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/lockdep.h>
30 
31 #include "gem/i915_gem_context_types.h"
32 #include "gt/intel_context_types.h"
33 #include "gt/intel_engine_types.h"
34 #include "gt/intel_timeline_types.h"
35 
36 #include "i915_gem.h"
37 #include "i915_scheduler.h"
38 #include "i915_selftest.h"
39 #include "i915_sw_fence.h"
40 
41 #include <uapi/drm/i915_drm.h>
42 
43 struct drm_file;
44 struct drm_i915_gem_object;
45 struct i915_request;
46 
47 struct i915_capture_list {
48 	struct i915_capture_list *next;
49 	struct i915_vma *vma;
50 };
51 
52 #define RQ_TRACE(rq, fmt, ...) do {					\
53 	const struct i915_request *rq__ = (rq);				\
54 	ENGINE_TRACE(rq__->engine, "fence %llx:%lld, current %d " fmt,	\
55 		     rq__->fence.context, rq__->fence.seqno,		\
56 		     hwsp_seqno(rq__), ##__VA_ARGS__);			\
57 } while (0)
58 
59 enum {
60 	/*
61 	 * I915_FENCE_FLAG_ACTIVE - this request is currently submitted to HW.
62 	 *
63 	 * Set by __i915_request_submit() on handing over to HW, and cleared
64 	 * by __i915_request_unsubmit() if we preempt this request.
65 	 *
66 	 * Finally cleared for consistency on retiring the request, when
67 	 * we know the HW is no longer running this request.
68 	 *
69 	 * See i915_request_is_active()
70 	 */
71 	I915_FENCE_FLAG_ACTIVE = DMA_FENCE_FLAG_USER_BITS,
72 
73 	/*
74 	 * I915_FENCE_FLAG_SIGNAL - this request is currently on signal_list
75 	 *
76 	 * Internal bookkeeping used by the breadcrumb code to track when
77 	 * a request is on the various signal_list.
78 	 */
79 	I915_FENCE_FLAG_SIGNAL,
80 
81 	/*
82 	 * I915_FENCE_FLAG_NOPREEMPT - this request should not be preempted
83 	 *
84 	 * The execution of some requests should not be interrupted. This is
85 	 * a sensitive operation as it makes the request super important,
86 	 * blocking other higher priority work. Abuse of this flag will
87 	 * lead to quality of service issues.
88 	 */
89 	I915_FENCE_FLAG_NOPREEMPT,
90 
91 	/*
92 	 * I915_FENCE_FLAG_SENTINEL - this request should be last in the queue
93 	 *
94 	 * A high priority sentinel request may be submitted to clear the
95 	 * submission queue. As it will be the only request in-flight, upon
96 	 * execution all other active requests will have been preempted and
97 	 * unsubmitted. This preemptive pulse is used to re-evaluate the
98 	 * in-flight requests, particularly in cases where an active context
99 	 * is banned and those active requests need to be cancelled.
100 	 */
101 	I915_FENCE_FLAG_SENTINEL,
102 
103 	/*
104 	 * I915_FENCE_FLAG_BOOST - upclock the gpu for this request
105 	 *
106 	 * Some requests are more important than others! In particular, a
107 	 * request that the user is waiting on is typically required for
108 	 * interactive latency, for which we want to minimise by upclocking
109 	 * the GPU. Here we track such boost requests on a per-request basis.
110 	 */
111 	I915_FENCE_FLAG_BOOST,
112 };
113 
114 /**
115  * Request queue structure.
116  *
117  * The request queue allows us to note sequence numbers that have been emitted
118  * and may be associated with active buffers to be retired.
119  *
120  * By keeping this list, we can avoid having to do questionable sequence
121  * number comparisons on buffer last_read|write_seqno. It also allows an
122  * emission time to be associated with the request for tracking how far ahead
123  * of the GPU the submission is.
124  *
125  * When modifying this structure be very aware that we perform a lockless
126  * RCU lookup of it that may race against reallocation of the struct
127  * from the slab freelist. We intentionally do not zero the structure on
128  * allocation so that the lookup can use the dangling pointers (and is
129  * cogniscent that those pointers may be wrong). Instead, everything that
130  * needs to be initialised must be done so explicitly.
131  *
132  * The requests are reference counted.
133  */
134 struct i915_request {
135 	struct dma_fence fence;
136 	spinlock_t lock;
137 
138 	/** On Which ring this request was generated */
139 	struct drm_i915_private *i915;
140 
141 	/**
142 	 * Context and ring buffer related to this request
143 	 * Contexts are refcounted, so when this request is associated with a
144 	 * context, we must increment the context's refcount, to guarantee that
145 	 * it persists while any request is linked to it. Requests themselves
146 	 * are also refcounted, so the request will only be freed when the last
147 	 * reference to it is dismissed, and the code in
148 	 * i915_request_free() will then decrement the refcount on the
149 	 * context.
