• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 
dbg_poison_ce(struct intel_context * ce)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, NULL))
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 
__engine_unpark(struct intel_wakeref * wf)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 
duration(struct dma_fence * fence,struct dma_fence_cb * cb)83 static void duration(struct dma_fence *fence, struct dma_fence_cb *cb)
84 {
85 	struct i915_request *rq = to_request(fence);
86 
87 	ewma__engine_latency_add(&rq->engine->latency,
88 				 ktime_us_delta(rq->fence.timestamp,
89 						rq->duration.emitted));
90 }
91 
92 static void
__queue_and_release_pm(struct i915_request * rq,struct intel_timeline * tl,struct intel_engine_cs * engine)93 __queue_and_release_pm(struct i915_request *rq,
94 		       struct intel_timeline *tl,
95 		       struct intel_engine_cs *engine)
96 {
97 	struct intel_gt_timelines *timelines = &engine->gt->timelines;
98 
99 	ENGINE_TRACE(engine, "parking\n");
100 
101 	/*
102 	 * We have to serialise all potential retirement paths with our
103 	 * submission, as we don't want to underflow either the
104 	 * engine->wakeref.counter or our timeline->active_count.
105 	 *
106 	 * Equally, we cannot allow a new submission to start until
107 	 * after we finish queueing, nor could we allow that submitter
108 	 * to retire us before we are ready!
109 	 */
110 	spin_lock(&timelines->lock);
111 
112 	/* Let intel_gt_retire_requests() retire us (acquired under lock) */
113 	if (!atomic_fetch_inc(&tl->active_count))
114 		list_add_tail(&tl->link, &timelines->active_list);
115 
116 	/* Hand the request over to HW and so engine_retire() */
117 	__i915_request_queue_bh(rq);
118 
119 	/* Let new submissions commence (and maybe retire this timeline) */
120 	__intel_wakeref_defer_park(&engine->wakeref);
121 
122 	spin_unlock(&timelines->lock);
123 }
124 
switch_to_kernel_context(struct intel_engine_cs * engine)125 static bool switch_to_kernel_context(struct intel_engine_cs *engine)
126 {
127 	struct intel_context *ce = engine->kernel_context;
128 	struct i915_request *rq;
129 	bool result = true;
130 
131 	/*
132 	 * This is execlist specific behaviour intended to ensure the GPU is
133 	 * idle by switching to a known 'safe' context. With GuC submission, the
134 	 * same idle guarantee is achieved by other means (disabling
135 	 * scheduling). Further, switching to a 'safe' context has no effect
136 	 * with GuC submission as the scheduler can just switch back again.
137 	 *
138 	 * FIXME: Move this backend scheduler specific behaviour into the
139 	 * scheduler backend.
140 	 */
141 	if (intel_engine_uses_guc(engine))
142 		return true;
143 
144 	/* GPU is pointing to the void, as good as in the kernel context. */
145 	if (intel_gt_is_wedged(engine->gt))
146 		return true;
147 
148 	GEM_BUG_ON(!intel_context_is_barrier(ce));
149 	GEM_BUG_ON(ce->timeline->hwsp_ggtt != engine->status_page.vma);
150 
151 	/* Already inside the kernel context, safe to power down. */
152 	if (engine->wakeref_serial == engine->serial)
153 		return true;
154 
155 	/*
156 	 * Note, we do this without taking the timeline->mutex. We cannot
157 	 * as we may be called while retiring the kernel context and so
158 	 * already underneath the timeline->mutex. Instead we rely on the
159 	 * exclusive property of the __engine_park that prevents anyone
160 	 * else from creating a request on this engine. This also requires
161 	 * that the ring is empty and we avoid any waits while constructing
162 	 * the context, as they assume protection by the timeline->mutex.
163 	 * This should hold true as we can only park the engine after
164 	 * retiring the last request, thus all rings should be empty and
165 	 * all timelines idle.
166 	 *
167 	 * For unlocking, there are 2 other parties and the GPU who have a
168 	 * stake here.
169 	 *
170 	 * A new gpu user will be waiting on the engine-pm to start their
171 	 * engine_unpark. New waiters are predicated on engine->wakeref.count
172 	 * and so intel_wakeref_defer_park() acts like a mutex_unlock of the
173 	 * engine->wakeref.
174 	 *
175 	 * The other party is intel_gt_retire_requests(), which is walking the
176 	 * list of active timelines looking for completions. Meanwhile as soon
177 	 * as we call __i915_request_queue(), the GPU may complete our request.
178 	 * Ergo, if we put ourselves on the timelines.active_list
179 	 * (se intel_timeline_enter()) before we increment the
180 	 * engine->wakeref.count, we may see the request completion and retire
181 	 * it causing an underflow of the engine->wakeref.
