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, 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(unsigned long data)
  78 {
  79         struct intel_engine_cs *engine = (struct intel_engine_cs *)data;
  80         struct intel_breadcrumbs *b = &engine->breadcrumbs;
  81
  82         if (!b->irq_armed)
  83                 return;
  84
  85         if (b->hangcheck_interrupts != atomic_read(&engine->irq_count)) {
  86                 b->hangcheck_interrupts = atomic_read(&engine->irq_count);
  87                 mod_timer(&b->hangcheck, wait_timeout());
  88                 return;
  89         }
  90
  91         /* We keep the hangcheck timer alive until we disarm the irq, even
  92          * if there are no waiters at present.
  93          *
  94          * If the waiter was currently running, assume it hasn't had a chance
  95          * to process the pending interrupt (e.g, low priority task on a loaded
  96          * system) and wait until it sleeps before declaring a missed interrupt.
  97          *
  98          * If the waiter was asleep (and not even pending a wakeup), then we
  99          * must have missed an interrupt as the GPU has stopped advancing
 100          * but we still have a waiter. Assuming all batches complete within
 101          * DRM_I915_HANGCHECK_JIFFIES [1.5s]!
 102          */
 103         if (intel_engine_wakeup(engine) & ENGINE_WAKEUP_ASLEEP) {
 104                 missed_breadcrumb(engine);
 105                 mod_timer(&engine->breadcrumbs.fake_irq, jiffies + 1);
 106         } else {
 107                 mod_timer(&b->hangcheck, wait_timeout());
 108         }
 109 }
 110
 111 static void intel_breadcrumbs_fake_irq(unsigned long data)
 112 {
 113         struct intel_engine_cs *engine = (struct intel_engine_cs *)data;
 114         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 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, better known as a "missed breadcrumb").
 119          * Here the worker will wake up every jiffie in order to kick the
 120          * oldest waiter to do the coherent seqno check.
 121          */
 122
 123         spin_lock_irq(&b->irq_lock);
 124         if (!__intel_breadcrumbs_wakeup(b))
 125                 __intel_engine_disarm_breadcrumbs(engine);
 126         spin_unlock_irq(&b->irq_lock);
 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         GEM_BUG_ON(b->irq_wait);
 172
 173         if (b->irq_enabled) {
 174                 irq_disable(engine);
 175                 b->irq_enabled = false;
 176         }
 177
 178         b->irq_armed = false;
 179 }
 180
 181 void intel_engine_disarm_breadcrumbs(struct intel_engine_cs *engine)
 182 {
 183         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 184         struct intel_wait *wait, *n, *first;
 185
 186         if (!b->irq_armed)
 187                 return;
 188
 189         /* We only disarm the irq when we are idle (all requests completed),
 190          * so if the bottom-half remains asleep, it missed the request
 191          * completion.
 192          */
 193
 194         spin_lock_irq(&b->rb_lock);
 195
 196         spin_lock(&b->irq_lock);
 197         first = fetch_and_zero(&b->irq_wait);
 198         __intel_engine_disarm_breadcrumbs(engine);
 199         spin_unlock(&b->irq_lock);
 200
 201         rbtree_postorder_for_each_entry_safe(wait, n, &b->waiters, node) {
 202                 RB_CLEAR_NODE(&wait->node);
 203                 if (wake_up_process(wait->tsk) && wait == first)
 204                         missed_breadcrumb(engine);
 205         }
 206         b->waiters = RB_ROOT;
 207
 208         spin_unlock_irq(&b->rb_lock);
 209 }
 210
 211 static bool use_fake_irq(const struct intel_breadcrumbs *b)
 212 {
 213         const struct intel_engine_cs *engine =
 214                 container_of(b, struct intel_engine_cs, breadcrumbs);
 215
 216         if (!test_bit(engine->id, &engine->i915->gpu_error.missed_irq_rings))
 217                 return false;
 218
 219         /* Only start with the heavy weight fake irq timer if we have not
 220          * seen any interrupts since enabling it the first time. If the
 221          * interrupts are still arriving, it means we made a mistake in our
 222          * engine->seqno_barrier(), a timing error that should be transient
 223          * and unlikely to reoccur.
