xref: /linux/drivers/gpu/drm/i915/display/intel_fb_pin.c (revision fcab107abe1ab5be9dbe874baa722372da8f4f73)
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_atomic_plane.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 
21 static struct i915_vma *
22 intel_fb_pin_to_dpt(const struct drm_framebuffer *fb,
23 		    const struct i915_gtt_view *view,
24 		    unsigned int alignment,
25 		    unsigned long *out_flags,
26 		    struct i915_address_space *vm)
27 {
28 	struct drm_device *dev = fb->dev;
29 	struct intel_display *display = to_intel_display(dev);
30 	struct drm_i915_private *dev_priv = to_i915(dev);
31 	struct drm_gem_object *_obj = intel_fb_bo(fb);
32 	struct drm_i915_gem_object *obj = to_intel_bo(_obj);
33 	struct i915_gem_ww_ctx ww;
34 	struct i915_vma *vma;
35 	int ret;
36 
37 	/*
38 	 * We are not syncing against the binding (and potential migrations)
39 	 * below, so this vm must never be async.
40 	 */
41 	if (drm_WARN_ON(&dev_priv->drm, vm->bind_async_flags))
42 		return ERR_PTR(-EINVAL);
43 
44 	if (WARN_ON(!i915_gem_object_is_framebuffer(obj)))
45 		return ERR_PTR(-EINVAL);
46 
47 	atomic_inc(&display->restore.pending_fb_pin);
48 
49 	for_i915_gem_ww(&ww, ret, true) {
50 		ret = i915_gem_object_lock(obj, &ww);
51 		if (ret)
52 			continue;
53 
54 		if (HAS_LMEM(dev_priv)) {
55 			unsigned int flags = obj->flags;
56 
57 			/*
58 			 * For this type of buffer we need to able to read from the CPU
59 			 * the clear color value found in the buffer, hence we need to
60 			 * ensure it is always in the mappable part of lmem, if this is
61 			 * a small-bar device.
62 			 */
63 			if (intel_fb_rc_ccs_cc_plane(fb) >= 0)
64 				flags &= ~I915_BO_ALLOC_GPU_ONLY;
65 			ret = __i915_gem_object_migrate(obj, &ww, INTEL_REGION_LMEM_0,
66 							flags);
67 			if (ret)
68 				continue;
69 		}
70 
71 		ret = i915_gem_object_set_cache_level(obj, I915_CACHE_NONE);
72 		if (ret)
73 			continue;
74 
75 		vma = i915_vma_instance(obj, vm, view);
76 		if (IS_ERR(vma)) {
77 			ret = PTR_ERR(vma);
78 			continue;
79 		}
80 
81 		if (i915_vma_misplaced(vma, 0, alignment, 0)) {
82 			ret = i915_vma_unbind(vma);
83 			if (ret)
84 				continue;
85 		}
86 
87 		ret = i915_vma_pin_ww(vma, &ww, 0, alignment, PIN_GLOBAL);
88 		if (ret)
89 			continue;
90 	}
91 	if (ret) {
92 		vma = ERR_PTR(ret);
93 		goto err;
94 	}
95 
96 	vma->display_alignment = max(vma->display_alignment, alignment);
97 
98 	i915_gem_object_flush_if_display(obj);
99 
100 	i915_vma_get(vma);
101 err:
102 	atomic_dec(&display->restore.pending_fb_pin);
103 
104 	return vma;
105 }
106 
107 struct i915_vma *
108 intel_fb_pin_to_ggtt(const struct drm_framebuffer *fb,
109 		     const struct i915_gtt_view *view,
110 		     unsigned int alignment,
111 		     unsigned int phys_alignment,
112 		     unsigned int vtd_guard,
113 		     bool uses_fence,
114 		     unsigned long *out_flags)
115 {
116 	struct drm_device *dev = fb->dev;
117 	struct intel_display *display = to_intel_display(dev);
118 	struct drm_i915_private *dev_priv = to_i915(dev);
119 	struct drm_gem_object *_obj = intel_fb_bo(fb);
120 	struct drm_i915_gem_object *obj = to_intel_bo(_obj);
121 	struct ref_tracker *wakeref;
122 	struct i915_gem_ww_ctx ww;
123 	struct i915_vma *vma;
124 	unsigned int pinctl;
125 	int ret;
126 
127 	if (drm_WARN_ON(dev, !i915_gem_object_is_framebuffer(obj)))
128 		return ERR_PTR(-EINVAL);
129 
130 	if (drm_WARN_ON(dev, alignment && !is_power_of_2(alignment)))
131 		return ERR_PTR(-EINVAL);
132 
133 	/*
134 	 * Global gtt pte registers are special registers which actually forward
135 	 * writes to a chunk of system memory. Which means that there is no risk
136 	 * that the register values disappear as soon as we call
137 	 * intel_runtime_pm_put(), so it is correct to wrap only the
138 	 * pin/unpin/fence and not more.
