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