1 /*
2  * Copyright © 2008-2018 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 #ifndef I915_REQUEST_H
26 #define I915_REQUEST_H
27 
28 #include <linux/dma-fence.h>
29 #include <linux/hrtimer.h>
30 #include <linux/irq_work.h>
31 #include <linux/llist.h>
32 #include <linux/lockdep.h>
33 
34 #include "gem/i915_gem_context_types.h"
35 #include "gt/intel_context_types.h"
36 #include "gt/intel_engine_types.h"
37 #include "gt/intel_timeline_types.h"
38 
39 #include "i915_gem.h"
40 #include "i915_scheduler.h"
41 #include "i915_selftest.h"
42 #include "i915_sw_fence.h"
43 #include "i915_vma_resource.h"
44 
45 #include <uapi/drm/i915_drm.h>
46 
47 struct drm_file;
48 struct drm_i915_gem_object;
49 struct drm_printer;
50 struct i915_deps;
51 struct i915_request;
52 
53 #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR)
54 struct i915_capture_list {
55 	struct i915_vma_resource *vma_res;
56 	struct i915_capture_list *next;
57 };
58 
59 void i915_request_free_capture_list(struct i915_capture_list *capture);
60 #else
61 #define i915_request_free_capture_list(_a) do {} while (0)
62 #endif
63 
64 #define RQ_TRACE(rq, fmt, ...) do {					\
65 	const struct i915_request *rq__ = (rq);				\
66 	ENGINE_TRACE(rq__->engine, "fence %llx:%lld, current %d " fmt,	\
67 		     rq__->fence.context, rq__->fence.seqno,		\
68 		     hwsp_seqno(rq__), ##__VA_ARGS__);			\
69 } while (0)
70 
71 enum {
72 	/*
73 	 * I915_FENCE_FLAG_ACTIVE - this request is currently submitted to HW.
74 	 *
75 	 * Set by __i915_request_submit() on handing over to HW, and cleared
76 	 * by __i915_request_unsubmit() if we preempt this request.
77 	 *
78 	 * Finally cleared for consistency on retiring the request, when
79 	 * we know the HW is no longer running this request.
80 	 *
81 	 * See i915_request_is_active()
82 	 */
83 	I915_FENCE_FLAG_ACTIVE = DMA_FENCE_FLAG_USER_BITS,
84 
85 	/*
86 	 * I915_FENCE_FLAG_PQUEUE - this request is ready for execution
87 	 *
88 	 * Using the scheduler, when a request is ready for execution it is put
89 	 * into the priority queue, and removed from that queue when transferred
90 	 * to the HW runlists. We want to track its membership within the
91 	 * priority queue so that we can easily check before rescheduling.
92 	 *
93 	 * See i915_request_in_priority_queue()
94 	 */
95 	I915_FENCE_FLAG_PQUEUE,
96 
97 	/*
98 	 * I915_FENCE_FLAG_HOLD - this request is currently on hold
99 	 *
100 	 * This request has been suspended, pending an ongoing investigation.
101 	 */
102 	I915_FENCE_FLAG_HOLD,
103 
104 	/*
105 	 * I915_FENCE_FLAG_INITIAL_BREADCRUMB - this request has the initial
106 	 * breadcrumb that marks the end of semaphore waits and start of the
107 	 * user payload.
108 	 */
109 	I915_FENCE_FLAG_INITIAL_BREADCRUMB,
110 
111 	/*
112 	 * I915_FENCE_FLAG_SIGNAL - this request is currently on signal_list
113 	 *
114 	 * Internal bookkeeping used by the breadcrumb code to track when
115 	 * a request is on the various signal_list.
116 	 */
117 	I915_FENCE_FLAG_SIGNAL,
118 
119 	/*
120 	 * I915_FENCE_FLAG_NOPREEMPT - this request should not be preempted
121 	 *
122 	 * The execution of some requests should not be interrupted. This is
123 	 * a sensitive operation as it makes the request super important,
124 	 * blocking other higher priority work. Abuse of this flag will
125 	 * lead to quality of service issues.