139 	 */
140 	wakeref = intel_display_rpm_get(display);
141 
142 	atomic_inc(&display->restore.pending_fb_pin);
143 
144 	/*
145 	 * Valleyview is definitely limited to scanning out the first
146 	 * 512MiB. Lets presume this behaviour was inherited from the
147 	 * g4x display engine and that all earlier gen are similarly
148 	 * limited. Testing suggests that it is a little more
149 	 * complicated than this. For example, Cherryview appears quite
150 	 * happy to scanout from anywhere within its global aperture.
151 	 */
152 	pinctl = 0;
153 	if (HAS_GMCH(dev_priv))
154 		pinctl |= PIN_MAPPABLE;
155 
156 	i915_gem_ww_ctx_init(&ww, true);
157 retry:
158 	ret = i915_gem_object_lock(obj, &ww);
159 	if (!ret && phys_alignment)
160 		ret = i915_gem_object_attach_phys(obj, phys_alignment);
161 	else if (!ret && HAS_LMEM(dev_priv))
162 		ret = i915_gem_object_migrate(obj, &ww, INTEL_REGION_LMEM_0);
163 	if (!ret)
164 		ret = i915_gem_object_pin_pages(obj);
165 	if (ret)
166 		goto err;
167 
168 	vma = i915_gem_object_pin_to_display_plane(obj, &ww, alignment,
169 						   vtd_guard, view, pinctl);
170 	if (IS_ERR(vma)) {
171 		ret = PTR_ERR(vma);
172 		goto err_unpin;
173 	}
174 
175 	if (uses_fence && i915_vma_is_map_and_fenceable(vma)) {
176 		/*
177 		 * Install a fence for tiled scan-out. Pre-i965 always needs a
178 		 * fence, whereas 965+ only requires a fence if using
179 		 * framebuffer compression.  For simplicity, we always, when
180 		 * possible, install a fence as the cost is not that onerous.
181 		 *
182 		 * If we fail to fence the tiled scanout, then either the
183 		 * modeset will reject the change (which is highly unlikely as
184 		 * the affected systems, all but one, do not have unmappable
185 		 * space) or we will not be able to enable full powersaving
186 		 * techniques (also likely not to apply due to various limits
187 		 * FBC and the like impose on the size of the buffer, which
188 		 * presumably we violated anyway with this unmappable buffer).
189 		 * Anyway, it is presumably better to stumble onwards with
190 		 * something and try to run the system in a "less than optimal"
191 		 * mode that matches the user configuration.
