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->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->irq_posted = false;
  66 }
  67
  68 static bool __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 false;
  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         return engine->irq_posted;
  98 }
  99
 100 static void __intel_breadcrumbs_disable_irq(struct intel_breadcrumbs *b)
 101 {
 102         struct intel_engine_cs *engine =
 103                 container_of(b, struct intel_engine_cs, breadcrumbs);
 104
 105         assert_spin_locked(&b->lock);
 106         if (!b->rpm_wakelock)
 107                 return;
 108
 109         if (b->irq_enabled) {
 110                 irq_disable(engine);
 111                 b->irq_enabled = false;
 112         }
 113
 114         intel_runtime_pm_put(engine->i915);
 115         b->rpm_wakelock = false;
 116 }
 117
 118 static inline struct intel_wait *to_wait(struct rb_node *node)
 119 {
 120         return container_of(node, struct intel_wait, node);
 121 }
 122
 123 static inline void __intel_breadcrumbs_finish(struct intel_breadcrumbs *b,
 124                                               struct intel_wait *wait)
 125 {
 126         assert_spin_locked(&b->lock);
 127
 128         /* This request is completed, so remove it from the tree, mark it as
 129          * complete, and *then* wake up the associated task.
 130          */
 131         rb_erase(&wait->node, &b->waiters);
 132         RB_CLEAR_NODE(&wait->node);
 133
 134         wake_up_process(wait->tsk); /* implicit smp_wmb() */
 135 }
 136
 137 static bool __intel_engine_add_wait(struct intel_engine_cs *engine,
 138                                     struct intel_wait *wait)
 139 {
 140         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 141         struct rb_node **p, *parent, *completed;
 142         bool first;
 143         u32 seqno;
 144
 145         /* Insert the request into the retirement ordered list
 146          * of waiters by walking the rbtree. If we are the oldest
 147          * seqno in the tree (the first to be retired), then
 148          * set ourselves as the bottom-half.
 149          *
 150          * As we descend the tree, prune completed branches since we hold the
 151          * spinlock we know that the first_waiter must be delayed and can
 152          * reduce some of the sequential wake up latency if we take action
 153          * ourselves and wake up the completed tasks in parallel. Also, by
 154          * removing stale elements in the tree, we may be able to reduce the
 155          * ping-pong between the old bottom-half and ourselves as first-waiter.
 156          */
 157         first = true;
 158         parent = NULL;
 159         completed = NULL;
 160         seqno = intel_engine_get_seqno(engine);
 161
 162          /* If the request completed before we managed to grab the spinlock,
 163           * return now before adding ourselves to the rbtree. We let the
 164           * current bottom-half handle any pending wakeups and instead
 165           * try and get out of the way quickly.
 166           */
 167         if (i915_seqno_passed(seqno, wait->seqno)) {
 168                 RB_CLEAR_NODE(&wait->node);
 169                 return first;
 170         }
 171
 172         p = &b->waiters.rb_node;
 173         while (*p) {
 174                 parent = *p;
 175                 if (wait->seqno == to_wait(parent)->seqno) {
 176                         /* We have multiple waiters on the same seqno, select
 177                          * the highest priority task (that with the smallest
 178                          * task->prio) to serve as the bottom-half for this
 179                          * group.
 180                          */
 181                         if (wait->tsk->prio > to_wait(parent)->tsk->prio) {
 182                                 p = &parent->rb_right;
 183                                 first = false;
 184                         } else {
 185                                 p = &parent->rb_left;
 186                         }
 187                 } else if (i915_seqno_passed(wait->seqno,
 188                                              to_wait(parent)->seqno)) {
 189                         p = &parent->rb_right;
 190                         if (i915_seqno_passed(seqno, to_wait(parent)->seqno))
 191                                 completed = parent;
 192                         else
 193                                 first = false;
 194                 } else {
 195                         p = &parent->rb_left;
 196                 }
 197         }
 198         rb_link_node(&wait->node, parent, p);
 199         rb_insert_color(&wait->node, &b->waiters);
 200         GEM_BUG_ON(!first && !b->tasklet);
 201
 202         if (completed) {
 203                 struct rb_node *next = rb_next(completed);
 204
 205                 GEM_BUG_ON(!next && !first);
 206                 if (next && next != &wait->node) {
 207                         GEM_BUG_ON(first);
 208                         b->first_wait = to_wait(next);
 209                         smp_store_mb(b->tasklet, b->first_wait->tsk);
 210                         /* As there is a delay between reading the current
 211                          * seqno, processing the completed tasks and selecting
 212                          * the next waiter, we may have missed the interrupt
 213                          * and so need for the next bottom-half to wakeup.
