xref: /linux/drivers/gpu/drm/xe/display/xe_fb_pin.c (revision 917b10d90990fd2138b5dbc2d22cfa428c070ade)
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_core.h"
10 #include "intel_display_types.h"
11 #include "intel_dpt.h"
12 #include "intel_fb.h"
13 #include "intel_fb_pin.h"
14 #include "intel_fbdev.h"
15 #include "xe_bo.h"
16 #include "xe_device.h"
17 #include "xe_ggtt.h"
18 #include "xe_pm.h"
19 
20 static void
21 write_dpt_rotated(struct xe_bo *bo, struct iosys_map *map, u32 *dpt_ofs, u32 bo_ofs,
22 		  u32 width, u32 height, u32 src_stride, u32 dst_stride)
23 {
24 	struct xe_device *xe = xe_bo_device(bo);
25 	struct xe_ggtt *ggtt = xe_device_get_root_tile(xe)->mem.ggtt;
26 	u32 column, row;
27 	u64 pte = xe_ggtt_encode_pte_flags(ggtt, bo, xe->pat.idx[XE_CACHE_NONE]);
28 
29 	/* TODO: Maybe rewrite so we can traverse the bo addresses sequentially,
30 	 * by writing dpt/ggtt in a different order?
31 	 */
32 
33 	for (column = 0; column < width; column++) {
34 		u32 src_idx = src_stride * (height - 1) + column + bo_ofs;
35 
36 		for (row = 0; row < height; row++) {
37 			u64 addr = xe_bo_addr(bo, src_idx * XE_PAGE_SIZE, XE_PAGE_SIZE);
38 
39 			iosys_map_wr(map, *dpt_ofs, u64, pte | addr);
40 			*dpt_ofs += 8;
41 			src_idx -= src_stride;
42 		}
43 
44 		/* The DE ignores the PTEs for the padding tiles */
45 		*dpt_ofs += (dst_stride - height) * 8;
46 	}
47 
48 	/* Align to next page */
49 	*dpt_ofs = ALIGN(*dpt_ofs, 4096);
50 }
51 
52 static void
53 write_dpt_remapped(struct xe_bo *bo, struct iosys_map *map, u32 *dpt_ofs,
54 		   u32 bo_ofs, u32 width, u32 height, u32 src_stride,
55 		   u32 dst_stride)
56 {
57 	struct xe_device *xe = xe_bo_device(bo);
58 	struct xe_ggtt *ggtt = xe_device_get_root_tile(xe)->mem.ggtt;
59 	u32 column, row;
60 	u64 pte = xe_ggtt_encode_pte_flags(ggtt, bo, xe->pat.idx[XE_CACHE_NONE]);
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 			u64 addr = xe_bo_addr(bo, src_idx * XE_PAGE_SIZE, XE_PAGE_SIZE);
67 			iosys_map_wr(map, *dpt_ofs, u64, pte | addr);
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 			       unsigned int alignment)
85 {
86 	struct xe_device *xe = to_xe_device(fb->base.dev);
87 	struct xe_tile *tile0 = xe_device_get_root_tile(xe);
88 	struct xe_ggtt *ggtt = tile0->mem.ggtt;
89 	struct drm_gem_object *obj = intel_fb_bo(&fb->base);
90 	struct xe_bo *bo = gem_to_xe_bo(obj), *dpt;
91 	u32 dpt_size, size = bo->ttm.base.size;
92 
93 	if (view->type == I915_GTT_VIEW_NORMAL)
94 		dpt_size = ALIGN(size / XE_PAGE_SIZE * 8, XE_PAGE_SIZE);
95 	else if (view->type == I915_GTT_VIEW_REMAPPED)
96 		dpt_size = ALIGN(intel_remapped_info_size(&fb->remapped_view.gtt.remapped) * 8,
97 				 XE_PAGE_SIZE);
98 	else
99 		/* display uses 4K tiles instead of bytes here, convert to entries.. */
100 		dpt_size = ALIGN(intel_rotation_info_size(&view->rotated) * 8,
