summaryrefslogtreecommitdiff
path: root/drivers/gpu/drm/i915/gt/intel_engine_pm.c
blob: a1334b48dde7be7f309ac4c599c0675079dc7866 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
// SPDX-License-Identifier: MIT
/*
 * Copyright © 2019 Intel Corporation
 */

#include "i915_drv.h"

#include "intel_breadcrumbs.h"
#include "intel_context.h"
#include "intel_engine.h"
#include "intel_engine_heartbeat.h"
#include "intel_engine_pm.h"
#include "intel_gt.h"
#include "intel_gt_pm.h"
#include "intel_rc6.h"
#include "intel_ring.h"
#include "shmem_utils.h"

static void dbg_poison_ce(struct intel_context *ce)
{
	if (!IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
		return;

	if (ce->state) {
		struct drm_i915_gem_object *obj = ce->state->obj;
		int type = i915_coherent_map_type(ce->engine->i915, obj, true);
		void *map;

		if (!i915_gem_object_trylock(obj))
			return;

		map = i915_gem_object_pin_map(obj, type);
		if (!IS_ERR(map)) {
			memset(map, CONTEXT_REDZONE, obj->base.size);
			i915_gem_object_flush_map(obj);
			i915_gem_object_unpin_map(obj);
		}
		i915_gem_object_unlock(obj);
	}
}

static int __engine_unpark(struct intel_wakeref *wf)
{
	struct intel_engine_cs *engine =
		container_of(wf, typeof(*engine), wakeref);
	struct intel_context *ce;

	ENGINE_TRACE(engine, "\n");

	intel_gt_pm_get(engine->gt);

	/* Discard stale context state from across idling */
	ce = engine->kernel_context;
	if (ce) {
		GEM_BUG_ON(test_bit(CONTEXT_VALID_BIT, &ce->flags));

		/* Flush all pending HW writes before we touch the context */
		while (unlikely(intel_context_inflight(ce)))
			intel_engine_flush_submission(engine);

		/* First poison the image to verify we never fully trust it */
		dbg_poison_ce(ce);

		/* Scrub the context image after our loss of control */
		ce->ops->reset(ce);

		CE_TRACE(ce, "reset { seqno:%x, *hwsp:%x, ring:%x }\n",
			 ce->timeline->seqno,
			 READ_ONCE(*ce->timeline->hwsp_seqno),
			 ce->ring->emit);
		GEM_BUG_ON(ce->timeline->seqno !=
			   READ_ONCE(*ce->timeline->hwsp_seqno));
	}

	if (engine->unpark)
		engine->unpark(engine);

	intel_breadcrumbs_unpark(engine->breadcrumbs);
	intel_engine_unpark_heartbeat(engine);
	return 0;
}

#if IS_ENABLED(CONFIG_LOCKDEP)

static unsigned long __timeline_mark_lock(struct intel_context *ce)
{
	unsigned long flags;

	local_irq_save(flags);
	mutex_acquire(&ce->timeline->mutex.dep_map, 2, 0, _THIS_IP_);

	return flags;
}

static void __timeline_mark_unlock(struct intel_context *ce,
				   unsigned long flags)
{
	mutex_release(&ce->timeline->mutex.dep_map, _THIS_IP_);
	local_irq_restore(flags);
}

#else

static unsigned long __timeline_mark_lock(struct intel_context *ce)
{
	return 0;
}

static void __timeline_mark_unlock(struct intel_context *ce,
				   unsigned long flags)
{
}

#endif /* !IS_ENABLED(CONFIG_LOCKDEP) */

static void duration(struct dma_fence *fence, struct dma_fence_cb *cb)
{
	struct i915_request *rq = to_request(fence);

	ewma__engine_latency_add(&rq->engine->latency,
				 ktime_us_delta(rq->fence.timestamp,
						rq->duration.emitted));
}

static void
__queue_and_release_pm(struct i915_request *rq,
		       struct intel_timeline *tl,
		       struct intel_engine_cs *engine)
{
	struct intel_gt_timelines *timelines = &engine->gt->timelines;

	ENGINE_TRACE(engine, "parking\n");

	/*
	 * We have to serialise all potential retirement paths with our
	 * submission, as we don't want to underflow either the
	 * engine->wakeref.counter or our timeline->active_count.
	 *
	 * Equally, we cannot allow a new submission to start until
	 * after we finish queueing, nor could we allow that submitter
	 * to retire us before we are ready!
	 */
	spin_lock(&timelines->lock);

	/* Let intel_gt_retire_requests() retire us (acquired under lock) */
	if (!atomic_fetch_inc(&tl->active_count))
		list_add_tail(&tl->link, &timelines->active_list);

	/* Hand the request over to HW and so engine_retire() */
	__i915_request_queue_bh(rq);

	/* Let new submissions commence (and maybe retire this timeline) */
	__intel_wakeref_defer_park(&engine->wakeref);

	spin_unlock(&timelines->lock);
}

static bool switch_to_kernel_context(struct intel_engine_cs *engine)
{
	struct intel_context *ce = engine->kernel_context;
	struct i915_request *rq;
	unsigned long flags;
	bool result = true;

	/*
	 * This is execlist specific behaviour intended to ensure the GPU is
	 * idle by switching to a known 'safe' context. With GuC submission, the
	 * same idle guarantee is achieved by other means (disabling
	 * scheduling). Further, switching to a 'safe' context has no effect
	 * with GuC submission as the scheduler can just switch back again.
	 *
	 * FIXME: Move this backend scheduler specific behaviour into the
	 * scheduler backend.
	 */
	if (intel_engine_uses_guc(engine))
		return true;