 214                          *
 215                          * Also as we enable the IRQ, we may miss the
 216                          * interrupt for that seqno, so we have to wake up
 217                          * the next bottom-half in order to do a coherent check
 218                          * in case the seqno passed.
 219                          */
 220                         __intel_breadcrumbs_enable_irq(b);
 221                         if (READ_ONCE(engine->irq_posted))
 222                                 wake_up_process(to_wait(next)->tsk);
 223                 }
 224
 225                 do {
 226                         struct intel_wait *crumb = to_wait(completed);
 227                         completed = rb_prev(completed);
 228                         __intel_breadcrumbs_finish(b, crumb);
 229                 } while (completed);
 230         }
 231
 232         if (first) {
 233                 GEM_BUG_ON(rb_first(&b->waiters) != &wait->node);
 234                 b->first_wait = wait;
 235                 smp_store_mb(b->tasklet, wait->tsk);
 236                 first = __intel_breadcrumbs_enable_irq(b);
 237         }
 238         GEM_BUG_ON(!b->tasklet);
 239         GEM_BUG_ON(!b->first_wait);
 240         GEM_BUG_ON(rb_first(&b->waiters) != &b->first_wait->node);
 241
 242         return first;
 243 }
 244
 245 bool intel_engine_add_wait(struct intel_engine_cs *engine,
 246                            struct intel_wait *wait)
 247 {
 248         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 249         bool first;
 250
 251         spin_lock(&b->lock);
 252         first = __intel_engine_add_wait(engine, wait);
 253         spin_unlock(&b->lock);
 254
 255         return first;
 256 }
 257
 258 void intel_engine_enable_fake_irq(struct intel_engine_cs *engine)
 259 {
 260         mod_timer(&engine->breadcrumbs.fake_irq, jiffies + 1);
 261 }
 262
 263 static inline bool chain_wakeup(struct rb_node *rb, int priority)
 264 {
 265         return rb && to_wait(rb)->tsk->prio <= priority;
 266 }
 267
 268 static inline int wakeup_priority(struct intel_breadcrumbs *b,
 269                                   struct task_struct *tsk)
 270 {
 271         if (tsk == b->signaler)
 272                 return INT_MIN;
 273         else
 274                 return tsk->prio;
 275 }
 276
 277 void intel_engine_remove_wait(struct intel_engine_cs *engine,
 278                               struct intel_wait *wait)
 279 {
 280         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 281
 282         /* Quick check to see if this waiter was already decoupled from
 283          * the tree by the bottom-half to avoid contention on the spinlock
 284          * by the herd.
 285          */
 286         if (RB_EMPTY_NODE(&wait->node))
 287                 return;
 288
 289         spin_lock(&b->lock);
 290
 291         if (RB_EMPTY_NODE(&wait->node))
 292                 goto out_unlock;
 293
 294         if (b->first_wait == wait) {
 295                 const int priority = wakeup_priority(b, wait->tsk);
 296                 struct rb_node *next;
 297
 298                 GEM_BUG_ON(b->tasklet != wait->tsk);
 299
 300                 /* We are the current bottom-half. Find the next candidate,
 301                  * the first waiter in the queue on the remaining oldest
 302                  * request. As multiple seqnos may complete in the time it
 303                  * takes us to wake up and find the next waiter, we have to
 304                  * wake up that waiter for it to perform its own coherent
 305                  * completion check.
 306                  */
 307                 next = rb_next(&wait->node);
 308                 if (chain_wakeup(next, priority)) {
 309                         /* If the next waiter is already complete,
 310                          * wake it up and continue onto the next waiter. So
 311                          * if have a small herd, they will wake up in parallel
 312                          * rather than sequentially, which should reduce
 313                          * the overall latency in waking all the completed
 314                          * clients.
 315                          *
 316                          * However, waking up a chain adds extra latency to
 317                          * the first_waiter. This is undesirable if that
 318                          * waiter is a high priority task.
 319                          */
 320                         u32 seqno = intel_engine_get_seqno(engine);
 321
 322                         while (i915_seqno_passed(seqno, to_wait(next)->seqno)) {
 323                                 struct rb_node *n = rb_next(next);
 324
 325                                 __intel_breadcrumbs_finish(b, to_wait(next));
 326                                 next = n;
 327                                 if (!chain_wakeup(next, priority))
 328                                         break;
 329                         }
 330                 }
 331
 332                 if (next) {
 333                         /* In our haste, we may have completed the first waiter
 334                          * before we enabled the interrupt. Do so now as we
 335                          * have a second waiter for a future seqno. Afterwards,
 336                          * we have to wake up that waiter in case we missed
 337                          * the interrupt, or if we have to handle an
 338                          * exception rather than a seqno completion.