192 		 */
193 		ret = i915_vma_pin_fence(vma);
194 		if (ret != 0 && DISPLAY_VER(dev_priv) < 4) {
195 			i915_vma_unpin(vma);
196 			goto err_unpin;
197 		}
198 		ret = 0;
199 
200 		if (vma->fence)
201 			*out_flags |= PLANE_HAS_FENCE;
202 	}
203 
204 	i915_vma_get(vma);
205 
206 err_unpin:
207 	i915_gem_object_unpin_pages(obj);
208 err:
209 	if (ret == -EDEADLK) {
210 		ret = i915_gem_ww_ctx_backoff(&ww);
211 		if (!ret)
212 			goto retry;
213 	}
214 	i915_gem_ww_ctx_fini(&ww);
215 	if (ret)
216 		vma = ERR_PTR(ret);
217 
218 	atomic_dec(&display->restore.pending_fb_pin);
219 	intel_display_rpm_put(display, wakeref);
220 	return vma;
221 }
222 
223 void intel_fb_unpin_vma(struct i915_vma *vma, unsigned long flags)
224 {
225 	if (flags & PLANE_HAS_FENCE)
226 		i915_vma_unpin_fence(vma);
227 	i915_vma_unpin(vma);
228 	i915_vma_put(vma);
229 }
230 
231 static unsigned int
232 intel_plane_fb_min_alignment(const struct intel_plane_state *plane_state)
233 {
234 	const struct intel_framebuffer *fb = to_intel_framebuffer(plane_state->hw.fb);
235 
236 	return fb->min_alignment;
237 }
238 
239 static unsigned int
240 intel_plane_fb_min_phys_alignment(const struct intel_plane_state *plane_state)
241 {
242 	struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
243 	const struct drm_framebuffer *fb = plane_state->hw.fb;
244 
245 	if (!intel_plane_needs_physical(plane))
246 		return 0;
247 
248 	return plane->min_alignment(plane, fb, 0);
249 }
250 
251 static unsigned int
252 intel_plane_fb_vtd_guard(const struct intel_plane_state *plane_state)
253 {
254 	return intel_fb_view_vtd_guard(plane_state->hw.fb,
255 				       &plane_state->view,
256 				       plane_state->hw.rotation);
257 }
258 
259 int intel_plane_pin_fb(struct intel_plane_state *plane_state,
260 		       const struct intel_plane_state *old_plane_state)
261 {
262 	struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
263 	const struct intel_framebuffer *fb =
264 		to_intel_framebuffer(plane_state->hw.fb);
265 	struct i915_vma *vma;
266 
267 	if (!intel_fb_uses_dpt(&fb->base)) {
268 		vma = intel_fb_pin_to_ggtt(&fb->base, &plane_state->view.gtt,
269 					   intel_plane_fb_min_alignment(plane_state),
270 					   intel_plane_fb_min_phys_alignment(plane_state),
271 					   intel_plane_fb_vtd_guard(plane_state),
272 					   intel_plane_uses_fence(plane_state),
273 					   &plane_state->flags);
274 		if (IS_ERR(vma))
275 			return PTR_ERR(vma);
276 
277 		plane_state->ggtt_vma = vma;
278 
279 		/*
280 		 * Pre-populate the dma address before we enter the vblank
281 		 * evade critical section as i915_gem_object_get_dma_address()
282 		 * will trigger might_sleep() even if it won't actually sleep,
283 		 * which is the case when the fb has already been pinned.
284 		 */
285 		if (intel_plane_needs_physical(plane)) {
286 			struct drm_i915_gem_object *obj = to_intel_bo(intel_fb_bo(&fb->base));
287 
288 			plane_state->phys_dma_addr = i915_gem_object_get_dma_address(obj, 0);
289 		}
290 	} else {
291 		unsigned int alignment = intel_plane_fb_min_alignment(plane_state);
292 
293 		vma = intel_dpt_pin_to_ggtt(fb->dpt_vm, alignment / 512);
294 		if (IS_ERR(vma))
295 			return PTR_ERR(vma);
296 
297 		plane_state->ggtt_vma = vma;
298 
299 		vma = intel_fb_pin_to_dpt(&fb->base, &plane_state->view.gtt,
300 					  alignment, &plane_state->flags,
301 					  fb->dpt_vm);
302 		if (IS_ERR(vma)) {
303 			intel_dpt_unpin_from_ggtt(fb->dpt_vm);
304 			plane_state->ggtt_vma = NULL;
305 			return PTR_ERR(vma);
306 		}
307 
308 		plane_state->dpt_vma = vma;
309 
310 		WARN_ON(plane_state->ggtt_vma == plane_state->dpt_vma);
311 	}
312 
313 	return 0;
314 }
315 
316 void intel_plane_unpin_fb(struct intel_plane_state *old_plane_state)
317 {
318 	const struct intel_framebuffer *fb =
319 		to_intel_framebuffer(old_plane_state->hw.fb);
320 	struct i915_vma *vma;
321 
322 	if (!intel_fb_uses_dpt(&fb->base)) {
323 		vma = fetch_and_zero(&old_plane_state->ggtt_vma);
324 		if (vma)
325 			intel_fb_unpin_vma(vma, old_plane_state->flags);
326 	} else {
327 		vma = fetch_and_zero(&old_plane_state->dpt_vma);
328 		if (vma)
329 			intel_fb_unpin_vma(vma, old_plane_state->flags);
330 
331 		vma = fetch_and_zero(&old_plane_state->ggtt_vma);
332 		if (vma)
333 			intel_dpt_unpin_from_ggtt(fb->dpt_vm);
334 	}
335 }
336