xref: /linux/drivers/gpu/drm/xe/display/xe_fb_pin.c (revision cca0d69baf950e5a82c21d09917b4cc654c83fe9)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2021 Intel Corporation
4  */
5 
6 #include <drm/ttm/ttm_bo.h>
7 
8 #include "intel_display_types.h"
9 #include "intel_dpt.h"
10 #include "intel_fb.h"
11 #include "intel_fb_pin.h"
12 #include "xe_bo.h"
13 #include "xe_device.h"
14 #include "xe_ggtt.h"
15 #include "xe_gt.h"
16 #include "xe_pm.h"
17 
18 static void
19 write_dpt_rotated(struct xe_bo *bo, struct iosys_map *map, u32 *dpt_ofs, u32 bo_ofs,
20 		  u32 width, u32 height, u32 src_stride, u32 dst_stride)
21 {
22 	struct xe_device *xe = xe_bo_device(bo);
23 	struct xe_ggtt *ggtt = xe_device_get_root_tile(xe)->mem.ggtt;
24 	u32 column, row;
25 
26 	/* TODO: Maybe rewrite so we can traverse the bo addresses sequentially,
27 	 * by writing dpt/ggtt in a different order?
28 	 */
29 
30 	for (column = 0; column < width; column++) {
31 		u32 src_idx = src_stride * (height - 1) + column + bo_ofs;
32 
33 		for (row = 0; row < height; row++) {
34 			u64 pte = ggtt->pt_ops->pte_encode_bo(bo, src_idx * XE_PAGE_SIZE,
35 							      xe->pat.idx[XE_CACHE_NONE]);
36 
37 			iosys_map_wr(map, *dpt_ofs, u64, pte);
38 			*dpt_ofs += 8;
39 			src_idx -= src_stride;
40 		}
41 
42 		/* The DE ignores the PTEs for the padding tiles */
43 		*dpt_ofs += (dst_stride - height) * 8;
44 	}
45 
46 	/* Align to next page */
47 	*dpt_ofs = ALIGN(*dpt_ofs, 4096);
48 }
49 
50 static void
51 write_dpt_remapped(struct xe_bo *bo, struct iosys_map *map, u32 *dpt_ofs,
52 		   u32 bo_ofs, u32 width, u32 height, u32 src_stride,
53 		   u32 dst_stride)
54 {
55 	struct xe_device *xe = xe_bo_device(bo);
56 	struct xe_ggtt *ggtt = xe_device_get_root_tile(xe)->mem.ggtt;
57 	u64 (*pte_encode_bo)(struct xe_bo *bo, u64 bo_offset, u16 pat_index)
58 		= ggtt->pt_ops->pte_encode_bo;
59 	u32 column, row;
60 
61 	for (row = 0; row < height; row++) {
62 		u32 src_idx = src_stride * row + bo_ofs;
63 
64 		for (column = 0; column < width; column++) {
65 			iosys_map_wr(map, *dpt_ofs, u64,
66 				     pte_encode_bo(bo, src_idx * XE_PAGE_SIZE,
67 				     xe->pat.idx[XE_CACHE_NONE]));
68 
69 			*dpt_ofs += 8;
70 			src_idx++;
71 		}
72 
73 		/* The DE ignores the PTEs for the padding tiles */
74 		*dpt_ofs += (dst_stride - width) * 8;
75 	}
76 
77 	/* Align to next page */
78 	*dpt_ofs = ALIGN(*dpt_ofs, 4096);
79 }
80 
81 static int __xe_pin_fb_vma_dpt(const struct intel_framebuffer *fb,
82 			       const struct i915_gtt_view *view,
83 			       struct i915_vma *vma)
84 {
85 	struct xe_device *xe = to_xe_device(fb->base.dev);
86 	struct xe_tile *tile0 = xe_device_get_root_tile(xe);
87 	struct xe_ggtt *ggtt = tile0->mem.ggtt;
88 	struct xe_bo *bo = intel_fb_obj(&fb->base), *dpt;
89 	u32 dpt_size, size = bo->ttm.base.size;
90 
91 	if (view->type == I915_GTT_VIEW_NORMAL)
92 		dpt_size = ALIGN(size / XE_PAGE_SIZE * 8, XE_PAGE_SIZE);
93 	else if (view->type == I915_GTT_VIEW_REMAPPED)
94 		dpt_size = ALIGN(intel_remapped_info_size(&fb->remapped_view.gtt.remapped) * 8,
95 				 XE_PAGE_SIZE);
96 	else
97 		/* display uses 4K tiles instead of bytes here, convert to entries.. */
98 		dpt_size = ALIGN(intel_rotation_info_size(&view->rotated) * 8,
