1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2019 Intel Corporation
4  */
5 
6 #include "i915_drv.h"
7 
8 #include "intel_breadcrumbs.h"
9 #include "intel_context.h"
10 #include "intel_engine.h"
11 #include "intel_engine_heartbeat.h"
12 #include "intel_engine_pm.h"
13 #include "intel_gt.h"
14 #include "intel_gt_pm.h"
15 #include "intel_rc6.h"
16 #include "intel_ring.h"
17 #include "shmem_utils.h"
18 
19 static void dbg_poison_ce(struct intel_context *ce)
20 {
21 	if (!IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
22 		return;
23 
24 	if (ce->state) {
25 		struct drm_i915_gem_object *obj = ce->state->obj;
26 		int type = i915_coherent_map_type(ce->engine->i915, obj, true);
27 		void *map;
28 
29 		if (!i915_gem_object_trylock(obj))
30 			return;
31 
32 		map = i915_gem_object_pin_map(obj, type);
33 		if (!IS_ERR(map)) {
34 			memset(map, CONTEXT_REDZONE, obj->base.size);
35 			i915_gem_object_flush_map(obj);
36 			i915_gem_object_unpin_map(obj);
37 		}
38 		i915_gem_object_unlock(obj);
39 	}
40 }
41 
42 static int __engine_unpark(struct intel_wakeref *wf)
43 {
44 	struct intel_engine_cs *engine =
45 		container_of(wf, typeof(*engine), wakeref);
46 	struct intel_context *ce;
47 
48 	ENGINE_TRACE(engine, "\n");
49 
50 	intel_gt_pm_get(engine->gt);
51 
52 	/* Discard stale context state from across idling */
53 	ce = engine->kernel_context;
54 	if (ce) {
55 		GEM_BUG_ON(test_bit(CONTEXT_VALID_BIT, &ce->flags));
56 
57 		/* Flush all pending HW writes before we touch the context */
58 		while (unlikely(intel_context_inflight(ce)))
59 			intel_engine_flush_submission(engine);
60 
61 		/* First poison the image to verify we never fully trust it */
62 		dbg_poison_ce(ce);
63 
64 		/* Scrub the context image after our loss of control */
65 		ce->ops->reset(ce);
66 
67 		CE_TRACE(ce, "reset { seqno:%x, *hwsp:%x, ring:%x }\n",
68 			 ce->timeline->seqno,
69 			 READ_ONCE(*ce->timeline->hwsp_seqno),
70 			 ce->ring->emit);
71 		GEM_BUG_ON(ce->timeline->seqno !=
72 			   READ_ONCE(*ce->timeline->hwsp_seqno));
73 	}
74 
75 	if (engine->unpark)
76 		engine->unpark(engine);
77 
78 	intel_breadcrumbs_unpark(engine->breadcrumbs);
79 	intel_engine_unpark_heartbeat(engine);
80 	return 0;
81 }
82 
83 #if IS_ENABLED(CONFIG_LOCKDEP)
84 
85 static unsigned long __timeline_mark_lock(struct intel_context *ce)
86 {
87 	unsigned long flags;
88 
89 	local_irq_save(flags);
90 	mutex_acquire(&ce->timeline->mutex.dep_map, 2, 0, _THIS_IP_);
91 
92 	return flags;
93 }
94 
95 static void __timeline_mark_unlock(struct intel_context *ce,
96 				   unsigned long flags)
97 {
98 	mutex_release(&ce->timeline->mutex.dep_map, _THIS_IP_);
99 	local_irq_restore(flags);
100 }
101 
102 #else
103 
104 static unsigned long __timeline_mark_lock(struct intel_context *ce)
105 {
106 	return 0;
107 }
108 
109 static void __timeline_mark_unlock(struct intel_context *ce,
110 				   unsigned long flags)
111 {
112 }
113 
114 #endif /* !IS_ENABLED(CONFIG_LOCKDEP) */
115 
116 static void duration(struct dma_fence *fence, struct dma_fence_cb *cb)
117 {
118 	struct i915_request *rq = to_request(fence);
119 
120 	ewma__engine_latency_add(&rq->engine->latency,
121 				 ktime_us_delta(rq->fence.timestamp,
122 						rq->duration.emitted));
123 }
124 
125 static void
126 __queue_and_release_pm(struct i915_request *rq,
127 		       struct intel_timeline *tl,
128 		       struct intel_engine_cs *engine)
129 {
130 	struct intel_gt_timelines *timelines = &engine->gt->timelines;
131 
132 	ENGINE_TRACE(engine, "parking\n");
133 
134 	/*
135 	 * We have to serialise all potential retirement paths with our
136 	 * submission, as we don't want to underflow either the
137 	 * engine->wakeref.counter or our timeline->active_count.