150 	 */
151 	struct intel_engine_cs *engine;
152 	struct intel_context *context;
153 	struct intel_ring *ring;
154 	struct intel_timeline __rcu *timeline;
155 	struct list_head signal_link;
156 
157 	/*
158 	 * The rcu epoch of when this request was allocated. Used to judiciously
159 	 * apply backpressure on future allocations to ensure that under
160 	 * mempressure there is sufficient RCU ticks for us to reclaim our
161 	 * RCU protected slabs.
162 	 */
163 	unsigned long rcustate;
164 
165 	/*
166 	 * We pin the timeline->mutex while constructing the request to
167 	 * ensure that no caller accidentally drops it during construction.
168 	 * The timeline->mutex must be held to ensure that only this caller
169 	 * can use the ring and manipulate the associated timeline during
170 	 * construction.
171 	 */
172 	struct pin_cookie cookie;
173 
174 	/*
175 	 * Fences for the various phases in the request's lifetime.
176 	 *
177 	 * The submit fence is used to await upon all of the request's
178 	 * dependencies. When it is signaled, the request is ready to run.
179 	 * It is used by the driver to then queue the request for execution.
180 	 */
181 	struct i915_sw_fence submit;
182 	union {
183 		wait_queue_entry_t submitq;
184 		struct i915_sw_dma_fence_cb dmaq;
185 		struct i915_request_duration_cb {
186 			struct dma_fence_cb cb;
187 			ktime_t emitted;
188 		} duration;
189 	};
190 	struct list_head execute_cb;
191 	struct i915_sw_fence semaphore;
192 
193 	/*
194 	 * A list of everyone we wait upon, and everyone who waits upon us.
195 	 * Even though we will not be submitted to the hardware before the
196 	 * submit fence is signaled (it waits for all external events as well
197 	 * as our own requests), the scheduler still needs to know the
198 	 * dependency tree for the lifetime of the request (from execbuf
199 	 * to retirement), i.e. bidirectional dependency information for the
200 	 * request not tied to individual fences.
201 	 */
202 	struct i915_sched_node sched;
203 	struct i915_dependency dep;
204 	intel_engine_mask_t execution_mask;
205 
206 	/*
207 	 * A convenience pointer to the current breadcrumb value stored in
208 	 * the HW status page (or our timeline's local equivalent). The full
209 	 * path would be rq->hw_context->ring->timeline->hwsp_seqno.
210 	 */
211 	const u32 *hwsp_seqno;
212 
213 	/*
214 	 * If we need to access the timeline's seqno for this request in
215 	 * another request, we need to keep a read reference to this associated
216 	 * cacheline, so that we do not free and recycle it before the foreign
217 	 * observers have completed. Hence, we keep a pointer to the cacheline
218 	 * inside the timeline's HWSP vma, but it is only valid while this
219 	 * request has not completed and guarded by the timeline mutex.
220 	 */
221 	struct intel_timeline_cacheline __rcu *hwsp_cacheline;
222 
223 	/** Position in the ring of the start of the request */
224 	u32 head;
225 
226 	/** Position in the ring of the start of the user packets */
227 	u32 infix;
228 
229 	/**
230 	 * Position in the ring of the start of the postfix.
231 	 * This is required to calculate the maximum available ring space
232 	 * without overwriting the postfix.
233 	 */
234 	u32 postfix;
235 
236 	/** Position in the ring of the end of the whole request */
237 	u32 tail;
238 
239 	/** Position in the ring of the end of any workarounds after the tail */
240 	u32 wa_tail;
241 
242 	/** Preallocate space in the ring for the emitting the request */
243 	u32 reserved_space;
244 
245 	/** Batch buffer related to this request if any (used for
246 	 * error state dump only).
247 	 */
248 	struct i915_vma *batch;
249 	/**
250 	 * Additional buffers requested by userspace to be captured upon
251 	 * a GPU hang. The vma/obj on this list are protected by their
252 	 * active reference - all objects on this list must also be
253 	 * on the active_list (of their final request).
