xref: /linux/drivers/gpu/drm/xe/display/xe_fb_pin.c (revision 2697b79a469b68e3ad3640f55284359c1396278d)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2021 Intel Corporation
4  */
5 
6 #include "i915_drv.h"
7 #include "intel_display_types.h"
8 #include "intel_dpt.h"
9 #include "intel_fb.h"
10 #include "intel_fb_pin.h"
11 #include "xe_ggtt.h"
12 #include "xe_gt.h"
13 #include "xe_pm.h"
14 
15 #include <drm/ttm/ttm_bo.h>
16 
17 static void
18 write_dpt_rotated(struct xe_bo *bo, struct iosys_map *map, u32 *dpt_ofs, u32 bo_ofs,
19 		  u32 width, u32 height, u32 src_stride, u32 dst_stride)
20 {
21 	struct xe_device *xe = xe_bo_device(bo);
22 	struct xe_ggtt *ggtt = xe_device_get_root_tile(xe)->mem.ggtt;
23 	u32 column, row;
24 
25 	/* TODO: Maybe rewrite so we can traverse the bo addresses sequentially,
26 	 * by writing dpt/ggtt in a different order?
27 	 */
28 
29 	for (column = 0; column < width; column++) {
30 		u32 src_idx = src_stride * (height - 1) + column + bo_ofs;
31 
32 		for (row = 0; row < height; row++) {
33 			u64 pte = ggtt->pt_ops->pte_encode_bo(bo, src_idx * XE_PAGE_SIZE,
34 							      xe->pat.idx[XE_CACHE_NONE]);
35 
36 			iosys_map_wr(map, *dpt_ofs, u64, pte);
37 			*dpt_ofs += 8;
38 			src_idx -= src_stride;
39 		}
40 
41 		/* The DE ignores the PTEs for the padding tiles */
42 		*dpt_ofs += (dst_stride - height) * 8;
43 	}
44 
45 	/* Align to next page */
46 	*dpt_ofs = ALIGN(*dpt_ofs, 4096);
47 }
48 
49 static void
50 write_dpt_remapped(struct xe_bo *bo, struct iosys_map *map, u32 *dpt_ofs,
51 		   u32 bo_ofs, u32 width, u32 height, u32 src_stride,
52 		   u32 dst_stride)
53 {
54 	struct xe_device *xe = xe_bo_device(bo);
55 	struct xe_ggtt *ggtt = xe_device_get_root_tile(xe)->mem.ggtt;
56 	u64 (*pte_encode_bo)(struct xe_bo *bo, u64 bo_offset, u16 pat_index)
57 		= ggtt->pt_ops->pte_encode_bo;
58 	u32 column, row;
59 
60 	for (row = 0; row < height; row++) {
61 		u32 src_idx = src_stride * row + bo_ofs;
62 
63 		for (column = 0; column < width; column++) {
64 			iosys_map_wr(map, *dpt_ofs, u64,
65 				     pte_encode_bo(bo, src_idx * XE_PAGE_SIZE,
66 				     xe->pat.idx[XE_CACHE_NONE]));
67 
68 			*dpt_ofs += 8;
69 			src_idx++;
70 		}
71 
72 		/* The DE ignores the PTEs for the padding tiles */
73 		*dpt_ofs += (dst_stride - width) * 8;
74 	}
75 
76 	/* Align to next page */
77 	*dpt_ofs = ALIGN(*dpt_ofs, 4096);
78 }
79 
80 static int __xe_pin_fb_vma_dpt(struct intel_framebuffer *fb,
81 			       const struct i915_gtt_view *view,
82 			       struct i915_vma *vma)
83 {
84 	struct xe_device *xe = to_xe_device(fb->base.dev);
85 	struct xe_tile *tile0 = xe_device_get_root_tile(xe);
86 	struct xe_ggtt *ggtt = tile0->mem.ggtt;
87 	struct xe_bo *bo = intel_fb_obj(&fb->base), *dpt;
88 	u32 dpt_size, size = bo->ttm.base.size;
89 
90 	if (view->type == I915_GTT_VIEW_NORMAL)
91 		dpt_size = ALIGN(size / XE_PAGE_SIZE * 8, XE_PAGE_SIZE);
92 	else if (view->type == I915_GTT_VIEW_REMAPPED)
93 		dpt_size = ALIGN(intel_remapped_info_size(&fb->remapped_view.gtt.remapped) * 8,
94 				 XE_PAGE_SIZE);
95 	else
96 		/* display uses 4K tiles instead of bytes here, convert to entries.. */
97 		dpt_size = ALIGN(intel_rotation_info_size(&view->rotated) * 8,
