root/drivers/gpu/drm/i915/i915_request.c

/* [<][>][^][v][top][bottom][index][help] */

DEFINITIONS

This source file includes following definitions.
  1. i915_fence_get_driver_name
  2. i915_fence_get_timeline_name
  3. i915_fence_signaled
  4. i915_fence_enable_signaling
  5. i915_fence_wait
  6. i915_fence_release
  7. irq_execute_cb
  8. irq_execute_cb_hook
  9. __notify_execute_cb
  10. remove_from_client
  11. free_capture_list
  12. remove_from_engine
  13. i915_request_retire
  14. i915_request_retire_upto
  15. __i915_request_await_execution
  16. __i915_request_submit
  17. i915_request_submit
  18. __i915_request_unsubmit
  19. i915_request_unsubmit
  20. submit_notify
  21. irq_semaphore_cb
  22. semaphore_notify
  23. retire_requests
  24. request_alloc_slow
  25. __i915_request_create
  26. i915_request_create
  27. i915_request_await_start
  28. already_busywaiting
  29. emit_semaphore_wait
  30. i915_request_await_request
  31. i915_request_await_dma_fence
  32. i915_request_await_execution
  33. i915_request_await_object
  34. i915_request_skip
  35. __i915_request_add_to_timeline
  36. __i915_request_commit
  37. __i915_request_queue
  38. i915_request_add
  39. local_clock_us
  40. busywait_stop
  41. __i915_spin_request
  42. request_wait_wake
  43. i915_request_wait
  44. i915_retire_requests
  45. i915_global_request_shrink
  46. i915_global_request_exit
  47. i915_global_request_init

   1 /*
   2  * Copyright © 2008-2015 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/dma-fence-array.h>
  26 #include <linux/irq_work.h>
  27 #include <linux/prefetch.h>
  28 #include <linux/sched.h>
  29 #include <linux/sched/clock.h>
  30 #include <linux/sched/signal.h>
  31 
  32 #include "gem/i915_gem_context.h"
  33 #include "gt/intel_context.h"
  34 
  35 #include "i915_active.h"
  36 #include "i915_drv.h"
  37 #include "i915_globals.h"
  38 #include "i915_trace.h"
  39 #include "intel_pm.h"
  40 
  41 struct execute_cb {
  42         struct list_head link;
  43         struct irq_work work;
  44         struct i915_sw_fence *fence;
  45         void (*hook)(struct i915_request *rq, struct dma_fence *signal);
  46         struct i915_request *signal;
  47 };
  48 
  49 static struct i915_global_request {
  50         struct i915_global base;
  51         struct kmem_cache *slab_requests;
  52         struct kmem_cache *slab_dependencies;
  53         struct kmem_cache *slab_execute_cbs;
  54 } global;
  55 
  56 static const char *i915_fence_get_driver_name(struct dma_fence *fence)
  57 {
  58         return "i915";
  59 }
  60 
  61 static const char *i915_fence_get_timeline_name(struct dma_fence *fence)
  62 {
  63         /*
  64          * The timeline struct (as part of the ppgtt underneath a context)
  65          * may be freed when the request is no longer in use by the GPU.
  66          * We could extend the life of a context to beyond that of all
  67          * fences, possibly keeping the hw resource around indefinitely,
  68          * or we just give them a false name. Since
  69          * dma_fence_ops.get_timeline_name is a debug feature, the occasional
  70          * lie seems justifiable.
  71          */
  72         if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
  73                 return "signaled";
  74 
  75         return to_request(fence)->gem_context->name ?: "[i915]";
  76 }
  77 
  78 static bool i915_fence_signaled(struct dma_fence *fence)
  79 {
  80         return i915_request_completed(to_request(fence));
  81 }
  82 
  83 static bool i915_fence_enable_signaling(struct dma_fence *fence)
  84 {
  85         return i915_request_enable_breadcrumb(to_request(fence));
  86 }
  87 
  88 static signed long i915_fence_wait(struct dma_fence *fence,
  89                                    bool interruptible,
  90                                    signed long timeout)
  91 {
  92         return i915_request_wait(to_request(fence),
  93                                  interruptible | I915_WAIT_PRIORITY,
  94                                  timeout);
  95 }
  96 
  97 static void i915_fence_release(struct dma_fence *fence)
  98 {
  99         struct i915_request *rq = to_request(fence);
 100 
 101         /*
 102          * The request is put onto a RCU freelist (i.e. the address
 103          * is immediately reused), mark the fences as being freed now.
 104          * Otherwise the debugobjects for the fences are only marked as
 105          * freed when the slab cache itself is freed, and so we would get
 106          * caught trying to reuse dead objects.
 107          */
 108         i915_sw_fence_fini(&rq->submit);
 109         i915_sw_fence_fini(&rq->semaphore);
 110 
 111         kmem_cache_free(global.slab_requests, rq);
 112 }
 113 
 114 const struct dma_fence_ops i915_fence_ops = {
 115         .get_driver_name = i915_fence_get_driver_name,
 116         .get_timeline_name = i915_fence_get_timeline_name,
 117         .enable_signaling = i915_fence_enable_signaling,
 118         .signaled = i915_fence_signaled,
 119         .wait = i915_fence_wait,
 120         .release = i915_fence_release,
 121 };
 122 
 123 static void irq_execute_cb(struct irq_work *wrk)
 124 {
 125         struct execute_cb *cb = container_of(wrk, typeof(*cb), work);
 126 
 127         i915_sw_fence_complete(cb->fence);
 128         kmem_cache_free(global.slab_execute_cbs, cb);
 129 }
 130 
 131 static void irq_execute_cb_hook(struct irq_work *wrk)
 132 {
 133         struct execute_cb *cb = container_of(wrk, typeof(*cb), work);
 134 
 135         cb->hook(container_of(cb->fence, struct i915_request, submit),
 136                  &cb->signal->fence);
 137         i915_request_put(cb->signal);
 138 
 139         irq_execute_cb(wrk);
 140 }
 141 
 142 static void __notify_execute_cb(struct i915_request *rq)
 143 {
 144         struct execute_cb *cb;
 145 
 146         lockdep_assert_held(&rq->lock);
 147 
 148         if (list_empty(&rq->execute_cb))
 149                 return;
 150 
 151         list_for_each_entry(cb, &rq->execute_cb, link)
 152                 irq_work_queue(&cb->work);
 153 
 154         /*
 155          * XXX Rollback on __i915_request_unsubmit()
 156          *
 157          * In the future, perhaps when we have an active time-slicing scheduler,
 158          * it will be interesting to unsubmit parallel execution and remove
 159          * busywaits from the GPU until their master is restarted. This is
 160          * quite hairy, we have to carefully rollback the fence and do a
 161          * preempt-to-idle cycle on the target engine, all the while the
 162          * master execute_cb may refire.
 163          */
 164         INIT_LIST_HEAD(&rq->execute_cb);
 165 }
 166 
 167 static inline void
 168 remove_from_client(struct i915_request *request)
 169 {
 170         struct drm_i915_file_private *file_priv;
 171 
 172         file_priv = READ_ONCE(request->file_priv);
 173         if (!file_priv)
 174                 return;
 175 
 176         spin_lock(&file_priv->mm.lock);
 177         if (request->file_priv) {
 178                 list_del(&request->client_link);
 179                 request->file_priv = NULL;
 180         }
 181         spin_unlock(&file_priv->mm.lock);
 182 }
 183 
 184 static void free_capture_list(struct i915_request *request)
 185 {
 186         struct i915_capture_list *capture;
 187 
 188         capture = request->capture_list;
 189         while (capture) {
 190                 struct i915_capture_list *next = capture->next;
 191 
 192                 kfree(capture);
 193                 capture = next;
 194         }
 195 }
 196 
 197 static void remove_from_engine(struct i915_request *rq)
 198 {
 199         struct intel_engine_cs *engine, *locked;
 200 
 201         /*
 202          * Virtual engines complicate acquiring the engine timeline lock,
 203          * as their rq->engine pointer is not stable until under that
 204          * engine lock. The simple ploy we use is to take the lock then
 205          * check that the rq still belongs to the newly locked engine.
