xref: /netbsd-src/sys/external/bsd/drm2/dist/drm/i915/display/intel_frontbuffer.c (revision d92071859d1bc8fc289928283e273e709e5be982)
1 /*	$NetBSD: intel_frontbuffer.c,v 1.3 2021/12/19 12:09:43 riastradh Exp $	*/
2 
3 /*
4  * Copyright © 2014 Intel Corporation
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the "Software"),
8  * to deal in the Software without restriction, including without limitation
9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10  * and/or sell copies of the Software, and to permit persons to whom the
11  * Software is furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice (including the next
14  * paragraph) shall be included in all copies or substantial portions of the
15  * Software.
16  *
17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
22  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
23  * DEALINGS IN THE SOFTWARE.
24  *
25  * Authors:
26  *	Daniel Vetter <daniel.vetter@ffwll.ch>
27  */
28 
29 /**
30  * DOC: frontbuffer tracking
31  *
32  * Many features require us to track changes to the currently active
33  * frontbuffer, especially rendering targeted at the frontbuffer.
34  *
35  * To be able to do so we track frontbuffers using a bitmask for all possible
36  * frontbuffer slots through intel_frontbuffer_track(). The functions in this
37  * file are then called when the contents of the frontbuffer are invalidated,
38  * when frontbuffer rendering has stopped again to flush out all the changes
39  * and when the frontbuffer is exchanged with a flip. Subsystems interested in
40  * frontbuffer changes (e.g. PSR, FBC, DRRS) should directly put their callbacks
41  * into the relevant places and filter for the frontbuffer slots that they are
42  * interested int.
43  *
44  * On a high level there are two types of powersaving features. The first one
45  * work like a special cache (FBC and PSR) and are interested when they should
46  * stop caching and when to restart caching. This is done by placing callbacks
47  * into the invalidate and the flush functions: At invalidate the caching must
48  * be stopped and at flush time it can be restarted. And maybe they need to know
49  * when the frontbuffer changes (e.g. when the hw doesn't initiate an invalidate
50  * and flush on its own) which can be achieved with placing callbacks into the
51  * flip functions.
52  *
53  * The other type of display power saving feature only cares about busyness
54  * (e.g. DRRS). In that case all three (invalidate, flush and flip) indicate
55  * busyness. There is no direct way to detect idleness. Instead an idle timer
56  * work delayed work should be started from the flush and flip functions and
57  * cancelled as soon as busyness is detected.
58  */
59 
60 #include <sys/cdefs.h>
61 __KERNEL_RCSID(0, "$NetBSD: intel_frontbuffer.c,v 1.3 2021/12/19 12:09:43 riastradh Exp $");
62 
63 #include "display/intel_dp.h"
64 
65 #include "i915_drv.h"
66 #include "intel_display_types.h"
67 #include "intel_fbc.h"
68 #include "intel_frontbuffer.h"
69 #include "intel_psr.h"
70 
71 /**
72  * frontbuffer_flush - flush frontbuffer
73  * @i915: i915 device
74  * @frontbuffer_bits: frontbuffer plane tracking bits
75  * @origin: which operation caused the flush
76  *
77  * This function gets called every time rendering on the given planes has
78  * completed and frontbuffer caching can be started again. Flushes will get
79  * delayed if they're blocked by some outstanding asynchronous rendering.
80  *
81  * Can be called without any locks held.
82  */
frontbuffer_flush(struct drm_i915_private * i915,unsigned int frontbuffer_bits,enum fb_op_origin origin)83 static void frontbuffer_flush(struct drm_i915_private *i915,
84 			      unsigned int frontbuffer_bits,
85 			      enum fb_op_origin origin)
86 {
87 	/* Delay flushing when rings are still busy.*/
88 	spin_lock(&i915->fb_tracking.lock);
89 	frontbuffer_bits &= ~i915->fb_tracking.busy_bits;
90 	spin_unlock(&i915->fb_tracking.lock);
91 
92 	if (!frontbuffer_bits)
93 		return;
94 
95 	might_sleep();
96 	intel_edp_drrs_flush(i915, frontbuffer_bits);
97 	intel_psr_flush(i915, frontbuffer_bits, origin);
98 	intel_fbc_flush(i915, frontbuffer_bits, origin);
99 }
100 
101 /**
102  * intel_frontbuffer_flip_prepare - prepare asynchronous frontbuffer flip
103  * @i915: i915 device
104  * @frontbuffer_bits: frontbuffer plane tracking bits
105  *
106  * This function gets called after scheduling a flip on @obj. The actual
107  * frontbuffer flushing will be delayed until completion is signalled with
108  * intel_frontbuffer_flip_complete. If an invalidate happens in between this
109  * flush will be cancelled.
110  *
111  * Can be called without any locks held.