98 				 XE_PAGE_SIZE);
99 
100 	if (IS_DGFX(xe))
101 		dpt = xe_bo_create_pin_map(xe, tile0, NULL, dpt_size,
102 					   ttm_bo_type_kernel,
103 					   XE_BO_FLAG_VRAM0 |
104 					   XE_BO_FLAG_GGTT |
105 					   XE_BO_FLAG_PAGETABLE);
106 	else
107 		dpt = xe_bo_create_pin_map(xe, tile0, NULL, dpt_size,
108 					   ttm_bo_type_kernel,
109 					   XE_BO_FLAG_STOLEN |
110 					   XE_BO_FLAG_GGTT |
111 					   XE_BO_FLAG_PAGETABLE);
112 	if (IS_ERR(dpt))
113 		dpt = xe_bo_create_pin_map(xe, tile0, NULL, dpt_size,
114 					   ttm_bo_type_kernel,
115 					   XE_BO_FLAG_SYSTEM |
116 					   XE_BO_FLAG_GGTT |
117 					   XE_BO_FLAG_PAGETABLE);
118 	if (IS_ERR(dpt))
119 		return PTR_ERR(dpt);
120 
121 	if (view->type == I915_GTT_VIEW_NORMAL) {
122 		u32 x;
123 
124 		for (x = 0; x < size / XE_PAGE_SIZE; x++) {
125 			u64 pte = ggtt->pt_ops->pte_encode_bo(bo, x * XE_PAGE_SIZE,
126 							      xe->pat.idx[XE_CACHE_NONE]);
127 
128 			iosys_map_wr(&dpt->vmap, x * 8, u64, pte);
129 		}
130 	} else if (view->type == I915_GTT_VIEW_REMAPPED) {
131 		const struct intel_remapped_info *remap_info = &view->remapped;
132 		u32 i, dpt_ofs = 0;
133 
134 		for (i = 0; i < ARRAY_SIZE(remap_info->plane); i++)
135 			write_dpt_remapped(bo, &dpt->vmap, &dpt_ofs,
136 					   remap_info->plane[i].offset,
137 					   remap_info->plane[i].width,
138 					   remap_info->plane[i].height,
139 					   remap_info->plane[i].src_stride,
140 					   remap_info->plane[i].dst_stride);
141 
142 	} else {
143 		const struct intel_rotation_info *rot_info = &view->rotated;
144 		u32 i, dpt_ofs = 0;
145 
146 		for (i = 0; i < ARRAY_SIZE(rot_info->plane); i++)
147 			write_dpt_rotated(bo, &dpt->vmap, &dpt_ofs,
148 					  rot_info->plane[i].offset,
149 					  rot_info->plane[i].width,
150 					  rot_info->plane[i].height,
151 					  rot_info->plane[i].src_stride,
152 					  rot_info->plane[i].dst_stride);
153 	}
154 
155 	vma->dpt = dpt;
156 	vma->node = dpt->ggtt_node;
157 	return 0;
158 }
159 
160 static void
161 write_ggtt_rotated(struct xe_bo *bo, struct xe_ggtt *ggtt, u32 *ggtt_ofs, u32 bo_ofs,
162 		   u32 width, u32 height, u32 src_stride, u32 dst_stride)
163 {
164 	struct xe_device *xe = xe_bo_device(bo);
165 	u32 column, row;
166 
167 	for (column = 0; column < width; column++) {
168 		u32 src_idx = src_stride * (height - 1) + column + bo_ofs;
169 
170 		for (row = 0; row < height; row++) {
171 			u64 pte = ggtt->pt_ops->pte_encode_bo(bo, src_idx * XE_PAGE_SIZE,
172 							      xe->pat.idx[XE_CACHE_NONE]);
173 
174 			xe_ggtt_set_pte(ggtt, *ggtt_ofs, pte);
175 			*ggtt_ofs += XE_PAGE_SIZE;
176 			src_idx -= src_stride;
177 		}
178 
179 		/* The DE ignores the PTEs for the padding tiles */
180 		*ggtt_ofs += (dst_stride - height) * XE_PAGE_SIZE;
181 	}
182 }
183 
184 static int __xe_pin_fb_vma_ggtt(struct intel_framebuffer *fb,
185 				const struct i915_gtt_view *view,
186 				struct i915_vma *vma)
187 {
188 	struct xe_bo *bo = intel_fb_obj(&fb->base);
189 	struct xe_device *xe = to_xe_device(fb->base.dev);
190 	struct xe_ggtt *ggtt = xe_device_get_root_tile(xe)->mem.ggtt;
191 	u32 align;
192 	int ret;
193 
194 	/* TODO: Consider sharing framebuffer mapping?