 206          */
 207         locked = READ_ONCE(rq->engine);
 208         spin_lock(&locked->active.lock);
 209         while (unlikely(locked != (engine = READ_ONCE(rq->engine)))) {
 210                 spin_unlock(&locked->active.lock);
 211                 spin_lock(&engine->active.lock);
 212                 locked = engine;
 213         }
 214         list_del(&rq->sched.link);
 215         spin_unlock(&locked->active.lock);
 216 }
 217 
 218 static bool i915_request_retire(struct i915_request *rq)
 219 {
 220         struct i915_active_request *active, *next;
 221 
 222         lockdep_assert_held(&rq->timeline->mutex);
 223         if (!i915_request_completed(rq))
 224                 return false;
 225 
 226         GEM_TRACE("%s fence %llx:%lld, current %d\n",
 227                   rq->engine->name,
 228                   rq->fence.context, rq->fence.seqno,
 229                   hwsp_seqno(rq));
 230 
 231         GEM_BUG_ON(!i915_sw_fence_signaled(&rq->submit));
 232         trace_i915_request_retire(rq);
 233 
 234         /*
 235          * We know the GPU must have read the request to have
 236          * sent us the seqno + interrupt, so use the position
 237          * of tail of the request to update the last known position
 238          * of the GPU head.
 239          *
 240          * Note this requires that we are always called in request
 241          * completion order.
 242          */
 243         GEM_BUG_ON(!list_is_first(&rq->link, &rq->timeline->requests));
 244         rq->ring->head = rq->postfix;
 245 
 246         /*
 247          * Walk through the active list, calling retire on each. This allows
 248          * objects to track their GPU activity and mark themselves as idle
 249          * when their *last* active request is completed (updating state
 250          * tracking lists for eviction, active references for GEM, etc).
 251          *
 252          * As the ->retire() may free the node, we decouple it first and
 253          * pass along the auxiliary information (to avoid dereferencing
 254          * the node after the callback).
 255          */
 256         list_for_each_entry_safe(active, next, &rq->active_list, link) {
 257                 /*
 258                  * In microbenchmarks or focusing upon time inside the kernel,
 259                  * we may spend an inordinate amount of time simply handling
 260                  * the retirement of requests and processing their callbacks.
 261                  * Of which, this loop itself is particularly hot due to the
 262                  * cache misses when jumping around the list of
 263                  * i915_active_request.  So we try to keep this loop as
 264                  * streamlined as possible and also prefetch the next
 265                  * i915_active_request to try and hide the likely cache miss.
 266                  */
 267                 prefetchw(next);
 268 
 269                 INIT_LIST_HEAD(&active->link);
 270                 RCU_INIT_POINTER(active->request, NULL);
 271 
 272                 active->retire(active, rq);
 273         }
 274 
 275         local_irq_disable();
 276 
 277         /*
 278          * We only loosely track inflight requests across preemption,
 279          * and so we may find ourselves attempting to retire a _completed_
 280          * request that we have removed from the HW and put back on a run
 281          * queue.
 282          */
 283         remove_from_engine(rq);
 284 
 285         spin_lock(&rq->lock);
 286         i915_request_mark_complete(rq);
 287         if (!i915_request_signaled(rq))
 288                 dma_fence_signal_locked(&rq->fence);
 289         if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &rq->fence.flags))
 290                 i915_request_cancel_breadcrumb(rq);
 291         if (i915_request_has_waitboost(rq)) {
 292                 GEM_BUG_ON(!atomic_read(&rq->i915->gt_pm.rps.num_waiters));
 293                 atomic_dec(&rq->i915->gt_pm.rps.num_waiters);
 294         }
 295         if (!test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags)) {
 296                 set_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags);
 297                 __notify_execute_cb(rq);
 298         }
 299         GEM_BUG_ON(!list_empty(&rq->execute_cb));
 300         spin_unlock(&rq->lock);
 301 
 302         local_irq_enable();
 303 
 304         remove_from_client(rq);
 305         list_del(&rq->link);
 306 
 307         intel_context_exit(rq->hw_context);
 308         intel_context_unpin(rq->hw_context);
 309 
 310         free_capture_list(rq);
 311         i915_sched_node_fini(&rq->sched);
 312         i915_request_put(rq);
 313 
 314         return true;
 315 }
 316 
 317 void i915_request_retire_upto(struct i915_request *rq)
 318 {
 319         struct intel_timeline * const tl = rq->timeline;
 320         struct i915_request *tmp;
 321 
 322         GEM_TRACE("%s fence %llx:%lld, current %d\n",
 323                   rq->engine->name,
 324                   rq->fence.context, rq->fence.seqno,
 325                   hwsp_seqno(rq));
 326 
 327         lockdep_assert_held(&tl->mutex);
 328         GEM_BUG_ON(!i915_request_completed(rq));
 329 
 330         do {
 331                 tmp = list_first_entry(&tl->requests, typeof(*tmp), link);
 332         } while (i915_request_retire(tmp) && tmp != rq);
 333 }
 334 
 335 static int
 336 __i915_request_await_execution(struct i915_request *rq,
 337                                struct i915_request *signal,
 338                                void (*hook)(struct i915_request *rq,
 339                                             struct dma_fence *signal),
 340                                gfp_t gfp)
 341 {
 342         struct execute_cb *cb;
 343 
 344         if (i915_request_is_active(signal)) {
 345                 if (hook)
 346                         hook(rq, &signal->fence);
 347                 return 0;
 348         }
 349 
 350         cb = kmem_cache_alloc(global.slab_execute_cbs, gfp);
 351         if (!cb)
 352                 return -ENOMEM;
 353 
 354         cb->fence = &rq->submit;
 355         i915_sw_fence_await(cb->fence);
 356         init_irq_work(&cb->work, irq_execute_cb);
 357 
 358         if (hook) {
 359                 cb->hook = hook;
 360                 cb->signal = i915_request_get(signal);
 361                 cb->work.func = irq_execute_cb_hook;
 362         }
 363 
 364         spin_lock_irq(&signal->lock);
 365         if (i915_request_is_active(signal)) {
 366                 if (hook) {
 367                         hook(rq, &signal->fence);
 368                         i915_request_put(signal);
 369                 }
 370                 i915_sw_fence_complete(cb->fence);
 371                 kmem_cache_free(global.slab_execute_cbs, cb);
 372         } else {
 373                 list_add_tail(&cb->link, &signal->execute_cb);
 374         }
 375         spin_unlock_irq(&signal->lock);
 376 
 377         return 0;
 378 }
 379 
 380 bool __i915_request_submit(struct i915_request *request)
 381 {
 382         struct intel_engine_cs *engine = request->engine;
 383         bool result = false;
 384 
 385         GEM_TRACE("%s fence %llx:%lld, current %d\n",
 386                   engine->name,
 387                   request->fence.context, request->fence.seqno,
 388                   hwsp_seqno(request));
 389 
 390         GEM_BUG_ON(!irqs_disabled());
 391         lockdep_assert_held(&engine->active.lock);
 392 
 393         /*
 394          * With the advent of preempt-to-busy, we frequently encounter
 395          * requests that we have unsubmitted from HW, but left running
 396          * until the next ack and so have completed in the meantime. On
 397          * resubmission of that completed request, we can skip
 398          * updating the payload, and execlists can even skip submitting
 399          * the request.
