xref: /linux/drivers/gpu/drm/xe/display/xe_fb_pin.c (revision 7abdafd2343ab199367c8243d6a5f06a9aa6976b)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2021 Intel Corporation
4  */
5 
6 #include <drm/ttm/ttm_bo.h>
7 
8 #include "i915_vma.h"
9 #include "intel_display_types.h"
10 #include "intel_dpt.h"
11 #include "intel_fb.h"
12 #include "intel_fb_pin.h"
13 #include "intel_fbdev.h"
14 #include "xe_bo.h"
15 #include "xe_device.h"
16 #include "xe_ggtt.h"
17 #include "xe_pm.h"
18 
19 static void
write_dpt_rotated(struct xe_bo * bo,struct iosys_map * map,u32 * dpt_ofs,u32 bo_ofs,u32 width,u32 height,u32 src_stride,u32 dst_stride)20 write_dpt_rotated(struct xe_bo *bo, struct iosys_map *map, u32 *dpt_ofs, u32 bo_ofs,
21 		  u32 width, u32 height, u32 src_stride, u32 dst_stride)
22 {
23 	struct xe_device *xe = xe_bo_device(bo);
24 	struct xe_ggtt *ggtt = xe_device_get_root_tile(xe)->mem.ggtt;
25 	u32 column, row;
26 
27 	/* TODO: Maybe rewrite so we can traverse the bo addresses sequentially,
28 	 * by writing dpt/ggtt in a different order?
29 	 */
30 
31 	for (column = 0; column < width; column++) {
32 		u32 src_idx = src_stride * (height - 1) + column + bo_ofs;
33 
34 		for (row = 0; row < height; row++) {
35 			u64 pte = ggtt->pt_ops->pte_encode_bo(bo, src_idx * XE_PAGE_SIZE,
36 							      xe->pat.idx[XE_CACHE_NONE]);
37 
38 			iosys_map_wr(map, *dpt_ofs, u64, pte);
39 			*dpt_ofs += 8;
40 			src_idx -= src_stride;
41 		}
42 
43 		/* The DE ignores the PTEs for the padding tiles */
44 		*dpt_ofs += (dst_stride - height) * 8;
45 	}
46 
47 	/* Align to next page */
48 	*dpt_ofs = ALIGN(*dpt_ofs, 4096);
49 }
50 
51 static void
write_dpt_remapped(struct xe_bo * bo,struct iosys_map * map,u32 * dpt_ofs,u32 bo_ofs,u32 width,u32 height,u32 src_stride,u32 dst_stride)52 write_dpt_remapped(struct xe_bo *bo, struct iosys_map *map, u32 *dpt_ofs,
53 		   u32 bo_ofs, u32 width, u32 height, u32 src_stride,
54 		   u32 dst_stride)
55 {
56 	struct xe_device *xe = xe_bo_device(bo);
57 	struct xe_ggtt *ggtt = xe_device_get_root_tile(xe)->mem.ggtt;
58 	u64 (*pte_encode_bo)(struct xe_bo *bo, u64 bo_offset, u16 pat_index)
59 		= ggtt->pt_ops->pte_encode_bo;
60 	u32 column, row;
61 
62 	for (row = 0; row < height; row++) {
63 		u32 src_idx = src_stride * row + bo_ofs;
64 
65 		for (column = 0; column < width; column++) {
66 			iosys_map_wr(map, *dpt_ofs, u64,
67 				     pte_encode_bo(bo, src_idx * XE_PAGE_SIZE,
68 				     xe->pat.idx[XE_CACHE_NONE]));
69 
70 			*dpt_ofs += 8;
71 			src_idx++;
72 		}
73 
74 		/* The DE ignores the PTEs for the padding tiles */
75 		*dpt_ofs += (dst_stride - width) * 8;
76 	}
77 
78 	/* Align to next page */
79 	*dpt_ofs = ALIGN(*dpt_ofs, 4096);
80 }
81 
__xe_pin_fb_vma_dpt(const struct intel_framebuffer * fb,const struct i915_gtt_view * view,struct i915_vma * vma,unsigned int alignment)82 static int __xe_pin_fb_vma_dpt(const struct intel_framebuffer *fb,
83 			       const struct i915_gtt_view *view,
84 			       struct i915_vma *vma,
85 			       unsigned int alignment)
86 {
87 	struct xe_device *xe = to_xe_device(fb->base.dev);
88 	struct xe_tile *tile0 = xe_device_get_root_tile(xe);
89 	struct xe_ggtt *ggtt = tile0->mem.ggtt;
90 	struct drm_gem_object *obj = intel_fb_bo(&fb->base);
