2 * Copyright © 2008-2018 Intel Corporation
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:
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
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
25 #ifndef I915_REQUEST_H
26 #define I915_REQUEST_H
28 #include <linux/dma-fence.h>
29 #include <linux/lockdep.h>
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"
37 #include "i915_scheduler.h"
38 #include "i915_selftest.h"
39 #include "i915_sw_fence.h"
41 #include <uapi/drm/i915_drm.h>
44 struct drm_i915_gem_object;
47 struct i915_capture_list {
48 struct i915_capture_list *next;
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__); \
61 * I915_FENCE_FLAG_ACTIVE - this request is currently submitted to HW.
63 * Set by __i915_request_submit() on handing over to HW, and cleared
64 * by __i915_request_unsubmit() if we preempt this request.
66 * Finally cleared for consistency on retiring the request, when
67 * we know the HW is no longer running this request.
69 * See i915_request_is_active()
71 I915_FENCE_FLAG_ACTIVE = DMA_FENCE_FLAG_USER_BITS,
74 * I915_FENCE_FLAG_PQUEUE - this request is ready for execution
76 * Using the scheduler, when a request is ready for execution it is put
77 * into the priority queue, and removed from that queue when transferred
78 * to the HW runlists. We want to track its membership within the
79 * priority queue so that we can easily check before rescheduling.
81 * See i915_request_in_priority_queue()
83 I915_FENCE_FLAG_PQUEUE,
86 * I915_FENCE_FLAG_SIGNAL - this request is currently on signal_list
88 * Internal bookkeeping used by the breadcrumb code to track when
89 * a request is on the various signal_list.
91 I915_FENCE_FLAG_SIGNAL,
94 * I915_FENCE_FLAG_HOLD - this request is currently on hold
96 * This request has been suspended, pending an ongoing investigation.
101 * I915_FENCE_FLAG_NOPREEMPT - this request should not be preempted
103 * The execution of some requests should not be interrupted. This is
104 * a sensitive operation as it makes the request super important,
105 * blocking other higher priority work. Abuse of this flag will
106 * lead to quality of service issues.
108 I915_FENCE_FLAG_NOPREEMPT,
111 * I915_FENCE_FLAG_SENTINEL - this request should be last in the queue
113 * A high priority sentinel request may be submitted to clear the
114 * submission queue. As it will be the only request in-flight, upon
115 * execution all other active requests will have been preempted and
116 * unsubmitted. This preemptive pulse is used to re-evaluate the
117 * in-flight requests, particularly in cases where an active context
118 * is banned and those active requests need to be cancelled.
120 I915_FENCE_FLAG_SENTINEL,
123 * I915_FENCE_FLAG_BOOST - upclock the gpu for this request
125 * Some requests are more important than others! In particular, a
126 * request that the user is waiting on is typically required for
127 * interactive latency, for which we want to minimise by upclocking
128 * the GPU. Here we track such boost requests on a per-request basis.
130 I915_FENCE_FLAG_BOOST,
134 * Request queue structure.
136 * The request queue allows us to note sequence numbers that have been emitted
137 * and may be associated with active buffers to be retired.
139 * By keeping this list, we can avoid having to do questionable sequence
140 * number comparisons on buffer last_read|write_seqno. It also allows an
141 * emission time to be associated with the request for tracking how far ahead
142 * of the GPU the submission is.
144 * When modifying this structure be very aware that we perform a lockless
145 * RCU lookup of it that may race against reallocation of the struct
146 * from the slab freelist. We intentionally do not zero the structure on
147 * allocation so that the lookup can use the dangling pointers (and is
148 * cogniscent that those pointers may be wrong). Instead, everything that
149 * needs to be initialised must be done so explicitly.
151 * The requests are reference counted.
