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 #include <uapi/linux/sched/types.h>
  27
  28 #include "i915_drv.h"
  29
  30 static unsigned int __intel_breadcrumbs_wakeup(struct intel_breadcrumbs *b)
  31 {
  32         struct intel_wait *wait;
  33         unsigned int result = 0;
  34
  35         lockdep_assert_held(&b->irq_lock);
  36
  37         wait = b->irq_wait;
  38         if (wait) {
  39                 result = ENGINE_WAKEUP_WAITER;
  40                 if (wake_up_process(wait->tsk))
  41                         result |= ENGINE_WAKEUP_ASLEEP;
  42         }
  43
  44         return result;
  45 }
  46
  47 unsigned int intel_engine_wakeup(struct intel_engine_cs *engine)
  48 {
  49         struct intel_breadcrumbs *b = &engine->breadcrumbs;
  50         unsigned long flags;
  51         unsigned int result;
  52
  53         spin_lock_irqsave(&b->irq_lock, flags);
  54         result = __intel_breadcrumbs_wakeup(b);
  55         spin_unlock_irqrestore(&b->irq_lock, flags);
  56
  57         return result;
  58 }
  59
  60 static unsigned long wait_timeout(void)
  61 {
  62         return round_jiffies_up(jiffies + DRM_I915_HANGCHECK_JIFFIES);
  63 }
  64
  65 static noinline void missed_breadcrumb(struct intel_engine_cs *engine)
  66 {
  67         DRM_DEBUG_DRIVER("%s missed breadcrumb at %pS, irq posted? %s, current seqno=%x, last=%x\n",
  68                          engine->name, __builtin_return_address(0),
  69                          yesno(test_bit(ENGINE_IRQ_BREADCRUMB,
  70                                         &engine->irq_posted)),
  71                          intel_engine_get_seqno(engine),
  72                          intel_engine_last_submit(engine));
  73
  74         set_bit(engine->id, &engine->i915->gpu_error.missed_irq_rings);
  75 }
  76
  77 static void intel_breadcrumbs_hangcheck(struct timer_list *t)
  78 {
  79         struct intel_engine_cs *engine = from_timer(engine, t,
  80                                                     breadcrumbs.hangcheck);
  81         struct intel_breadcrumbs *b = &engine->breadcrumbs;
  82
  83         if (!b->irq_armed)
  84                 return;
  85
  86         if (b->hangcheck_interrupts != atomic_read(&engine->irq_count)) {
  87                 b->hangcheck_interrupts = atomic_read(&engine->irq_count);
  88                 mod_timer(&b->hangcheck, wait_timeout());
  89                 return;
  90         }
  91
  92         /* We keep the hangcheck timer alive until we disarm the irq, even
  93          * if there are no waiters at present.
  94          *
  95          * If the waiter was currently running, assume it hasn't had a chance
  96          * to process the pending interrupt (e.g, low priority task on a loaded
  97          * system) and wait until it sleeps before declaring a missed interrupt.
  98          *
  99          * If the waiter was asleep (and not even pending a wakeup), then we
 100          * must have missed an interrupt as the GPU has stopped advancing
 101          * but we still have a waiter. Assuming all batches complete within
 102          * DRM_I915_HANGCHECK_JIFFIES [1.5s]!
 103          */
 104         if (intel_engine_wakeup(engine) & ENGINE_WAKEUP_ASLEEP) {
 105                 missed_breadcrumb(engine);
 106                 mod_timer(&engine->breadcrumbs.fake_irq, jiffies + 1);
 107         } else {
 108                 mod_timer(&b->hangcheck, wait_timeout());
 109         }
 110 }
 111
 112 static void intel_breadcrumbs_fake_irq(struct timer_list *t)
 113 {
 114         struct intel_engine_cs *engine = from_timer(engine, t,
 115                                                     breadcrumbs.fake_irq);
 116         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 117
 118         /* The timer persists in case we cannot enable interrupts,
 119          * or if we have previously seen seqno/interrupt incoherency
 120          * ("missed interrupt" syndrome, better known as a "missed breadcrumb").
 121          * Here the worker will wake up every jiffie in order to kick the
 122          * oldest waiter to do the coherent seqno check.