126 	 */
127 	I915_FENCE_FLAG_NOPREEMPT,
128 
129 	/*
130 	 * I915_FENCE_FLAG_SENTINEL - this request should be last in the queue
131 	 *
132 	 * A high priority sentinel request may be submitted to clear the
133 	 * submission queue. As it will be the only request in-flight, upon
134 	 * execution all other active requests will have been preempted and
135 	 * unsubmitted. This preemptive pulse is used to re-evaluate the
136 	 * in-flight requests, particularly in cases where an active context
137 	 * is banned and those active requests need to be cancelled.
138 	 */
139 	I915_FENCE_FLAG_SENTINEL,
140 
141 	/*
142 	 * I915_FENCE_FLAG_BOOST - upclock the gpu for this request
143 	 *
144 	 * Some requests are more important than others! In particular, a
145 	 * request that the user is waiting on is typically required for
146 	 * interactive latency, for which we want to minimise by upclocking
147 	 * the GPU. Here we track such boost requests on a per-request basis.
148 	 */
149 	I915_FENCE_FLAG_BOOST,
150 
151 	/*
152 	 * I915_FENCE_FLAG_SUBMIT_PARALLEL - request with a context in a
153 	 * parent-child relationship (parallel submission, multi-lrc) should
154 	 * trigger a submission to the GuC rather than just moving the context
155 	 * tail.
156 	 */
157 	I915_FENCE_FLAG_SUBMIT_PARALLEL,
158 
159 	/*
160 	 * I915_FENCE_FLAG_SKIP_PARALLEL - request with a context in a
161 	 * parent-child relationship (parallel submission, multi-lrc) that
162 	 * hit an error while generating requests in the execbuf IOCTL.
163 	 * Indicates this request should be skipped as another request in
164 	 * submission / relationship encoutered an error.
165 	 */
166 	I915_FENCE_FLAG_SKIP_PARALLEL,
167 
168 	/*
169 	 * I915_FENCE_FLAG_COMPOSITE - Indicates fence is part of a composite
170 	 * fence (dma_fence_array) and i915 generated for parallel submission.
171 	 */
172 	I915_FENCE_FLAG_COMPOSITE,
173 };
174 
175 /**
176  * Request queue structure.
177  *
178  * The request queue allows us to note sequence numbers that have been emitted
179  * and may be associated with active buffers to be retired.
180  *
181  * By keeping this list, we can avoid having to do questionable sequence
182  * number comparisons on buffer last_read|write_seqno. It also allows an
183  * emission time to be associated with the request for tracking how far ahead
184  * of the GPU the submission is.
185  *
186  * When modifying this structure be very aware that we perform a lockless
187  * RCU lookup of it that may race against reallocation of the struct
188  * from the slab freelist. We intentionally do not zero the structure on
189  * allocation so that the lookup can use the dangling pointers (and is
190  * cogniscent that those pointers may be wrong). Instead, everything that
191  * needs to be initialised must be done so explicitly.
192  *
193  * The requests are reference counted.
194  */
195 struct i915_request {
196 	struct dma_fence fence;
197 	spinlock_t lock;
198 
199 	/**
200 	 * Context and ring buffer related to this request
201 	 * Contexts are refcounted, so when this request is associated with a
202 	 * context, we must increment the context's refcount, to guarantee that
203 	 * it persists while any request is linked to it. Requests themselves
204 	 * are also refcounted, so the request will only be freed when the last
205 	 * reference to it is dismissed, and the code in
206 	 * i915_request_free() will then decrement the refcount on the
207 	 * context.
208 	 */
209 	struct intel_engine_cs *engine;
210 	struct intel_context *context;
211 	struct intel_ring *ring;
212 	struct intel_timeline __rcu *timeline;
213 
214 	struct list_head signal_link;
215 	struct llist_node signal_node;
216 
217 	/*
218 	 * The rcu epoch of when this request was allocated. Used to judiciously
219 	 * apply backpressure on future allocations to ensure that under
220 	 * mempressure there is sufficient RCU ticks for us to reclaim our
221 	 * RCU protected slabs.