153 struct i915_request {
154 struct dma_fence fence;
157 /** On Which ring this request was generated */
158 struct drm_i915_private *i915;
161 * Context and ring buffer related to this request
162 * Contexts are refcounted, so when this request is associated with a
163 * context, we must increment the context's refcount, to guarantee that
164 * it persists while any request is linked to it. Requests themselves
165 * are also refcounted, so the request will only be freed when the last
166 * reference to it is dismissed, and the code in
167 * i915_request_free() will then decrement the refcount on the
170 struct intel_engine_cs *engine;
171 struct intel_context *context;
172 struct intel_ring *ring;
173 struct intel_timeline __rcu *timeline;
174 struct list_head signal_link;
177 * The rcu epoch of when this request was allocated. Used to judiciously
178 * apply backpressure on future allocations to ensure that under
179 * mempressure there is sufficient RCU ticks for us to reclaim our
180 * RCU protected slabs.
182 unsigned long rcustate;
185 * We pin the timeline->mutex while constructing the request to
186 * ensure that no caller accidentally drops it during construction.
187 * The timeline->mutex must be held to ensure that only this caller
188 * can use the ring and manipulate the associated timeline during
191 struct pin_cookie cookie;
194 * Fences for the various phases in the request's lifetime.
196 * The submit fence is used to await upon all of the request's
197 * dependencies. When it is signaled, the request is ready to run.
198 * It is used by the driver to then queue the request for execution.
200 struct i915_sw_fence submit;
202 wait_queue_entry_t submitq;
203 struct i915_sw_dma_fence_cb dmaq;
204 struct i915_request_duration_cb {
205 struct dma_fence_cb cb;
209 struct list_head execute_cb;
210 struct i915_sw_fence semaphore;
213 * A list of everyone we wait upon, and everyone who waits upon us.
214 * Even though we will not be submitted to the hardware before the
215 * submit fence is signaled (it waits for all external events as well
216 * as our own requests), the scheduler still needs to know the
217 * dependency tree for the lifetime of the request (from execbuf
218 * to retirement), i.e. bidirectional dependency information for the
219 * request not tied to individual fences.
221 struct i915_sched_node sched;
222 struct i915_dependency dep;
223 intel_engine_mask_t execution_mask;
226 * A convenience pointer to the current breadcrumb value stored in
227 * the HW status page (or our timeline's local equivalent). The full
228 * path would be rq->hw_context->ring->timeline->hwsp_seqno.
230 const u32 *hwsp_seqno;
233 * If we need to access the timeline's seqno for this request in
234 * another request, we need to keep a read reference to this associated
235 * cacheline, so that we do not free and recycle it before the foreign
236 * observers have completed. Hence, we keep a pointer to the cacheline
237 * inside the timeline's HWSP vma, but it is only valid while this
238 * request has not completed and guarded by the timeline mutex.
240 struct intel_timeline_cacheline __rcu *hwsp_cacheline;
242 /** Position in the ring of the start of the request */
245 /** Position in the ring of the start of the user packets */
249 * Position in the ring of the start of the postfix.
250 * This is required to calculate the maximum available ring space
251 * without overwriting the postfix.
255 /** Position in the ring of the end of the whole request */
258 /** Position in the ring of the end of any workarounds after the tail */
261 /** Preallocate space in the ring for the emitting the request */
264 /** Batch buffer related to this request if any (used for
265 * error state dump only).
267 struct i915_vma *batch;
269 * Additional buffers requested by userspace to be captured upon
270 * a GPU hang. The vma/obj on this list are protected by their
271 * active reference - all objects on this list must also be
272 * on the active_list (of their final request).