101 				 XE_PAGE_SIZE);
102 
103 	if (IS_DGFX(xe))
104 		dpt = xe_bo_create_pin_map_at_aligned(xe, tile0, NULL,
105 						      dpt_size, ~0ull,
106 						      ttm_bo_type_kernel,
107 						      XE_BO_FLAG_VRAM0 |
108 						      XE_BO_FLAG_GGTT |
109 						      XE_BO_FLAG_PAGETABLE,
110 						      alignment);
111 	else
112 		dpt = xe_bo_create_pin_map_at_aligned(xe, tile0, NULL,
113 						      dpt_size,  ~0ull,
114 						      ttm_bo_type_kernel,
115 						      XE_BO_FLAG_STOLEN |
116 						      XE_BO_FLAG_GGTT |
117 						      XE_BO_FLAG_PAGETABLE,
118 						      alignment);
119 	if (IS_ERR(dpt))
120 		dpt = xe_bo_create_pin_map_at_aligned(xe, tile0, NULL,
121 						      dpt_size,  ~0ull,
122 						      ttm_bo_type_kernel,
123 						      XE_BO_FLAG_SYSTEM |
124 						      XE_BO_FLAG_GGTT |
125 						      XE_BO_FLAG_PAGETABLE,
126 						      alignment);
127 	if (IS_ERR(dpt))
128 		return PTR_ERR(dpt);
129 
130 	if (view->type == I915_GTT_VIEW_NORMAL) {
131 		u64 pte = xe_ggtt_encode_pte_flags(ggtt, bo, xe->pat.idx[XE_CACHE_NONE]);
132 		u32 x;
133 
134 		for (x = 0; x < size / XE_PAGE_SIZE; x++) {
135 			u64 addr = xe_bo_addr(bo, x * XE_PAGE_SIZE, XE_PAGE_SIZE);
136 
137 			iosys_map_wr(&dpt->vmap, x * 8, u64, pte | addr);
138 		}
139 	} else if (view->type == I915_GTT_VIEW_REMAPPED) {
140 		const struct intel_remapped_info *remap_info = &view->remapped;
141 		u32 i, dpt_ofs = 0;
142 
143 		for (i = 0; i < ARRAY_SIZE(remap_info->plane); i++)
144 			write_dpt_remapped(bo, &dpt->vmap, &dpt_ofs,
145 					   remap_info->plane[i].offset,
146 					   remap_info->plane[i].width,
147 					   remap_info->plane[i].height,
148 					   remap_info->plane[i].src_stride,
149 					   remap_info->plane[i].dst_stride);
150 
151 	} else {
152 		const struct intel_rotation_info *rot_info = &view->rotated;
153 		u32 i, dpt_ofs = 0;
154 
155 		for (i = 0; i < ARRAY_SIZE(rot_info->plane); i++)
156 			write_dpt_rotated(bo, &dpt->vmap, &dpt_ofs,
157 					  rot_info->plane[i].offset,
158 					  rot_info->plane[i].width,
159 					  rot_info->plane[i].height,
160 					  rot_info->plane[i].src_stride,
161 					  rot_info->plane[i].dst_stride);
162 	}
163 
164 	vma->dpt = dpt;
165 	vma->node = dpt->ggtt_node[tile0->id];
166 	return 0;
167 }
168 
169 static void
170 write_ggtt_rotated(struct xe_bo *bo, struct xe_ggtt *ggtt, u32 *ggtt_ofs, u32 bo_ofs,
171 		   u32 width, u32 height, u32 src_stride, u32 dst_stride)
172 {
173 	struct xe_device *xe = xe_bo_device(bo);
174 	u32 column, row;
175 	u64 pte = ggtt->pt_ops->pte_encode_flags(bo, xe->pat.idx[XE_CACHE_NONE]);
176 
177 	for (column = 0; column < width; column++) {
178 		u32 src_idx = src_stride * (height - 1) + column + bo_ofs;
179 
180 		for (row = 0; row < height; row++) {
181 			u64 addr = xe_bo_addr(bo, src_idx * XE_PAGE_SIZE, XE_PAGE_SIZE);
182 