91 	struct xe_bo *bo = gem_to_xe_bo(obj), *dpt;
92 	u32 dpt_size, size = bo->ttm.base.size;
93 
94 	if (view->type == I915_GTT_VIEW_NORMAL)
95 		dpt_size = ALIGN(size / XE_PAGE_SIZE * 8, XE_PAGE_SIZE);
96 	else if (view->type == I915_GTT_VIEW_REMAPPED)
97 		dpt_size = ALIGN(intel_remapped_info_size(&fb->remapped_view.gtt.remapped) * 8,
98 				 XE_PAGE_SIZE);
99 	else
100 		/* display uses 4K tiles instead of bytes here, convert to entries.. */
101 		dpt_size = ALIGN(intel_rotation_info_size(&view->rotated) * 8,
102 				 XE_PAGE_SIZE);
103 
104 	if (IS_DGFX(xe))
105 		dpt = xe_bo_create_pin_map_at_aligned(xe, tile0, NULL,
106 						      dpt_size, ~0ull,
107 						      ttm_bo_type_kernel,
108 						      XE_BO_FLAG_VRAM0 |
109 						      XE_BO_FLAG_GGTT |
110 						      XE_BO_FLAG_PAGETABLE,
111 						      alignment);
112 	else
113 		dpt = xe_bo_create_pin_map_at_aligned(xe, tile0, NULL,
114 						      dpt_size,  ~0ull,
115 						      ttm_bo_type_kernel,
116 						      XE_BO_FLAG_STOLEN |
117 						      XE_BO_FLAG_GGTT |
118 						      XE_BO_FLAG_PAGETABLE,
119 						      alignment);
120 	if (IS_ERR(dpt))
121 		dpt = xe_bo_create_pin_map_at_aligned(xe, tile0, NULL,
122 						      dpt_size,  ~0ull,
123 						      ttm_bo_type_kernel,
124 						      XE_BO_FLAG_SYSTEM |
125 						      XE_BO_FLAG_GGTT |
126 						      XE_BO_FLAG_PAGETABLE,
127 						      alignment);
128 	if (IS_ERR(dpt))
129 		return PTR_ERR(dpt);
130 
131 	if (view->type == I915_GTT_VIEW_NORMAL) {
132 		u32 x;
133 
134 		for (x = 0; x < size / XE_PAGE_SIZE; x++) {
135 			u64 pte = ggtt->pt_ops->pte_encode_bo(bo, x * XE_PAGE_SIZE,
136 							      xe->pat.idx[XE_CACHE_NONE]);
137 
138 			iosys_map_wr(&dpt->vmap, x * 8, u64, pte);
139 		}
140 	} else if (view->type == I915_GTT_VIEW_REMAPPED) {
141 		const struct intel_remapped_info *remap_info = &view->remapped;
142 		u32 i, dpt_ofs = 0;
143 
144 		for (i = 0; i < ARRAY_SIZE(remap_info->plane); i++)
145 			write_dpt_remapped(bo, &dpt->vmap, &dpt_ofs,
146 					   remap_info->plane[i].offset,
147 					   remap_info->plane[i].width,
148 					   remap_info->plane[i].height,
149 					   remap_info->plane[i].src_stride,
150 					   remap_info->plane[i].dst_stride);
151 
152 	} else {
153 		const struct intel_rotation_info *rot_info = &view->rotated;
154 		u32 i, dpt_ofs = 0;
155 
156 		for (i = 0; i < ARRAY_SIZE(rot_info->plane); i++)
157 			write_dpt_rotated(bo, &dpt->vmap, &dpt_ofs,
158 					  rot_info->plane[i].offset,
159 					  rot_info->plane[i].width,
160 					  rot_info->plane[i].height,
161 					  rot_info->plane[i].src_stride,
162 					  rot_info->plane[i].dst_stride);
163 	}
164 
165 	vma->dpt = dpt;
166 	vma->node = dpt->ggtt_node[tile0->id];
167 
168 	/* Ensure DPT writes are flushed */
169 	xe_device_l2_flush(xe);
170 	return 0;
171 }
172 
173 static void
write_ggtt_rotated(struct xe_bo * bo,struct xe_ggtt * ggtt,u32 * ggtt_ofs,u32 bo_ofs,u32 width,u32 height,u32 src_stride,u32 dst_stride)174 write_ggtt_rotated(struct xe_bo *bo, struct xe_ggtt *ggtt, u32 *ggtt_ofs, u32 bo_ofs,
175 		   u32 width, u32 height, u32 src_stride, u32 dst_stride)
176 {
177 	struct xe_device *xe = xe_bo_device(bo);
178 	u32 column, row;
179 
180 	for (column = 0; column < width; column++) {