 400          *
 401          * We must remove the request from the caller's priority queue,
 402          * and the caller must only call us when the request is in their
 403          * priority queue, under the active.lock. This ensures that the
 404          * request has *not* yet been retired and we can safely move
 405          * the request into the engine->active.list where it will be
 406          * dropped upon retiring. (Otherwise if resubmit a *retired*
 407          * request, this would be a horrible use-after-free.)
 408          */
 409         if (i915_request_completed(request))
 410                 goto xfer;
 411 
 412         if (i915_gem_context_is_banned(request->gem_context))
 413                 i915_request_skip(request, -EIO);
 414 
 415         /*
 416          * Are we using semaphores when the gpu is already saturated?
 417          *
 418          * Using semaphores incurs a cost in having the GPU poll a
 419          * memory location, busywaiting for it to change. The continual
 420          * memory reads can have a noticeable impact on the rest of the
 421          * system with the extra bus traffic, stalling the cpu as it too
 422          * tries to access memory across the bus (perf stat -e bus-cycles).
 423          *
 424          * If we installed a semaphore on this request and we only submit
 425          * the request after the signaler completed, that indicates the
 426          * system is overloaded and using semaphores at this time only
 427          * increases the amount of work we are doing. If so, we disable
 428          * further use of semaphores until we are idle again, whence we
 429          * optimistically try again.
 430          */
 431         if (request->sched.semaphores &&
 432             i915_sw_fence_signaled(&request->semaphore))
 433                 engine->saturated |= request->sched.semaphores;
 434 
 435         engine->emit_fini_breadcrumb(request,
 436                                      request->ring->vaddr + request->postfix);
 437 
 438         trace_i915_request_execute(request);
 439         engine->serial++;
 440         result = true;
 441 
 442 xfer:   /* We may be recursing from the signal callback of another i915 fence */
 443         spin_lock_nested(&request->lock, SINGLE_DEPTH_NESTING);
 444 
 445         if (!test_and_set_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags))
 446                 list_move_tail(&request->sched.link, &engine->active.requests);
 447 
 448         if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags) &&
 449             !test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &request->fence.flags) &&
 450             !i915_request_enable_breadcrumb(request))
 451                 intel_engine_queue_breadcrumbs(engine);
 452 
 453         __notify_execute_cb(request);
 454 
 455         spin_unlock(&request->lock);
 456 
 457         return result;
 458 }
 459 
 460 void i915_request_submit(struct i915_request *request)
 461 {
 462         struct intel_engine_cs *engine = request->engine;
 463         unsigned long flags;
 464 
 465         /* Will be called from irq-context when using foreign fences. */
 466         spin_lock_irqsave(&engine->active.lock, flags);
 467 
 468         __i915_request_submit(request);
 469 
 470         spin_unlock_irqrestore(&engine->active.lock, flags);
 471 }
 472 
 473 void __i915_request_unsubmit(struct i915_request *request)
 474 {
 475         struct intel_engine_cs *engine = request->engine;
 476 
 477         GEM_TRACE("%s fence %llx:%lld, current %d\n",
 478                   engine->name,
 479                   request->fence.context, request->fence.seqno,
 480                   hwsp_seqno(request));
 481 
 482         GEM_BUG_ON(!irqs_disabled());
 483         lockdep_assert_held(&engine->active.lock);
 484 
 485         /*
 486          * Only unwind in reverse order, required so that the per-context list
 487          * is kept in seqno/ring order.
 488          */
 489 
 490         /* We may be recursing from the signal callback of another i915 fence */
 491         spin_lock_nested(&request->lock, SINGLE_DEPTH_NESTING);
 492 
 493         if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags))
 494                 i915_request_cancel_breadcrumb(request);
 495 
 496         GEM_BUG_ON(!test_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags));
 497         clear_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags);
 498 
 499         spin_unlock(&request->lock);
 500 
 501         /* We've already spun, don't charge on resubmitting. */
 502         if (request->sched.semaphores && i915_request_started(request)) {
 503                 request->sched.attr.priority |= I915_PRIORITY_NOSEMAPHORE;
 504                 request->sched.semaphores = 0;
 505         }
 506 
 507         /*
 508          * We don't need to wake_up any waiters on request->execute, they
 509          * will get woken by any other event or us re-adding this request
 510          * to the engine timeline (__i915_request_submit()). The waiters
 511          * should be quite adapt at finding that the request now has a new
 512          * global_seqno to the one they went to sleep on.
 513          */
 514 }
 515 
 516 void i915_request_unsubmit(struct i915_request *request)
 517 {
 518         struct intel_engine_cs *engine = request->engine;
 519         unsigned long flags;
 520 
 521         /* Will be called from irq-context when using foreign fences. */
 522         spin_lock_irqsave(&engine->active.lock, flags);
 523 
 524         __i915_request_unsubmit(request);
 525 
 526         spin_unlock_irqrestore(&engine->active.lock, flags);
 527 }
 528 
 529 static int __i915_sw_fence_call
 530 submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
 531 {
 532         struct i915_request *request =
 533                 container_of(fence, typeof(*request), submit);
 534 
 535         switch (state) {
 536         case FENCE_COMPLETE:
 537                 trace_i915_request_submit(request);
 538 
 539                 if (unlikely(fence->error))
 540                         i915_request_skip(request, fence->error);
 541 
 542                 /*
 543                  * We need to serialize use of the submit_request() callback
 544                  * with its hotplugging performed during an emergency
 545                  * i915_gem_set_wedged().  We use the RCU mechanism to mark the
 546                  * critical section in order to force i915_gem_set_wedged() to
 547                  * wait until the submit_request() is completed before
 548                  * proceeding.
 549                  */
 550                 rcu_read_lock();
 551                 request->engine->submit_request(request);
 552                 rcu_read_unlock();
 553                 break;
 554 
 555         case FENCE_FREE:
 556                 i915_request_put(request);
 557                 break;
 558         }
 559 
 560         return NOTIFY_DONE;
 561 }
 562 
 563 static void irq_semaphore_cb(struct irq_work *wrk)
 564 {
 565         struct i915_request *rq =
 566                 container_of(wrk, typeof(*rq), semaphore_work);
 567 
 568         i915_schedule_bump_priority(rq, I915_PRIORITY_NOSEMAPHORE);
 569         i915_request_put(rq);
 570 }
 571 
 572 static int __i915_sw_fence_call
 573 semaphore_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
 574 {
 575         struct i915_request *rq = container_of(fence, typeof(*rq), semaphore);
 576 
 577         switch (state) {
 578         case FENCE_COMPLETE:
 579                 if (!(READ_ONCE(rq->sched.attr.priority) & I915_PRIORITY_NOSEMAPHORE)) {
 580                         i915_request_get(rq);
 581                         init_irq_work(&rq->semaphore_work, irq_semaphore_cb);
 582                         irq_work_queue(&rq->semaphore_work);
 583                 }
 584                 break;
 585 
 586         case FENCE_FREE:
 587                 i915_request_put(rq);
 588                 break;
 589         }
 590 
 591         return NOTIFY_DONE;
 592 }
 593 
 594 static void retire_requests(struct intel_timeline *tl)
 595 {
 596         struct i915_request *rq, *rn;
 597 
 598         list_for_each_entry_safe(rq, rn, &tl->requests, link)
 599                 if (!i915_request_retire(rq))
 600                         break;
 601 }
 602 
 603 static noinline struct i915_request *
 604 request_alloc_slow(struct intel_timeline *tl, gfp_t gfp)
 605 {
 606         struct i915_request *rq;
 607 
 608         if (list_empty(&tl->requests))
 609                 goto out;
 610 
 611         if (!gfpflags_allow_blocking(gfp))
 612                 goto out;
 613 
 614         /* Move our oldest request to the slab-cache (if not in use!) */
 615         rq = list_first_entry(&tl->requests, typeof(*rq), link);
 616         i915_request_retire(rq);
 617 
 618         rq = kmem_cache_alloc(global.slab_requests,
 619                               gfp | __GFP_RETRY_MAYFAIL | __GFP_NOWARN);
 620         if (rq)
 621                 return rq;
 622 
 623         /* Ratelimit ourselves to prevent oom from malicious clients */
 624         rq = list_last_entry(&tl->requests, typeof(*rq), link);
 625         cond_synchronize_rcu(rq->rcustate);
 626 
 627         /* Retire our old requests in the hope that we free some */
 628         retire_requests(tl);
 629 
 630 out:
 631         return kmem_cache_alloc(global.slab_requests, gfp);
 632 }
 633 
 634 struct i915_request *
 635 __i915_request_create(struct intel_context *ce, gfp_t gfp)
 636 {
 637         struct intel_timeline *tl = ce->timeline;
 638         struct i915_request *rq;
 639         u32 seqno;
 640         int ret;
 641 
 642         might_sleep_if(gfpflags_allow_blocking(gfp));
 643 
 644         /* Check that the caller provided an already pinned context */
 645         __intel_context_pin(ce);
 646 
 647         /*
 648          * Beware: Dragons be flying overhead.