 224          */
 225         return atomic_read(&engine->irq_count) == b->hangcheck_interrupts;
 226 }
 227
 228 static void enable_fake_irq(struct intel_breadcrumbs *b)
 229 {
 230         /* Ensure we never sleep indefinitely */
 231         if (!b->irq_enabled || use_fake_irq(b))
 232                 mod_timer(&b->fake_irq, jiffies + 1);
 233         else
 234                 mod_timer(&b->hangcheck, wait_timeout());
 235 }
 236
 237 static bool __intel_breadcrumbs_enable_irq(struct intel_breadcrumbs *b)
 238 {
 239         struct intel_engine_cs *engine =
 240                 container_of(b, struct intel_engine_cs, breadcrumbs);
 241         struct drm_i915_private *i915 = engine->i915;
 242
 243         lockdep_assert_held(&b->irq_lock);
 244         if (b->irq_armed)
 245                 return false;
 246
 247         /* The breadcrumb irq will be disarmed on the interrupt after the
 248          * waiters are signaled. This gives us a single interrupt window in
 249          * which we can add a new waiter and avoid the cost of re-enabling
 250          * the irq.
 251          */
 252         b->irq_armed = true;
 253         GEM_BUG_ON(b->irq_enabled);
 254
 255         if (I915_SELFTEST_ONLY(b->mock)) {
 256                 /* For our mock objects we want to avoid interaction
 257                  * with the real hardware (which is not set up). So
 258                  * we simply pretend we have enabled the powerwell
 259                  * and the irq, and leave it up to the mock
 260                  * implementation to call intel_engine_wakeup()
 261                  * itself when it wants to simulate a user interrupt,
 262                  */
 263                 return true;
 264         }
 265
 266         /* Since we are waiting on a request, the GPU should be busy
 267          * and should have its own rpm reference. This is tracked
 268          * by i915->gt.awake, we can forgo holding our own wakref
 269          * for the interrupt as before i915->gt.awake is released (when
 270          * the driver is idle) we disarm the breadcrumbs.
 271          */
 272
 273         /* No interrupts? Kick the waiter every jiffie! */
 274         if (intel_irqs_enabled(i915)) {
 275                 if (!test_bit(engine->id, &i915->gpu_error.test_irq_rings))
 276                         irq_enable(engine);
 277                 b->irq_enabled = true;
 278         }
 279
 280         enable_fake_irq(b);
 281         return true;
 282 }
 283
 284 static inline struct intel_wait *to_wait(struct rb_node *node)
 285 {
 286         return rb_entry(node, struct intel_wait, node);
 287 }
 288
 289 static inline void __intel_breadcrumbs_finish(struct intel_breadcrumbs *b,
 290                                               struct intel_wait *wait)
 291 {
 292         lockdep_assert_held(&b->rb_lock);
 293         GEM_BUG_ON(b->irq_wait == wait);
 294
 295         /* This request is completed, so remove it from the tree, mark it as
 296          * complete, and *then* wake up the associated task. N.B. when the
 297          * task wakes up, it will find the empty rb_node, discern that it
 298          * has already been removed from the tree and skip the serialisation
 299          * of the b->rb_lock and b->irq_lock. This means that the destruction
 300          * of the intel_wait is not serialised with the interrupt handler
 301          * by the waiter - it must instead be serialised by the caller.
 302          */
 303         rb_erase(&wait->node, &b->waiters);
 304         RB_CLEAR_NODE(&wait->node);
 305
 306         wake_up_process(wait->tsk); /* implicit smp_wmb() */
 307 }
 308
 309 static inline void __intel_breadcrumbs_next(struct intel_engine_cs *engine,
 310                                             struct rb_node *next)
 311 {
 312         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 313
 314         spin_lock(&b->irq_lock);
 315         GEM_BUG_ON(!b->irq_armed);
 316         GEM_BUG_ON(!b->irq_wait);
 317         b->irq_wait = to_wait(next);
 318         spin_unlock(&b->irq_lock);
 319
 320         /* We always wake up the next waiter that takes over as the bottom-half
 321          * as we may delegate not only the irq-seqno barrier to the next waiter
 322          * but also the task of waking up concurrent waiters.