99 				 XE_PAGE_SIZE);
100 
101 	if (IS_DGFX(xe))
102 		dpt = xe_bo_create_pin_map(xe, tile0, NULL, dpt_size,
103 					   ttm_bo_type_kernel,
104 					   XE_BO_FLAG_VRAM0 |
105 					   XE_BO_FLAG_GGTT |
106 					   XE_BO_FLAG_PAGETABLE);
107 	else
108 		dpt = xe_bo_create_pin_map(xe, tile0, NULL, dpt_size,
109 					   ttm_bo_type_kernel,
110 					   XE_BO_FLAG_STOLEN |
111 					   XE_BO_FLAG_GGTT |
112 					   XE_BO_FLAG_PAGETABLE);
113 	if (IS_ERR(dpt))
114 		dpt = xe_bo_create_pin_map(xe, tile0, NULL, dpt_size,
115 					   ttm_bo_type_kernel,
116 					   XE_BO_FLAG_SYSTEM |
117 					   XE_BO_FLAG_GGTT |
118 					   XE_BO_FLAG_PAGETABLE);
119 	if (IS_ERR(dpt))
120 		return PTR_ERR(dpt);
121 
122 	if (view->type == I915_GTT_VIEW_NORMAL) {
123 		u32 x;
124 
125 		for (x = 0; x < size / XE_PAGE_SIZE; x++) {
126 			u64 pte = ggtt->pt_ops->pte_encode_bo(bo, x * XE_PAGE_SIZE,
127 							      xe->pat.idx[XE_CACHE_NONE]);
128 
129 			iosys_map_wr(&dpt->vmap, x * 8, u64, pte);
130 		}
131 	} else if (view->type == I915_GTT_VIEW_REMAPPED) {
132 		const struct intel_remapped_info *remap_info = &view->remapped;
133 		u32 i, dpt_ofs = 0;
134 
135 		for (i = 0; i < ARRAY_SIZE(remap_info->plane); i++)
136 			write_dpt_remapped(bo, &dpt->vmap, &dpt_ofs,
137 					   remap_info->plane[i].offset,
138 					   remap_info->plane[i].width,
139 					   remap_info->plane[i].height,
140 					   remap_info->plane[i].src_stride,
141 					   remap_info->plane[i].dst_stride);
142 
143 	} else {
144 		const struct intel_rotation_info *rot_info = &view->rotated;
145 		u32 i, dpt_ofs = 0;
146 
147 		for (i = 0; i < ARRAY_SIZE(rot_info->plane); i++)
148 			write_dpt_rotated(bo, &dpt->vmap, &dpt_ofs,
149 					  rot_info->plane[i].offset,
150 					  rot_info->plane[i].width,
151 					  rot_info->plane[i].height,
152 					  rot_info->plane[i].src_stride,
153 					  rot_info->plane[i].dst_stride);
154 	}
155 
156 	vma->dpt = dpt;
157 	vma->node = dpt->ggtt_node;
158 	return 0;
159 }
160 
161 static void
162 write_ggtt_rotated(struct xe_bo *bo, struct xe_ggtt *ggtt, u32 *ggtt_ofs, u32 bo_ofs,
163 		   u32 width, u32 height, u32 src_stride, u32 dst_stride)
164 {
165 	struct xe_device *xe = xe_bo_device(bo);
166 	u32 column, row;
167 
168 	for (column = 0; column < width; column++) {
169 		u32 src_idx = src_stride * (height - 1) + column + bo_ofs;
170 
171 		for (row = 0; row < height; row++) {
172 			u64 pte = ggtt->pt_ops->pte_encode_bo(bo, src_idx * XE_PAGE_SIZE,
173 							      xe->pat.idx[XE_CACHE_NONE]);
174 
175 			ggtt->pt_ops->ggtt_set_pte(ggtt, *ggtt_ofs, pte);
176 			*ggtt_ofs += XE_PAGE_SIZE;
177 			src_idx -= src_stride;
178 		}
179 
180 		/* The DE ignores the PTEs for the padding tiles */
181 		*ggtt_ofs += (dst_stride - height) * XE_PAGE_SIZE;
182 	}
183 }
184 
185 static int __xe_pin_fb_vma_ggtt(const struct intel_framebuffer *fb,
186 				const struct i915_gtt_view *view,
187 				struct i915_vma *vma)
188 {
189 	struct xe_bo *bo = intel_fb_obj(&fb->base);
190 	struct xe_device *xe = to_xe_device(fb->base.dev);
191 	struct xe_ggtt *ggtt = xe_device_get_root_tile(xe)->mem.ggtt;
192 	u32 align;
193 	int ret;
194 
195 	/* TODO: Consider sharing framebuffer mapping?