195 	 * embed i915_vma inside intel_framebuffer
196 	 */
197 	xe_pm_runtime_get_noresume(tile_to_xe(ggtt->tile));
198 	ret = mutex_lock_interruptible(&ggtt->lock);
199 	if (ret)
200 		goto out;
201 
202 	align = XE_PAGE_SIZE;
203 	if (xe_bo_is_vram(bo) && ggtt->flags & XE_GGTT_FLAGS_64K)
204 		align = max_t(u32, align, SZ_64K);
205 
206 	if (bo->ggtt_node.size && view->type == I915_GTT_VIEW_NORMAL) {
207 		vma->node = bo->ggtt_node;
208 	} else if (view->type == I915_GTT_VIEW_NORMAL) {
209 		u32 x, size = bo->ttm.base.size;
210 
211 		ret = xe_ggtt_insert_special_node_locked(ggtt, &vma->node, size,
212 							 align, 0);
213 		if (ret)
214 			goto out_unlock;
215 
216 		for (x = 0; x < size; x += XE_PAGE_SIZE) {
217 			u64 pte = ggtt->pt_ops->pte_encode_bo(bo, x,
218 							      xe->pat.idx[XE_CACHE_NONE]);
219 
220 			xe_ggtt_set_pte(ggtt, vma->node.start + x, pte);
221 		}
222 	} else {
223 		u32 i, ggtt_ofs;
224 		const struct intel_rotation_info *rot_info = &view->rotated;
225 
226 		/* display seems to use tiles instead of bytes here, so convert it back.. */
227 		u32 size = intel_rotation_info_size(rot_info) * XE_PAGE_SIZE;
228 
229 		ret = xe_ggtt_insert_special_node_locked(ggtt, &vma->node, size,
230 							 align, 0);
231 		if (ret)
232 			goto out_unlock;
233 
234 		ggtt_ofs = vma->node.start;
235 
236 		for (i = 0; i < ARRAY_SIZE(rot_info->plane); i++)
237 			write_ggtt_rotated(bo, ggtt, &ggtt_ofs,
238 					   rot_info->plane[i].offset,
239 					   rot_info->plane[i].width,
240 					   rot_info->plane[i].height,
241 					   rot_info->plane[i].src_stride,
242 					   rot_info->plane[i].dst_stride);
243 	}
244 
245 out_unlock:
246 	mutex_unlock(&ggtt->lock);
247 out:
248 	xe_pm_runtime_put(tile_to_xe(ggtt->tile));
249 	return ret;
250 }
251 
252 static struct i915_vma *__xe_pin_fb_vma(struct intel_framebuffer *fb,
253 					const struct i915_gtt_view *view)
254 {
255 	struct drm_device *dev = fb->base.dev;
256 	struct xe_device *xe = to_xe_device(dev);
257 	struct i915_vma *vma = kzalloc(sizeof(*vma), GFP_KERNEL);
258 	struct xe_bo *bo = intel_fb_obj(&fb->base);
259 	int ret;
260 
261 	if (!vma)
262 		return ERR_PTR(-ENODEV);
263 
264 	if (IS_DGFX(to_xe_device(bo->ttm.base.dev)) &&
265 	    intel_fb_rc_ccs_cc_plane(&fb->base) >= 0 &&
266 	    !(bo->flags & XE_BO_FLAG_NEEDS_CPU_ACCESS)) {
267 		struct xe_tile *tile = xe_device_get_root_tile(xe);
268 
269 		/*
270 		 * If we need to able to access the clear-color value stored in
271 		 * the buffer, then we require that such buffers are also CPU
272 		 * accessible.  This is important on small-bar systems where
273 		 * only some subset of VRAM is CPU accessible.