 323          */
 324         if (next)
 325                 wake_up_process(to_wait(next)->tsk);
 326 }
 327
 328 static bool __intel_engine_add_wait(struct intel_engine_cs *engine,
 329                                     struct intel_wait *wait)
 330 {
 331         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 332         struct rb_node **p, *parent, *completed;
 333         bool first, armed;
 334         u32 seqno;
 335
 336         /* Insert the request into the retirement ordered list
 337          * of waiters by walking the rbtree. If we are the oldest
 338          * seqno in the tree (the first to be retired), then
 339          * set ourselves as the bottom-half.
 340          *
 341          * As we descend the tree, prune completed branches since we hold the
 342          * spinlock we know that the first_waiter must be delayed and can
 343          * reduce some of the sequential wake up latency if we take action
 344          * ourselves and wake up the completed tasks in parallel. Also, by
 345          * removing stale elements in the tree, we may be able to reduce the
 346          * ping-pong between the old bottom-half and ourselves as first-waiter.
 347          */
 348         armed = false;
 349         first = true;
 350         parent = NULL;
 351         completed = NULL;
 352         seqno = intel_engine_get_seqno(engine);
 353
 354          /* If the request completed before we managed to grab the spinlock,
 355           * return now before adding ourselves to the rbtree. We let the
 356           * current bottom-half handle any pending wakeups and instead
 357           * try and get out of the way quickly.
 358           */
 359         if (i915_seqno_passed(seqno, wait->seqno)) {
 360                 RB_CLEAR_NODE(&wait->node);
 361                 return first;
 362         }
 363
 364         p = &b->waiters.rb_node;
 365         while (*p) {
 366                 parent = *p;
 367                 if (wait->seqno == to_wait(parent)->seqno) {
 368                         /* We have multiple waiters on the same seqno, select
 369                          * the highest priority task (that with the smallest
 370                          * task->prio) to serve as the bottom-half for this
 371                          * group.
 372                          */
 373                         if (wait->tsk->prio > to_wait(parent)->tsk->prio) {
 374                                 p = &parent->rb_right;
 375                                 first = false;
 376                         } else {
 377                                 p = &parent->rb_left;
 378                         }
 379                 } else if (i915_seqno_passed(wait->seqno,
 380                                              to_wait(parent)->seqno)) {
 381                         p = &parent->rb_right;
 382                         if (i915_seqno_passed(seqno, to_wait(parent)->seqno))
 383                                 completed = parent;
 384                         else
 385                                 first = false;
 386                 } else {
 387                         p = &parent->rb_left;
 388                 }
 389         }
 390         rb_link_node(&wait->node, parent, p);
 391         rb_insert_color(&wait->node, &b->waiters);
 392
 393         if (first) {
 394                 spin_lock(&b->irq_lock);
 395                 b->irq_wait = wait;
 396                 /* After assigning ourselves as the new bottom-half, we must
 397                  * perform a cursory check to prevent a missed interrupt.
 398                  * Either we miss the interrupt whilst programming the hardware,
 399                  * or if there was a previous waiter (for a later seqno) they
 400                  * may be woken instead of us (due to the inherent race
 401                  * in the unlocked read of b->irq_seqno_bh in the irq handler)
 402                  * and so we miss the wake up.
 403                  */
 404                 armed = __intel_breadcrumbs_enable_irq(b);
 405                 spin_unlock(&b->irq_lock);
 406         }
 407
 408         if (completed) {
 409                 /* Advance the bottom-half (b->irq_wait) before we wake up
 410                  * the waiters who may scribble over their intel_wait
 411                  * just as the interrupt handler is dereferencing it via
 412                  * b->irq_wait.