 123          */
 124
 125         spin_lock_irq(&b->irq_lock);
 126         if (!__intel_breadcrumbs_wakeup(b))
 127                 __intel_engine_disarm_breadcrumbs(engine);
 128         spin_unlock_irq(&b->irq_lock);
 129         if (!b->irq_armed)
 130                 return;
 131
 132         mod_timer(&b->fake_irq, jiffies + 1);
 133
 134         /* Ensure that even if the GPU hangs, we get woken up.
 135          *
 136          * However, note that if no one is waiting, we never notice
 137          * a gpu hang. Eventually, we will have to wait for a resource
 138          * held by the GPU and so trigger a hangcheck. In the most
 139          * pathological case, this will be upon memory starvation! To
 140          * prevent this, we also queue the hangcheck from the retire
 141          * worker.
 142          */
 143         i915_queue_hangcheck(engine->i915);
 144 }
 145
 146 static void irq_enable(struct intel_engine_cs *engine)
 147 {
 148         /* Enabling the IRQ may miss the generation of the interrupt, but
 149          * we still need to force the barrier before reading the seqno,
 150          * just in case.
 151          */
 152         set_bit(ENGINE_IRQ_BREADCRUMB, &engine->irq_posted);
 153
 154         /* Caller disables interrupts */
 155         spin_lock(&engine->i915->irq_lock);
 156         engine->irq_enable(engine);
 157         spin_unlock(&engine->i915->irq_lock);
 158 }
 159
 160 static void irq_disable(struct intel_engine_cs *engine)
 161 {
 162         /* Caller disables interrupts */
 163         spin_lock(&engine->i915->irq_lock);
 164         engine->irq_disable(engine);
 165         spin_unlock(&engine->i915->irq_lock);
 166 }
 167
 168 void __intel_engine_disarm_breadcrumbs(struct intel_engine_cs *engine)
 169 {
 170         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 171
 172         lockdep_assert_held(&b->irq_lock);
 173         GEM_BUG_ON(b->irq_wait);
 174
 175         if (b->irq_enabled) {
 176                 irq_disable(engine);
 177                 b->irq_enabled = false;
 178         }
 179
 180         b->irq_armed = false;
 181 }
 182
 183 void intel_engine_disarm_breadcrumbs(struct intel_engine_cs *engine)
 184 {
 185         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 186         struct intel_wait *wait, *n, *first;
 187
 188         if (!b->irq_armed)
 189                 return;
 190
 191         /* We only disarm the irq when we are idle (all requests completed),
 192          * so if the bottom-half remains asleep, it missed the request
 193          * completion.
 194          */
 195
 196         spin_lock_irq(&b->rb_lock);
 197
 198         spin_lock(&b->irq_lock);
 199         first = fetch_and_zero(&b->irq_wait);
 200         __intel_engine_disarm_breadcrumbs(engine);
 201         spin_unlock(&b->irq_lock);
 202
 203         rbtree_postorder_for_each_entry_safe(wait, n, &b->waiters, node) {
 204                 RB_CLEAR_NODE(&wait->node);
 205                 if (wake_up_process(wait->tsk) && wait == first)
 206                         missed_breadcrumb(engine);
 207         }
 208         b->waiters = RB_ROOT;
 209
 210         spin_unlock_irq(&b->rb_lock);
 211 }
 212
 213 static bool use_fake_irq(const struct intel_breadcrumbs *b)
 214 {
 215         const struct intel_engine_cs *engine =
 216                 container_of(b, struct intel_engine_cs, breadcrumbs);
 217
 218         if (!test_bit(engine->id, &engine->i915->gpu_error.missed_irq_rings))
 219                 return false;
 220
 221         /* Only start with the heavy weight fake irq timer if we have not
 222          * seen any interrupts since enabling it the first time. If the
 223          * interrupts are still arriving, it means we made a mistake in our
 224          * engine->seqno_barrier(), a timing error that should be transient
 225          * and unlikely to reoccur.