	/* GPU is pointing to the void, as good as in the kernel context. */
	if (intel_gt_is_wedged(engine->gt))
		return true;

	GEM_BUG_ON(!intel_context_is_barrier(ce));
	GEM_BUG_ON(ce->timeline->hwsp_ggtt != engine->status_page.vma);

	/* Already inside the kernel context, safe to power down. */
	if (engine->wakeref_serial == engine->serial)
		return true;

	/*
	 * Note, we do this without taking the timeline->mutex. We cannot
	 * as we may be called while retiring the kernel context and so
	 * already underneath the timeline->mutex. Instead we rely on the
	 * exclusive property of the __engine_park that prevents anyone
	 * else from creating a request on this engine. This also requires
	 * that the ring is empty and we avoid any waits while constructing
	 * the context, as they assume protection by the timeline->mutex.
	 * This should hold true as we can only park the engine after
	 * retiring the last request, thus all rings should be empty and
	 * all timelines idle.
	 *
	 * For unlocking, there are 2 other parties and the GPU who have a
	 * stake here.
	 *
	 * A new gpu user will be waiting on the engine-pm to start their
	 * engine_unpark. New waiters are predicated on engine->wakeref.count
	 * and so intel_wakeref_defer_park() acts like a mutex_unlock of the
	 * engine->wakeref.
	 *
	 * The other party is intel_gt_retire_requests(), which is walking the
	 * list of active timelines looking for completions. Meanwhile as soon
	 * as we call __i915_request_queue(), the GPU may complete our request.
	 * Ergo, if we put ourselves on the timelines.active_list
	 * (se intel_timeline_enter()) before we increment the
	 * engine->wakeref.count, we may see the request completion and retire
	 * it causing an underflow of the engine->wakeref.
	 */
	flags = __timeline_mark_lock(ce);
	GEM_BUG_ON(atomic_read(&ce->timeline->active_count) < 0);

	rq = __i915_request_create(ce, GFP_NOWAIT);
	if (IS_ERR(rq))
		/* Context switch failed, hope for the best! Maybe reset? */
		goto out_unlock;

	/* Check again on the next retirement. */
	engine->wakeref_serial = engine->serial + 1;
	i915_request_add_active_barriers(rq);

	/* Install ourselves as a preemption barrier */
	rq->sched.attr.priority = I915_PRIORITY_BARRIER;
	if (likely(!__i915_request_commit(rq))) { /* engine should be idle! */
		/*
		 * Use an interrupt for precise measurement of duration,
		 * otherwise we rely on someone else retiring all the requests
		 * which may delay the signaling (i.e. we will likely wait
		 * until the background request retirement running every
		 * second or two).
		 */
		BUILD_BUG_ON(sizeof(rq->duration) > sizeof(rq->submitq));
		dma_fence_add_callback(&rq->fence, &rq->duration.cb, duration);
		rq->duration.emitted = ktime_get();
	}

	/* Expose ourselves to the world */
	__queue_and_release_pm(rq, ce->timeline, engine);

	result = false;
out_unlock:
	__timeline_mark_unlock(ce, flags);
	return result;
}

static void call_idle_barriers(struct intel_engine_cs *engine)
{
	struct llist_node *node, *next;

	llist_for_each_safe(node, next, llist_del_all(&engine->barrier_tasks)) {
		struct dma_fence_cb *cb =
			container_of((struct list_head *)node,
				     typeof(*cb), node);

		cb->func(ERR_PTR(-EAGAIN), cb);
	}
}

static int __engine_park(struct intel_wakeref *wf)
{
	struct intel_engine_cs *engine =
		container_of(wf, typeof(*engine), wakeref);

	engine->saturated = 0;

	/*
	 * If one and only one request is completed between pm events,
	 * we know that we are inside the kernel context and it is
	 * safe to power down. (We are paranoid in case that runtime
	 * suspend causes corruption to the active context image, and
	 * want to avoid that impacting userspace.)
	 */
	if (!switch_to_kernel_context(engine))
		return -EBUSY;

	ENGINE_TRACE(engine, "parked\n");

	call_idle_barriers(engine); /* cleanup after wedging */

	intel_engine_park_heartbeat(engine);
	intel_breadcrumbs_park(engine->breadcrumbs);

	/* Must be reset upon idling, or we may miss the busy wakeup. */
	GEM_BUG_ON(engine->sched_engine->queue_priority_hint != INT_MIN);

	if (engine->park)
		engine->park(engine);

	/* While gt calls i915_vma_parked(), we have to break the lock cycle */
	intel_gt_pm_put_async(engine->gt);
	return 0;
}

static const struct intel_wakeref_ops wf_ops = {
	.get = __engine_unpark,
	.put = __engine_park,
};

void intel_engine_init__pm(struct intel_engine_cs *engine)
{
	struct intel_runtime_pm *rpm = engine->uncore->rpm;

	intel_wakeref_init(&engine->wakeref, rpm, &wf_ops);
	intel_engine_init_heartbeat(engine);
}

/**
 * intel_engine_reset_pinned_contexts - Reset the pinned contexts of
 * an engine.
 * @engine: The engine whose pinned contexts we want to reset.
 *
 * Typically the pinned context LMEM images lose or get their content
 * corrupted on suspend. This function resets their images.
 */
void intel_engine_reset_pinned_contexts(struct intel_engine_cs *engine)
{
	struct intel_context *ce;

	list_for_each_entry(ce, &engine->pinned_contexts_list,
			    pinned_contexts_link) {
		/* kernel context gets reset at __engine_unpark() */
		if (ce == engine->kernel_context)
			continue;

		dbg_poison_ce(ce);
		ce->ops->reset(ce);
	}
}

#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
#include "selftest_engine_pm.c"
#endif