.
 321                  */
 322                 next = rb_next(&wait->node);
 323                 if (chain_wakeup(next, priority)) {
 324                         /* If the next waiter is already complete,
 325                          * wake it up and continue onto the next waiter. So
 326                          * if have a small herd, they will wake up in parallel
 327                          * rather than sequentially, which should reduce
 328                          * the overall latency in waking all the completed
 329                          * clients.
 330                          *
 331                          * However, waking up a chain adds extra latency to
 332                          * the first_waiter. This is undesirable if that
 333                          * waiter is a high priority task.
 334                          */
 335                         u32 seqno = intel_engine_get_seqno(engine);
 336
 337                         while (i915_seqno_passed(seqno, to_wait(next)->seqno)) {
 338                                 struct rb_node *n = rb_next(next);
 339
 340                                 __intel_breadcrumbs_finish(b, to_wait(next));
 341                                 next = n;
 342                                 if (!chain_wakeup(next, priority))
 343                                         break;
 344                         }
 345                 }
 346
 347                 if (next) {
 348                         /* In our haste, we may have completed the first waiter
 349                          * before we enabled the interrupt. Do so now as we
 350                          * have a second waiter for a future seqno. Afterwards,
 351                          * we have to wake up that waiter in case we missed
 352                          * the interrupt, or if we have to handle an
 353                          * exception rather than a seqno completion.
 354                          */
 355                         b->first_wait = to_wait(next);
 356                         smp_store_mb(b->irq_seqno_bh, b->first_wait->tsk);
 357                         if (b->first_wait->seqno != wait->seqno)
 358                                 __intel_breadcrumbs_enable_irq(b);
 359                         wake_up_process(b->irq_seqno_bh);
 360                 } else {
 361                         b->first_wait = NULL;
 362                         WRITE_ONCE(b->irq_seqno_bh, NULL);
 363                         __intel_breadcrumbs_disable_irq(b);
 364                 }
 365         } else {
 366                 GEM_BUG_ON(rb_first(&b->waiters) == &wait->node);
 367         }
 368
 369         GEM_BUG_ON(RB_EMPTY_NODE(&wait->node));
 370         rb_erase(&wait->node, &b->waiters);
 371
 372 out_unlock:
 373         GEM_BUG_ON(b->first_wait == wait);
 374         GEM_BUG_ON(rb_first(&b->waiters) !=
 375                    (b->first_wait ? &b->first_wait->node : NULL));
 376         GEM_BUG_ON(!b->irq_seqno_bh ^ RB_EMPTY_ROOT(&b->waiters));
 377         spin_unlock(&b->lock);
 378 }
 379
 380 static bool signal_complete(struct drm_i915_gem_request *request)
 381 {
 382         if (!request)
 383                 return false;
 384
 385         /* If another process served as the bottom-half it may have already
 386          * signalled that this wait is already completed.
 387          */
 388         if (intel_wait_complete(&request->signaling.wait))
 389                 return true;
 390
 391         /* Carefully check if the request is complete, giving time for the
 392          * seqno to be visible or if the GPU hung.
 393          */
 394         if (__i915_request_irq_complete(request))
 395                 return true;
 396
 397         return false;
 398 }
 399
 400 static struct drm_i915_gem_request *to_signaler(struct rb_node *rb)
 401 {
 402         return container_of(rb, struct drm_i915_gem_request, signaling.node);
 403 }
 404
 405 static void signaler_set_rtpriority(void)
 406 {
 407          struct sched_param param = { .sched_priority = 1 };
 408
 409          sched_setscheduler_nocheck(current, SCHED_FIFO, &param);
 410 }
 411
 412 static int intel_breadcrumbs_signaler(void *arg)
 413 {
 414         struct intel_engine_cs *engine = arg;
 415         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 416         struct drm_i915_gem_request *request;
 417
 418         /* Install ourselves with high priority to reduce signalling latency */
 419         signaler_set_rtpriority();
 420
 421         do {
 422                 set_current_state(TASK_INTERRUPTIBLE);
 423
 424                 /* We are either woken up by the interrupt bottom-half,
 425                  * or by a client adding a new signaller. In both cases,
 426                  * the GPU seqno may have advanced beyond our oldest signal.
 427                  * If it has, propagate the signal, remove the waiter and
 428                  * check again with the next oldest signal. Otherwise we
 429                  * need to wait for a new interrupt from the GPU or for
 430                  * a new client.
 431                  */
 432                 request = READ_ONCE(b->first_signal);
 433                 if (signal_complete(request)) {
 434                         /* Wake up all other completed waiters and select the
 435                          * next bottom-half for the next user interrupt.
 436                          */
 437                         intel_engine_remove_wait(engine,
 438                                                  &request->signaling.wait);
 439
 440                         /* Find the next oldest signal. Note that as we have
 441                          * not been holding the lock, another client may
 442                          * have installed an even older signal than the one
 443                          * we just completed - so double check we are still
 444                          * the oldest before picking the next one.