196 	 * embed i915_vma inside intel_framebuffer
197 	 */
198 	xe_pm_runtime_get_noresume(tile_to_xe(ggtt->tile));
199 	ret = mutex_lock_interruptible(&ggtt->lock);
200 	if (ret)
201 		goto out;
202 
203 	align = XE_PAGE_SIZE;
204 	if (xe_bo_is_vram(bo) && ggtt->flags & XE_GGTT_FLAGS_64K)
205 		align = max_t(u32, align, SZ_64K);
206 
207 	if (bo->ggtt_node.size && view->type == I915_GTT_VIEW_NORMAL) {
208 		vma->node = bo->ggtt_node;
209 	} else if (view->type == I915_GTT_VIEW_NORMAL) {
210 		u32 x, size = bo->ttm.base.size;
211 
212 		ret = xe_ggtt_insert_special_node_locked(ggtt, &vma->node, size,
213 							 align, 0);
214 		if (ret)
215 			goto out_unlock;
216 
217 		for (x = 0; x < size; x += XE_PAGE_SIZE) {
218 			u64 pte = ggtt->pt_ops->pte_encode_bo(bo, x,
219 							      xe->pat.idx[XE_CACHE_NONE]);
220 
221 			ggtt->pt_ops->ggtt_set_pte(ggtt, vma->node.start + x, pte);
222 		}
223 	} else {
224 		u32 i, ggtt_ofs;
225 		const struct intel_rotation_info *rot_info = &view->rotated;
226 
227 		/* display seems to use tiles instead of bytes here, so convert it back.. */
228 		u32 size = intel_rotation_info_size(rot_info) * XE_PAGE_SIZE;
229 
230 		ret = xe_ggtt_insert_special_node_locked(ggtt, &vma->node, size,
231 							 align, 0);
232 		if (ret)
233 			goto out_unlock;
234 
235 		ggtt_ofs = vma->node.start;
236 
237 		for (i = 0; i < ARRAY_SIZE(rot_info->plane); i++)
238 			write_ggtt_rotated(bo, ggtt, &ggtt_ofs,
239 					   rot_info->plane[i].offset,
240 					   rot_info->plane[i].width,
241 					   rot_info->plane[i].height,
242 					   rot_info->plane[i].src_stride,
243 					   rot_info->plane[i].dst_stride);
244 	}
245 
246 out_unlock:
247 	mutex_unlock(&ggtt->lock);
248 out:
249 	xe_pm_runtime_put(tile_to_xe(ggtt->tile));
250 	return ret;
251 }
252 
253 static struct i915_vma *__xe_pin_fb_vma(const struct intel_framebuffer *fb,
254 					const struct i915_gtt_view *view)
255 {
256 	struct drm_device *dev = fb->base.dev;
257 	struct xe_device *xe = to_xe_device(dev);
258 	struct i915_vma *vma = kzalloc(sizeof(*vma), GFP_KERNEL);
259 	struct xe_bo *bo = intel_fb_obj(&fb->base);
260 	int ret;
261 
262 	if (!vma)
263 		return ERR_PTR(-ENODEV);
264 
265 	if (IS_DGFX(to_xe_device(bo->ttm.base.dev)) &&
266 	    intel_fb_rc_ccs_cc_plane(&fb->base) >= 0 &&
267 	    !(bo->flags & XE_BO_FLAG_NEEDS_CPU_ACCESS)) {
268 		struct xe_tile *tile = xe_device_get_root_tile(xe);
269 
270 		/*
271 		 * If we need to able to access the clear-color value stored in
272 		 * the buffer, then we require that such buffers are also CPU
273 		 * accessible.  This is important on small-bar systems where
274 		 * only some subset of VRAM is CPU accessible.