 226          */
 227         return atomic_read(&engine->irq_count) == b->hangcheck_interrupts;
 228 }
 229
 230 static void enable_fake_irq(struct intel_breadcrumbs *b)
 231 {
 232         /* Ensure we never sleep indefinitely */
 233         if (!b->irq_enabled || use_fake_irq(b))
 234                 mod_timer(&b->fake_irq, jiffies + 1);
 235         else
 236                 mod_timer(&b->hangcheck, wait_timeout());
 237 }
 238
 239 static bool __intel_breadcrumbs_enable_irq(struct intel_breadcrumbs *b)
 240 {
 241         struct intel_engine_cs *engine =
 242                 container_of(b, struct intel_engine_cs, breadcrumbs);
 243         struct drm_i915_private *i915 = engine->i915;
 244
 245         lockdep_assert_held(&b->irq_lock);
 246         if (b->irq_armed)
 247                 return false;
 248
 249         /* The breadcrumb irq will be disarmed on the interrupt after the
 250          * waiters are signaled. This gives us a single interrupt window in
 251          * which we can add a new waiter and avoid the cost of re-enabling
 252          * the irq.
 253          */
 254         b->irq_armed = true;
 255         GEM_BUG_ON(b->irq_enabled);
 256
 257         if (I915_SELFTEST_ONLY(b->mock)) {
 258                 /* For our mock objects we want to avoid interaction
 259                  * with the real hardware (which is not set up). So
 260                  * we simply pretend we have enabled the powerwell
 261                  * and the irq, and leave it up to the mock
 262                  * implementation to call intel_engine_wakeup()
 263                  * itself when it wants to simulate a user interrupt,
 264                  */
 265                 return true;
 266         }
 267
 268         /* Since we are waiting on a request, the GPU should be busy
 269          * and should have its own rpm reference. This is tracked
 270          * by i915->gt.awake, we can forgo holding our own wakref
 271          * for the interrupt as before i915->gt.awake is released (when
 272          * the driver is idle) we disarm the breadcrumbs.
 273          */
 274
 275         /* No interrupts? Kick the waiter every jiffie! */
 276         if (intel_irqs_enabled(i915)) {
 277                 if (!test_bit(engine->id, &i915->gpu_error.test_irq_rings))
 278                         irq_enable(engine);
 279                 b->irq_enabled = true;
 280         }
 281
 282         enable_fake_irq(b);
 283         return true;
 284 }
 285
 286 static inline struct intel_wait *to_wait(struct rb_node *node)
 287 {
 288         return rb_entry(node, struct intel_wait, node);
 289 }
 290
 291 static inline void __intel_breadcrumbs_finish(struct intel_breadcrumbs *b,
 292                                               struct intel_wait *wait)
 293 {
 294         lockdep_assert_held(&b->rb_lock);
 295         GEM_BUG_ON(b->irq_wait == wait);
 296
 297         /* This request is completed, so remove it from the tree, mark it as
 298          * complete, and *then* wake up the associated task. N.B. when the
 299          * task wakes up, it will find the empty rb_node, discern that it
 300          * has already been removed from the tree and skip the serialisation
 301          * of the b->rb_lock and b->irq_lock. This means that the destruction
 302          * of the intel_wait is not serialised with the interrupt handler
 303          * by the waiter - it must instead be serialised by the caller.
 304          */
 305         rb_erase(&wait->node, &b->waiters);
 306         RB_CLEAR_NODE(&wait->node);
 307
 308         wake_up_process(wait->tsk); /* implicit smp_wmb() */
 309 }
 310
 311 static inline void __intel_breadcrumbs_next(struct intel_engine_cs *engine,
 312                                             struct rb_node *next)
 313 {
 314         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 315
 316         spin_lock(&b->irq_lock);
 317         GEM_BUG_ON(!b->irq_armed);
 318         GEM_BUG_ON(!b->irq_wait);
 319         b->irq_wait = to_wait(next);
 320         spin_unlock(&b->irq_lock);
 321
 322         /* We always wake up the next waiter that takes over as the bottom-half
 323          * as we may delegate not only the irq-seqno barrier to the next waiter
 324          * but also the task of waking up concurrent waiters.