 649          *
 650          * We use RCU to look up requests in flight. The lookups may
 651          * race with the request being allocated from the slab freelist.
 652          * That is the request we are writing to here, may be in the process
 653          * of being read by __i915_active_request_get_rcu(). As such,
 654          * we have to be very careful when overwriting the contents. During
 655          * the RCU lookup, we change chase the request->engine pointer,
 656          * read the request->global_seqno and increment the reference count.
 657          *
 658          * The reference count is incremented atomically. If it is zero,
 659          * the lookup knows the request is unallocated and complete. Otherwise,
 660          * it is either still in use, or has been reallocated and reset
 661          * with dma_fence_init(). This increment is safe for release as we
 662          * check that the request we have a reference to and matches the active
 663          * request.
 664          *
 665          * Before we increment the refcount, we chase the request->engine
 666          * pointer. We must not call kmem_cache_zalloc() or else we set
 667          * that pointer to NULL and cause a crash during the lookup. If
 668          * we see the request is completed (based on the value of the
 669          * old engine and seqno), the lookup is complete and reports NULL.
 670          * If we decide the request is not completed (new engine or seqno),
 671          * then we grab a reference and double check that it is still the
 672          * active request - which it won't be and restart the lookup.
 673          *
 674          * Do not use kmem_cache_zalloc() here!
 675          */
 676         rq = kmem_cache_alloc(global.slab_requests,
 677                               gfp | __GFP_RETRY_MAYFAIL | __GFP_NOWARN);
 678         if (unlikely(!rq)) {
 679                 rq = request_alloc_slow(tl, gfp);
 680                 if (!rq) {
 681                         ret = -ENOMEM;
 682                         goto err_unreserve;
 683                 }
 684         }
 685 
 686         ret = intel_timeline_get_seqno(tl, rq, &seqno);
 687         if (ret)
 688                 goto err_free;
 689 
 690         rq->i915 = ce->engine->i915;
 691         rq->hw_context = ce;
 692         rq->gem_context = ce->gem_context;
 693         rq->engine = ce->engine;
 694         rq->ring = ce->ring;
 695         rq->timeline = tl;
 696         rq->hwsp_seqno = tl->hwsp_seqno;
 697         rq->hwsp_cacheline = tl->hwsp_cacheline;
 698         rq->rcustate = get_state_synchronize_rcu(); /* acts as smp_mb() */
 699 
 700         spin_lock_init(&rq->lock);
 701         dma_fence_init(&rq->fence, &i915_fence_ops, &rq->lock,
 702                        tl->fence_context, seqno);
 703 
 704         /* We bump the ref for the fence chain */
 705         i915_sw_fence_init(&i915_request_get(rq)->submit, submit_notify);
 706         i915_sw_fence_init(&i915_request_get(rq)->semaphore, semaphore_notify);
 707 
 708         i915_sched_node_init(&rq->sched);
 709 
 710         /* No zalloc, must clear what we need by hand */
 711         rq->file_priv = NULL;
 712         rq->batch = NULL;
 713         rq->capture_list = NULL;
 714         rq->flags = 0;
 715         rq->execution_mask = ALL_ENGINES;
 716 
 717         INIT_LIST_HEAD(&rq->active_list);
 718         INIT_LIST_HEAD(&rq->execute_cb);
 719 
 720         /*
 721          * Reserve space in the ring buffer for all the commands required to
 722          * eventually emit this request. This is to guarantee that the
 723          * i915_request_add() call can't fail. Note that the reserve may need
 724          * to be redone if the request is not actually submitted straight
 725          * away, e.g. because a GPU scheduler has deferred it.
 726          *
 727          * Note that due to how we add reserved_space to intel_ring_begin()
 728          * we need to double our request to ensure that if we need to wrap
 729          * around inside i915_request_add() there is sufficient space at
 730          * the beginning of the ring as well.
 731          */
 732         rq->reserved_space =
 733                 2 * rq->engine->emit_fini_breadcrumb_dw * sizeof(u32);
 734 
 735         /*
 736          * Record the position of the start of the request so that
 737          * should we detect the updated seqno part-way through the
 738          * GPU processing the request, we never over-estimate the
 739          * position of the head.
 740          */
 741         rq->head = rq->ring->emit;
 742 
 743         ret = rq->engine->request_alloc(rq);
 744         if (ret)
 745                 goto err_unwind;
 746 
 747         rq->infix = rq->ring->emit; /* end of header; start of user payload */
 748 
 749         intel_context_mark_active(ce);
 750         return rq;
 751 
 752 err_unwind:
 753         ce->ring->emit = rq->head;
 754 
 755         /* Make sure we didn't add ourselves to external state before freeing */
 756         GEM_BUG_ON(!list_empty(&rq->active_list));
 757         GEM_BUG_ON(!list_empty(&rq->sched.signalers_list));
 758         GEM_BUG_ON(!list_empty(&rq->sched.waiters_list));
 759 
 760 err_free:
 761         kmem_cache_free(global.slab_requests, rq);
 762 err_unreserve:
 763         intel_context_unpin(ce);
 764         return ERR_PTR(ret);
 765 }
 766 
 767 struct i915_request *
 768 i915_request_create(struct intel_context *ce)
 769 {
 770         struct i915_request *rq;
 771         struct intel_timeline *tl;
 772 
 773         tl = intel_context_timeline_lock(ce);
 774         if (IS_ERR(tl))
 775                 return ERR_CAST(tl);
 776 
 777         /* Move our oldest request to the slab-cache (if not in use!) */
 778         rq = list_first_entry(&tl->requests, typeof(*rq), link);
 779         if (!list_is_last(&rq->link, &tl->requests))
 780                 i915_request_retire(rq);
 781 
 782         intel_context_enter(ce);
 783         rq = __i915_request_create(ce, GFP_KERNEL);
 784         intel_context_exit(ce); /* active reference transferred to request */
 785         if (IS_ERR(rq))
 786                 goto err_unlock;
 787 
 788         /* Check that we do not interrupt ourselves with a new request */
 789         rq->cookie = lockdep_pin_lock(&tl->mutex);
 790 
 791         return rq;
 792 
 793 err_unlock:
 794         intel_context_timeline_unlock(tl);
 795         return rq;
 796 }
 797 
 798 static int
 799 i915_request_await_start(struct i915_request *rq, struct i915_request *signal)
 800 {
 801         if (list_is_first(&signal->link, &signal->timeline->requests))
 802                 return 0;
 803 
 804         signal = list_prev_entry(signal, link);
 805         if (intel_timeline_sync_is_later(rq->timeline, &signal->fence))
 806                 return 0;
 807 
 808         return i915_sw_fence_await_dma_fence(&rq->submit,
 809                                              &signal->fence, 0,
 810                                              I915_FENCE_GFP);
 811 }
 812 
 813 static intel_engine_mask_t
 814 already_busywaiting(struct i915_request *rq)
 815 {
 816         /*
 817          * Polling a semaphore causes bus traffic, delaying other users of
 818          * both the GPU and CPU. We want to limit the impact on others,
 819          * while taking advantage of early submission to reduce GPU
 820          * latency. Therefore we restrict ourselves to not using more
 821          * than one semaphore from each source, and not using a semaphore
 822          * if we have detected the engine is saturated (i.e. would not be
 823          * submitted early and cause bus traffic reading an already passed
 824          * semaphore).