183 			ggtt->pt_ops->ggtt_set_pte(ggtt, *ggtt_ofs, pte | addr);
184 			*ggtt_ofs += XE_PAGE_SIZE;
185 			src_idx -= src_stride;
186 		}
187 
188 		/* The DE ignores the PTEs for the padding tiles */
189 		*ggtt_ofs += (dst_stride - height) * XE_PAGE_SIZE;
190 	}
191 }
192 
193 static int __xe_pin_fb_vma_ggtt(const struct intel_framebuffer *fb,
194 				const struct i915_gtt_view *view,
195 				struct i915_vma *vma,
196 				unsigned int alignment)
197 {
198 	struct drm_gem_object *obj = intel_fb_bo(&fb->base);
199 	struct xe_bo *bo = gem_to_xe_bo(obj);
200 	struct xe_device *xe = to_xe_device(fb->base.dev);
201 	struct xe_tile *tile0 = xe_device_get_root_tile(xe);
202 	struct xe_ggtt *ggtt = tile0->mem.ggtt;
203 	u32 align;
204 	int ret;
205 
206 	/* TODO: Consider sharing framebuffer mapping?
207 	 * embed i915_vma inside intel_framebuffer
208 	 */
209 	xe_pm_runtime_get_noresume(xe);
210 	ret = mutex_lock_interruptible(&ggtt->lock);
211 	if (ret)
212 		goto out;
213 
214 	align = XE_PAGE_SIZE;
215 	if (xe_bo_is_vram(bo) && ggtt->flags & XE_GGTT_FLAGS_64K)
216 		align = max_t(u32, align, SZ_64K);
217 
218 	if (bo->ggtt_node[tile0->id] && view->type == I915_GTT_VIEW_NORMAL) {
219 		vma->node = bo->ggtt_node[tile0->id];
220 	} else if (view->type == I915_GTT_VIEW_NORMAL) {
221 		vma->node = xe_ggtt_node_init(ggtt);
222 		if (IS_ERR(vma->node)) {
223 			ret = PTR_ERR(vma->node);
224 			goto out_unlock;
225 		}
226 
227 		ret = xe_ggtt_node_insert_locked(vma->node, bo->size, align, 0);
228 		if (ret) {
229 			xe_ggtt_node_fini(vma->node);
230 			goto out_unlock;
231 		}
232 
233 		xe_ggtt_map_bo(ggtt, vma->node, bo, xe->pat.idx[XE_CACHE_NONE]);
234 	} else {
235 		u32 i, ggtt_ofs;
236 		const struct intel_rotation_info *rot_info = &view->rotated;
237 
238 		/* display seems to use tiles instead of bytes here, so convert it back.. */
239 		u32 size = intel_rotation_info_size(rot_info) * XE_PAGE_SIZE;
240 
241 		vma->node = xe_ggtt_node_init(ggtt);
242 		if (IS_ERR(vma->node)) {
243 			ret = PTR_ERR(vma->node);
244 			goto out_unlock;
245 		}
246 
247 		ret = xe_ggtt_node_insert_locked(vma->node, size, align, 0);
248 		if (ret) {
249 			xe_ggtt_node_fini(vma->node);
250 			goto out_unlock;
251 		}
252 
253 		ggtt_ofs = vma->node->base.start;
254 
255 		for (i = 0; i < ARRAY_SIZE(rot_info->plane); i++)
256 			write_ggtt_rotated(bo, ggtt, &ggtt_ofs,
257 					   rot_info->plane[i].offset,
258 					   rot_info->plane[i].width,
259 					   rot_info->plane[i].height,
260 					   rot_info->plane[i].src_stride,
261 					   rot_info->plane[i].dst_stride);
262 	}
263 
264 out_unlock:
265 	mutex_unlock(&ggtt->lock);
266 out:
267 	xe_pm_runtime_put(xe);
268 	return ret;
269 }
270 
271 static struct i915_vma *__xe_pin_fb_vma(const struct intel_framebuffer *fb,
272 					const struct i915_gtt_view *view,
273 					unsigned int alignment)
274 {