112  */
intel_frontbuffer_flip_prepare(struct drm_i915_private * i915,unsigned frontbuffer_bits)113 void intel_frontbuffer_flip_prepare(struct drm_i915_private *i915,
114 				    unsigned frontbuffer_bits)
115 {
116 	spin_lock(&i915->fb_tracking.lock);
117 	i915->fb_tracking.flip_bits |= frontbuffer_bits;
118 	/* Remove stale busy bits due to the old buffer. */
119 	i915->fb_tracking.busy_bits &= ~frontbuffer_bits;
120 	spin_unlock(&i915->fb_tracking.lock);
121 }
122 
123 /**
124  * intel_frontbuffer_flip_complete - complete asynchronous frontbuffer flip
125  * @i915: i915 device
126  * @frontbuffer_bits: frontbuffer plane tracking bits
127  *
128  * This function gets called after the flip has been latched and will complete
129  * on the next vblank. It will execute the flush if it hasn't been cancelled yet.
130  *
131  * Can be called without any locks held.
132  */
intel_frontbuffer_flip_complete(struct drm_i915_private * i915,unsigned frontbuffer_bits)133 void intel_frontbuffer_flip_complete(struct drm_i915_private *i915,
134 				     unsigned frontbuffer_bits)
135 {
136 	spin_lock(&i915->fb_tracking.lock);
137 	/* Mask any cancelled flips. */
138 	frontbuffer_bits &= i915->fb_tracking.flip_bits;
139 	i915->fb_tracking.flip_bits &= ~frontbuffer_bits;
140 	spin_unlock(&i915->fb_tracking.lock);
141 
142 	if (frontbuffer_bits)
143 		frontbuffer_flush(i915, frontbuffer_bits, ORIGIN_FLIP);
144 }
145 
146 /**
147  * intel_frontbuffer_flip - synchronous frontbuffer flip
148  * @i915: i915 device
149  * @frontbuffer_bits: frontbuffer plane tracking bits
150  *
151  * This function gets called after scheduling a flip on @obj. This is for
152  * synchronous plane updates which will happen on the next vblank and which will
153  * not get delayed by pending gpu rendering.
154  *
155  * Can be called without any locks held.
156  */
intel_frontbuffer_flip(struct drm_i915_private * i915,unsigned frontbuffer_bits)157 void intel_frontbuffer_flip(struct drm_i915_private *i915,
158 			    unsigned frontbuffer_bits)
159 {
160 	spin_lock(&i915->fb_tracking.lock);
161 	/* Remove stale busy bits due to the old buffer. */
162 	i915->fb_tracking.busy_bits &= ~frontbuffer_bits;
163 	spin_unlock(&i915->fb_tracking.lock);
164 
165 	frontbuffer_flush(i915, frontbuffer_bits, ORIGIN_FLIP);
166 }
167 
__intel_fb_invalidate(struct intel_frontbuffer * front,enum fb_op_origin origin,unsigned int frontbuffer_bits)168 void __intel_fb_invalidate(struct intel_frontbuffer *front,
169 			   enum fb_op_origin origin,
170 			   unsigned int frontbuffer_bits)
171 {
172 	struct drm_i915_private *i915 = to_i915(front->obj->base.dev);
173 
174 	if (origin == ORIGIN_CS) {
175 		spin_lock(&i915->fb_tracking.lock);
176 		i915->fb_tracking.busy_bits |= frontbuffer_bits;
177 		i915->fb_tracking.flip_bits &= ~frontbuffer_bits;
178 		spin_unlock(&i915->fb_tracking.lock);
179 	}
180 
181 	might_sleep();
182 	intel_psr_invalidate(i915, frontbuffer_bits, origin);
183 	intel_edp_drrs_invalidate(i915, frontbuffer_bits);
184 	intel_fbc_invalidate(i915, frontbuffer_bits, origin);
185 }
186 
__intel_fb_flush(struct intel_frontbuffer * front,enum fb_op_origin origin,unsigned int frontbuffer_bits)187 void __intel_fb_flush(struct intel_frontbuffer *front,
188 		      enum fb_op_origin origin,
189 		      unsigned int frontbuffer_bits)
190 {
191 	struct drm_i915_private *i915 = to_i915(front->obj->base.dev);
192 
193 	if (origin == ORIGIN_CS) {
194 		spin_lock(&i915->fb_tracking.lock);
195 		/* Filter out new bits since rendering started. */
196 		frontbuffer_bits &= i915->fb_tracking.busy_bits;
197 		i915->fb_tracking.busy_bits &= ~frontbuffer_bits;
198 		spin_unlock(&i915->fb_tracking.lock);
199 	}
200 
201 	if (frontbuffer_bits)
202 		frontbuffer_flush(i915, frontbuffer_bits, origin);
203 }
204 
frontbuffer_active(struct i915_active * ref)205 static int frontbuffer_active(struct i915_active *ref)