 413                  */
 414                 if (!first) {
 415                         struct rb_node *next = rb_next(completed);
 416                         GEM_BUG_ON(next == &wait->node);
 417                         __intel_breadcrumbs_next(engine, next);
 418                 }
 419
 420                 do {
 421                         struct intel_wait *crumb = to_wait(completed);
 422                         completed = rb_prev(completed);
 423                         __intel_breadcrumbs_finish(b, crumb);
 424                 } while (completed);
 425         }
 426
 427         GEM_BUG_ON(!b->irq_wait);
 428         GEM_BUG_ON(!b->irq_armed);
 429         GEM_BUG_ON(rb_first(&b->waiters) != &b->irq_wait->node);
 430
 431         return armed;
 432 }
 433
 434 bool intel_engine_add_wait(struct intel_engine_cs *engine,
 435                            struct intel_wait *wait)
 436 {
 437         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 438         bool armed;
 439
 440         spin_lock_irq(&b->rb_lock);
 441         armed = __intel_engine_add_wait(engine, wait);
 442         spin_unlock_irq(&b->rb_lock);
 443         if (armed)
 444                 return armed;
 445
 446         /* Make the caller recheck if its request has already started. */
 447         return i915_seqno_passed(intel_engine_get_seqno(engine),
 448                                  wait->seqno - 1);
 449 }
 450
 451 static inline bool chain_wakeup(struct rb_node *rb, int priority)
 452 {
 453         return rb && to_wait(rb)->tsk->prio <= priority;
 454 }
 455
 456 static inline int wakeup_priority(struct intel_breadcrumbs *b,
 457                                   struct task_struct *tsk)
 458 {
 459         if (tsk == b->signaler)
 460                 return INT_MIN;
 461         else
 462                 return tsk->prio;
 463 }
 464
 465 static void __intel_engine_remove_wait(struct intel_engine_cs *engine,
 466                                        struct intel_wait *wait)
 467 {
 468         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 469
 470         lockdep_assert_held(&b->rb_lock);
 471
 472         if (RB_EMPTY_NODE(&wait->node))
 473                 goto out;
 474
 475         if (b->irq_wait == wait) {
 476                 const int priority = wakeup_priority(b, wait->tsk);
 477                 struct rb_node *next;
 478
 479                 /* We are the current bottom-half. Find the next candidate,
 480                  * the first waiter in the queue on the remaining oldest
 481                  * request. As multiple seqnos may complete in the time it
 482                  * takes us to wake up and find the next waiter, we have to
 483                  * wake up that waiter for it to perform its own coherent
 484                  * completion check.
 485                  */
 486                 next = rb_next(&wait->node);
 487                 if (chain_wakeup(next, priority)) {
 488                         /* If the next waiter is already complete,
 489                          * wake it up and continue onto the next waiter. So
 490                          * if have a small herd, they will wake up in parallel
 491                          * rather than sequentially, which should reduce
 492                          * the overall latency in waking all the completed
 493                          * clients.
 494                          *
 495                          * However, waking up a chain adds extra latency to
 496                          * the first_waiter. This is undesirable if that
 497                          * waiter is a high priority task.
 498                          */
 499                         u32 seqno = intel_engine_get_seqno(engine);
 500
 501                         while (i915_seqno_passed(seqno, to_wait(next)->seqno)) {
 502                                 struct rb_node *n = rb_next(next);
 503
 504                                 __intel_breadcrumbs_finish(b, to_wait(next));
 505                                 next = n;
 506                                 if (!chain_wakeup(next, priority))
 507                                         break;
 508                         }
 509                 }
 510
 511                 __intel_breadcrumbs_next(engine, next);
 512         } else {
 513                 GEM_BUG_ON(rb_first(&b->waiters) == &wait->node);
 514         }
 515
 516         GEM_BUG_ON(RB_EMPTY_NODE(&wait->node));
 517         rb_erase(&wait->node, &b->waiters);
 518
 519 out:
 520         GEM_BUG_ON(b->irq_wait == wait);
 521         GEM_BUG_ON(rb_first(&b->waiters) !=
 522                    (b->irq_wait ? &b->irq_wait->node : NULL));
 523 }
 524
 525 void intel_engine_remove_wait(struct intel_engine_cs *engine,
 526                               struct intel_wait *wait)
 527 {
 528         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 529
 530         /* Quick check to see if this waiter was already decoupled from
 531          * the tree by the bottom-half to avoid contention on the spinlock
 532          * by the herd.