222 	 */
223 	unsigned long rcustate;
224 
225 	/*
226 	 * We pin the timeline->mutex while constructing the request to
227 	 * ensure that no caller accidentally drops it during construction.
228 	 * The timeline->mutex must be held to ensure that only this caller
229 	 * can use the ring and manipulate the associated timeline during
230 	 * construction.
231 	 */
232 	struct pin_cookie cookie;
233 
234 	/*
235 	 * Fences for the various phases in the request's lifetime.
236 	 *
237 	 * The submit fence is used to await upon all of the request's
238 	 * dependencies. When it is signaled, the request is ready to run.
239 	 * It is used by the driver to then queue the request for execution.
240 	 */
241 	struct i915_sw_fence submit;
242 	union {
243 		wait_queue_entry_t submitq;
244 		struct i915_sw_dma_fence_cb dmaq;
245 		struct i915_request_duration_cb {
246 			struct dma_fence_cb cb;
247 			ktime_t emitted;
248 		} duration;
249 	};
250 	struct llist_head execute_cb;
251 	struct i915_sw_fence semaphore;
252 	/**
253 	 * @submit_work: complete submit fence from an IRQ if needed for
254 	 * locking hierarchy reasons.
255 	 */
256 	struct irq_work submit_work;
257 
258 	/*
259 	 * A list of everyone we wait upon, and everyone who waits upon us.
260 	 * Even though we will not be submitted to the hardware before the
261 	 * submit fence is signaled (it waits for all external events as well
262 	 * as our own requests), the scheduler still needs to know the
263 	 * dependency tree for the lifetime of the request (from execbuf
264 	 * to retirement), i.e. bidirectional dependency information for the
265 	 * request not tied to individual fences.
266 	 */
267 	struct i915_sched_node sched;
268 	struct i915_dependency dep;
269 	intel_engine_mask_t execution_mask;
270 
271 	/*
272 	 * A convenience pointer to the current breadcrumb value stored in
273 	 * the HW status page (or our timeline's local equivalent). The full
274 	 * path would be rq->hw_context->ring->timeline->hwsp_seqno.
275 	 */
276 	const u32 *hwsp_seqno;
277 
278 	/** Position in the ring of the start of the request */
279 	u32 head;
280 
281 	/** Position in the ring of the start of the user packets */
282 	u32 infix;
283 
284 	/**
285 	 * Position in the ring of the start of the postfix.
286 	 * This is required to calculate the maximum available ring space
287 	 * without overwriting the postfix.
288 	 */
289 	u32 postfix;
290 
291 	/** Position in the ring of the end of the whole request */
292 	u32 tail;
293 
294 	/** Position in the ring of the end of any workarounds after the tail */
295 	u32 wa_tail;
296 
297 	/** Preallocate space in the ring for the emitting the request */
298 	u32 reserved_space;
299 
300 	/** Batch buffer pointer for selftest internal use. */
301 	I915_SELFTEST_DECLARE(struct i915_vma *batch);
302 
303 	struct i915_vma_resource *batch_res;
304 
305 #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR)
306 	/**
307 	 * Additional buffers requested by userspace to be captured upon
308 	 * a GPU hang. The vma/obj on this list are protected by their
309 	 * active reference - all objects on this list must also be
310 	 * on the active_list (of their final request).
311 	 */
312 	struct i915_capture_list *capture_list;
313 #endif
314 
315 	/** Time at which this request was emitted, in jiffies. */
316 	unsigned long emitted_jiffies;
317 
318 	/** timeline->request entry for this request */
319 	struct list_head link;
320 
321 	/** Watchdog support fields. */
322 	struct i915_request_watchdog {
323 		struct llist_node link;
324 		struct hrtimer timer;
325 	} watchdog;
326 
327 	/**
328 	 * @guc_fence_link: Requests may need to be stalled when using GuC
329 	 * submission waiting for certain GuC operations to complete. If that is
330 	 * the case, stalled requests are added to a per context list of stalled
331 	 * requests. The below list_head is the link in that list. Protected by
332 	 * ce->guc_state.lock.