138 	 *
139 	 * Equally, we cannot allow a new submission to start until
140 	 * after we finish queueing, nor could we allow that submitter
141 	 * to retire us before we are ready!
142 	 */
143 	spin_lock(&timelines->lock);
144 
145 	/* Let intel_gt_retire_requests() retire us (acquired under lock) */
146 	if (!atomic_fetch_inc(&tl->active_count))
147 		list_add_tail(&tl->link, &timelines->active_list);
148 
149 	/* Hand the request over to HW and so engine_retire() */
150 	__i915_request_queue_bh(rq);
151 
152 	/* Let new submissions commence (and maybe retire this timeline) */
153 	__intel_wakeref_defer_park(&engine->wakeref);
154 
155 	spin_unlock(&timelines->lock);
156 }
157 
158 static bool switch_to_kernel_context(struct intel_engine_cs *engine)
159 {
160 	struct intel_context *ce = engine->kernel_context;
161 	struct i915_request *rq;
162 	unsigned long flags;
163 	bool result = true;
164 
165 	/*
166 	 * This is execlist specific behaviour intended to ensure the GPU is
167 	 * idle by switching to a known 'safe' context. With GuC submission, the
168 	 * same idle guarantee is achieved by other means (disabling
169 	 * scheduling). Further, switching to a 'safe' context has no effect
170 	 * with GuC submission as the scheduler can just switch back again.
171 	 *
172 	 * FIXME: Move this backend scheduler specific behaviour into the
173 	 * scheduler backend.
174 	 */
175 	if (intel_engine_uses_guc(engine))
176 		return true;
177 
178 	/* GPU is pointing to the void, as good as in the kernel context. */
179 	if (intel_gt_is_wedged(engine->gt))
180 		return true;
181 
182 	GEM_BUG_ON(!intel_context_is_barrier(ce));
183 	GEM_BUG_ON(ce->timeline->hwsp_ggtt != engine->status_page.vma);
184 
185 	/* Already inside the kernel context, safe to power down. */
186 	if (engine->wakeref_serial == engine->serial)
187 		return true;
188 
189 	/*
190 	 * Note, we do this without taking the timeline->mutex. We cannot
191 	 * as we may be called while retiring the kernel context and so
192 	 * already underneath the timeline->mutex. Instead we rely on the
193 	 * exclusive property of the __engine_park that prevents anyone
194 	 * else from creating a request on this engine. This also requires
195 	 * that the ring is empty and we avoid any waits while constructing
196 	 * the context, as they assume protection by the timeline->mutex.
197 	 * This should hold true as we can only park the engine after
198 	 * retiring the last request, thus all rings should be empty and
199 	 * all timelines idle.
200 	 *
201 	 * For unlocking, there are 2 other parties and the GPU who have a
202 	 * stake here.
203 	 *
204 	 * A new gpu user will be waiting on the engine-pm to start their
205 	 * engine_unpark. New waiters are predicated on engine->wakeref.count
206 	 * and so intel_wakeref_defer_park() acts like a mutex_unlock of the
207 	 * engine->wakeref.
208 	 *
209 	 * The other party is intel_gt_retire_requests(), which is walking the
210 	 * list of active timelines looking for completions. Meanwhile as soon
211 	 * as we call __i915_request_queue(), the GPU may complete our request.
212 	 * Ergo, if we put ourselves on the timelines.active_list
213 	 * (se intel_timeline_enter()) before we increment the
214 	 * engine->wakeref.count, we may see the request completion and retire
215 	 * it causing an underflow of the engine->wakeref.