 825          *
 826          * See the are-we-too-late? check in __i915_request_submit().
 827          */
 828         return rq->sched.semaphores | rq->engine->saturated;
 829 }
 830 
 831 static int
 832 emit_semaphore_wait(struct i915_request *to,
 833                     struct i915_request *from,
 834                     gfp_t gfp)
 835 {
 836         u32 hwsp_offset;
 837         u32 *cs;
 838         int err;
 839 
 840         GEM_BUG_ON(!from->timeline->has_initial_breadcrumb);
 841         GEM_BUG_ON(INTEL_GEN(to->i915) < 8);
 842 
 843         /* Just emit the first semaphore we see as request space is limited. */
 844         if (already_busywaiting(to) & from->engine->mask)
 845                 return i915_sw_fence_await_dma_fence(&to->submit,
 846                                                      &from->fence, 0,
 847                                                      I915_FENCE_GFP);
 848 
 849         err = i915_request_await_start(to, from);
 850         if (err < 0)
 851                 return err;
 852 
 853         /* Only submit our spinner after the signaler is running! */
 854         err = __i915_request_await_execution(to, from, NULL, gfp);
 855         if (err)
 856                 return err;
 857 
 858         /* We need to pin the signaler's HWSP until we are finished reading. */
 859         err = intel_timeline_read_hwsp(from, to, &hwsp_offset);
 860         if (err)
 861                 return err;
 862 
 863         cs = intel_ring_begin(to, 4);
 864         if (IS_ERR(cs))
 865                 return PTR_ERR(cs);
 866 
 867         /*
 868          * Using greater-than-or-equal here means we have to worry
 869          * about seqno wraparound. To side step that issue, we swap
 870          * the timeline HWSP upon wrapping, so that everyone listening
 871          * for the old (pre-wrap) values do not see the much smaller
 872          * (post-wrap) values than they were expecting (and so wait
 873          * forever).
 874          */
 875         *cs++ = MI_SEMAPHORE_WAIT |
 876                 MI_SEMAPHORE_GLOBAL_GTT |
 877                 MI_SEMAPHORE_POLL |
 878                 MI_SEMAPHORE_SAD_GTE_SDD;
 879         *cs++ = from->fence.seqno;
 880         *cs++ = hwsp_offset;
 881         *cs++ = 0;
 882 
 883         intel_ring_advance(to, cs);
 884         to->sched.semaphores |= from->engine->mask;
 885         to->sched.flags |= I915_SCHED_HAS_SEMAPHORE_CHAIN;
 886         return 0;
 887 }
 888 
 889 static int
 890 i915_request_await_request(struct i915_request *to, struct i915_request *from)
 891 {
 892         int ret;
 893 
 894         GEM_BUG_ON(to == from);
 895         GEM_BUG_ON(to->timeline == from->timeline);
 896 
 897         if (i915_request_completed(from)) {
 898                 i915_sw_fence_set_error_once(&to->submit, from->fence.error);
 899                 return 0;
 900         }
 901 
 902         if (to->engine->schedule) {
 903                 ret = i915_sched_node_add_dependency(&to->sched, &from->sched);
 904                 if (ret < 0)
 905                         return ret;
 906         }
 907 
 908         if (to->engine == from->engine) {
 909                 ret = i915_sw_fence_await_sw_fence_gfp(&to->submit,
 910                                                        &from->submit,
 911                                                        I915_FENCE_GFP);
 912         } else if (intel_engine_has_semaphores(to->engine) &&
 913                    to->gem_context->sched.priority >= I915_PRIORITY_NORMAL) {
 914                 ret = emit_semaphore_wait(to, from, I915_FENCE_GFP);
 915         } else {
 916                 ret = i915_sw_fence_await_dma_fence(&to->submit,
 917                                                     &from->fence, 0,
 918                                                     I915_FENCE_GFP);
 919         }
 920         if (ret < 0)
 921                 return ret;
 922 
 923         if (to->sched.flags & I915_SCHED_HAS_SEMAPHORE_CHAIN) {
 924                 ret = i915_sw_fence_await_dma_fence(&to->semaphore,
 925                                                     &from->fence, 0,
 926                                                     I915_FENCE_GFP);
 927                 if (ret < 0)
 928                         return ret;
 929         }
 930 
 931         return 0;
 932 }
 933 
 934 int
 935 i915_request_await_dma_fence(struct i915_request *rq, struct dma_fence *fence)
 936 {
 937         struct dma_fence **child = &fence;
 938         unsigned int nchild = 1;
 939         int ret;
 940 
 941         /*
 942          * Note that if the fence-array was created in signal-on-any mode,
 943          * we should *not* decompose it into its individual fences. However,
 944          * we don't currently store which mode the fence-array is operating
 945          * in. Fortunately, the only user of signal-on-any is private to
 946          * amdgpu and we should not see any incoming fence-array from
 947          * sync-file being in signal-on-any mode.
 948          */
 949         if (dma_fence_is_array(fence)) {
 950                 struct dma_fence_array *array = to_dma_fence_array(fence);
 951 
 952                 child = array->fences;
 953                 nchild = array->num_fences;
 954                 GEM_BUG_ON(!nchild);
 955         }
 956 
 957         do {
 958                 fence = *child++;
 959                 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
 960                         continue;
 961 
 962                 /*
 963                  * Requests on the same timeline are explicitly ordered, along
 964                  * with their dependencies, by i915_request_add() which ensures
 965                  * that requests are submitted in-order through each ring.
 966                  */
 967                 if (fence->context == rq->fence.context)
 968                         continue;
 969 
 970                 /* Squash repeated waits to the same timelines */
 971                 if (fence->context &&
 972                     intel_timeline_sync_is_later(rq->timeline, fence))
 973                         continue;
 974 
 975                 if (dma_fence_is_i915(fence))
 976                         ret = i915_request_await_request(rq, to_request(fence));
 977                 else
 978                         ret = i915_sw_fence_await_dma_fence(&rq->submit, fence,
 979                                                             I915_FENCE_TIMEOUT,
 980                                                             I915_FENCE_GFP);
 981                 if (ret < 0)
 982                         return ret;
 983 
 984                 /* Record the latest fence used against each timeline */
 985                 if (fence->context)
 986                         intel_timeline_sync_set(rq->timeline, fence);
 987         } while (--nchild);
 988 
 989         return 0;
 990 }
 991 
 992 int
 993 i915_request_await_execution(struct i915_request *rq,
 994                              struct dma_fence *fence,
 995                              void (*hook)(struct i915_request *rq,
 996                                           struct dma_fence *signal))
 997 {
 998         struct dma_fence **child = &fence;
 999         unsigned int nchild = 1;
1000         int ret;
1001 
1002         if (dma_fence_is_array(fence)) {
1003                 struct dma_fence_array *array = to_dma_fence_array(fence);
1004 
1005                 /* XXX Error for signal-on-any fence arrays */
1006 
1007                 child = array->fences;
1008                 nchild = array->num_fences;
1009                 GEM_BUG_ON(!nchild);
1010         }
1011 
1012         do {
1013                 fence = *child++;
1014                 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
1015                         continue;
1016 
1017                 /*
1018                  * We don't squash repeated fence dependencies here as we
1019                  * want to run our callback in all cases.