333 	 */
334 	struct list_head guc_fence_link;
335 
336 	/**
337 	 * @guc_prio: Priority level while the request is in flight. Differs
338 	 * from i915 scheduler priority. See comment above
339 	 * I915_SCHEDULER_CAP_STATIC_PRIORITY_MAP for details. Protected by
340 	 * ce->guc_active.lock. Two special values (GUC_PRIO_INIT and
341 	 * GUC_PRIO_FINI) outside the GuC priority range are used to indicate
342 	 * if the priority has not been initialized yet or if no more updates
343 	 * are possible because the request has completed.
344 	 */
345 #define	GUC_PRIO_INIT	0xff
346 #define	GUC_PRIO_FINI	0xfe
347 	u8 guc_prio;
348 
349 	I915_SELFTEST_DECLARE(struct {
350 		struct list_head link;
351 		unsigned long delay;
352 	} mock;)
353 };
354 
355 #define I915_FENCE_GFP (GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN)
356 
357 extern const struct dma_fence_ops i915_fence_ops;
358 
dma_fence_is_i915(const struct dma_fence * fence)359 static inline bool dma_fence_is_i915(const struct dma_fence *fence)
360 {
361 	return fence->ops == &i915_fence_ops;
362 }
363 
364 struct kmem_cache *i915_request_slab_cache(void);
365 
366 struct i915_request * __must_check
367 __i915_request_create(struct intel_context *ce, gfp_t gfp);
368 struct i915_request * __must_check
369 i915_request_create(struct intel_context *ce);
370 
371 void __i915_request_skip(struct i915_request *rq);
372 bool i915_request_set_error_once(struct i915_request *rq, int error);
373 struct i915_request *i915_request_mark_eio(struct i915_request *rq);
374 
375 struct i915_request *__i915_request_commit(struct i915_request *request);
376 void __i915_request_queue(struct i915_request *rq,
377 			  const struct i915_sched_attr *attr);
378 void __i915_request_queue_bh(struct i915_request *rq);
379 
380 bool i915_request_retire(struct i915_request *rq);
381 void i915_request_retire_upto(struct i915_request *rq);
382 
383 static inline struct i915_request *
to_request(struct dma_fence * fence)384 to_request(struct dma_fence *fence)
385 {
386 	/* We assume that NULL fence/request are interoperable */
387 	BUILD_BUG_ON(offsetof(struct i915_request, fence) != 0);
388 	GEM_BUG_ON(fence && !dma_fence_is_i915(fence));
389 	return container_of(fence, struct i915_request, fence);
390 }
391 
392 static inline struct i915_request *
i915_request_get(struct i915_request * rq)393 i915_request_get(struct i915_request *rq)
394 {
395 	return to_request(dma_fence_get(&rq->fence));
396 }
397 
398 static inline struct i915_request *
i915_request_get_rcu(struct i915_request * rq)399 i915_request_get_rcu(struct i915_request *rq)
400 {
401 	return to_request(dma_fence_get_rcu(&rq->fence));
402 }
403 
404 static inline void
i915_request_put(struct i915_request * rq)405 i915_request_put(struct i915_request *rq)
406 {
407 	dma_fence_put(&rq->fence);
408 }
409 
410 int i915_request_await_object(struct i915_request *to,
411 			      struct drm_i915_gem_object *obj,
412 			      bool write);
413 int i915_request_await_dma_fence(struct i915_request *rq,
414 				 struct dma_fence *fence);
415 int i915_request_await_deps(struct i915_request *rq, const struct i915_deps *deps);
416 int i915_request_await_execution(struct i915_request *rq,
417 				 struct dma_fence *fence);
418 
419 void i915_request_add(struct i915_request *rq);
420 
421 bool __i915_request_submit(struct i915_request *request);
422 void i915_request_submit(struct i915_request *request);
423 
424 void __i915_request_unsubmit(struct i915_request *request);
425 void i915_request_unsubmit(struct i915_request *request);
426 
427 void i915_request_cancel(struct i915_request *rq, int error);
428 
429 long i915_request_wait_timeout(struct i915_request *rq,
430 			       unsigned int flags,
431 			       long timeout)
432 	__attribute__((nonnull(1)));
433 
434 long i915_request_wait(struct i915_request *rq,
435 		       unsigned int flags,
