xref: /linux/drivers/gpu/drm/i915/display/intel_fb_pin.c (revision 22c55fb9eb92395d999b8404d73e58540d11bdd8)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2021 Intel Corporation
4  */
5 
6 /**
7  * DOC: display pinning helpers
8  */
9 
10 #include "gem/i915_gem_domain.h"
11 #include "gem/i915_gem_object.h"
12 
13 #include "i915_drv.h"
14 #include "intel_display_core.h"
15 #include "intel_display_rpm.h"
16 #include "intel_display_types.h"
17 #include "intel_dpt.h"
18 #include "intel_fb.h"
19 #include "intel_fb_pin.h"
20 #include "intel_plane.h"
21 
22 static struct i915_vma *
23 intel_fb_pin_to_dpt(const struct drm_framebuffer *fb,
24 		    const struct i915_gtt_view *view,
25 		    unsigned int alignment,
26 		    unsigned long *out_flags,
27 		    struct i915_address_space *vm)
28 {
29 	struct drm_device *dev = fb->dev;
30 	struct intel_display *display = to_intel_display(dev);
31 	struct drm_i915_private *dev_priv = to_i915(dev);
32 	struct drm_gem_object *_obj = intel_fb_bo(fb);
33 	struct drm_i915_gem_object *obj = to_intel_bo(_obj);
34 	struct i915_gem_ww_ctx ww;
35 	struct i915_vma *vma;
36 	int ret;
37 
38 	/*
39 	 * We are not syncing against the binding (and potential migrations)
40 	 * below, so this vm must never be async.
41 	 */
42 	if (drm_WARN_ON(&dev_priv->drm, vm->bind_async_flags))
43 		return ERR_PTR(-EINVAL);
44 
45 	if (WARN_ON(!i915_gem_object_is_framebuffer(obj)))
46 		return ERR_PTR(-EINVAL);
47 
48 	atomic_inc(&display->restore.pending_fb_pin);
49 
50 	for_i915_gem_ww(&ww, ret, true) {
51 		ret = i915_gem_object_lock(obj, &ww);
52 		if (ret)
53 			continue;
54 
55 		if (HAS_LMEM(dev_priv)) {
56 			unsigned int flags = obj->flags;
57 
58 			/*
59 			 * For this type of buffer we need to able to read from the CPU
60 			 * the clear color value found in the buffer, hence we need to
61 			 * ensure it is always in the mappable part of lmem, if this is
62 			 * a small-bar device.
63 			 */
64 			if (intel_fb_rc_ccs_cc_plane(fb) >= 0)
65 				flags &= ~I915_BO_ALLOC_GPU_ONLY;
66 			ret = __i915_gem_object_migrate(obj, &ww, INTEL_REGION_LMEM_0,
67 							flags);
68 			if (ret)
69 				continue;
70 		}
71 
72 		ret = i915_gem_object_set_cache_level(obj, I915_CACHE_NONE);
73 		if (ret)
74 			continue;
75 
76 		vma = i915_vma_instance(obj, vm, view);
77 		if (IS_ERR(vma)) {
78 			ret = PTR_ERR(vma);
79 			continue;
80 		}
81 
82 		if (i915_vma_misplaced(vma, 0, alignment, 0)) {
83 			ret = i915_vma_unbind(vma);
84 			if (ret)
85 				continue;
86 		}
87 
88 		ret = i915_vma_pin_ww(vma, &ww, 0, alignment, PIN_GLOBAL);
89 		if (ret)
90 			continue;
91 	}
92 	if (ret) {
93 		vma = ERR_PTR(ret);
94 		goto err;
95 	}
96 
97 	vma->display_alignment = max(vma->display_alignment, alignment);
98 
99 	i915_gem_object_flush_if_display(obj);
100 
101 	i915_vma_get(vma);
102 err:
103 	atomic_dec(&display->restore.pending_fb_pin);
104 
105 	return vma;
106 }
107 
108 struct i915_vma *
109 intel_fb_pin_to_ggtt(const struct drm_framebuffer *fb,
110 		     const struct i915_gtt_view *view,
111 		     unsigned int alignment,
112 		     unsigned int phys_alignment,
113 		     unsigned int vtd_guard,
114 		     bool uses_fence,
115 		     unsigned long *out_flags)
116 {
117 	struct drm_device *dev = fb->dev;
118 	struct intel_display *display = to_intel_display(dev);
119 	struct drm_i915_private *dev_priv = to_i915(dev);
120 	struct drm_gem_object *_obj = intel_fb_bo(fb);
121 	struct drm_i915_gem_object *obj = to_intel_bo(_obj);
122 	struct ref_tracker *wakeref;
123 	struct i915_gem_ww_ctx ww;
124 	struct i915_vma *vma;
125 	unsigned int pinctl;
126 	int ret;
127 
128 	if (drm_WARN_ON(dev, !i915_gem_object_is_framebuffer(obj)))
129 		return ERR_PTR(-EINVAL);
130 
131 	if (drm_WARN_ON(dev, alignment && !is_power_of_2(alignment)))
132 		return ERR_PTR(-EINVAL);
133 
134 	/*
135 	 * Global gtt pte registers are special registers which actually forward
136 	 * writes to a chunk of system memory. Which means that there is no risk
137 	 * that the register values disappear as soon as we call
138 	 * intel_runtime_pm_put(), so it is correct to wrap only the
139 	 * pin/unpin/fence and not more.