 339                          */
 340                         b->first_wait = to_wait(next);
 341                         smp_store_mb(b->tasklet, b->first_wait->tsk);
 342                         if (b->first_wait->seqno != wait->seqno)
 343                                 __intel_breadcrumbs_enable_irq(b);
 344                         wake_up_process(b->tasklet);
 345                 } else {
 346                         b->first_wait = NULL;
 347                         WRITE_ONCE(b->tasklet, NULL);
 348                         __intel_breadcrumbs_disable_irq(b);
 349                 }
 350         } else {
 351                 GEM_BUG_ON(rb_first(&b->waiters) == &wait->node);
 352         }
 353
 354         GEM_BUG_ON(RB_EMPTY_NODE(&wait->node));
 355         rb_erase(&wait->node, &b->waiters);
 356
 357 out_unlock:
 358         GEM_BUG_ON(b->first_wait == wait);
 359         GEM_BUG_ON(rb_first(&b->waiters) !=
 360                    (b->first_wait ? &b->first_wait->node : NULL));
 361         GEM_BUG_ON(!b->tasklet ^ RB_EMPTY_ROOT(&b->waiters));
 362         spin_unlock(&b->lock);
 363 }
 364
 365 static bool signal_complete(struct drm_i915_gem_request *request)
 366 {
 367         if (!request)
 368                 return false;
 369
 370         /* If another process served as the bottom-half it may have already
 371          * signalled that this wait is already completed.
 372          */
 373         if (intel_wait_complete(&request->signaling.wait))
 374                 return true;
 375
 376         /* Carefully check if the request is complete, giving time for the
 377          * seqno to be visible or if the GPU hung.
 378          */
 379         if (__i915_request_irq_complete(request))
 380                 return true;
 381
 382         return false;
 383 }
 384
 385 static struct drm_i915_gem_request *to_signaler(struct rb_node *rb)
 386 {
 387         return container_of(rb, struct drm_i915_gem_request, signaling.node);
 388 }
 389
 390 static void signaler_set_rtpriority(void)
 391 {
 392          struct sched_param param = { .sched_priority = 1 };
 393
 394          sched_setscheduler_nocheck(current, SCHED_FIFO, &param);
 395 }
 396
 397 static int intel_breadcrumbs_signaler(void *arg)
 398 {
 399         struct intel_engine_cs *engine = arg;
 400         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 401         struct drm_i915_gem_request *request;
 402
 403         /* Install ourselves with high priority to reduce signalling latency */
 404         signaler_set_rtpriority();
 405
 406         do {
 407                 set_current_state(TASK_INTERRUPTIBLE);
 408
 409                 /* We are either woken up by the interrupt bottom-half,
 410                  * or by a client adding a new signaller. In both cases,
 411                  * the GPU seqno may have advanced beyond our oldest signal.
 412                  * If it has, propagate the signal, remove the waiter and
 413                  * check again with the next oldest signal. Otherwise we
 414                  * need to wait for a new interrupt from the GPU or for
 415                  * a new client.
 416                  */
 417                 request = READ_ONCE(b->first_signal);
 418                 if (signal_complete(request)) {
 419                         /* Wake up all other completed waiters and select the
 420                          * next bottom-half for the next user interrupt.
 421                          */
 422                         intel_engine_remove_wait(engine,
 423                                                  &request->signaling.wait);
 424
 425                         /* Find the next oldest signal. Note that as we have
 426                          * not been holding the lock, another client may
 427                          * have installed an even older signal than the one
 428                          * we just completed - so double check we are still
 429                          * the oldest before picking the next one.