274 		 */
275 		if (tile->mem.vram.io_size < tile->mem.vram.usable_size) {
276 			ret = -EINVAL;
277 			goto err;
278 		}
279 	}
280 
281 	/*
282 	 * Pin the framebuffer, we can't use xe_bo_(un)pin functions as the
283 	 * assumptions are incorrect for framebuffers
284 	 */
285 	ret = ttm_bo_reserve(&bo->ttm, false, false, NULL);
286 	if (ret)
287 		goto err;
288 
289 	if (IS_DGFX(xe))
290 		ret = xe_bo_migrate(bo, XE_PL_VRAM0);
291 	else
292 		ret = xe_bo_validate(bo, NULL, true);
293 	if (!ret)
294 		ttm_bo_pin(&bo->ttm);
295 	ttm_bo_unreserve(&bo->ttm);
296 	if (ret)
297 		goto err;
298 
299 	vma->bo = bo;
300 	if (intel_fb_uses_dpt(&fb->base))
301 		ret = __xe_pin_fb_vma_dpt(fb, view, vma);
302 	else
303 		ret = __xe_pin_fb_vma_ggtt(fb, view, vma);
304 	if (ret)
305 		goto err_unpin;
306 
307 	return vma;
308 
309 err_unpin:
310 	ttm_bo_reserve(&bo->ttm, false, false, NULL);
311 	ttm_bo_unpin(&bo->ttm);
312 	ttm_bo_unreserve(&bo->ttm);
313 err:
314 	kfree(vma);
315 	return ERR_PTR(ret);
316 }
317 
318 static void __xe_unpin_fb_vma(struct i915_vma *vma)
319 {
320 	struct xe_device *xe = to_xe_device(vma->bo->ttm.base.dev);
321 	struct xe_ggtt *ggtt = xe_device_get_root_tile(xe)->mem.ggtt;
322 
323 	if (vma->dpt)
324 		xe_bo_unpin_map_no_vm(vma->dpt);
325 	else if (!drm_mm_node_allocated(&vma->bo->ggtt_node) ||
326 		 vma->bo->ggtt_node.start != vma->node.start)
327 		xe_ggtt_remove_node(ggtt, &vma->node, false);
328 
329 	ttm_bo_reserve(&vma->bo->ttm, false, false, NULL);
330 	ttm_bo_unpin(&vma->bo->ttm);
331 	ttm_bo_unreserve(&vma->bo->ttm);
332 	kfree(vma);
333 }
334 
335 struct i915_vma *
336 intel_pin_and_fence_fb_obj(struct drm_framebuffer *fb,
337 			   bool phys_cursor,
338 			   const struct i915_gtt_view *view,
339 			   bool uses_fence,
340 			   unsigned long *out_flags)
341 {
342 	*out_flags = 0;
343 
344 	return __xe_pin_fb_vma(to_intel_framebuffer(fb), view);
345 }
346 
347 void intel_unpin_fb_vma(struct i915_vma *vma, unsigned long flags)
348 {
349 	__xe_unpin_fb_vma(vma);
350 }
351 
352 int intel_plane_pin_fb(struct intel_plane_state *plane_state)
353 {
354 	struct drm_framebuffer *fb = plane_state->hw.fb;
355 	struct xe_bo *bo = intel_fb_obj(fb);
356 	struct i915_vma *vma;
357 
358 	/* We reject creating !SCANOUT fb's, so this is weird.. */
359 	drm_WARN_ON(bo->ttm.base.dev, !(bo->flags & XE_BO_FLAG_SCANOUT));
360 
361 	vma = __xe_pin_fb_vma(to_intel_framebuffer(fb), &plane_state->view.gtt);
362 	if (IS_ERR(vma))
363 		return PTR_ERR(vma);
364 
365 	plane_state->ggtt_vma = vma;
366 	return 0;
367 }
368 
369 void intel_plane_unpin_fb(struct intel_plane_state *old_plane_state)
370 {
371 	__xe_unpin_fb_vma(old_plane_state->ggtt_vma);
372 	old_plane_state->ggtt_vma = NULL;
373 }
374 
375 /*
376  * For Xe introduce dummy intel_dpt_create which just return NULL and
377  * intel_dpt_destroy which does nothing.
378  */
379 struct i915_address_space *intel_dpt_create(struct intel_framebuffer *fb)
380 {
381 	return NULL;
382 }
383 
384 void intel_dpt_destroy(struct i915_address_space *vm)
385 {
386 	return;
387 }
388