1020                  */
1021 
1022                 if (dma_fence_is_i915(fence))
1023                         ret = __i915_request_await_execution(rq,
1024                                                              to_request(fence),
1025                                                              hook,
1026                                                              I915_FENCE_GFP);
1027                 else
1028                         ret = i915_sw_fence_await_dma_fence(&rq->submit, fence,
1029                                                             I915_FENCE_TIMEOUT,
1030                                                             GFP_KERNEL);
1031                 if (ret < 0)
1032                         return ret;
1033         } while (--nchild);
1034 
1035         return 0;
1036 }
1037 
1038 /**
1039  * i915_request_await_object - set this request to (async) wait upon a bo
1040  * @to: request we are wishing to use
1041  * @obj: object which may be in use on another ring.
1042  * @write: whether the wait is on behalf of a writer
1043  *
1044  * This code is meant to abstract object synchronization with the GPU.
1045  * Conceptually we serialise writes between engines inside the GPU.
1046  * We only allow one engine to write into a buffer at any time, but
1047  * multiple readers. To ensure each has a coherent view of memory, we must:
1048  *
1049  * - If there is an outstanding write request to the object, the new
1050  *   request must wait for it to complete (either CPU or in hw, requests
1051  *   on the same ring will be naturally ordered).
1052  *
1053  * - If we are a write request (pending_write_domain is set), the new
1054  *   request must wait for outstanding read requests to complete.
1055  *
1056  * Returns 0 if successful, else propagates up the lower layer error.
1057  */
1058 int
1059 i915_request_await_object(struct i915_request *to,
1060                           struct drm_i915_gem_object *obj,
1061                           bool write)
1062 {
1063         struct dma_fence *excl;
1064         int ret = 0;
1065 
1066         if (write) {
1067                 struct dma_fence **shared;
1068                 unsigned int count, i;
1069 
1070                 ret = dma_resv_get_fences_rcu(obj->base.resv,
1071                                                         &excl, &count, &shared);
1072                 if (ret)
1073                         return ret;
1074 
1075                 for (i = 0; i < count; i++) {
1076                         ret = i915_request_await_dma_fence(to, shared[i]);
1077                         if (ret)
1078                                 break;
1079 
1080                         dma_fence_put(shared[i]);
1081                 }
1082 
1083                 for (; i < count; i++)
1084                         dma_fence_put(shared[i]);
1085                 kfree(shared);
1086         } else {
1087                 excl = dma_resv_get_excl_rcu(obj->base.resv);
1088         }
1089 
1090         if (excl) {
1091                 if (ret == 0)
1092                         ret = i915_request_await_dma_fence(to, excl);
1093 
1094                 dma_fence_put(excl);
1095         }
1096 
1097         return ret;
1098 }
1099 
1100 void i915_request_skip(struct i915_request *rq, int error)
1101 {
1102         void *vaddr = rq->ring->vaddr;
1103         u32 head;
1104 
1105         GEM_BUG_ON(!IS_ERR_VALUE((long)error));
1106         dma_fence_set_error(&rq->fence, error);
1107 
1108         if (rq->infix == rq->postfix)
1109                 return;
1110 
1111         /*
1112          * As this request likely depends on state from the lost
1113          * context, clear out all the user operations leaving the
1114          * breadcrumb at the end (so we get the fence notifications).
1115          */
1116         head = rq->infix;
1117         if (rq->postfix < head) {
1118                 memset(vaddr + head, 0, rq->ring->size - head);
1119                 head = 0;
1120         }
1121         memset(vaddr + head, 0, rq->postfix - head);
1122         rq->infix = rq->postfix;
1123 }
1124 
1125 static struct i915_request *
1126 __i915_request_add_to_timeline(struct i915_request *rq)
1127 {
1128         struct intel_timeline *timeline = rq->timeline;
1129         struct i915_request *prev;
1130 
1131         /*
1132          * Dependency tracking and request ordering along the timeline
1133          * is special cased so that we can eliminate redundant ordering
1134          * operations while building the request (we know that the timeline
1135          * itself is ordered, and here we guarantee it).
1136          *
1137          * As we know we will need to emit tracking along the timeline,
1138          * we embed the hooks into our request struct -- at the cost of
1139          * having to have specialised no-allocation interfaces (which will
1140          * be beneficial elsewhere).
1141          *
1142          * A second benefit to open-coding i915_request_await_request is
1143          * that we can apply a slight variant of the rules specialised
1144          * for timelines that jump between engines (such as virtual engines).
1145          * If we consider the case of virtual engine, we must emit a dma-fence
1146          * to prevent scheduling of the second request until the first is
1147          * complete (to maximise our greedy late load balancing) and this
1148          * precludes optimising to use semaphores serialisation of a single
1149          * timeline across engines.
1150          */
1151         prev = rcu_dereference_protected(timeline->last_request.request,
1152                                          lockdep_is_held(&timeline->mutex));
1153         if (prev && !i915_request_completed(prev)) {
1154                 if (is_power_of_2(prev->engine->mask | rq->engine->mask))
1155                         i915_sw_fence_await_sw_fence(&rq->submit,
1156                                                      &prev->submit,
1157                                                      &rq->submitq);
1158                 else
1159                         __i915_sw_fence_await_dma_fence(&rq->submit,
1160                                                         &prev->fence,
1161                                                         &rq->dmaq);
1162                 if (rq->engine->schedule)
1163                         __i915_sched_node_add_dependency(&rq->sched,
1164                                                          &prev->sched,
1165                                                          &rq->dep,
1166                                                          0);
1167         }
1168 
1169         list_add_tail(&rq->link, &timeline->requests);
1170 
1171         /*
1172          * Make sure that no request gazumped us - if it was allocated after
1173          * our i915_request_alloc() and called __i915_request_add() before
1174          * us, the timeline will hold its seqno which is later than ours.
1175          */
1176         GEM_BUG_ON(timeline->seqno != rq->fence.seqno);
1177         __i915_active_request_set(&timeline->last_request, rq);
1178 
1179         return prev;
1180 }
1181 
1182 /*
1183  * NB: This function is not allowed to fail. Doing so would mean the the
1184  * request is not being tracked for completion but the work itself is
1185  * going to happen on the hardware. This would be a Bad Thing(tm).
1186  */
1187 struct i915_request *__i915_request_commit(struct i915_request *rq)
1188 {
1189         struct intel_engine_cs *engine = rq->engine;
1190         struct intel_ring *ring = rq->ring;
1191         u32 *cs;
1192 
1193         GEM_TRACE("%s fence %llx:%lld\n",
1194                   engine->name, rq->fence.context, rq->fence.seqno);
1195 
1196         /*
1197          * To ensure that this call will not fail, space for its emissions
1198          * should already have been reserved in the ring buffer. Let the ring
1199          * know that it is time to use that space up.
1200          */
1201         GEM_BUG_ON(rq->reserved_space > ring->space);
1202         rq->reserved_space = 0;
1203         rq->emitted_jiffies = jiffies;
1204 
1205         /*
1206          * Record the position of the start of the breadcrumb so that
1207          * should we detect the updated seqno part-way through the
1208          * GPU processing the request, we never over-estimate the
1209          * position of the ring's HEAD.
1210          */
1211         cs = intel_ring_begin(rq, engine->emit_fini_breadcrumb_dw);
1212         GEM_BUG_ON(IS_ERR(cs));
1213         rq->postfix = intel_ring_offset(rq, cs);
1214 
1215         return __i915_request_add_to_timeline(rq);
1216 }
1217 
1218 void __i915_request_queue(struct i915_request *rq,
1219                           const struct i915_sched_attr *attr)
1220 {
1221         /*
1222          * Let the backend know a new request has arrived that may need
1223          * to adjust the existing execution schedule due to a high priority
1224          * request - i.e. we may want to preempt the current request in order
1225          * to run a high priority dependency chain *before* we can execute this
1226          * request.
