xref: /netbsd-src/sys/external/bsd/drm2/dist/drm/i915/i915_buddy.c (revision e555192603a4ddcdb47375a0c00636be87870e26)
1 /*	$NetBSD: i915_buddy.c,v 1.5 2021/12/19 11:13:36 riastradh Exp $	*/
2 
3 // SPDX-License-Identifier: MIT
4 /*
5  * Copyright © 2019 Intel Corporation
6  */
7 
8 #include <sys/cdefs.h>
9 __KERNEL_RCSID(0, "$NetBSD: i915_buddy.c,v 1.5 2021/12/19 11:13:36 riastradh Exp $");
10 
11 #include <linux/err.h>
12 #include <linux/kmemleak.h>
13 #include <linux/slab.h>
14 
15 #include "i915_buddy.h"
16 
17 #include "i915_gem.h"
18 #include "i915_globals.h"
19 #include "i915_utils.h"
20 
21 static struct i915_global_block {
22 	struct i915_global base;
23 	struct kmem_cache *slab_blocks;
24 } global;
25 
i915_global_buddy_shrink(void)26 static void i915_global_buddy_shrink(void)
27 {
28 	kmem_cache_shrink(global.slab_blocks);
29 }
30 
i915_global_buddy_exit(void)31 static void i915_global_buddy_exit(void)
32 {
33 	kmem_cache_destroy(global.slab_blocks);
34 }
35 
36 static struct i915_global_block global = { {
37 	.shrink = i915_global_buddy_shrink,
38 	.exit = i915_global_buddy_exit,
39 } };
40 
41 #ifdef __NetBSD__
42 #define	__init	/* called from i915_module.c */
43 #endif
44 
i915_global_buddy_init(void)45 int __init i915_global_buddy_init(void)
46 {
47 	global.slab_blocks = KMEM_CACHE(i915_buddy_block, SLAB_HWCACHE_ALIGN);
48 	if (!global.slab_blocks)
49 		return -ENOMEM;
50 
51 	i915_global_register(&global.base);
52 	return 0;
53 }
54 
i915_block_alloc(struct i915_buddy_block * parent,unsigned int order,u64 offset)55 static struct i915_buddy_block *i915_block_alloc(struct i915_buddy_block *parent,
56 						 unsigned int order,
57 						 u64 offset)
58 {
59 	struct i915_buddy_block *block;
60 
61 	block = kmem_cache_zalloc(global.slab_blocks, GFP_KERNEL);
62 	if (!block)
63 		return NULL;
64 
65 	block->header = offset;
66 	block->header |= order;
67 	block->parent = parent;
68 
69 	return block;
70 }
71 
i915_block_free(struct i915_buddy_block * block)72 static void i915_block_free(struct i915_buddy_block *block)
73 {
74 	kmem_cache_free(global.slab_blocks, block);
75 }
76 
mark_allocated(struct i915_buddy_block * block)77 static void mark_allocated(struct i915_buddy_block *block)
78 {
79 	block->header &= ~I915_BUDDY_HEADER_STATE;
80 	block->header |= I915_BUDDY_ALLOCATED;
81 
82 	list_del(&block->link);
83 }
84 
mark_free(struct i915_buddy_mm * mm,struct i915_buddy_block * block)85 static void mark_free(struct i915_buddy_mm *mm,
86 		      struct i915_buddy_block *block)
87 {
88 	block->header &= ~I915_BUDDY_HEADER_STATE;
89 	block->header |= I915_BUDDY_FREE;
90 
91 	list_add(&block->link,
92 		 &mm->free_list[i915_buddy_block_order(block)]);
93 }
94 
mark_split(struct i915_buddy_block * block)95 static void mark_split(struct i915_buddy_block *block)
96 {
97 	block->header &= ~I915_BUDDY_HEADER_STATE;
98 	block->header |= I915_BUDDY_SPLIT;
99 
100 	list_del(&block->link);
101 }
102 
i915_buddy_init(struct i915_buddy_mm * mm,u64 size,u64 chunk_size)103 int i915_buddy_init(struct i915_buddy_mm *mm, u64 size, u64 chunk_size)
104 {
105 	unsigned int i;
106 	u64 offset;
107 
108 	if (size < chunk_size)
109 		return -EINVAL;
110 
111 	if (chunk_size < PAGE_SIZE)
112 		return -EINVAL;
113 
114 	if (!is_power_of_2(chunk_size))
115 		return -EINVAL;
116 
117 	size = round_down(size, chunk_size);
118 
119 	mm->size = size;
120 	mm->chunk_size = chunk_size;
121 	mm->max_order = ilog2(size) - ilog2(chunk_size);
122 
123 	GEM_BUG_ON(mm->max_order > I915_BUDDY_MAX_ORDER);
124 
125 	mm->free_list = kmalloc_array(mm->max_order + 1,
126 				      sizeof(struct list_head),
127 				      GFP_KERNEL);
128 	if (!mm->free_list)
129 		return -ENOMEM;
130 
131 	for (i = 0; i <= mm->max_order; ++i)
132 		INIT_LIST_HEAD(&mm->free_list[i]);
133 
134 	mm->n_roots = hweight64(size);
135 
136 	mm->roots = kmalloc_array(mm->n_roots,
137 				  sizeof(struct i915_buddy_block *),
138 				  GFP_KERNEL);
139 	if (!mm->roots)
140 		goto out_free_list;
141 
142 	offset = 0;
143 	i = 0;
144 
145 	/*
146 	 * Split into power-of-two blocks, in case we are given a size that is
147 	 * not itself a power-of-two.