 325          */
 326         if (next)
 327                 wake_up_process(to_wait(next)->tsk);
 328 }
 329
 330 static bool __intel_engine_add_wait(struct intel_engine_cs *engine,
 331                                     struct intel_wait *wait)
 332 {
 333         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 334         struct rb_node **p, *parent, *completed;
 335         bool first, armed;
 336         u32 seqno;
 337
 338         /* Insert the request into the retirement ordered list
 339          * of waiters by walking the rbtree. If we are the oldest
 340          * seqno in the tree (the first to be retired), then
 341          * set ourselves as the bottom-half.
 342          *
 343          * As we descend the tree, prune completed branches since we hold the
 344          * spinlock we know that the first_waiter must be delayed and can
 345          * reduce some of the sequential wake up latency if we take action
 346          * ourselves and wake up the completed tasks in parallel. Also, by
 347          * removing stale elements in the tree, we may be able to reduce the
 348          * ping-pong between the old bottom-half and ourselves as first-waiter.
 349          */
 350         armed = false;
 351         first = true;
 352         parent = NULL;
 353         completed = NULL;
 354         seqno = intel_engine_get_seqno(engine);
 355
 356          /* If the request completed before we managed to grab the spinlock,
 357           * return now before adding ourselves to the rbtree. We let the
 358           * current bottom-half handle any pending wakeups and instead
 359           * try and get out of the way quickly.
 360           */
 361         if (i915_seqno_passed(seqno, wait->seqno)) {
 362                 RB_CLEAR_NODE(&wait->node);
 363                 return first;
 364         }
 365
 366         p = &b->waiters.rb_node;
 367         while (*p) {
 368                 parent = *p;
 369                 if (wait->seqno == to_wait(parent)->seqno) {
 370                         /* We have multiple waiters on the same seqno, select
 371                          * the highest priority task (that with the smallest
 372                          * task->prio) to serve as the bottom-half for this
 373                          * group.
 374                          */
 375                         if (wait->tsk->prio > to_wait(parent)->tsk->prio) {
 376                                 p = &parent->rb_right;
 377                                 first = false;
 378                         } else {
 379                                 p = &parent->rb_left;
 380                         }
 381                 } else if (i915_seqno_passed(wait->seqno,
 382                                              to_wait(parent)->seqno)) {
 383                         p = &parent->rb_right;
 384                         if (i915_seqno_passed(seqno, to_wait(parent)->seqno))
 385                                 completed = parent;
 386                         else
 387                                 first = false;
 388                 } else {
 389                         p = &parent->rb_left;
 390                 }
 391         }
 392         rb_link_node(&wait->node, parent, p);
 393         rb_insert_color(&wait->node, &b->waiters);
 394
 395         if (first) {
 396                 spin_lock(&b->irq_lock);
 397                 b->irq_wait = wait;
 398                 /* After assigning ourselves as the new bottom-half, we must
 399                  * perform a cursory check to prevent a missed interrupt.
 400                  * Either we miss the interrupt whilst programming the hardware,
 401                  * or if there was a previous waiter (for a later seqno) they
 402                  * may be woken instead of us (due to the inherent race
 403                  * in the unlocked read of b->irq_seqno_bh in the irq handler)
 404                  * and so we miss the wake up.
 405                  */
 406                 armed = __intel_breadcrumbs_enable_irq(b);
 407                 spin_unlock(&b->irq_lock);
 408         }
 409
 410         if (completed) {
 411                 /* Advance the bottom-half (b->irq_wait) before we wake up
 412                  * the waiters who may scribble over their intel_wait
 413                  * just as the interrupt handler is dereferencing it via
 414                  * b->irq_wait.