1227          *
1228          * This is called before the request is ready to run so that we can
1229          * decide whether to preempt the entire chain so that it is ready to
1230          * run at the earliest possible convenience.
1231          */
1232         if (attr && rq->engine->schedule)
1233                 rq->engine->schedule(rq, attr);
1234         i915_sw_fence_commit(&rq->semaphore);
1235         i915_sw_fence_commit(&rq->submit);
1236 }
1237 
1238 void i915_request_add(struct i915_request *rq)
1239 {
1240         struct i915_sched_attr attr = rq->gem_context->sched;
1241         struct intel_timeline * const tl = rq->timeline;
1242         struct i915_request *prev;
1243 
1244         lockdep_assert_held(&tl->mutex);
1245         lockdep_unpin_lock(&tl->mutex, rq->cookie);
1246 
1247         trace_i915_request_add(rq);
1248 
1249         prev = __i915_request_commit(rq);
1250 
1251         /*
1252          * Boost actual workloads past semaphores!
1253          *
1254          * With semaphores we spin on one engine waiting for another,
1255          * simply to reduce the latency of starting our work when
1256          * the signaler completes. However, if there is any other
1257          * work that we could be doing on this engine instead, that
1258          * is better utilisation and will reduce the overall duration
1259          * of the current work. To avoid PI boosting a semaphore
1260          * far in the distance past over useful work, we keep a history
1261          * of any semaphore use along our dependency chain.
1262          */
1263         if (!(rq->sched.flags & I915_SCHED_HAS_SEMAPHORE_CHAIN))
1264                 attr.priority |= I915_PRIORITY_NOSEMAPHORE;
1265 
1266         /*
1267          * Boost priorities to new clients (new request flows).
1268          *
1269          * Allow interactive/synchronous clients to jump ahead of
1270          * the bulk clients. (FQ_CODEL)
1271          */
1272         if (list_empty(&rq->sched.signalers_list))
1273                 attr.priority |= I915_PRIORITY_WAIT;
1274 
1275         local_bh_disable();
1276         __i915_request_queue(rq, &attr);
1277         local_bh_enable(); /* Kick the execlists tasklet if just scheduled */
1278 
1279         /*
1280          * In typical scenarios, we do not expect the previous request on
1281          * the timeline to be still tracked by timeline->last_request if it
1282          * has been completed. If the completed request is still here, that
1283          * implies that request retirement is a long way behind submission,
1284          * suggesting that we haven't been retiring frequently enough from
1285          * the combination of retire-before-alloc, waiters and the background
1286          * retirement worker. So if the last request on this timeline was
1287          * already completed, do a catch up pass, flushing the retirement queue
1288          * up to this client. Since we have now moved the heaviest operations
1289          * during retirement onto secondary workers, such as freeing objects
1290          * or contexts, retiring a bunch of requests is mostly list management
1291          * (and cache misses), and so we should not be overly penalizing this
1292          * client by performing excess work, though we may still performing
1293          * work on behalf of others -- but instead we should benefit from
1294          * improved resource management. (Well, that's the theory at least.)
1295          */
1296         if (prev && i915_request_completed(prev) && prev->timeline == tl)
1297                 i915_request_retire_upto(prev);
1298 
1299         mutex_unlock(&tl->mutex);
1300 }
1301 
1302 static unsigned long local_clock_us(unsigned int *cpu)
1303 {
1304         unsigned long t;
1305 
1306         /*
1307          * Cheaply and approximately convert from nanoseconds to microseconds.
1308          * The result and subsequent calculations are also defined in the same
1309          * approximate microseconds units. The principal source of timing
1310          * error here is from the simple truncation.
1311          *
1312          * Note that local_clock() is only defined wrt to the current CPU;
1313          * the comparisons are no longer valid if we switch CPUs. Instead of
1314          * blocking preemption for the entire busywait, we can detect the CPU
1315          * switch and use that as indicator of system load and a reason to
1316          * stop busywaiting, see busywait_stop().
1317          */
1318         *cpu = get_cpu();
1319         t = local_clock() >> 10;
1320         put_cpu();
1321 
1322         return t;
1323 }
1324 
1325 static bool busywait_stop(unsigned long timeout, unsigned int cpu)
1326 {
1327         unsigned int this_cpu;
1328 
1329         if (time_after(local_clock_us(&this_cpu), timeout))
1330                 return true;
1331 
1332         return this_cpu != cpu;
1333 }
1334 
1335 static bool __i915_spin_request(const struct i915_request * const rq,
1336                                 int state, unsigned long timeout_us)
1337 {
1338         unsigned int cpu;
1339 
1340         /*
1341          * Only wait for the request if we know it is likely to complete.
1342          *
1343          * We don't track the timestamps around requests, nor the average
1344          * request length, so we do not have a good indicator that this
1345          * request will complete within the timeout. What we do know is the
1346          * order in which requests are executed by the context and so we can
1347          * tell if the request has been started. If the request is not even
1348          * running yet, it is a fair assumption that it will not complete
1349          * within our relatively short timeout.
1350          */
1351         if (!i915_request_is_running(rq))
1352                 return false;
1353 
1354         /*
1355          * When waiting for high frequency requests, e.g. during synchronous
1356          * rendering split between the CPU and GPU, the finite amount of time
1357          * required to set up the irq and wait upon it limits the response
1358          * rate. By busywaiting on the request completion for a short while we
1359          * can service the high frequency waits as quick as possible. However,
1360          * if it is a slow request, we want to sleep as quickly as possible.
1361          * The tradeoff between waiting and sleeping is roughly the time it
1362          * takes to sleep on a request, on the order of a microsecond.
1363          */
1364 
1365         timeout_us += local_clock_us(&cpu);
1366         do {
1367                 if (i915_request_completed(rq))
1368                         return true;
1369 
1370                 if (signal_pending_state(state, current))
1371                         break;
1372 
1373                 if (busywait_stop(timeout_us, cpu))
1374                         break;
1375 
1376                 cpu_relax();
1377         } while (!need_resched());
1378 
1379         return false;
1380 }
1381 
1382 struct request_wait {
1383         struct dma_fence_cb cb;
1384         struct task_struct *tsk;
1385 };
1386 
1387 static void request_wait_wake(struct dma_fence *fence, struct dma_fence_cb *cb)
1388 {
1389         struct request_wait *wait = container_of(cb, typeof(*wait), cb);
1390 
1391         wake_up_process(wait->tsk);
1392 }
1393 
1394 /**
1395  * i915_request_wait - wait until execution of request has finished
1396  * @rq: the request to wait upon
1397  * @flags: how to wait
1398  * @timeout: how long to wait in jiffies
1399  *
1400  * i915_request_wait() waits for the request to be completed, for a
1401  * maximum of @timeout jiffies (with MAX_SCHEDULE_TIMEOUT implying an
1402  * unbounded wait).
1403  *
1404  * Returns the remaining time (in jiffies) if the request completed, which may
1405  * be zero or -ETIME if the request is unfinished after the timeout expires.
1406  * May return -EINTR is called with I915_WAIT_INTERRUPTIBLE and a signal is
1407  * pending before the request completes.
1408  */
1409 long i915_request_wait(struct i915_request *rq,
1410                        unsigned int flags,
1411                        long timeout)
1412 {
1413         const int state = flags & I915_WAIT_INTERRUPTIBLE ?