436 		       long timeout)
437 	__attribute__((nonnull(1)));
438 #define I915_WAIT_INTERRUPTIBLE	BIT(0)
439 #define I915_WAIT_PRIORITY	BIT(1) /* small priority bump for the request */
440 #define I915_WAIT_ALL		BIT(2) /* used by i915_gem_object_wait() */
441 
442 void i915_request_show(struct drm_printer *m,
443 		       const struct i915_request *rq,
444 		       const char *prefix,
445 		       int indent);
446 
i915_request_signaled(const struct i915_request * rq)447 static inline bool i915_request_signaled(const struct i915_request *rq)
448 {
449 	/* The request may live longer than its HWSP, so check flags first! */
450 	return test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &rq->fence.flags);
451 }
452 
i915_request_is_active(const struct i915_request * rq)453 static inline bool i915_request_is_active(const struct i915_request *rq)
454 {
455 	return test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags);
456 }
457 
i915_request_in_priority_queue(const struct i915_request * rq)458 static inline bool i915_request_in_priority_queue(const struct i915_request *rq)
459 {
460 	return test_bit(I915_FENCE_FLAG_PQUEUE, &rq->fence.flags);
461 }
462 
463 static inline bool
i915_request_has_initial_breadcrumb(const struct i915_request * rq)464 i915_request_has_initial_breadcrumb(const struct i915_request *rq)
465 {
466 	return test_bit(I915_FENCE_FLAG_INITIAL_BREADCRUMB, &rq->fence.flags);
467 }
468 
469 /**
470  * Returns true if seq1 is later than seq2.
471  */
i915_seqno_passed(u32 seq1,u32 seq2)472 static inline bool i915_seqno_passed(u32 seq1, u32 seq2)
473 {
474 	return (s32)(seq1 - seq2) >= 0;
475 }
476 
__hwsp_seqno(const struct i915_request * rq)477 static inline u32 __hwsp_seqno(const struct i915_request *rq)
478 {
479 	const u32 *hwsp = READ_ONCE(rq->hwsp_seqno);
480 
481 	return READ_ONCE(*hwsp);
482 }
483 
484 /**
485  * hwsp_seqno - the current breadcrumb value in the HW status page
486  * @rq: the request, to chase the relevant HW status page
487  *
488  * The emphasis in naming here is that hwsp_seqno() is not a property of the
489  * request, but an indication of the current HW state (associated with this
490  * request). Its value will change as the GPU executes more requests.
491  *
492  * Returns the current breadcrumb value in the associated HW status page (or
493  * the local timeline's equivalent) for this request. The request itself
494  * has the associated breadcrumb value of rq->fence.seqno, when the HW
495  * status page has that breadcrumb or later, this request is complete.
496  */
hwsp_seqno(const struct i915_request * rq)497 static inline u32 hwsp_seqno(const struct i915_request *rq)
498 {
499 	u32 seqno;
500 
501 	rcu_read_lock(); /* the HWSP may be freed at runtime */
502 	seqno = __hwsp_seqno(rq);
503 	rcu_read_unlock();
504 
505 	return seqno;
506 }
507 
__i915_request_has_started(const struct i915_request * rq)508 static inline bool __i915_request_has_started(const struct i915_request *rq)
509 {
510 	return i915_seqno_passed(__hwsp_seqno(rq), rq->fence.seqno - 1);
511 }
512 
513 /**
514  * i915_request_started - check if the request has begun being executed
515  * @rq: the request
516  *
517  * If the timeline is not using initial breadcrumbs, a request is
518  * considered started if the previous request on its timeline (i.e.
519  * context) has been signaled.
520  *
521  * If the timeline is using semaphores, it will also be emitting an
522  * "initial breadcrumb" after the semaphores are complete and just before
523  * it began executing the user payload. A request can therefore be active
524  * on the HW and not yet started as it is still busywaiting on its
525  * dependencies (via HW semaphores).
526  *
527  * If the request has started, its dependencies will have been signaled
528  * (either by fences or by semaphores) and it will have begun processing
529  * the user payload.