275 	struct drm_device *dev = fb->base.dev;
276 	struct xe_device *xe = to_xe_device(dev);
277 	struct i915_vma *vma = kzalloc(sizeof(*vma), GFP_KERNEL);
278 	struct drm_gem_object *obj = intel_fb_bo(&fb->base);
279 	struct xe_bo *bo = gem_to_xe_bo(obj);
280 	int ret;
281 
282 	if (!vma)
283 		return ERR_PTR(-ENODEV);
284 
285 	refcount_set(&vma->ref, 1);
286 	if (IS_DGFX(to_xe_device(bo->ttm.base.dev)) &&
287 	    intel_fb_rc_ccs_cc_plane(&fb->base) >= 0 &&
288 	    !(bo->flags & XE_BO_FLAG_NEEDS_CPU_ACCESS)) {
289 		struct xe_tile *tile = xe_device_get_root_tile(xe);
290 
291 		/*
292 		 * If we need to able to access the clear-color value stored in
293 		 * the buffer, then we require that such buffers are also CPU
294 		 * accessible.  This is important on small-bar systems where
295 		 * only some subset of VRAM is CPU accessible.
296 		 */
297 		if (tile->mem.vram.io_size < tile->mem.vram.usable_size) {
298 			ret = -EINVAL;
299 			goto err;
300 		}
301 	}
302 
303 	/*
304 	 * Pin the framebuffer, we can't use xe_bo_(un)pin functions as the
305 	 * assumptions are incorrect for framebuffers
306 	 */
307 	ret = ttm_bo_reserve(&bo->ttm, false, false, NULL);
308 	if (ret)
309 		goto err;
310 
311 	if (IS_DGFX(xe))
312 		ret = xe_bo_migrate(bo, XE_PL_VRAM0);
313 	else
314 		ret = xe_bo_validate(bo, NULL, true);
315 	if (!ret)
316 		ttm_bo_pin(&bo->ttm);
317 	ttm_bo_unreserve(&bo->ttm);
318 	if (ret)
319 		goto err;
320 
321 	vma->bo = bo;
322 	if (intel_fb_uses_dpt(&fb->base))
323 		ret = __xe_pin_fb_vma_dpt(fb, view, vma, alignment);
324 	else
325 		ret = __xe_pin_fb_vma_ggtt(fb, view, vma,  alignment);
326 	if (ret)
327 		goto err_unpin;
328 
329 	/* Ensure DPT writes are flushed */
330 	xe_device_l2_flush(xe);
331 	return vma;
332 
333 err_unpin:
334 	ttm_bo_reserve(&bo->ttm, false, false, NULL);
335 	ttm_bo_unpin(&bo->ttm);
336 	ttm_bo_unreserve(&bo->ttm);
337 err:
338 	kfree(vma);
339 	return ERR_PTR(ret);
340 }
341 
342 static void __xe_unpin_fb_vma(struct i915_vma *vma)
343 {
344 	u8 tile_id = xe_device_get_root_tile(xe_bo_device(vma->bo))->id;
345 
346 	if (!refcount_dec_and_test(&vma->ref))
347 		return;
348 
349 	if (vma->dpt)
350 		xe_bo_unpin_map_no_vm(vma->dpt);
351 	else if (!xe_ggtt_node_allocated(vma->bo->ggtt_node[tile_id]) ||
352 		 vma->bo->ggtt_node[tile_id]->base.start != vma->node->base.start)
353 		xe_ggtt_node_remove(vma->node, false);
354 
355 	ttm_bo_reserve(&vma->bo->ttm, false, false, NULL);
356 	ttm_bo_unpin(&vma->bo->ttm);
357 	ttm_bo_unreserve(&vma->bo->ttm);
358 	kfree(vma);
359 }
360 
361 struct i915_vma *
362 intel_fb_pin_to_ggtt(const struct drm_framebuffer *fb,
363 		     const struct i915_gtt_view *view,
364 		     unsigned int alignment,
365 		     unsigned int phys_alignment,
366 		     unsigned int vtd_guard,
367 		     bool uses_fence,
368 		     unsigned long *out_flags)