181 		u32 src_idx = src_stride * (height - 1) + column + bo_ofs;
182 
183 		for (row = 0; row < height; row++) {
184 			u64 pte = ggtt->pt_ops->pte_encode_bo(bo, src_idx * XE_PAGE_SIZE,
185 							      xe->pat.idx[XE_CACHE_NONE]);
186 
187 			ggtt->pt_ops->ggtt_set_pte(ggtt, *ggtt_ofs, pte);
188 			*ggtt_ofs += XE_PAGE_SIZE;
189 			src_idx -= src_stride;
190 		}
191 
192 		/* The DE ignores the PTEs for the padding tiles */
193 		*ggtt_ofs += (dst_stride - height) * XE_PAGE_SIZE;
194 	}
195 }
196 
__xe_pin_fb_vma_ggtt(const struct intel_framebuffer * fb,const struct i915_gtt_view * view,struct i915_vma * vma,unsigned int alignment)197 static int __xe_pin_fb_vma_ggtt(const struct intel_framebuffer *fb,
198 				const struct i915_gtt_view *view,
199 				struct i915_vma *vma,
200 				unsigned int alignment)
201 {
202 	struct drm_gem_object *obj = intel_fb_bo(&fb->base);
203 	struct xe_bo *bo = gem_to_xe_bo(obj);
204 	struct xe_device *xe = to_xe_device(fb->base.dev);
205 	struct xe_ggtt *ggtt = xe_device_get_root_tile(xe)->mem.ggtt;
206 	u32 align;
207 	int ret;
208 
209 	/* TODO: Consider sharing framebuffer mapping?
210 	 * embed i915_vma inside intel_framebuffer
211 	 */
212 	xe_pm_runtime_get_noresume(tile_to_xe(ggtt->tile));
213 	ret = mutex_lock_interruptible(&ggtt->lock);
214 	if (ret)
215 		goto out;
216 
217 	align = XE_PAGE_SIZE;
218 	if (xe_bo_is_vram(bo) && ggtt->flags & XE_GGTT_FLAGS_64K)
219 		align = max_t(u32, align, SZ_64K);
220 
221 	if (bo->ggtt_node[ggtt->tile->id] && view->type == I915_GTT_VIEW_NORMAL) {
222 		vma->node = bo->ggtt_node[ggtt->tile->id];
223 	} else if (view->type == I915_GTT_VIEW_NORMAL) {
224 		u32 x, size = bo->ttm.base.size;
225 
226 		vma->node = xe_ggtt_node_init(ggtt);
227 		if (IS_ERR(vma->node)) {
228 			ret = PTR_ERR(vma->node);
229 			goto out_unlock;
230 		}
231 
232 		ret = xe_ggtt_node_insert_locked(vma->node, size, align, 0);
233 		if (ret) {
234 			xe_ggtt_node_fini(vma->node);
235 			goto out_unlock;
236 		}
237 
238 		for (x = 0; x < size; x += XE_PAGE_SIZE) {
239 			u64 pte = ggtt->pt_ops->pte_encode_bo(bo, x,
240 							      xe->pat.idx[XE_CACHE_NONE]);
241 
242 			ggtt->pt_ops->ggtt_set_pte(ggtt, vma->node->base.start + x, pte);
243 		}
244 	} else {
245 		u32 i, ggtt_ofs;
246 		const struct intel_rotation_info *rot_info = &view->rotated;
247 
248 		/* display seems to use tiles instead of bytes here, so convert it back.. */
249 		u32 size = intel_rotation_info_size(rot_info) * XE_PAGE_SIZE;
250 
251 		vma->node = xe_ggtt_node_init(ggtt);
252 		if (IS_ERR(vma->node)) {
253 			ret = PTR_ERR(vma->node);
254 			goto out_unlock;
255 		}
256 
257 		ret = xe_ggtt_node_insert_locked(vma->node, size, align, 0);
258 		if (ret) {
259 			xe_ggtt_node_fini(vma->node);
260 			goto out_unlock;
261 		}
262 
263 		ggtt_ofs = vma->node->base.start;
264 
265 		for (i = 0; i < ARRAY_SIZE(rot_info->plane); i++)
266 			write_ggtt_rotated(bo, ggtt, &ggtt_ofs,
267 					   rot_info->plane[i].offset,
268 					   rot_info->plane[i].width,
269 					   rot_info->plane[i].height,
270 					   rot_info->plane[i].src_stride,
271 					   rot_info->plane[i].dst_stride);
272 	}
273 
274 out_unlock:
275 	mutex_unlock(&ggtt->lock);
276 out:
277 	xe_pm_runtime_put(tile_to_xe(ggtt->tile));
278 	return ret;
279 }
280 
__xe_pin_fb_vma(const struct intel_framebuffer * fb,const struct i915_gtt_view * view,unsigned int alignment)281 static struct i915_vma *__xe_pin_fb_vma(const struct intel_framebuffer *fb,
282 					const struct i915_gtt_view *view,
283 					unsigned int alignment)
284 {
285 	struct drm_device *dev = fb->base.dev;
286 	struct xe_device *xe = to_xe_device(dev);
287 	struct i915_vma *vma = kzalloc(sizeof(*vma), GFP_KERNEL);
288 	struct drm_gem_object *obj = intel_fb_bo(&fb->base);
289 	struct xe_bo *bo = gem_to_xe_bo(obj);
290 	int ret;
291 
292 	if (!vma)
293 		return ERR_PTR(-ENODEV);
294 
295 	refcount_set(&vma->ref, 1);
296 	if (IS_DGFX(to_xe_device(bo->ttm.base.dev)) &&
297 	    intel_fb_rc_ccs_cc_plane(&fb->base) >= 0 &&
298 	    !(bo->flags & XE_BO_FLAG_NEEDS_CPU_ACCESS)) {
299 		struct xe_tile *tile = xe_device_get_root_tile(xe);
300 
301 		/*
302 		 * If we need to able to access the clear-color value stored in
303 		 * the buffer, then we require that such buffers are also CPU
304 		 * accessible.  This is important on small-bar systems where
305 		 * only some subset of VRAM is CPU accessible.
306 		 */
307 		if (tile->mem.vram.io_size < tile->mem.vram.usable_size) {
308 			ret = -EINVAL;
309 			goto err;
310 		}
311 	}
312 
313 	/*
314 	 * Pin the framebuffer, we can't use xe_bo_(un)pin functions as the
315 	 * assumptions are incorrect for framebuffers
316 	 */
317 	ret = ttm_bo_reserve(&bo->ttm, false, false, NULL);
318 	if (ret)
319 		goto err;
320 
321 	if (IS_DGFX(xe))
322 		ret = xe_bo_migrate(bo, XE_PL_VRAM0);
323 	else
324 		ret = xe_bo_validate(bo, NULL, true);
325 	if (!ret)
326 		ttm_bo_pin(&bo->ttm);
327 	ttm_bo_unreserve(&bo->ttm);
328 	if (ret)
329 		goto err;
330 
331 	vma->bo = bo;
332 	if (intel_fb_uses_dpt(&fb->base))
333 		ret = __xe_pin_fb_vma_dpt(fb, view, vma, alignment);
334 	else
335 		ret = __xe_pin_fb_vma_ggtt(fb, view, vma,  alignment);
336 	if (ret)
337 		goto err_unpin;
338 
339 	return vma;
340 
341 err_unpin:
342 	ttm_bo_reserve(&bo->ttm, false, false, NULL);
343 	ttm_bo_unpin(&bo->ttm);
344 	ttm_bo_unreserve(&bo->ttm);
345 err:
346 	kfree(vma);
347 	return ERR_PTR(ret);
348 }
349 
__xe_unpin_fb_vma(struct i915_vma * vma)350 static void __xe_unpin_fb_vma(struct i915_vma *vma)
351 {
352 	u8 tile_id = vma->node->ggtt->tile->id;
353 
354 	if (!refcount_dec_and_test(&vma->ref))
355 		return;
356 
357 	if (vma->dpt)
358 		xe_bo_unpin_map_no_vm(vma->dpt);
359 	else if (!xe_ggtt_node_allocated(vma->bo->ggtt_node[tile_id]) ||
360 		 vma->bo->ggtt_node[tile_id]->base.start != vma->node->base.start)
361 		xe_ggtt_node_remove(vma->node, false);
362 
363 	ttm_bo_reserve(&vma->bo->ttm, false, false, NULL);
364 	ttm_bo_unpin(&vma->bo->ttm);
365 	ttm_bo_unreserve(&vma->bo->ttm);
366 	kfree(vma);
367 }
368 
369 struct i915_vma *
intel_fb_pin_to_ggtt(const struct drm_framebuffer * fb,const struct i915_gtt_view * view,unsigned int alignment,unsigned int phys_alignment,unsigned int vtd_guard,bool uses_fence,unsigned long * out_flags)370 intel_fb_pin_to_ggtt(const struct drm_framebuffer *fb,
371 		     const struct i915_gtt_view *view,
372 		     unsigned int alignment,
373 		     unsigned int phys_alignment,