530  *
531  * However, even if a request has started, it may have been preempted and
532  * so no longer active, or it may have already completed.
533  *
534  * See also i915_request_is_active().
535  *
536  * Returns true if the request has begun executing the user payload, or
537  * has completed:
538  */
i915_request_started(const struct i915_request * rq)539 static inline bool i915_request_started(const struct i915_request *rq)
540 {
541 	bool result;
542 
543 	if (i915_request_signaled(rq))
544 		return true;
545 
546 	result = true;
547 	rcu_read_lock(); /* the HWSP may be freed at runtime */
548 	if (likely(!i915_request_signaled(rq)))
549 		/* Remember: started but may have since been preempted! */
550 		result = __i915_request_has_started(rq);
551 	rcu_read_unlock();
552 
553 	return result;
554 }
555 
556 /**
557  * i915_request_is_running - check if the request may actually be executing
558  * @rq: the request
559  *
560  * Returns true if the request is currently submitted to hardware, has passed
561  * its start point (i.e. the context is setup and not busywaiting). Note that
562  * it may no longer be running by the time the function returns!
563  */
i915_request_is_running(const struct i915_request * rq)564 static inline bool i915_request_is_running(const struct i915_request *rq)
565 {
566 	bool result;
567 
568 	if (!i915_request_is_active(rq))
569 		return false;
570 
571 	rcu_read_lock();
572 	result = __i915_request_has_started(rq) && i915_request_is_active(rq);
573 	rcu_read_unlock();
574 
575 	return result;
576 }
577 
578 /**
579  * i915_request_is_ready - check if the request is ready for execution
580  * @rq: the request
581  *
582  * Upon construction, the request is instructed to wait upon various
583  * signals before it is ready to be executed by the HW. That is, we do
584  * not want to start execution and read data before it is written. In practice,
585  * this is controlled with a mixture of interrupts and semaphores. Once
586  * the submit fence is completed, the backend scheduler will place the
587  * request into its queue and from there submit it for execution. So we
588  * can detect when a request is eligible for execution (and is under control
589  * of the scheduler) by querying where it is in any of the scheduler's lists.
590  *
591  * Returns true if the request is ready for execution (it may be inflight),
592  * false otherwise.
593  */
i915_request_is_ready(const struct i915_request * rq)594 static inline bool i915_request_is_ready(const struct i915_request *rq)
595 {
596 	return !list_empty(&rq->sched.link);
597 }
598 
__i915_request_is_complete(const struct i915_request * rq)599 static inline bool __i915_request_is_complete(const struct i915_request *rq)
600 {
601 	return i915_seqno_passed(__hwsp_seqno(rq), rq->fence.seqno);
602 }
603 
i915_request_completed(const struct i915_request * rq)604 static inline bool i915_request_completed(const struct i915_request *rq)
605 {
606 	bool result;
607 
608 	if (i915_request_signaled(rq))
609 		return true;
610 
611 	result = true;
612 	rcu_read_lock(); /* the HWSP may be freed at runtime */
613 	if (likely(!i915_request_signaled(rq)))
614 		result = __i915_request_is_complete(rq);
615 	rcu_read_unlock();
616 
617 	return result;
618 }
619 
i915_request_mark_complete(struct i915_request * rq)620 static inline void i915_request_mark_complete(struct i915_request *rq)
621 {
622 	WRITE_ONCE(rq->hwsp_seqno, /* decouple from HWSP */
623 		   (u32 *)&rq->fence.seqno);
624 }
625 
i915_request_has_waitboost(const struct i915_request * rq)626 static inline bool i915_request_has_waitboost(const struct i915_request *rq)
627 {
628 	return test_bit(I915_FENCE_FLAG_BOOST, &rq->fence.flags);
629 }
630 