254 	 */
255 	struct i915_capture_list *capture_list;
256 
257 	/** Time at which this request was emitted, in jiffies. */
258 	unsigned long emitted_jiffies;
259 
260 	/** timeline->request entry for this request */
261 	struct list_head link;
262 
263 	struct drm_i915_file_private *file_priv;
264 	/** file_priv list entry for this request */
265 	struct list_head client_link;
266 
267 	I915_SELFTEST_DECLARE(struct {
268 		struct list_head link;
269 		unsigned long delay;
270 	} mock;)
271 };
272 
273 #define I915_FENCE_GFP (GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN)
274 
275 extern const struct dma_fence_ops i915_fence_ops;
276 
277 static inline bool dma_fence_is_i915(const struct dma_fence *fence)
278 {
279 	return fence->ops == &i915_fence_ops;
280 }
281 
282 struct i915_request * __must_check
283 __i915_request_create(struct intel_context *ce, gfp_t gfp);
284 struct i915_request * __must_check
285 i915_request_create(struct intel_context *ce);
286 
287 struct i915_request *__i915_request_commit(struct i915_request *request);
288 void __i915_request_queue(struct i915_request *rq,
289 			  const struct i915_sched_attr *attr);
290 
291 bool i915_request_retire(struct i915_request *rq);
292 void i915_request_retire_upto(struct i915_request *rq);
293 
294 static inline struct i915_request *
295 to_request(struct dma_fence *fence)
296 {
297 	/* We assume that NULL fence/request are interoperable */
298 	BUILD_BUG_ON(offsetof(struct i915_request, fence) != 0);
299 	GEM_BUG_ON(fence && !dma_fence_is_i915(fence));
300 	return container_of(fence, struct i915_request, fence);
301 }
302 
303 static inline struct i915_request *
304 i915_request_get(struct i915_request *rq)
305 {
306 	return to_request(dma_fence_get(&rq->fence));
307 }
308 
309 static inline struct i915_request *
310 i915_request_get_rcu(struct i915_request *rq)
311 {
312 	return to_request(dma_fence_get_rcu(&rq->fence));
313 }
314 
315 static inline void
316 i915_request_put(struct i915_request *rq)
317 {
318 	dma_fence_put(&rq->fence);
319 }
320 
321 int i915_request_await_object(struct i915_request *to,
322 			      struct drm_i915_gem_object *obj,
323 			      bool write);
324 int i915_request_await_dma_fence(struct i915_request *rq,
325 				 struct dma_fence *fence);
326 int i915_request_await_execution(struct i915_request *rq,
327 				 struct dma_fence *fence,
328 				 void (*hook)(struct i915_request *rq,
329 					      struct dma_fence *signal));
330 
331 void i915_request_add(struct i915_request *rq);
332 
333 bool __i915_request_submit(struct i915_request *request);
334 void i915_request_submit(struct i915_request *request);
335 
336 void i915_request_skip(struct i915_request *request, int error);
337 
338 void __i915_request_unsubmit(struct i915_request *request);
339 void i915_request_unsubmit(struct i915_request *request);
340 
341 /* Note: part of the intel_breadcrumbs family */
342 bool i915_request_enable_breadcrumb(struct i915_request *request);
343 void i915_request_cancel_breadcrumb(struct i915_request *request);
344 
345 long i915_request_wait(struct i915_request *rq,
346 		       unsigned int flags,
347 		       long timeout)
348 	__attribute__((nonnull(1)));
349 #define I915_WAIT_INTERRUPTIBLE	BIT(0)
350 #define I915_WAIT_PRIORITY	BIT(1) /* small priority bump for the request */
351 #define I915_WAIT_ALL		BIT(2) /* used by i915_gem_object_wait() */
352 
353 static inline bool i915_request_signaled(const struct i915_request *rq)
354 {
355 	/* The request may live longer than its HWSP, so check flags first! */
356 	return test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &rq->fence.flags);
357 }
358 
359 static inline bool i915_request_is_active(const struct i915_request *rq)
360 {
361 	return test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags);
362 }
363 
364 /**
365  * Returns true if seq1 is later than seq2.
366  */
367 static inline bool i915_seqno_passed(u32 seq1, u32 seq2)
368 {
369 	return (s32)(seq1 - seq2) >= 0;
370 }
371 
372 static inline u32 __hwsp_seqno(const struct i915_request *rq)
373 {
374 	return READ_ONCE(*rq->hwsp_seqno);
375 }
376 
377 /**
378  * hwsp_seqno - the current breadcrumb value in the HW status page
379  * @rq: the request, to chase the relevant HW status page
380  *
381  * The emphasis in naming here is that hwsp_seqno() is not a property of the
382  * request, but an indication of the current HW state (associated with this
383  * request). Its value will change as the GPU executes more requests.
384  *
385  * Returns the current breadcrumb value in the associated HW status page (or
386  * the local timeline's equivalent) for this request. The request itself
387  * has the associated breadcrumb value of rq->fence.seqno, when the HW
388  * status page has that breadcrumb or later, this request is complete.