140 	 */
141 	wakeref = intel_display_rpm_get(display);
142 
143 	atomic_inc(&display->restore.pending_fb_pin);
144 
145 	/*
146 	 * Valleyview is definitely limited to scanning out the first
147 	 * 512MiB. Lets presume this behaviour was inherited from the
148 	 * g4x display engine and that all earlier gen are similarly
149 	 * limited. Testing suggests that it is a little more
150 	 * complicated than this. For example, Cherryview appears quite
151 	 * happy to scanout from anywhere within its global aperture.
152 	 */
153 	pinctl = 0;
154 	if (HAS_GMCH(dev_priv))
155 		pinctl |= PIN_MAPPABLE;
156 
157 	i915_gem_ww_ctx_init(&ww, true);
158 retry:
159 	ret = i915_gem_object_lock(obj, &ww);
160 	if (!ret && phys_alignment)
161 		ret = i915_gem_object_attach_phys(obj, phys_alignment);
162 	else if (!ret && HAS_LMEM(dev_priv))
163 		ret = i915_gem_object_migrate(obj, &ww, INTEL_REGION_LMEM_0);
164 	if (!ret)
165 		ret = i915_gem_object_pin_pages(obj);
166 	if (ret)
167 		goto err;
168 
169 	vma = i915_gem_object_pin_to_display_plane(obj, &ww, alignment,
170 						   vtd_guard, view, pinctl);
171 	if (IS_ERR(vma)) {
172 		ret = PTR_ERR(vma);
173 		goto err_unpin;
174 	}
175 
176 	if (uses_fence && i915_vma_is_map_and_fenceable(vma)) {
177 		/*
178 		 * Install a fence for tiled scan-out. Pre-i965 always needs a
179 		 * fence, whereas 965+ only requires a fence if using
180 		 * framebuffer compression.  For simplicity, we always, when
181 		 * possible, install a fence as the cost is not that onerous.
182 		 *
183 		 * If we fail to fence the tiled scanout, then either the
184 		 * modeset will reject the change (which is highly unlikely as
185 		 * the affected systems, all but one, do not have unmappable
186 		 * space) or we will not be able to enable full powersaving
187 		 * techniques (also likely not to apply due to various limits
188 		 * FBC and the like impose on the size of the buffer, which
189 		 * presumably we violated anyway with this unmappable buffer).
190 		 * Anyway, it is presumably better to stumble onwards with
191 		 * something and try to run the system in a "less than optimal"
192 		 * mode that matches the user configuration.