i915_request_has_nopreempt(const struct i915_request * rq)631 static inline bool i915_request_has_nopreempt(const struct i915_request *rq)
632 {
633 	/* Preemption should only be disabled very rarely */
634 	return unlikely(test_bit(I915_FENCE_FLAG_NOPREEMPT, &rq->fence.flags));
635 }
636 
i915_request_has_sentinel(const struct i915_request * rq)637 static inline bool i915_request_has_sentinel(const struct i915_request *rq)
638 {
639 	return unlikely(test_bit(I915_FENCE_FLAG_SENTINEL, &rq->fence.flags));
640 }
641 
i915_request_on_hold(const struct i915_request * rq)642 static inline bool i915_request_on_hold(const struct i915_request *rq)
643 {
644 	return unlikely(test_bit(I915_FENCE_FLAG_HOLD, &rq->fence.flags));
645 }
646 
i915_request_set_hold(struct i915_request * rq)647 static inline void i915_request_set_hold(struct i915_request *rq)
648 {
649 	set_bit(I915_FENCE_FLAG_HOLD, &rq->fence.flags);
650 }
651 
i915_request_clear_hold(struct i915_request * rq)652 static inline void i915_request_clear_hold(struct i915_request *rq)
653 {
654 	clear_bit(I915_FENCE_FLAG_HOLD, &rq->fence.flags);
655 }
656 
657 static inline struct intel_timeline *
i915_request_timeline(const struct i915_request * rq)658 i915_request_timeline(const struct i915_request *rq)
659 {
660 	/* Valid only while the request is being constructed (or retired). */
661 	return rcu_dereference_protected(rq->timeline,
662 					 lockdep_is_held(&rcu_access_pointer(rq->timeline)->mutex) ||
663 					 test_bit(CONTEXT_IS_PARKING, &rq->context->flags));
664 }
665 
666 static inline struct i915_gem_context *
i915_request_gem_context(const struct i915_request * rq)667 i915_request_gem_context(const struct i915_request *rq)
668 {
669 	/* Valid only while the request is being constructed (or retired). */
670 	return rcu_dereference_protected(rq->context->gem_context, true);
671 }
672 
673 static inline struct intel_timeline *
i915_request_active_timeline(const struct i915_request * rq)674 i915_request_active_timeline(const struct i915_request *rq)
675 {
676 	/*
677 	 * When in use during submission, we are protected by a guarantee that
678 	 * the context/timeline is pinned and must remain pinned until after
679 	 * this submission.
680 	 */
681 	return rcu_dereference_protected(rq->timeline,
682 					 lockdep_is_held(&rq->engine->sched_engine->lock));
683 }
684 
685 static inline u32
i915_request_active_seqno(const struct i915_request * rq)686 i915_request_active_seqno(const struct i915_request *rq)
687 {
688 	u32 hwsp_phys_base =
689 		page_mask_bits(i915_request_active_timeline(rq)->hwsp_offset);
690 	u32 hwsp_relative_offset = offset_in_page(rq->hwsp_seqno);
691 
692 	/*
693 	 * Because of wraparound, we cannot simply take tl->hwsp_offset,
694 	 * but instead use the fact that the relative for vaddr is the
695 	 * offset as for hwsp_offset. Take the top bits from tl->hwsp_offset
696 	 * and combine them with the relative offset in rq->hwsp_seqno.
697 	 *
698 	 * As rw->hwsp_seqno is rewritten when signaled, this only works
699 	 * when the request isn't signaled yet, but at that point you
700 	 * no longer need the offset.
701 	 */
702 
703 	return hwsp_phys_base + hwsp_relative_offset;
704 }
705 
706 bool
707 i915_request_active_engine(struct i915_request *rq,
708 			   struct intel_engine_cs **active);
709 
710 void i915_request_notify_execute_cb_imm(struct i915_request *rq);
711 
712 enum i915_request_state {
713 	I915_REQUEST_UNKNOWN = 0,
714 	I915_REQUEST_COMPLETE,
715 	I915_REQUEST_PENDING,
716 	I915_REQUEST_QUEUED,
717 	I915_REQUEST_ACTIVE,
718 };
719 
720 enum i915_request_state i915_test_request_state(struct i915_request *rq);
721 
722 void i915_request_module_exit(void);
723 int i915_request_module_init(void);
724 
725 #endif /* I915_REQUEST_H */
726