 533          */
 534         if (RB_EMPTY_NODE(&wait->node)) {
 535                 GEM_BUG_ON(READ_ONCE(b->irq_wait) == wait);
 536                 return;
 537         }
 538
 539         spin_lock_irq(&b->rb_lock);
 540         __intel_engine_remove_wait(engine, wait);
 541         spin_unlock_irq(&b->rb_lock);
 542 }
 543
 544 static bool signal_complete(const struct drm_i915_gem_request *request)
 545 {
 546         if (!request)
 547                 return false;
 548
 549         /*
 550          * Carefully check if the request is complete, giving time for the
 551          * seqno to be visible or if the GPU hung.
 552          */
 553         return __i915_request_irq_complete(request);
 554 }
 555
 556 static struct drm_i915_gem_request *to_signaler(struct rb_node *rb)
 557 {
 558         return rb_entry(rb, struct drm_i915_gem_request, signaling.node);
 559 }
 560
 561 static void signaler_set_rtpriority(void)
 562 {
 563          struct sched_param param = { .sched_priority = 1 };
 564
 565          sched_setscheduler_nocheck(current, SCHED_FIFO, &param);
 566 }
 567
 568 static int intel_breadcrumbs_signaler(void *arg)
 569 {
 570         struct intel_engine_cs *engine = arg;
 571         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 572         struct drm_i915_gem_request *request;
 573
 574         /* Install ourselves with high priority to reduce signalling latency */
 575         signaler_set_rtpriority();
 576
 577         do {
 578                 bool do_schedule = true;
 579
 580                 set_current_state(TASK_INTERRUPTIBLE);
 581
 582                 /* We are either woken up by the interrupt bottom-half,
 583                  * or by a client adding a new signaller. In both cases,
 584                  * the GPU seqno may have advanced beyond our oldest signal.
 585                  * If it has, propagate the signal, remove the waiter and
 586                  * check again with the next oldest signal. Otherwise we
 587                  * need to wait for a new interrupt from the GPU or for
 588                  * a new client.
 589                  */
 590                 rcu_read_lock();
 591                 request = rcu_dereference(b->first_signal);
 592                 if (request)
 593                         request = i915_gem_request_get_rcu(request);
 594                 rcu_read_unlock();
 595                 if (signal_complete(request)) {
 596                         if (!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
 597                                       &request->fence.flags)) {
 598                                 local_bh_disable();
 599                                 dma_fence_signal(&request->fence);
 600                                 GEM_BUG_ON(!i915_gem_request_completed(request));
 601                                 local_bh_enable(); /* kick start the tasklets */
 602                         }
 603
 604                         spin_lock_irq(&b->rb_lock);
 605
 606                         /* Wake up all other completed waiters and select the
 607                          * next bottom-half for the next user interrupt.
 608                          */
 609                         __intel_engine_remove_wait(engine,
 610                                                    &request->signaling.wait);
 611
 612                         /* Find the next oldest signal. Note that as we have
 613                          * not been holding the lock, another client may
 614                          * have installed an even older signal than the one
 615                          * we just completed - so double check we are still
 616                          * the oldest before picking the next one.
 617                          */
 618                         if (request == rcu_access_pointer(b->first_signal)) {
 619                                 struct rb_node *rb =
 620                                         rb_next(&request->signaling.node);
 621                                 rcu_assign_pointer(b->first_signal,
 622                                                    rb ? to_signaler(rb) : NULL);
 623                         }
 624                         rb_erase(&request->signaling.node, &b->signals);
 625                         RB_CLEAR_NODE(&request->signaling.node);
 626
 627                         spin_unlock_irq(&b->rb_lock);
 628
 629                         i915_gem_request_put(request);
 630
 631                         /* If the engine is saturated we may be continually
 632                          * processing completed requests. This angers the
 633                          * NMI watchdog if we never let anything else
 634                          * have access to the CPU. Let's pretend to be nice
 635                          * and relinquish the CPU if we burn through the
 636                          * entire RT timeslice!