274 struct i915_capture_list *capture_list;
276 /** Time at which this request was emitted, in jiffies. */
277 unsigned long emitted_jiffies;
279 /** timeline->request entry for this request */
280 struct list_head link;
282 struct drm_i915_file_private *file_priv;
283 /** file_priv list entry for this request */
284 struct list_head client_link;
286 I915_SELFTEST_DECLARE(struct {
287 struct list_head link;
292 #define I915_FENCE_GFP (GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN)
294 extern const struct dma_fence_ops i915_fence_ops;
296 static inline bool dma_fence_is_i915(const struct dma_fence *fence)
298 return fence->ops == &i915_fence_ops;
301 struct i915_request * __must_check
302 __i915_request_create(struct intel_context *ce, gfp_t gfp);
303 struct i915_request * __must_check
304 i915_request_create(struct intel_context *ce);
306 struct i915_request *__i915_request_commit(struct i915_request *request);
307 void __i915_request_queue(struct i915_request *rq,
308 const struct i915_sched_attr *attr);
310 bool i915_request_retire(struct i915_request *rq);
311 void i915_request_retire_upto(struct i915_request *rq);
313 static inline struct i915_request *
314 to_request(struct dma_fence *fence)
316 /* We assume that NULL fence/request are interoperable */
317 BUILD_BUG_ON(offsetof(struct i915_request, fence) != 0);
318 GEM_BUG_ON(fence && !dma_fence_is_i915(fence));
319 return container_of(fence, struct i915_request, fence);
322 static inline struct i915_request *
323 i915_request_get(struct i915_request *rq)
325 return to_request(dma_fence_get(&rq->fence));
328 static inline struct i915_request *
329 i915_request_get_rcu(struct i915_request *rq)
331 return to_request(dma_fence_get_rcu(&rq->fence));
335 i915_request_put(struct i915_request *rq)
337 dma_fence_put(&rq->fence);
340 int i915_request_await_object(struct i915_request *to,
341 struct drm_i915_gem_object *obj,
343 int i915_request_await_dma_fence(struct i915_request *rq,
344 struct dma_fence *fence);
345 int i915_request_await_execution(struct i915_request *rq,
346 struct dma_fence *fence,
347 void (*hook)(struct i915_request *rq,
348 struct dma_fence *signal));
350 void i915_request_add(struct i915_request *rq);
352 bool __i915_request_submit(struct i915_request *request);
353 void i915_request_submit(struct i915_request *request);
355 void i915_request_skip(struct i915_request *request, int error);
357 void __i915_request_unsubmit(struct i915_request *request);
358 void i915_request_unsubmit(struct i915_request *request);
360 /* Note: part of the intel_breadcrumbs family */
361 bool i915_request_enable_breadcrumb(struct i915_request *request);
362 void i915_request_cancel_breadcrumb(struct i915_request *request);
364 long i915_request_wait(struct i915_request *rq,
367 __attribute__((nonnull(1)));
368 #define I915_WAIT_INTERRUPTIBLE BIT(0)
369 #define I915_WAIT_PRIORITY BIT(1) /* small priority bump for the request */
370 #define I915_WAIT_ALL BIT(2) /* used by i915_gem_object_wait() */
372 static inline bool i915_request_signaled(const struct i915_request *rq)
374 /* The request may live longer than its HWSP, so check flags first! */
375 return test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &rq->fence.flags);
378 static inline bool i915_request_is_active(const struct i915_request *rq)
380 return test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags);
383 static inline bool i915_request_in_priority_queue(const struct i915_request *rq)
385 return test_bit(I915_FENCE_FLAG_PQUEUE, &rq->fence.flags);
389 * Returns true if seq1 is later than seq2.
391 static inline bool i915_seqno_passed(u32 seq1, u32 seq2)
393 return (s32)(seq1 - seq2) >= 0;
396 static inline u32 __hwsp_seqno(const struct i915_request *rq)
398 return READ_ONCE(*rq->hwsp_seqno);
402 * hwsp_seqno - the current breadcrumb value in the HW status page
403 * @rq: the request, to chase the relevant HW status page
405 * The emphasis in naming here is that hwsp_seqno() is not a property of the
406 * request, but an indication of the current HW state (associated with this
407 * request). Its value will change as the GPU executes more requests.
409 * Returns the current breadcrumb value in the associated HW status page (or
410 * the local timeline's equivalent) for this request. The request itself
411 * has the associated breadcrumb value of rq->fence.seqno, when the HW
412 * status page has that breadcrumb or later, this request is complete.
414 static inline u32 hwsp_seqno(const struct i915_request *rq)
418 rcu_read_lock(); /* the HWSP may be freed at runtime */
419 seqno = __hwsp_seqno(rq);
425 static inline bool __i915_request_has_started(const struct i915_request *rq)
427 return i915_seqno_passed(hwsp_seqno(rq), rq->fence.seqno - 1);
431 * i915_request_started - check if the request has begun being executed
434 * If the timeline is not using initial breadcrumbs, a request is
435 * considered started if the previous request on its timeline (i.e.
436 * context) has been signaled.