275 		 */
276 		if (tile->mem.vram.io_size < tile->mem.vram.usable_size) {
277 			ret = -EINVAL;
278 			goto err;
279 		}
280 	}
281 
282 	/*
283 	 * Pin the framebuffer, we can't use xe_bo_(un)pin functions as the
284 	 * assumptions are incorrect for framebuffers
285 	 */
286 	ret = ttm_bo_reserve(&bo->ttm, false, false, NULL);
287 	if (ret)
288 		goto err;
289 
290 	if (IS_DGFX(xe))
291 		ret = xe_bo_migrate(bo, XE_PL_VRAM0);
292 	else
293 		ret = xe_bo_validate(bo, NULL, true);
294 	if (!ret)
295 		ttm_bo_pin(&bo->ttm);
296 	ttm_bo_unreserve(&bo->ttm);
297 	if (ret)
298 		goto err;
299 
300 	vma->bo = bo;
301 	if (intel_fb_uses_dpt(&fb->base))
302 		ret = __xe_pin_fb_vma_dpt(fb, view, vma);
303 	else
304 		ret = __xe_pin_fb_vma_ggtt(fb, view, vma);
305 	if (ret)
306 		goto err_unpin;
307 
308 	/* Ensure DPT writes are flushed */
309 	xe_device_l2_flush(xe);
310 	return vma;
311 
312 err_unpin:
313 	ttm_bo_reserve(&bo->ttm, false, false, NULL);
314 	ttm_bo_unpin(&bo->ttm);
315 	ttm_bo_unreserve(&bo->ttm);
316 err:
317 	kfree(vma);
318 	return ERR_PTR(ret);
319 }
320 
321 static void __xe_unpin_fb_vma(struct i915_vma *vma)
322 {
323 	struct xe_device *xe = to_xe_device(vma->bo->ttm.base.dev);
324 	struct xe_ggtt *ggtt = xe_device_get_root_tile(xe)->mem.ggtt;
325 
326 	if (vma->dpt)
327 		xe_bo_unpin_map_no_vm(vma->dpt);
328 	else if (!drm_mm_node_allocated(&vma->bo->ggtt_node) ||
329 		 vma->bo->ggtt_node.start != vma->node.start)
330 		xe_ggtt_remove_node(ggtt, &vma->node, false);
331 
332 	ttm_bo_reserve(&vma->bo->ttm, false, false, NULL);
333 	ttm_bo_unpin(&vma->bo->ttm);
334 	ttm_bo_unreserve(&vma->bo->ttm);
335 	kfree(vma);
336 }
337 
338 struct i915_vma *
339 intel_fb_pin_to_ggtt(const struct drm_framebuffer *fb,
340 		     const struct i915_gtt_view *view,
341 		     unsigned int alignment,
342 		     unsigned int phys_alignment,
343 		     bool uses_fence,
344 		     unsigned long *out_flags)
345 {
346 	*out_flags = 0;
347 
348 	return __xe_pin_fb_vma(to_intel_framebuffer(fb), view);
349 }
350 
351 void intel_fb_unpin_vma(struct i915_vma *vma, unsigned long flags)
352 {
353 	__xe_unpin_fb_vma(vma);
354 }
355 
356 int intel_plane_pin_fb(struct intel_plane_state *plane_state)
357 {
358 	struct drm_framebuffer *fb = plane_state->hw.fb;
359 	struct xe_bo *bo = intel_fb_obj(fb);
360 	struct i915_vma *vma;
361 
362 	/* We reject creating !SCANOUT fb's, so this is weird.. */
363 	drm_WARN_ON(bo->ttm.base.dev, !(bo->flags & XE_BO_FLAG_SCANOUT));
364 
365 	vma = __xe_pin_fb_vma(to_intel_framebuffer(fb), &plane_state->view.gtt);
366 	if (IS_ERR(vma))
367 		return PTR_ERR(vma);
368 
369 	plane_state->ggtt_vma = vma;
370 	return 0;
371 }
372 
373 void intel_plane_unpin_fb(struct intel_plane_state *old_plane_state)
374 {
375 	__xe_unpin_fb_vma(old_plane_state->ggtt_vma);
376 	old_plane_state->ggtt_vma = NULL;
377 }
378 
379 /*
380  * For Xe introduce dummy intel_dpt_create which just return NULL and
381  * intel_dpt_destroy which does nothing.
382  */
383 struct i915_address_space *intel_dpt_create(struct intel_framebuffer *fb)
384 {
385 	return NULL;
386 }
387 
388 void intel_dpt_destroy(struct i915_address_space *vm)
389 {
390 	return;
391 }
392