 415                  */
 416                 if (!first) {
 417                         struct rb_node *next = rb_next(completed);
 418                         GEM_BUG_ON(next == &wait->node);
 419                         __intel_breadcrumbs_next(engine, next);
 420                 }
 421
 422                 do {
 423                         struct intel_wait *crumb = to_wait(completed);
 424                         completed = rb_prev(completed);
 425                         __intel_breadcrumbs_finish(b, crumb);
 426                 } while (completed);
 427         }
 428
 429         GEM_BUG_ON(!b->irq_wait);
 430         GEM_BUG_ON(!b->irq_armed);
 431         GEM_BUG_ON(rb_first(&b->waiters) != &b->irq_wait->node);
 432
 433         return armed;
 434 }
 435
 436 bool intel_engine_add_wait(struct intel_engine_cs *engine,
 437                            struct intel_wait *wait)
 438 {
 439         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 440         bool armed;
 441
 442         spin_lock_irq(&b->rb_lock);
 443         armed = __intel_engine_add_wait(engine, wait);
 444         spin_unlock_irq(&b->rb_lock);
 445         if (armed)
 446                 return armed;
 447
 448         /* Make the caller recheck if its request has already started. */
 449         return i915_seqno_passed(intel_engine_get_seqno(engine),
 450                                  wait->seqno - 1);
 451 }
 452
 453 static inline bool chain_wakeup(struct rb_node *rb, int priority)
 454 {
 455         return rb && to_wait(rb)->tsk->prio <= priority;
 456 }
 457
 458 static inline int wakeup_priority(struct intel_breadcrumbs *b,
 459                                   struct task_struct *tsk)
 460 {
 461         if (tsk == b->signaler)
 462                 return INT_MIN;
 463         else
 464                 return tsk->prio;
 465 }
 466
 467 static void __intel_engine_remove_wait(struct intel_engine_cs *engine,
 468                                        struct intel_wait *wait)
 469 {
 470         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 471
 472         lockdep_assert_held(&b->rb_lock);
 473
 474         if (RB_EMPTY_NODE(&wait->node))
 475                 goto out;
 476
 477         if (b->irq_wait == wait) {
 478                 const int priority = wakeup_priority(b, wait->tsk);
 479                 struct rb_node *next;
 480
 481                 /* We are the current bottom-half. Find the next candidate,
 482                  * the first waiter in the queue on the remaining oldest
 483                  * request. As multiple seqnos may complete in the time it
 484                  * takes us to wake up and find the next waiter, we have to
 485                  * wake up that waiter for it to perform its own coherent
 486                  * completion check.
 487                  */
 488                 next = rb_next(&wait->node);
 489                 if (chain_wakeup(next, priority)) {
 490                         /* If the next waiter is already complete,
 491                          * wake it up and continue onto the next waiter. So
 492                          * if have a small herd, they will wake up in parallel
 493                          * rather than sequentially, which should reduce
 494                          * the overall latency in waking all the completed
 495                          * clients.
 496                          *
 497                          * However, waking up a chain adds extra latency to
 498                          * the first_waiter. This is undesirable if that
 499                          * waiter is a high priority task.
 500                          */
 501                         u32 seqno = intel_engine_get_seqno(engine);
 502
 503                         while (i915_seqno_passed(seqno, to_wait(next)->seqno)) {
 504                                 struct rb_node *n = rb_next(next);
 505
 506                                 __intel_breadcrumbs_finish(b, to_wait(next));
 507                                 next = n;
 508                                 if (!chain_wakeup(next, priority))
 509                                         break;
 510                         }
 511                 }
 512
 513                 __intel_breadcrumbs_next(engine, next);
 514         } else {
 515                 GEM_BUG_ON(rb_first(&b->waiters) == &wait->node);
 516         }
 517
 518         GEM_BUG_ON(RB_EMPTY_NODE(&wait->node));
 519         rb_erase(&wait->node, &b->waiters);
 520
 521 out:
 522         GEM_BUG_ON(b->irq_wait == wait);
 523         GEM_BUG_ON(rb_first(&b->waiters) !=
 524                    (b->irq_wait ? &b->irq_wait->node : NULL));
 525 }
 526
 527 void intel_engine_remove_wait(struct intel_engine_cs *engine,
 528                               struct intel_wait *wait)
 529 {
 530         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 531
 532         /* Quick check to see if this waiter was already decoupled from
 533          * the tree by the bottom-half to avoid contention on the spinlock
 534          * by the herd.
 535          */
 536         if (RB_EMPTY_NODE(&wait->node)) {
 537                 GEM_BUG_ON(READ_ONCE(b->irq_wait) == wait);
 538                 return;
 539         }
 540
 541         spin_lock_irq(&b->rb_lock);
 542         __intel_engine_remove_wait(engine, wait);
 543         spin_unlock_irq(&b->rb_lock);
 544 }
 545
 546 static bool signal_valid(const struct drm_i915_gem_request *request)
 547 {
 548         return intel_wait_check_request(&request->signaling.wait, request);
 549 }
 550
 551 static bool signal_complete(const struct drm_i915_gem_request *request)
 552 {
 553         if (!request)
 554                 return false;
 555
 556         /* If another process served as the bottom-half it may have already
 557          * signalled that this wait is already completed.