193 		 */
194 		ret = i915_vma_pin_fence(vma);
195 		if (ret != 0 && DISPLAY_VER(dev_priv) < 4) {
196 			i915_vma_unpin(vma);
197 			goto err_unpin;
198 		}
199 		ret = 0;
200 
201 		if (vma->fence)
202 			*out_flags |= PLANE_HAS_FENCE;
203 	}
204 
205 	i915_vma_get(vma);
206 
207 err_unpin:
208 	i915_gem_object_unpin_pages(obj);
209 err:
210 	if (ret == -EDEADLK) {
211 		ret = i915_gem_ww_ctx_backoff(&ww);
212 		if (!ret)
213 			goto retry;
214 	}
215 	i915_gem_ww_ctx_fini(&ww);
216 	if (ret)
217 		vma = ERR_PTR(ret);
218 
219 	atomic_dec(&display->restore.pending_fb_pin);
220 	intel_display_rpm_put(display, wakeref);
221 	return vma;
222 }
223 
224 void intel_fb_unpin_vma(struct i915_vma *vma, unsigned long flags)
225 {
226 	if (flags & PLANE_HAS_FENCE)
227 		i915_vma_unpin_fence(vma);
228 	i915_vma_unpin(vma);
229 	i915_vma_put(vma);
230 }
231 
232 static unsigned int
233 intel_plane_fb_min_alignment(const struct intel_plane_state *plane_state)
234 {
235 	const struct intel_framebuffer *fb = to_intel_framebuffer(plane_state->hw.fb);
236 
237 	return fb->min_alignment;
238 }
239 
240 static unsigned int
241 intel_plane_fb_min_phys_alignment(const struct intel_plane_state *plane_state)
242 {
243 	struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
244 	const struct drm_framebuffer *fb = plane_state->hw.fb;
245 
246 	if (!intel_plane_needs_physical(plane))
247 		return 0;
248 
249 	return plane->min_alignment(plane, fb, 0);
250 }
251 
252 static unsigned int
253 intel_plane_fb_vtd_guard(const struct intel_plane_state *plane_state)
254 {
255 	return intel_fb_view_vtd_guard(plane_state->hw.fb,
256 				       &plane_state->view,
257 				       plane_state->hw.rotation);
258 }
259 
260 int intel_plane_pin_fb(struct intel_plane_state *plane_state,
261 		       const struct intel_plane_state *old_plane_state)
262 {
263 	struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
264 	const struct intel_framebuffer *fb =
265 		to_intel_framebuffer(plane_state->hw.fb);
266 	struct i915_vma *vma;
267 
268 	if (!intel_fb_uses_dpt(&fb->base)) {
269 		vma = intel_fb_pin_to_ggtt(&fb->base, &plane_state->view.gtt,
270 					   intel_plane_fb_min_alignment(plane_state),
271 					   intel_plane_fb_min_phys_alignment(plane_state),
272 					   intel_plane_fb_vtd_guard(plane_state),
273 					   intel_plane_uses_fence(plane_state),
274 					   &plane_state->flags);
275 		if (IS_ERR(vma))
276 			return PTR_ERR(vma);
277 
278 		plane_state->ggtt_vma = vma;
279 
280 		/*
281 		 * Pre-populate the dma address before we enter the vblank
282 		 * evade critical section as i915_gem_object_get_dma_address()
283 		 * will trigger might_sleep() even if it won't actually sleep,
284 		 * which is the case when the fb has already been pinned.
285 		 */
286 		if (intel_plane_needs_physical(plane)) {
287 			struct drm_i915_gem_object *obj = to_intel_bo(intel_fb_bo(&fb->base));
288 
289 			plane_state->phys_dma_addr = i915_gem_object_get_dma_address(obj, 0);
290 		}
291 	} else {
292 		unsigned int alignment = intel_plane_fb_min_alignment(plane_state);
293 
294 		vma = intel_dpt_pin_to_ggtt(fb->dpt_vm, alignment / 512);
295 		if (IS_ERR(vma))
296 			return PTR_ERR(vma);
297 
298 		plane_state->ggtt_vma = vma;
299 
300 		vma = intel_fb_pin_to_dpt(&fb->base, &plane_state->view.gtt,
301 					  alignment, &plane_state->flags,
302 					  fb->dpt_vm);
303 		if (IS_ERR(vma)) {
304 			intel_dpt_unpin_from_ggtt(fb->dpt_vm);
305 			plane_state->ggtt_vma = NULL;
306 			return PTR_ERR(vma);
307 		}
308 
309 		plane_state->dpt_vma = vma;
310 
311 		WARN_ON(plane_state->ggtt_vma == plane_state->dpt_vma);
312 	}
313 
314 	return 0;
315 }
316 
317 void intel_plane_unpin_fb(struct intel_plane_state *old_plane_state)
318 {
319 	const struct intel_framebuffer *fb =
320 		to_intel_framebuffer(old_plane_state->hw.fb);
321 	struct i915_vma *vma;
322 
323 	if (!intel_fb_uses_dpt(&fb->base)) {
324 		vma = fetch_and_zero(&old_plane_state->ggtt_vma);
325 		if (vma)
326 			intel_fb_unpin_vma(vma, old_plane_state->flags);
327 	} else {
328 		vma = fetch_and_zero(&old_plane_state->dpt_vma);
329 		if (vma)
330 			intel_fb_unpin_vma(vma, old_plane_state->flags);
331 
332 		vma = fetch_and_zero(&old_plane_state->ggtt_vma);
333 		if (vma)
334 			intel_dpt_unpin_from_ggtt(fb->dpt_vm);
335 	}
336 }
337 
338 void intel_fb_get_map(struct i915_vma *vma, struct iosys_map *map)
339 {
340 	iosys_map_set_vaddr_iomem(map, i915_vma_get_iomap(vma));
341 }
342