 430                          */
 431                         spin_lock(&b->lock);
 432                         if (request == b->first_signal) {
 433                                 struct rb_node *rb =
 434                                         rb_next(&request->signaling.node);
 435                                 b->first_signal = rb ? to_signaler(rb) : NULL;
 436                         }
 437                         rb_erase(&request->signaling.node, &b->signals);
 438                         spin_unlock(&b->lock);
 439
 440                         i915_gem_request_unreference(request);
 441                 } else {
 442                         if (kthread_should_stop())
 443                                 break;
 444
 445                         schedule();
 446                 }
 447         } while (1);
 448         __set_current_state(TASK_RUNNING);
 449
 450         return 0;
 451 }
 452
 453 void intel_engine_enable_signaling(struct drm_i915_gem_request *request)
 454 {
 455         struct intel_engine_cs *engine = request->engine;
 456         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 457         struct rb_node *parent, **p;
 458         bool first, wakeup;
 459
 460         if (unlikely(READ_ONCE(request->signaling.wait.tsk)))
 461                 return;
 462
 463         spin_lock(&b->lock);
 464         if (unlikely(request->signaling.wait.tsk)) {
 465                 wakeup = false;
 466                 goto unlock;
 467         }
 468
 469         request->signaling.wait.tsk = b->signaler;
 470         request->signaling.wait.seqno = request->seqno;
 471         i915_gem_request_reference(request);
 472
 473         /* First add ourselves into the list of waiters, but register our
 474          * bottom-half as the signaller thread. As per usual, only the oldest
 475          * waiter (not just signaller) is tasked as the bottom-half waking
 476          * up all completed waiters after the user interrupt.
 477          *
 478          * If we are the oldest waiter, enable the irq (after which we
 479          * must double check that the seqno did not complete).
 480          */
 481         wakeup = __intel_engine_add_wait(engine, &request->signaling.wait);
 482
 483         /* Now insert ourselves into the retirement ordered list of signals
 484          * on this engine. We track the oldest seqno as that will be the
 485          * first signal to complete.
 486          */
 487         parent = NULL;
 488         first = true;
 489         p = &b->signals.rb_node;
 490         while (*p) {
 491                 parent = *p;
 492                 if (i915_seqno_passed(request->seqno,
 493                                       to_signaler(parent)->seqno)) {
 494                         p = &parent->rb_right;
 495                         first = false;
 496                 } else {
 497                         p = &parent->rb_left;
 498                 }
 499         }
 500         rb_link_node(&request->signaling.node, parent, p);
 501         rb_insert_color(&request->signaling.node, &b->signals);
 502         if (first)
 503                 smp_store_mb(b->first_signal, request);
 504
 505 unlock:
 506         spin_unlock(&b->lock);
 507
 508         if (wakeup)
 509                 wake_up_process(b->signaler);
 510 }
 511
 512 int intel_engine_init_breadcrumbs(struct intel_engine_cs *engine)
 513 {
 514         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 515         struct task_struct *tsk;
 516
 517         spin_lock_init(&b->lock);
 518         setup_timer(&b->fake_irq,
 519                     intel_breadcrumbs_fake_irq,
 520                     (unsigned long)engine);
 521
 522         /* Spawn a thread to provide a common bottom-half for all signals.
 523          * As this is an asynchronous interface we cannot steal the current
 524          * task for handling the bottom-half to the user interrupt, therefore
 525          * we create a thread to do the coherent seqno dance after the
 526          * interrupt and then signal the waitqueue (via the dma-buf/fence).
 527          */
 528         tsk = kthread_run(intel_breadcrumbs_signaler, engine,
 529                           "i915/signal:%d", engine->id);
 530         if (IS_ERR(tsk))
 531                 return PTR_ERR(tsk);
 532
 533         b->signaler = tsk;
 534
 535         return 0;
 536 }
 537
 538 void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine)
 539 {
 540         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 541
 542         if (!IS_ERR_OR_NULL(b->signaler))
 543                 kthread_stop(b->signaler);
 544
 545         del_timer_sync(&b->fake_irq);
 546 }
 547
 548 unsigned int intel_kick_waiters(struct drm_i915_private *i915)
 549 {
 550         struct intel_engine_cs *engine;
 551         unsigned int mask = 0;
 552
 553         /* To avoid the task_struct disappearing beneath us as we wake up
 554          * the process, we must first inspect the task_struct->state under the
 555          * RCU lock, i.e. as we call wake_up_process() we must be holding the
 556          * rcu_read_lock().
 557          */
 558         rcu_read_lock();
 559         for_each_engine(engine, i915)
 560                 if (unlikely(intel_engine_wakeup(engine)))
 561                         mask |= intel_engine_flag(engine);
 562         rcu_read_unlock();
 563
 564         return mask;
 565 }
 566
 567 unsigned int intel_kick_signalers(struct drm_i915_private *i915)
 568 {
 569         struct intel_engine_cs *engine;
 570         unsigned int mask = 0;
 571
 572         for_each_engine(engine, i915) {
 573                 if (unlikely(READ_ONCE(engine->breadcrumbs.first_signal))) {
 574                         wake_up_process(engine->breadcrumbs.signaler);
 575                         mask |= intel_engine_flag(engine);
 576                 }
 577         }
 578
 579         return mask;
 580 }