216 	 */
217 	flags = __timeline_mark_lock(ce);
218 	GEM_BUG_ON(atomic_read(&ce->timeline->active_count) < 0);
219 
220 	rq = __i915_request_create(ce, GFP_NOWAIT);
221 	if (IS_ERR(rq))
222 		/* Context switch failed, hope for the best! Maybe reset? */
223 		goto out_unlock;
224 
225 	/* Check again on the next retirement. */
226 	engine->wakeref_serial = engine->serial + 1;
227 	i915_request_add_active_barriers(rq);
228 
229 	/* Install ourselves as a preemption barrier */
230 	rq->sched.attr.priority = I915_PRIORITY_BARRIER;
231 	if (likely(!__i915_request_commit(rq))) { /* engine should be idle! */
232 		/*
233 		 * Use an interrupt for precise measurement of duration,
234 		 * otherwise we rely on someone else retiring all the requests
235 		 * which may delay the signaling (i.e. we will likely wait
236 		 * until the background request retirement running every
237 		 * second or two).
238 		 */
239 		BUILD_BUG_ON(sizeof(rq->duration) > sizeof(rq->submitq));
240 		dma_fence_add_callback(&rq->fence, &rq->duration.cb, duration);
241 		rq->duration.emitted = ktime_get();
242 	}
243 
244 	/* Expose ourselves to the world */
245 	__queue_and_release_pm(rq, ce->timeline, engine);
246 
247 	result = false;
248 out_unlock:
249 	__timeline_mark_unlock(ce, flags);
250 	return result;
251 }
252 
253 static void call_idle_barriers(struct intel_engine_cs *engine)
254 {
255 	struct llist_node *node, *next;
256 
257 	llist_for_each_safe(node, next, llist_del_all(&engine->barrier_tasks)) {
258 		struct dma_fence_cb *cb =
259 			container_of((struct list_head *)node,
260 				     typeof(*cb), node);
261 
262 		cb->func(ERR_PTR(-EAGAIN), cb);
263 	}
264 }
265 
266 static int __engine_park(struct intel_wakeref *wf)
267 {
268 	struct intel_engine_cs *engine =
269 		container_of(wf, typeof(*engine), wakeref);
270 
271 	engine->saturated = 0;
272 
273 	/*
274 	 * If one and only one request is completed between pm events,
275 	 * we know that we are inside the kernel context and it is
276 	 * safe to power down. (We are paranoid in case that runtime
277 	 * suspend causes corruption to the active context image, and
278 	 * want to avoid that impacting userspace.)
279 	 */
280 	if (!switch_to_kernel_context(engine))
281 		return -EBUSY;
282 
283 	ENGINE_TRACE(engine, "parked\n");
284 
285 	call_idle_barriers(engine); /* cleanup after wedging */
286 
287 	intel_engine_park_heartbeat(engine);
288 	intel_breadcrumbs_park(engine->breadcrumbs);
289 
290 	/* Must be reset upon idling, or we may miss the busy wakeup. */
291 	GEM_BUG_ON(engine->sched_engine->queue_priority_hint != INT_MIN);
292 
293 	if (engine->park)
294 		engine->park(engine);
295 
296 	/* While gt calls i915_vma_parked(), we have to break the lock cycle */
297 	intel_gt_pm_put_async(engine->gt);
298 	return 0;
299 }
300 
301 static const struct intel_wakeref_ops wf_ops = {
302 	.get = __engine_unpark,
303 	.put = __engine_park,
304 };
305 
306 void intel_engine_init__pm(struct intel_engine_cs *engine)
307 {
308 	struct intel_runtime_pm *rpm = engine->uncore->rpm;
309 
310 	intel_wakeref_init(&engine->wakeref, rpm, &wf_ops);
311 	intel_engine_init_heartbeat(engine);
312 }
313 
314 /**
315  * intel_engine_reset_pinned_contexts - Reset the pinned contexts of
316  * an engine.
317  * @engine: The engine whose pinned contexts we want to reset.
318  *
319  * Typically the pinned context LMEM images lose or get their content
320  * corrupted on suspend. This function resets their images.
321  */
322 void intel_engine_reset_pinned_contexts(struct intel_engine_cs *engine)
323 {
324 	struct intel_context *ce;
325 
326 	list_for_each_entry(ce, &engine->pinned_contexts_list,
327 			    pinned_contexts_link) {
328 		/* kernel context gets reset at __engine_unpark() */
329 		if (ce == engine->kernel_context)
330 			continue;
331 
332 		dbg_poison_ce(ce);
333 		ce->ops->reset(ce);
334 	}
335 }
336 
337 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
338 #include "selftest_engine_pm.c"
339 #endif
340