 637                          */
 638                         do_schedule = need_resched();
 639                 }
 640
 641                 if (unlikely(do_schedule)) {
 642                         DEFINE_WAIT(exec);
 643
 644                         if (kthread_should_park())
 645                                 kthread_parkme();
 646
 647                         if (kthread_should_stop()) {
 648                                 GEM_BUG_ON(request);
 649                                 break;
 650                         }
 651
 652                         if (request)
 653                                 add_wait_queue(&request->execute, &exec);
 654
 655                         schedule();
 656
 657                         if (request)
 658                                 remove_wait_queue(&request->execute, &exec);
 659                 }
 660                 i915_gem_request_put(request);
 661         } while (1);
 662         __set_current_state(TASK_RUNNING);
 663
 664         return 0;
 665 }
 666
 667 void intel_engine_enable_signaling(struct drm_i915_gem_request *request,
 668                                    bool wakeup)
 669 {
 670         struct intel_engine_cs *engine = request->engine;
 671         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 672         u32 seqno;
 673
 674         /* Note that we may be called from an interrupt handler on another
 675          * device (e.g. nouveau signaling a fence completion causing us
 676          * to submit a request, and so enable signaling). As such,
 677          * we need to make sure that all other users of b->rb_lock protect
 678          * against interrupts, i.e. use spin_lock_irqsave.
 679          */
 680
 681         /* locked by dma_fence_enable_sw_signaling() (irqsafe fence->lock) */
 682         GEM_BUG_ON(!irqs_disabled());
 683         lockdep_assert_held(&request->lock);
 684
 685         seqno = i915_gem_request_global_seqno(request);
 686         if (!seqno)
 687                 return;
 688
 689         request->signaling.wait.tsk = b->signaler;
 690         request->signaling.wait.request = request;
 691         request->signaling.wait.seqno = seqno;
 692         i915_gem_request_get(request);
 693
 694         spin_lock(&b->rb_lock);
 695
 696         /* First add ourselves into the list of waiters, but register our
 697          * bottom-half as the signaller thread. As per usual, only the oldest
 698          * waiter (not just signaller) is tasked as the bottom-half waking
 699          * up all completed waiters after the user interrupt.
 700          *
 701          * If we are the oldest waiter, enable the irq (after which we
 702          * must double check that the seqno did not complete).
 703          */
 704         wakeup &= __intel_engine_add_wait(engine, &request->signaling.wait);
 705
 706         if (!__i915_gem_request_completed(request, seqno)) {
 707                 struct rb_node *parent, **p;
 708                 bool first;
 709
 710                 /* Now insert ourselves into the retirement ordered list of
 711                  * signals on this engine. We track the oldest seqno as that
 712                  * will be the first signal to complete.
 713                  */
 714                 parent = NULL;
 715                 first = true;
 716                 p = &b->signals.rb_node;
 717                 while (*p) {
 718                         parent = *p;
 719                         if (i915_seqno_passed(seqno,
 720                                               to_signaler(parent)->signaling.wait.seqno)) {
 721                                 p = &parent->rb_right;
 722                                 first = false;
 723                         } else {
 724                                 p = &parent->rb_left;
 725                         }
 726                 }
 727                 rb_link_node(&request->signaling.node, parent, p);
 728                 rb_insert_color(&request->signaling.node, &b->signals);
 729                 if (first)
 730                         rcu_assign_pointer(b->first_signal, request);
 731         } else {
 732                 __intel_engine_remove_wait(engine, &request->signaling.wait);
 733                 i915_gem_request_put(request);
 734                 wakeup = false;
 735         }
 736
 737         spin_unlock(&b->rb_lock);
 738
 739         if (wakeup)
 740                 wake_up_process(b->signaler);
 741 }
 742
 743 void intel_engine_cancel_signaling(struct drm_i915_gem_request *request)
 744 {
 745         struct intel_engine_cs *engine = request->engine;
 746         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 747
 748         GEM_BUG_ON(!irqs_disabled());
 749         lockdep_assert_held(&request->lock);
 750         GEM_BUG_ON(!request->signaling.wait.seqno);
 751
 752         spin_lock(&b->rb_lock);
 753
 754         if (!RB_EMPTY_NODE(&request->signaling.node)) {
 755                 if (request == rcu_access_pointer(b->first_signal)) {
 756                         struct rb_node *rb =
 757                                 rb_next(&request->signaling.node);
 758                         rcu_assign_pointer(b->first_signal,
 759                                            rb ? to_signaler(rb) : NULL);
 760                 }
 761                 rb_erase(&request->signaling.node, &b->signals);
 762                 RB_CLEAR_NODE(&request->signaling.node);
 763                 i915_gem_request_put(request);
 764         }
 765
 766         __intel_engine_remove_wait(engine, &request->signaling.wait);
 767
 768         spin_unlock(&b->rb_lock);
 769
 770         request->signaling.wait.seqno = 0;
 771 }
 772
 773 int intel_engine_init_breadcrumbs(struct intel_engine_cs *engine)
 774 {
 775         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 776         struct task_struct *tsk;
 777
 778         spin_lock_init(&b->rb_lock);
 779         spin_lock_init(&b->irq_lock);
 780
 781         setup_timer(&b->fake_irq,
 782                     intel_breadcrumbs_fake_irq,
 783                     (unsigned long)engine);
 784         setup_timer(&b->hangcheck,
 785                     intel_breadcrumbs_hangcheck,
 786                     (unsigned long)engine);
 787
 788         /* Spawn a thread to provide a common bottom-half for all signals.