389  */
390 static inline u32 hwsp_seqno(const struct i915_request *rq)
391 {
392 	u32 seqno;
393 
394 	rcu_read_lock(); /* the HWSP may be freed at runtime */
395 	seqno = __hwsp_seqno(rq);
396 	rcu_read_unlock();
397 
398 	return seqno;
399 }
400 
401 static inline bool __i915_request_has_started(const struct i915_request *rq)
402 {
403 	return i915_seqno_passed(hwsp_seqno(rq), rq->fence.seqno - 1);
404 }
405 
406 /**
407  * i915_request_started - check if the request has begun being executed
408  * @rq: the request
409  *
410  * If the timeline is not using initial breadcrumbs, a request is
411  * considered started if the previous request on its timeline (i.e.
412  * context) has been signaled.
413  *
414  * If the timeline is using semaphores, it will also be emitting an
415  * "initial breadcrumb" after the semaphores are complete and just before
416  * it began executing the user payload. A request can therefore be active
417  * on the HW and not yet started as it is still busywaiting on its
418  * dependencies (via HW semaphores).
419  *
420  * If the request has started, its dependencies will have been signaled
421  * (either by fences or by semaphores) and it will have begun processing
422  * the user payload.
423  *
424  * However, even if a request has started, it may have been preempted and
425  * so no longer active, or it may have already completed.
426  *
427  * See also i915_request_is_active().
428  *
429  * Returns true if the request has begun executing the user payload, or
430  * has completed:
431  */
432 static inline bool i915_request_started(const struct i915_request *rq)
433 {
434 	if (i915_request_signaled(rq))
435 		return true;
436 
437 	/* Remember: started but may have since been preempted! */
438 	return __i915_request_has_started(rq);
439 }
440 
441 /**
442  * i915_request_is_running - check if the request may actually be executing
443  * @rq: the request
444  *
445  * Returns true if the request is currently submitted to hardware, has passed
446  * its start point (i.e. the context is setup and not busywaiting). Note that
447  * it may no longer be running by the time the function returns!
448  */
449 static inline bool i915_request_is_running(const struct i915_request *rq)
450 {
451 	if (!i915_request_is_active(rq))
452 		return false;
453 
454 	return __i915_request_has_started(rq);
455 }
456 
457 static inline bool i915_request_completed(const struct i915_request *rq)
458 {
459 	if (i915_request_signaled(rq))
460 		return true;
461 
462 	return i915_seqno_passed(hwsp_seqno(rq), rq->fence.seqno);
463 }
464 
465 static inline void i915_request_mark_complete(struct i915_request *rq)
466 {
467 	rq->hwsp_seqno = (u32 *)&rq->fence.seqno; /* decouple from HWSP */
468 }
469 
470 static inline bool i915_request_has_waitboost(const struct i915_request *rq)
471 {
472 	return test_bit(I915_FENCE_FLAG_BOOST, &rq->fence.flags);
473 }
474 
475 static inline bool i915_request_has_nopreempt(const struct i915_request *rq)
476 {
477 	/* Preemption should only be disabled very rarely */
478 	return unlikely(test_bit(I915_FENCE_FLAG_NOPREEMPT, &rq->fence.flags));
479 }
480 
481 static inline bool i915_request_has_sentinel(const struct i915_request *rq)
482 {
483 	return unlikely(test_bit(I915_FENCE_FLAG_SENTINEL, &rq->fence.flags));
484 }
485 
486 static inline struct intel_timeline *
487 i915_request_timeline(struct i915_request *rq)
488 {
489 	/* Valid only while the request is being constructed (or retired). */
490 	return rcu_dereference_protected(rq->timeline,
491 					 lockdep_is_held(&rcu_access_pointer(rq->timeline)->mutex));
492 }
493 
494 static inline struct i915_gem_context *
495 i915_request_gem_context(struct i915_request *rq)
496 {
497 	/* Valid only while the request is being constructed (or retired). */
498 	return rcu_dereference_protected(rq->context->gem_context, true);
499 }
500 
501 static inline struct intel_timeline *
502 i915_request_active_timeline(struct i915_request *rq)
503 {
504 	/*
505 	 * When in use during submission, we are protected by a guarantee that
506 	 * the context/timeline is pinned and must remain pinned until after
507 	 * this submission.
508 	 */
509 	return rcu_dereference_protected(rq->timeline,
510 					 lockdep_is_held(&rq->engine->active.lock));
511 }
512 
513 #endif /* I915_REQUEST_H */
514