 558          */
 559         if (intel_wait_complete(&request->signaling.wait))
 560                 return signal_valid(request);
 561
 562         /* Carefully check if the request is complete, giving time for the
 563          * seqno to be visible or if the GPU hung.
 564          */
 565         if (__i915_request_irq_complete(request))
 566                 return true;
 567
 568         return false;
 569 }
 570
 571 static struct drm_i915_gem_request *to_signaler(struct rb_node *rb)
 572 {
 573         return rb_entry(rb, struct drm_i915_gem_request, signaling.node);
 574 }
 575
 576 static void signaler_set_rtpriority(void)
 577 {
 578          struct sched_param param = { .sched_priority = 1 };
 579
 580          sched_setscheduler_nocheck(current, SCHED_FIFO, &param);
 581 }
 582
 583 static int intel_breadcrumbs_signaler(void *arg)
 584 {
 585         struct intel_engine_cs *engine = arg;
 586         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 587         struct drm_i915_gem_request *request;
 588
 589         /* Install ourselves with high priority to reduce signalling latency */
 590         signaler_set_rtpriority();
 591
 592         do {
 593                 bool do_schedule = true;
 594
 595                 set_current_state(TASK_INTERRUPTIBLE);
 596
 597                 /* We are either woken up by the interrupt bottom-half,
 598                  * or by a client adding a new signaller. In both cases,
 599                  * the GPU seqno may have advanced beyond our oldest signal.
 600                  * If it has, propagate the signal, remove the waiter and
 601                  * check again with the next oldest signal. Otherwise we
 602                  * need to wait for a new interrupt from the GPU or for
 603                  * a new client.
 604                  */
 605                 rcu_read_lock();
 606                 request = rcu_dereference(b->first_signal);
 607                 if (request)
 608                         request = i915_gem_request_get_rcu(request);
 609                 rcu_read_unlock();
 610                 if (signal_complete(request)) {
 611                         local_bh_disable();
 612                         dma_fence_signal(&request->fence);
 613                         local_bh_enable(); /* kick start the tasklets */
 614
 615                         spin_lock_irq(&b->rb_lock);
 616
 617                         /* Wake up all other completed waiters and select the
 618                          * next bottom-half for the next user interrupt.
 619                          */
 620                         __intel_engine_remove_wait(engine,
 621                                                    &request->signaling.wait);
 622
 623                         /* Find the next oldest signal. Note that as we have
 624                          * not been holding the lock, another client may
 625                          * have installed an even older signal than the one
 626                          * we just completed - so double check we are still
 627                          * the oldest before picking the next one.
 628                          */
 629                         if (request == rcu_access_pointer(b->first_signal)) {
 630                                 struct rb_node *rb =
 631                                         rb_next(&request->signaling.node);
 632                                 rcu_assign_pointer(b->first_signal,
 633                                                    rb ? to_signaler(rb) : NULL);
 634                         }
 635                         rb_erase(&request->signaling.node, &b->signals);
 636                         RB_CLEAR_NODE(&request->signaling.node);
 637
 638                         spin_unlock_irq(&b->rb_lock);
 639
 640                         i915_gem_request_put(request);
 641
 642                         /* If the engine is saturated we may be continually
 643                          * processing completed requests. This angers the
 644                          * NMI watchdog if we never let anything else
 645                          * have access to the CPU. Let's pretend to be nice
 646                          * and relinquish the CPU if we burn through the
 647                          * entire RT timeslice!