 789          * As this is an asynchronous interface we cannot steal the current
 790          * task for handling the bottom-half to the user interrupt, therefore
 791          * we create a thread to do the coherent seqno dance after the
 792          * interrupt and then signal the waitqueue (via the dma-buf/fence).
 793          */
 794         tsk = kthread_run(intel_breadcrumbs_signaler, engine,
 795                           "i915/signal:%d", engine->id);
 796         if (IS_ERR(tsk))
 797                 return PTR_ERR(tsk);
 798
 799         b->signaler = tsk;
 800
 801         return 0;
 802 }
 803
 804 static void cancel_fake_irq(struct intel_engine_cs *engine)
 805 {
 806         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 807
 808         del_timer_sync(&b->hangcheck);
 809         del_timer_sync(&b->fake_irq);
 810         clear_bit(engine->id, &engine->i915->gpu_error.missed_irq_rings);
 811 }
 812
 813 void intel_engine_reset_breadcrumbs(struct intel_engine_cs *engine)
 814 {
 815         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 816
 817         cancel_fake_irq(engine);
 818         spin_lock_irq(&b->irq_lock);
 819
 820         if (b->irq_enabled)
 821                 irq_enable(engine);
 822         else
 823                 irq_disable(engine);
 824
 825         /* We set the IRQ_BREADCRUMB bit when we enable the irq presuming the
 826          * GPU is active and may have already executed the MI_USER_INTERRUPT
 827          * before the CPU is ready to receive. However, the engine is currently
 828          * idle (we haven't started it yet), there is no possibility for a
 829          * missed interrupt as we enabled the irq and so we can clear the
 830          * immediate wakeup (until a real interrupt arrives for the waiter).
 831          */
 832         clear_bit(ENGINE_IRQ_BREADCRUMB, &engine->irq_posted);
 833
 834         if (b->irq_armed)
 835                 enable_fake_irq(b);
 836
 837         spin_unlock_irq(&b->irq_lock);
 838 }
 839
 840 void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine)
 841 {
 842         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 843
 844         /* The engines should be idle and all requests accounted for! */
 845         WARN_ON(READ_ONCE(b->irq_wait));
 846         WARN_ON(!RB_EMPTY_ROOT(&b->waiters));
 847         WARN_ON(rcu_access_pointer(b->first_signal));
 848         WARN_ON(!RB_EMPTY_ROOT(&b->signals));
 849
 850         if (!IS_ERR_OR_NULL(b->signaler))
 851                 kthread_stop(b->signaler);
 852
 853         cancel_fake_irq(engine);
 854 }
 855
 856 bool intel_breadcrumbs_busy(struct intel_engine_cs *engine)
 857 {
 858         struct intel_breadcrumbs *b = &engine->breadcrumbs;
 859         bool busy = false;
 860
 861         spin_lock_irq(&b->rb_lock);
 862
 863         if (b->irq_wait) {
 864                 wake_up_process(b->irq_wait->tsk);
 865                 busy = true;
 866         }
 867
 868         if (rcu_access_pointer(b->first_signal)) {
 869                 wake_up_process(b->signaler);
 870                 busy = true;
 871         }
 872
 873         spin_unlock_irq(&b->rb_lock);
 874
 875         return busy;
 876 }
 877
 878 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
 879 #include "selftests/intel_breadcrumbs.c"
 880 #endif