 445                          */
 446                         spin_lock(&b->lock);
 447                         if (request == b->first_signal) {
 448                                 struct rb_node *rb =
 449                                         rb_next(&request->signaling.node);
 450                                 b->first_signal = rb ? to_signaler(rb) : NULL;
 451                         }
 452                         rb_erase(&request->signaling.node, &b->signals);
 453                         spin_unlock(&b->lock);
 454
 455                         i915_gem_request_unreference(request);
 456                 } else {
 457                         if (kthread_should_stop())
 458                                 break;
 459
 460                         schedule();
 461                 }
 462         } while (1);
 463         __set_current_state(TASK_RUNNING);
 464
 465         return 0;
 466 }
 467
 468 void intel_engine_enable_signaling(struct drm_i915_gem_request *request)
 469 {
 470         struct intel_engine_cs *engine = request->engine;
 471         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 472         struct rb_node *parent, **p;
 473         bool first, wakeup;
 474
 475         if (unlikely(READ_ONCE(request->signaling.wait.tsk)))
 476                 return;
 477
 478         spin_lock(&b->lock);
 479         if (unlikely(request->signaling.wait.tsk)) {
 480                 wakeup = false;
 481                 goto unlock;
 482         }
 483
 484         request->signaling.wait.tsk = b->signaler;
 485         request->signaling.wait.seqno = request->seqno;
 486         i915_gem_request_reference(request);
 487
 488         /* First add ourselves into the list of waiters, but register our
 489          * bottom-half as the signaller thread. As per usual, only the oldest
 490          * waiter (not just signaller) is tasked as the bottom-half waking
 491          * up all completed waiters after the user interrupt.
 492          *
 493          * If we are the oldest waiter, enable the irq (after which we
 494          * must double check that the seqno did not complete).
 495          */
 496         wakeup = __intel_engine_add_wait(engine, &request->signaling.wait);
 497
 498         /* Now insert ourselves into the retirement ordered list of signals
 499          * on this engine. We track the oldest seqno as that will be the
 500          * first signal to complete.
 501          */
 502         parent = NULL;
 503         first = true;
 504         p = &b->signals.rb_node;
 505         while (*p) {
 506                 parent = *p;
 507                 if (i915_seqno_passed(request->seqno,
 508                                       to_signaler(parent)->seqno)) {
 509                         p = &parent->rb_right;
 510                         first = false;
 511                 } else {
 512                         p = &parent->rb_left;
 513                 }
 514         }
 515         rb_link_node(&request->signaling.node, parent, p);
 516         rb_insert_color(&request->signaling.node, &b->signals);
 517         if (first)
 518                 smp_store_mb(b->first_signal, request);
 519
 520 unlock:
 521         spin_unlock(&b->lock);
 522
 523         if (wakeup)
 524                 wake_up_process(b->signaler);
 525 }
 526
 527 int intel_engine_init_breadcrumbs(struct intel_engine_cs *engine)
 528 {
 529         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 530         struct task_struct *tsk;
 531
 532         spin_lock_init(&b->lock);
 533         setup_timer(&b->fake_irq,
 534                     intel_breadcrumbs_fake_irq,
 535                     (unsigned long)engine);
 536
 537         /* Spawn a thread to provide a common bottom-half for all signals.
 538          * As this is an asynchronous interface we cannot steal the current
 539          * task for handling the bottom-half to the user interrupt, therefore
 540          * we create a thread to do the coherent seqno dance after the
 541          * interrupt and then signal the waitqueue (via the dma-buf/fence).
 542          */
 543         tsk = kthread_run(intel_breadcrumbs_signaler, engine,
 544                           "i915/signal:%d", engine->id);
 545         if (IS_ERR(tsk))
 546                 return PTR_ERR(tsk);
 547
 548         b->signaler = tsk;
 549
 550         return 0;
 551 }
 552
 553 void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine)
 554 {
 555         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 556
 557         if (!IS_ERR_OR_NULL(b->signaler))
 558                 kthread_stop(b->signaler);
 559
 560         del_timer_sync(&b->fake_irq);
 561 }
 562
 563 unsigned int intel_kick_waiters(struct drm_i915_private *i915)
 564 {
 565         struct intel_engine_cs *engine;
 566         unsigned int mask = 0;
 567
 568         /* To avoid the task_struct disappearing beneath us as we wake up
 569          * the process, we must first inspect the task_struct->state under the
 570          * RCU lock, i.e. as we call wake_up_process() we must be holding the
 571          * rcu_read_lock().
 572          */
 573         rcu_read_lock();
 574         for_each_engine(engine, i915)
 575                 if (unlikely(intel_engine_wakeup(engine)))
 576                         mask |= intel_engine_flag(engine);
 577         rcu_read_unlock();
 578
 579         return mask;
 580 }
 581
 582 unsigned int intel_kick_signalers(struct drm_i915_private *i915)
 583 {
 584         struct intel_engine_cs *engine;
 585         unsigned int mask = 0;
 586
 587         for_each_engine(engine, i915) {
 588                 if (unlikely(READ_ONCE(engine->breadcrumbs.first_signal))) {
 589                         wake_up_process(engine->breadcrumbs.signaler);
 590                         mask |= intel_engine_flag(engine);
 591                 }
 592         }
 593
 594         return mask;
 595 }