369 {
370 	*out_flags = 0;
371 
372 	return __xe_pin_fb_vma(to_intel_framebuffer(fb), view, phys_alignment);
373 }
374 
375 void intel_fb_unpin_vma(struct i915_vma *vma, unsigned long flags)
376 {
377 	__xe_unpin_fb_vma(vma);
378 }
379 
380 static bool reuse_vma(struct intel_plane_state *new_plane_state,
381 		      const struct intel_plane_state *old_plane_state)
382 {
383 	struct intel_framebuffer *fb = to_intel_framebuffer(new_plane_state->hw.fb);
384 	struct xe_device *xe = to_xe_device(fb->base.dev);
385 	struct intel_display *display = xe->display;
386 	struct i915_vma *vma;
387 
388 	if (old_plane_state->hw.fb == new_plane_state->hw.fb &&
389 	    !memcmp(&old_plane_state->view.gtt,
390 		    &new_plane_state->view.gtt,
391 		    sizeof(new_plane_state->view.gtt))) {
392 		vma = old_plane_state->ggtt_vma;
393 		goto found;
394 	}
395 
396 	if (fb == intel_fbdev_framebuffer(display->fbdev.fbdev)) {
397 		vma = intel_fbdev_vma_pointer(display->fbdev.fbdev);
398 		if (vma)
399 			goto found;
400 	}
401 
402 	return false;
403 
404 found:
405 	refcount_inc(&vma->ref);
406 	new_plane_state->ggtt_vma = vma;
407 	return true;
408 }
409 
410 int intel_plane_pin_fb(struct intel_plane_state *new_plane_state,
411 		       const struct intel_plane_state *old_plane_state)
412 {
413 	struct drm_framebuffer *fb = new_plane_state->hw.fb;
414 	struct drm_gem_object *obj = intel_fb_bo(fb);
415 	struct xe_bo *bo = gem_to_xe_bo(obj);
416 	struct i915_vma *vma;
417 	struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
418 	struct intel_plane *plane = to_intel_plane(new_plane_state->uapi.plane);
419 	unsigned int alignment = plane->min_alignment(plane, fb, 0);
420 
421 	if (reuse_vma(new_plane_state, old_plane_state))
422 		return 0;
423 
424 	/* We reject creating !SCANOUT fb's, so this is weird.. */
425 	drm_WARN_ON(bo->ttm.base.dev, !(bo->flags & XE_BO_FLAG_SCANOUT));
426 
427 	vma = __xe_pin_fb_vma(intel_fb, &new_plane_state->view.gtt, alignment);
428 
429 	if (IS_ERR(vma))
430 		return PTR_ERR(vma);
431 
432 	new_plane_state->ggtt_vma = vma;
433 	return 0;
434 }
435 
436 void intel_plane_unpin_fb(struct intel_plane_state *old_plane_state)
437 {
438 	__xe_unpin_fb_vma(old_plane_state->ggtt_vma);
439 	old_plane_state->ggtt_vma = NULL;
440 }
441 
442 /*
443  * For Xe introduce dummy intel_dpt_create which just return NULL,
444  * intel_dpt_destroy which does nothing, and fake intel_dpt_ofsset returning 0;
445  */
446 struct i915_address_space *intel_dpt_create(struct intel_framebuffer *fb)
447 {
448 	return NULL;
449 }
450 
451 void intel_dpt_destroy(struct i915_address_space *vm)
452 {
453 	return;
454 }
455 
456 u64 intel_dpt_offset(struct i915_vma *dpt_vma)
457 {
458 	return 0;
459 }
460 
461 void intel_fb_get_map(struct i915_vma *vma, struct iosys_map *map)
462 {
463 	*map = vma->bo->vmap;
464 }
465