 648                          */
 649                         do_schedule = need_resched();
 650                 }
 651
 652                 if (unlikely(do_schedule)) {
 653                         DEFINE_WAIT(exec);
 654
 655                         if (kthread_should_park())
 656                                 kthread_parkme();
 657
 658                         if (kthread_should_stop()) {
 659                                 GEM_BUG_ON(request);
 660                                 break;
 661                         }
 662
 663                         if (request)
 664                                 add_wait_queue(&request->execute, &exec);
 665
 666                         schedule();
 667
 668                         if (request)
 669                                 remove_wait_queue(&request->execute, &exec);
 670                 }
 671                 i915_gem_request_put(request);
 672         } while (1);
 673         __set_current_state(TASK_RUNNING);
 674
 675         return 0;
 676 }
 677
 678 void intel_engine_enable_signaling(struct drm_i915_gem_request *request,
 679                                    bool wakeup)
 680 {
 681         struct intel_engine_cs *engine = request->engine;
 682         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 683         u32 seqno;
 684
 685         /* Note that we may be called from an interrupt handler on another
 686          * device (e.g. nouveau signaling a fence completion causing us
 687          * to submit a request, and so enable signaling). As such,
 688          * we need to make sure that all other users of b->rb_lock protect
 689          * against interrupts, i.e. use spin_lock_irqsave.
 690          */
 691
 692         /* locked by dma_fence_enable_sw_signaling() (irqsafe fence->lock) */
 693         GEM_BUG_ON(!irqs_disabled());
 694         lockdep_assert_held(&request->lock);
 695
 696         seqno = i915_gem_request_global_seqno(request);
 697         if (!seqno)
 698                 return;
 699
 700         request->signaling.wait.tsk = b->signaler;
 701         request->signaling.wait.request = request;
 702         request->signaling.wait.seqno = seqno;
 703         i915_gem_request_get(request);
 704
 705         spin_lock(&b->rb_lock);
 706
 707         /* First add ourselves into the list of waiters, but register our
 708          * bottom-half as the signaller thread. As per usual, only the oldest
 709          * waiter (not just signaller) is tasked as the bottom-half waking
 710          * up all completed waiters after the user interrupt.
 711          *
 712          * If we are the oldest waiter, enable the irq (after which we
 713          * must double check that the seqno did not complete).
 714          */
 715         wakeup &= __intel_engine_add_wait(engine, &request->signaling.wait);
 716
 717         if (!__i915_gem_request_completed(request, seqno)) {
 718                 struct rb_node *parent, **p;
 719                 bool first;
 720
 721                 /* Now insert ourselves into the retirement ordered list of
 722                  * signals on this engine. We track the oldest seqno as that
 723                  * will be the first signal to complete.
 724                  */
 725                 parent = NULL;
 726                 first = true;
 727                 p = &b->signals.rb_node;
 728                 while (*p) {
 729                         parent = *p;
 730                         if (i915_seqno_passed(seqno,
 731                                               to_signaler(parent)->signaling.wait.seqno)) {
 732                                 p = &parent->rb_right;
 733                                 first = false;
 734                         } else {
 735                                 p = &parent->rb_left;
 736                         }
 737                 }
 738                 rb_link_node(&request->signaling.node, parent, p);
 739                 rb_insert_color(&request->signaling.node, &b->signals);
 740                 if (first)
 741                         rcu_assign_pointer(b->first_signal, request);
 742         } else {
 743                 __intel_engine_remove_wait(engine, &request->signaling.wait);
 744                 i915_gem_request_put(request);
 745                 wakeup = false;
 746         }
 747
 748         spin_unlock(&b->rb_lock);
 749
 750         if (wakeup)
 751                 wake_up_process(b->signaler);
 752 }
 753
 754 void intel_engine_cancel_signaling(struct drm_i915_gem_request *request)
 755 {
 756         struct intel_engine_cs *engine = request->engine;
 757         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 758
 759         GEM_BUG_ON(!irqs_disabled());
 760         lockdep_assert_held(&request->lock);
 761         GEM_BUG_ON(!request->signaling.wait.seqno);
 762
 763         spin_lock(&b->rb_lock);
 764
 765         if (!RB_EMPTY_NODE(&request->signaling.node)) {
 766                 if (request == rcu_access_pointer(b->first_signal)) {
 767                         struct rb_node *rb =
 768                                 rb_next(&request->signaling.node);
 769                         rcu_assign_pointer(b->first_signal,
 770                                            rb ? to_signaler(rb) : NULL);
 771                 }
 772                 rb_erase(&request->signaling.node, &b->signals);
 773                 RB_CLEAR_NODE(&request->signaling.node);
 774                 i915_gem_request_put(request);
 775         }
 776
 777         __intel_engine_remove_wait(engine, &request->signaling.wait);
 778
 779         spin_unlock(&b->rb_lock);
 780
 781         request->signaling.wait.seqno = 0;
 782 }
 783
 784 int intel_engine_init_breadcrumbs(struct intel_engine_cs *engine)
 785 {
 786         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 787         struct task_struct *tsk;
 788
 789         spin_lock_init(&b->rb_lock);
 790         spin_lock_init(&b->irq_lock);
 791
 792         timer_setup(&b->fake_irq, intel_breadcrumbs_fake_irq, 0);
 793         timer_setup(&b->hangcheck, intel_breadcrumbs_hangcheck, 0);
 794
 795         /* Spawn a thread to provide a common bottom-half for all signals.