182 	 */
183 	set_bit(CONTEXT_IS_PARKING, &ce->flags);
184 	GEM_BUG_ON(atomic_read(&ce->timeline->active_count) < 0);
185 
186 	rq = __i915_request_create(ce, GFP_NOWAIT);
187 	if (IS_ERR(rq))
188 		/* Context switch failed, hope for the best! Maybe reset? */
189 		goto out_unlock;
190 
191 	/* Check again on the next retirement. */
192 	engine->wakeref_serial = engine->serial + 1;
193 	i915_request_add_active_barriers(rq);
194 
195 	/* Install ourselves as a preemption barrier */
196 	rq->sched.attr.priority = I915_PRIORITY_BARRIER;
197 	if (likely(!__i915_request_commit(rq))) { /* engine should be idle! */
198 		/*
199 		 * Use an interrupt for precise measurement of duration,
200 		 * otherwise we rely on someone else retiring all the requests
201 		 * which may delay the signaling (i.e. we will likely wait
202 		 * until the background request retirement running every
203 		 * second or two).
204 		 */
205 		BUILD_BUG_ON(sizeof(rq->duration) > sizeof(rq->submitq));
206 		dma_fence_add_callback(&rq->fence, &rq->duration.cb, duration);
207 		rq->duration.emitted = ktime_get();
208 	}
209 
210 	/* Expose ourselves to the world */
211 	__queue_and_release_pm(rq, ce->timeline, engine);
212 
213 	result = false;
214 out_unlock:
215 	clear_bit(CONTEXT_IS_PARKING, &ce->flags);
216 	return result;
217 }
218 
call_idle_barriers(struct intel_engine_cs * engine)219 static void call_idle_barriers(struct intel_engine_cs *engine)
220 {
221 	struct llist_node *node, *next;
222 
223 	llist_for_each_safe(node, next, llist_del_all(&engine->barrier_tasks)) {
224 		struct dma_fence_cb *cb =
225 			container_of((struct list_head *)node,
226 				     typeof(*cb), node);
227 
228 		cb->func(ERR_PTR(-EAGAIN), cb);
229 	}
230 }
231 
__engine_park(struct intel_wakeref * wf)232 static int __engine_park(struct intel_wakeref *wf)
233 {
234 	struct intel_engine_cs *engine =
235 		container_of(wf, typeof(*engine), wakeref);
236 
237 	engine->saturated = 0;
238 
239 	/*
240 	 * If one and only one request is completed between pm events,
241 	 * we know that we are inside the kernel context and it is
242 	 * safe to power down. (We are paranoid in case that runtime
243 	 * suspend causes corruption to the active context image, and
244 	 * want to avoid that impacting userspace.)
245 	 */
246 	if (!switch_to_kernel_context(engine))
247 		return -EBUSY;
248 
249 	ENGINE_TRACE(engine, "parked\n");
250 
251 	call_idle_barriers(engine); /* cleanup after wedging */
252 
253 	intel_engine_park_heartbeat(engine);
254 	intel_breadcrumbs_park(engine->breadcrumbs);
255 
256 	/* Must be reset upon idling, or we may miss the busy wakeup. */
257 	GEM_BUG_ON(engine->sched_engine->queue_priority_hint != INT_MIN);
258 
259 	if (engine->park)
260 		engine->park(engine);
261 
262 	/* While gt calls i915_vma_parked(), we have to break the lock cycle */
263 	intel_gt_pm_put_async(engine->gt);
264 	return 0;
265 }
266 
267 static const struct intel_wakeref_ops wf_ops = {
268 	.get = __engine_unpark,
269 	.put = __engine_park,
270 };
271 
intel_engine_init__pm(struct intel_engine_cs * engine)272 void intel_engine_init__pm(struct intel_engine_cs *engine)
273 {
274 	struct intel_runtime_pm *rpm = engine->uncore->rpm;
275 
276 	intel_wakeref_init(&engine->wakeref, rpm, &wf_ops);
277 	intel_engine_init_heartbeat(engine);
278 }
279 
280 /**
281  * intel_engine_reset_pinned_contexts - Reset the pinned contexts of
282  * an engine.
283  * @engine: The engine whose pinned contexts we want to reset.
284  *
285  * Typically the pinned context LMEM images lose or get their content
286  * corrupted on suspend. This function resets their images.
287  */
intel_engine_reset_pinned_contexts(struct intel_engine_cs * engine)288 void intel_engine_reset_pinned_contexts(struct intel_engine_cs *engine)
289 {
290 	struct intel_context *ce;
291 
292 	list_for_each_entry(ce, &engine->pinned_contexts_list,
293 			    pinned_contexts_link) {
294 		/* kernel context gets reset at __engine_unpark() */
295 		if (ce == engine->kernel_context)
296 			continue;
297 
298 		dbg_poison_ce(ce);
299 		ce->ops->reset(ce);
300 	}
301 }
302 
303 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
304 #include "selftest_engine_pm.c"
305 #endif
306