1414                 TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE;
1415         struct request_wait wait;
1416 
1417         might_sleep();
1418         GEM_BUG_ON(timeout < 0);
1419 
1420         if (dma_fence_is_signaled(&rq->fence))
1421                 return timeout;
1422 
1423         if (!timeout)
1424                 return -ETIME;
1425 
1426         trace_i915_request_wait_begin(rq, flags);
1427 
1428         /*
1429          * We must never wait on the GPU while holding a lock as we
1430          * may need to perform a GPU reset. So while we don't need to
1431          * serialise wait/reset with an explicit lock, we do want
1432          * lockdep to detect potential dependency cycles.
1433          */
1434         mutex_acquire(&rq->engine->gt->reset.mutex.dep_map, 0, 0, _THIS_IP_);
1435 
1436         /*
1437          * Optimistic spin before touching IRQs.
1438          *
1439          * We may use a rather large value here to offset the penalty of
1440          * switching away from the active task. Frequently, the client will
1441          * wait upon an old swapbuffer to throttle itself to remain within a
1442          * frame of the gpu. If the client is running in lockstep with the gpu,
1443          * then it should not be waiting long at all, and a sleep now will incur
1444          * extra scheduler latency in producing the next frame. To try to
1445          * avoid adding the cost of enabling/disabling the interrupt to the
1446          * short wait, we first spin to see if the request would have completed
1447          * in the time taken to setup the interrupt.
1448          *
1449          * We need upto 5us to enable the irq, and upto 20us to hide the
1450          * scheduler latency of a context switch, ignoring the secondary
1451          * impacts from a context switch such as cache eviction.
1452          *
1453          * The scheme used for low-latency IO is called "hybrid interrupt
1454          * polling". The suggestion there is to sleep until just before you
1455          * expect to be woken by the device interrupt and then poll for its
1456          * completion. That requires having a good predictor for the request
1457          * duration, which we currently lack.
1458          */
1459         if (CONFIG_DRM_I915_SPIN_REQUEST &&
1460             __i915_spin_request(rq, state, CONFIG_DRM_I915_SPIN_REQUEST)) {
1461                 dma_fence_signal(&rq->fence);
1462                 goto out;
1463         }
1464 
1465         /*
1466          * This client is about to stall waiting for the GPU. In many cases
1467          * this is undesirable and limits the throughput of the system, as
1468          * many clients cannot continue processing user input/output whilst
1469          * blocked. RPS autotuning may take tens of milliseconds to respond
1470          * to the GPU load and thus incurs additional latency for the client.
1471          * We can circumvent that by promoting the GPU frequency to maximum
1472          * before we sleep. This makes the GPU throttle up much more quickly
1473          * (good for benchmarks and user experience, e.g. window animations),
1474          * but at a cost of spending more power processing the workload
1475          * (bad for battery).
1476          */
1477         if (flags & I915_WAIT_PRIORITY) {
1478                 if (!i915_request_started(rq) && INTEL_GEN(rq->i915) >= 6)
1479                         gen6_rps_boost(rq);
1480                 i915_schedule_bump_priority(rq, I915_PRIORITY_WAIT);
1481         }
1482 
1483         wait.tsk = current;
1484         if (dma_fence_add_callback(&rq->fence, &wait.cb, request_wait_wake))
1485                 goto out;
1486 
1487         for (;;) {
1488                 set_current_state(state);
1489 
1490                 if (i915_request_completed(rq)) {
1491                         dma_fence_signal(&rq->fence);
1492                         break;
1493                 }
1494 
1495                 if (signal_pending_state(state, current)) {
1496                         timeout = -ERESTARTSYS;
1497                         break;
1498                 }
1499 
1500                 if (!timeout) {
1501                         timeout = -ETIME;
1502                         break;
1503                 }
1504 
1505                 timeout = io_schedule_timeout(timeout);
1506         }
1507         __set_current_state(TASK_RUNNING);
1508 
1509         dma_fence_remove_callback(&rq->fence, &wait.cb);
1510 
1511 out:
1512         mutex_release(&rq->engine->gt->reset.mutex.dep_map, 0, _THIS_IP_);
1513         trace_i915_request_wait_end(rq);
1514         return timeout;
1515 }
1516 
1517 bool i915_retire_requests(struct drm_i915_private *i915)
1518 {
1519         struct intel_gt_timelines *timelines = &i915->gt.timelines;
1520         struct intel_timeline *tl, *tn;
1521         unsigned long flags;
1522         LIST_HEAD(free);
1523 
1524         spin_lock_irqsave(&timelines->lock, flags);
1525         list_for_each_entry_safe(tl, tn, &timelines->active_list, link) {
1526                 if (!mutex_trylock(&tl->mutex))
1527                         continue;
1528 
1529                 intel_timeline_get(tl);
1530                 GEM_BUG_ON(!tl->active_count);
1531                 tl->active_count++; /* pin the list element */
1532                 spin_unlock_irqrestore(&timelines->lock, flags);
1533 
1534                 retire_requests(tl);
1535 
1536                 spin_lock_irqsave(&timelines->lock, flags);
1537 
1538                 /* Resume iteration after dropping lock */
1539                 list_safe_reset_next(tl, tn, link);
1540                 if (!--tl->active_count)
1541                         list_del(&tl->link);
1542 
1543                 mutex_unlock(&tl->mutex);
1544 
1545                 /* Defer the final release to after the spinlock */
1546                 if (refcount_dec_and_test(&tl->kref.refcount)) {
1547                         GEM_BUG_ON(tl->active_count);
1548                         list_add(&tl->link, &free);
1549                 }
1550         }
1551         spin_unlock_irqrestore(&timelines->lock, flags);
1552 
1553         list_for_each_entry_safe(tl, tn, &free, link)
1554                 __intel_timeline_free(&tl->kref);
1555 
1556         return !list_empty(&timelines->active_list);
1557 }
1558 
1559 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1560 #include "selftests/mock_request.c"
1561 #include "selftests/i915_request.c"
1562 #endif
1563 
1564 static void i915_global_request_shrink(void)
1565 {
1566         kmem_cache_shrink(global.slab_dependencies);
1567         kmem_cache_shrink(global.slab_execute_cbs);
1568         kmem_cache_shrink(global.slab_requests);
1569 }
1570 
1571 static void i915_global_request_exit(void)
1572 {
1573         kmem_cache_destroy(global.slab_dependencies);
1574         kmem_cache_destroy(global.slab_execute_cbs);
1575         kmem_cache_destroy(global.slab_requests);
1576 }
1577 
1578 static struct i915_global_request global = { {
1579         .shrink = i915_global_request_shrink,
1580         .exit = i915_global_request_exit,
1581 } };
1582 
1583 int __init i915_global_request_init(void)
1584 {
1585         global.slab_requests = KMEM_CACHE(i915_request,
1586                                           SLAB_HWCACHE_ALIGN |
1587                                           SLAB_RECLAIM_ACCOUNT |
1588                                           SLAB_TYPESAFE_BY_RCU);
1589         if (!global.slab_requests)
1590                 return -ENOMEM;
1591 
1592         global.slab_execute_cbs = KMEM_CACHE(execute_cb,
1593                                              SLAB_HWCACHE_ALIGN |
1594                                              SLAB_RECLAIM_ACCOUNT |
1595                                              SLAB_TYPESAFE_BY_RCU);
1596         if (!global.slab_execute_cbs)
1597                 goto err_requests;
1598 
1599         global.slab_dependencies = KMEM_CACHE(i915_dependency,
1600                                               SLAB_HWCACHE_ALIGN |
1601                                               SLAB_RECLAIM_ACCOUNT);
1602         if (!global.slab_dependencies)
1603                 goto err_execute_cbs;
1604 
1605         i915_global_register(&global.base);
1606         return 0;
1607 
1608 err_execute_cbs:
1609         kmem_cache_destroy(global.slab_execute_cbs);
1610 err_requests:
1611         kmem_cache_destroy(global.slab_requests);
1612         return -ENOMEM;
1613 }

/* [<][>][^][v][top][bottom][index][help] */