 796          * As this is an asynchronous interface we cannot steal the current
 797          * task for handling the bottom-half to the user interrupt, therefore
 798          * we create a thread to do the coherent seqno dance after the
 799          * interrupt and then signal the waitqueue (via the dma-buf/fence).
 800          */
 801         tsk = kthread_run(intel_breadcrumbs_signaler, engine,
 802                           "i915/signal:%d", engine->id);
 803         if (IS_ERR(tsk))
 804                 return PTR_ERR(tsk);
 805
 806         b->signaler = tsk;
 807
 808         return 0;
 809 }
 810
 811 static void cancel_fake_irq(struct intel_engine_cs *engine)
 812 {
 813         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 814
 815         del_timer_sync(&b->hangcheck);
 816         del_timer_sync(&b->fake_irq);
 817         clear_bit(engine->id, &engine->i915->gpu_error.missed_irq_rings);
 818 }
 819
 820 void intel_engine_reset_breadcrumbs(struct intel_engine_cs *engine)
 821 {
 822         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 823
 824         cancel_fake_irq(engine);
 825         spin_lock_irq(&b->irq_lock);
 826
 827         if (b->irq_enabled)
 828                 irq_enable(engine);
 829         else
 830                 irq_disable(engine);
 831
 832         /* We set the IRQ_BREADCRUMB bit when we enable the irq presuming the
 833          * GPU is active and may have already executed the MI_USER_INTERRUPT
 834          * before the CPU is ready to receive. However, the engine is currently
 835          * idle (we haven't started it yet), there is no possibility for a
 836          * missed interrupt as we enabled the irq and so we can clear the
 837          * immediate wakeup (until a real interrupt arrives for the waiter).
 838          */
 839         clear_bit(ENGINE_IRQ_BREADCRUMB, &engine->irq_posted);
 840
 841         if (b->irq_armed)
 842                 enable_fake_irq(b);
 843
 844         spin_unlock_irq(&b->irq_lock);
 845 }
 846
 847 void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine)
 848 {
 849         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 850
 851         /* The engines should be idle and all requests accounted for! */
 852         WARN_ON(READ_ONCE(b->irq_wait));
 853         WARN_ON(!RB_EMPTY_ROOT(&b->waiters));
 854         WARN_ON(rcu_access_pointer(b->first_signal));
 855         WARN_ON(!RB_EMPTY_ROOT(&b->signals));
 856
 857         if (!IS_ERR_OR_NULL(b->signaler))
 858                 kthread_stop(b->signaler);
 859
 860         cancel_fake_irq(engine);
 861 }
 862
 863 bool intel_breadcrumbs_busy(struct intel_engine_cs *engine)
 864 {
 865         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 866         bool busy = false;
 867
 868         spin_lock_irq(&b->rb_lock);
 869
 870         if (b->irq_wait) {
 871                 wake_up_process(b->irq_wait->tsk);
 872                 busy = true;
 873         }
 874
 875         if (rcu_access_pointer(b->first_signal)) {
 876                 wake_up_process(b->signaler);
 877                 busy = true;
 878         }
 879
 880         spin_unlock_irq(&b->rb_lock);
 881
 882         return busy;
 883 }
 884
 885 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
 886 #include "selftests/intel_breadcrumbs.c"
 887 #endif