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