206 {
207 	struct intel_frontbuffer *front =
208 		container_of(ref, typeof(*front), write);
209 
210 	kref_get(&front->ref);
211 	return 0;
212 }
213 
214 __i915_active_call
frontbuffer_retire(struct i915_active * ref)215 static void frontbuffer_retire(struct i915_active *ref)
216 {
217 	struct intel_frontbuffer *front =
218 		container_of(ref, typeof(*front), write);
219 
220 	intel_frontbuffer_flush(front, ORIGIN_CS);
221 	intel_frontbuffer_put(front);
222 }
223 
frontbuffer_release(struct kref * ref)224 static void frontbuffer_release(struct kref *ref)
225 	__releases(&to_i915(front->obj->base.dev)->fb_tracking.lock)
226 {
227 	struct intel_frontbuffer *front =
228 		container_of(ref, typeof(*front), ref);
229 	struct drm_i915_gem_object *obj = front->obj;
230 	struct i915_vma *vma;
231 
232 	spin_lock(&obj->vma.lock);
233 	for_each_ggtt_vma(vma, obj)
234 		vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
235 	spin_unlock(&obj->vma.lock);
236 
237 	RCU_INIT_POINTER(obj->frontbuffer, NULL);
238 	spin_unlock(&to_i915(obj->base.dev)->fb_tracking.lock);
239 
240 	i915_gem_object_put(obj);
241 	i915_active_fini(&front->write);
242 	kfree_rcu(front, rcu);
243 }
244 
245 struct intel_frontbuffer *
intel_frontbuffer_get(struct drm_i915_gem_object * obj)246 intel_frontbuffer_get(struct drm_i915_gem_object *obj)
247 {
248 	struct drm_i915_private *i915 = to_i915(obj->base.dev);
249 	struct intel_frontbuffer *front;
250 
251 	front = __intel_frontbuffer_get(obj);
252 	if (front)
253 		return front;
254 
255 	front = kmalloc(sizeof(*front), GFP_KERNEL);
256 	if (!front)
257 		return NULL;
258 
259 	front->obj = obj;
260 	kref_init(&front->ref);
261 	atomic_set(&front->bits, 0);
262 	i915_active_init(&front->write,
263 			 frontbuffer_active,
264 			 i915_active_may_sleep(frontbuffer_retire));
265 
266 	spin_lock(&i915->fb_tracking.lock);
267 	if (rcu_access_pointer(obj->frontbuffer)) {
268 		kfree(front);
269 		front = rcu_dereference_protected(obj->frontbuffer, true);
270 		kref_get(&front->ref);
271 	} else {
272 		i915_gem_object_get(obj);
273 		rcu_assign_pointer(obj->frontbuffer, front);
274 	}
275 	spin_unlock(&i915->fb_tracking.lock);
276 
277 	return front;
278 }
279 
intel_frontbuffer_put(struct intel_frontbuffer * front)280 void intel_frontbuffer_put(struct intel_frontbuffer *front)
281 {
282 	kref_put_lock(&front->ref,
283 		      frontbuffer_release,
284 		      &to_i915(front->obj->base.dev)->fb_tracking.lock);
285 }
286 
287 /**
288  * intel_frontbuffer_track - update frontbuffer tracking
289  * @old: current buffer for the frontbuffer slots
290  * @new: new buffer for the frontbuffer slots
291  * @frontbuffer_bits: bitmask of frontbuffer slots
292  *
293  * This updates the frontbuffer tracking bits @frontbuffer_bits by clearing them
294  * from @old and setting them in @new. Both @old and @new can be NULL.
295  */
intel_frontbuffer_track(struct intel_frontbuffer * old,struct intel_frontbuffer * new,unsigned int frontbuffer_bits)296 void intel_frontbuffer_track(struct intel_frontbuffer *old,
297 			     struct intel_frontbuffer *new,
298 			     unsigned int frontbuffer_bits)
299 {
300 	/*
301 	 * Control of individual bits within the mask are guarded by
302 	 * the owning plane->mutex, i.e. we can never see concurrent
303 	 * manipulation of individual bits. But since the bitfield as a whole
304 	 * is updated using RMW, we need to use atomics in order to update
305 	 * the bits.
306 	 */
307 	BUILD_BUG_ON(INTEL_FRONTBUFFER_BITS_PER_PIPE * I915_MAX_PIPES >
308 		     BITS_PER_TYPE(atomic_t));
309 
310 	if (old) {
311 		WARN_ON(!(atomic_read(&old->bits) & frontbuffer_bits));
312 		atomic_andnot(frontbuffer_bits, &old->bits);
313 	}
314 
315 	if (new) {
316 		WARN_ON(atomic_read(&new->bits) & frontbuffer_bits);
317 		atomic_or(frontbuffer_bits, &new->bits);
318 	}
319 }
320