374 		     unsigned int vtd_guard,
375 		     bool uses_fence,
376 		     unsigned long *out_flags)
377 {
378 	*out_flags = 0;
379 
380 	return __xe_pin_fb_vma(to_intel_framebuffer(fb), view, phys_alignment);
381 }
382 
intel_fb_unpin_vma(struct i915_vma * vma,unsigned long flags)383 void intel_fb_unpin_vma(struct i915_vma *vma, unsigned long flags)
384 {
385 	__xe_unpin_fb_vma(vma);
386 }
387 
reuse_vma(struct intel_plane_state * new_plane_state,const struct intel_plane_state * old_plane_state)388 static bool reuse_vma(struct intel_plane_state *new_plane_state,
389 		      const struct intel_plane_state *old_plane_state)
390 {
391 	struct intel_framebuffer *fb = to_intel_framebuffer(new_plane_state->hw.fb);
392 	struct xe_device *xe = to_xe_device(fb->base.dev);
393 	struct i915_vma *vma;
394 
395 	if (old_plane_state->hw.fb == new_plane_state->hw.fb &&
396 	    !memcmp(&old_plane_state->view.gtt,
397 		    &new_plane_state->view.gtt,
398 		    sizeof(new_plane_state->view.gtt))) {
399 		vma = old_plane_state->ggtt_vma;
400 		goto found;
401 	}
402 
403 	if (fb == intel_fbdev_framebuffer(xe->display.fbdev.fbdev)) {
404 		vma = intel_fbdev_vma_pointer(xe->display.fbdev.fbdev);
405 		if (vma)
406 			goto found;
407 	}
408 
409 	return false;
410 
411 found:
412 	refcount_inc(&vma->ref);
413 	new_plane_state->ggtt_vma = vma;
414 	return true;
415 }
416 
intel_plane_pin_fb(struct intel_plane_state * new_plane_state,const struct intel_plane_state * old_plane_state)417 int intel_plane_pin_fb(struct intel_plane_state *new_plane_state,
418 		       const struct intel_plane_state *old_plane_state)
419 {
420 	struct drm_framebuffer *fb = new_plane_state->hw.fb;
421 	struct drm_gem_object *obj = intel_fb_bo(fb);
422 	struct xe_bo *bo = gem_to_xe_bo(obj);
423 	struct i915_vma *vma;
424 	struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
425 	struct intel_plane *plane = to_intel_plane(new_plane_state->uapi.plane);
426 	unsigned int alignment = plane->min_alignment(plane, fb, 0);
427 
428 	if (reuse_vma(new_plane_state, old_plane_state))
429 		return 0;
430 
431 	/* We reject creating !SCANOUT fb's, so this is weird.. */
432 	drm_WARN_ON(bo->ttm.base.dev, !(bo->flags & XE_BO_FLAG_SCANOUT));
433 
434 	vma = __xe_pin_fb_vma(intel_fb, &new_plane_state->view.gtt, alignment);
435 
436 	if (IS_ERR(vma))
437 		return PTR_ERR(vma);
438 
439 	new_plane_state->ggtt_vma = vma;
440 	return 0;
441 }
442 
intel_plane_unpin_fb(struct intel_plane_state * old_plane_state)443 void intel_plane_unpin_fb(struct intel_plane_state *old_plane_state)
444 {
445 	__xe_unpin_fb_vma(old_plane_state->ggtt_vma);
446 	old_plane_state->ggtt_vma = NULL;
447 }
448 
449 /*
450  * For Xe introduce dummy intel_dpt_create which just return NULL,
451  * intel_dpt_destroy which does nothing, and fake intel_dpt_ofsset returning 0;
452  */
intel_dpt_create(struct intel_framebuffer * fb)453 struct i915_address_space *intel_dpt_create(struct intel_framebuffer *fb)
454 {
455 	return NULL;
456 }
457 
intel_dpt_destroy(struct i915_address_space * vm)458 void intel_dpt_destroy(struct i915_address_space *vm)
459 {
460 	return;
461 }
462 
intel_dpt_offset(struct i915_vma * dpt_vma)463 u64 intel_dpt_offset(struct i915_vma *dpt_vma)
464 {
465 	return 0;
466 }
467