438 * If the timeline is using semaphores, it will also be emitting an
439 * "initial breadcrumb" after the semaphores are complete and just before
440 * it began executing the user payload. A request can therefore be active
441 * on the HW and not yet started as it is still busywaiting on its
442 * dependencies (via HW semaphores).
444 * If the request has started, its dependencies will have been signaled
445 * (either by fences or by semaphores) and it will have begun processing
448 * However, even if a request has started, it may have been preempted and
449 * so no longer active, or it may have already completed.
451 * See also i915_request_is_active().
453 * Returns true if the request has begun executing the user payload, or
456 static inline bool i915_request_started(const struct i915_request *rq)
458 if (i915_request_signaled(rq))
461 /* Remember: started but may have since been preempted! */
462 return __i915_request_has_started(rq);
466 * i915_request_is_running - check if the request may actually be executing
469 * Returns true if the request is currently submitted to hardware, has passed
470 * its start point (i.e. the context is setup and not busywaiting). Note that
471 * it may no longer be running by the time the function returns!
473 static inline bool i915_request_is_running(const struct i915_request *rq)
475 if (!i915_request_is_active(rq))
478 return __i915_request_has_started(rq);
482 * i915_request_is_running - check if the request is ready for execution
485 * Upon construction, the request is instructed to wait upon various
486 * signals before it is ready to be executed by the HW. That is, we do
487 * not want to start execution and read data before it is written. In practice,
488 * this is controlled with a mixture of interrupts and semaphores. Once
489 * the submit fence is completed, the backend scheduler will place the
490 * request into its queue and from there submit it for execution. So we
491 * can detect when a request is eligible for execution (and is under control
492 * of the scheduler) by querying where it is in any of the scheduler's lists.
494 * Returns true if the request is ready for execution (it may be inflight),
497 static inline bool i915_request_is_ready(const struct i915_request *rq)
499 return !list_empty(&rq->sched.link);
502 static inline bool i915_request_completed(const struct i915_request *rq)
504 if (i915_request_signaled(rq))
507 return i915_seqno_passed(hwsp_seqno(rq), rq->fence.seqno);
510 static inline void i915_request_mark_complete(struct i915_request *rq)
512 rq->hwsp_seqno = (u32 *)&rq->fence.seqno; /* decouple from HWSP */
515 static inline bool i915_request_has_waitboost(const struct i915_request *rq)
517 return test_bit(I915_FENCE_FLAG_BOOST, &rq->fence.flags);
520 static inline bool i915_request_has_nopreempt(const struct i915_request *rq)
522 /* Preemption should only be disabled very rarely */
523 return unlikely(test_bit(I915_FENCE_FLAG_NOPREEMPT, &rq->fence.flags));
526 static inline bool i915_request_has_sentinel(const struct i915_request *rq)
528 return unlikely(test_bit(I915_FENCE_FLAG_SENTINEL, &rq->fence.flags));
531 static inline bool i915_request_on_hold(const struct i915_request *rq)
533 return unlikely(test_bit(I915_FENCE_FLAG_HOLD, &rq->fence.flags));
536 static inline void i915_request_set_hold(struct i915_request *rq)
538 set_bit(I915_FENCE_FLAG_HOLD, &rq->fence.flags);
541 static inline void i915_request_clear_hold(struct i915_request *rq)
543 clear_bit(I915_FENCE_FLAG_HOLD, &rq->fence.flags);
546 static inline struct intel_timeline *
547 i915_request_timeline(struct i915_request *rq)
549 /* Valid only while the request is being constructed (or retired). */
550 return rcu_dereference_protected(rq->timeline,
551 lockdep_is_held(&rcu_access_pointer(rq->timeline)->mutex));
554 static inline struct i915_gem_context *
555 i915_request_gem_context(struct i915_request *rq)
557 /* Valid only while the request is being constructed (or retired). */
558 return rcu_dereference_protected(rq->context->gem_context, true);
561 static inline struct intel_timeline *
562 i915_request_active_timeline(struct i915_request *rq)
565 * When in use during submission, we are protected by a guarantee that
566 * the context/timeline is pinned and must remain pinned until after
569 return rcu_dereference_protected(rq->timeline,
570 lockdep_is_held(&rq->engine->active.lock));
573 #endif /* I915_REQUEST_H */