148 	 */
149 	do {
150 		struct i915_buddy_block *root;
151 		unsigned int order;
152 		u64 root_size;
153 
154 		root_size = rounddown_pow_of_two(size);
155 		order = ilog2(root_size) - ilog2(chunk_size);
156 
157 		root = i915_block_alloc(NULL, order, offset);
158 		if (!root)
159 			goto out_free_roots;
160 
161 		mark_free(mm, root);
162 
163 		GEM_BUG_ON(i > mm->max_order);
164 		GEM_BUG_ON(i915_buddy_block_size(mm, root) < chunk_size);
165 
166 		mm->roots[i] = root;
167 
168 		offset += root_size;
169 		size -= root_size;
170 		i++;
171 	} while (size);
172 
173 	return 0;
174 
175 out_free_roots:
176 	while (i--)
177 		i915_block_free(mm->roots[i]);
178 	kfree(mm->roots);
179 out_free_list:
180 	kfree(mm->free_list);
181 	return -ENOMEM;
182 }
183 
i915_buddy_fini(struct i915_buddy_mm * mm)184 void i915_buddy_fini(struct i915_buddy_mm *mm)
185 {
186 	int i;
187 
188 	for (i = 0; i < mm->n_roots; ++i) {
189 		GEM_WARN_ON(!i915_buddy_block_is_free(mm->roots[i]));
190 		i915_block_free(mm->roots[i]);
191 	}
192 
193 	kfree(mm->roots);
194 	kfree(mm->free_list);
195 }
196 
split_block(struct i915_buddy_mm * mm,struct i915_buddy_block * block)197 static int split_block(struct i915_buddy_mm *mm,
198 		       struct i915_buddy_block *block)
199 {
200 	unsigned int block_order = i915_buddy_block_order(block) - 1;
201 	u64 offset = i915_buddy_block_offset(block);
202 
203 	GEM_BUG_ON(!i915_buddy_block_is_free(block));
204 	GEM_BUG_ON(!i915_buddy_block_order(block));
205 
206 	block->left = i915_block_alloc(block, block_order, offset);
207 	if (!block->left)
208 		return -ENOMEM;
209 
210 	block->right = i915_block_alloc(block, block_order,
211 					offset + (mm->chunk_size << block_order));
212 	if (!block->right) {
213 		i915_block_free(block->left);
214 		return -ENOMEM;
215 	}
216 
217 	mark_free(mm, block->left);
218 	mark_free(mm, block->right);
219 
220 	mark_split(block);
221 
222 	return 0;
223 }
224 
225 static struct i915_buddy_block *
get_buddy(struct i915_buddy_block * block)226 get_buddy(struct i915_buddy_block *block)
227 {
228 	struct i915_buddy_block *parent;
229 
230 	parent = block->parent;
231 	if (!parent)
232 		return NULL;
233 
234 	if (parent->left == block)
235 		return parent->right;
236 
237 	return parent->left;
238 }
239 
__i915_buddy_free(struct i915_buddy_mm * mm,struct i915_buddy_block * block)240 static void __i915_buddy_free(struct i915_buddy_mm *mm,
241 			      struct i915_buddy_block *block)
242 {
243 	struct i915_buddy_block *parent;
244 
245 	while ((parent = block->parent)) {
246 		struct i915_buddy_block *buddy;
247 
248 		buddy = get_buddy(block);
249 
250 		if (!i915_buddy_block_is_free(buddy))
251 			break;
252 
253 		list_del(&buddy->link);
254 
255 		i915_block_free(block);
256 		i915_block_free(buddy);
257 
258 		block = parent;
259 	}
260 
261 	mark_free(mm, block);
262 }
263 
i915_buddy_free(struct i915_buddy_mm * mm,struct i915_buddy_block * block)264 void i915_buddy_free(struct i915_buddy_mm *mm,
265 		     struct i915_buddy_block *block)
266 {
267 	GEM_BUG_ON(!i915_buddy_block_is_allocated(block));
268 	__i915_buddy_free(mm, block);
269 }
270 
i915_buddy_free_list(struct i915_buddy_mm * mm,struct list_head * objects)271 void i915_buddy_free_list(struct i915_buddy_mm *mm, struct list_head *objects)
272 {
273 	struct i915_buddy_block *block, *on;
274 
275 	list_for_each_entry_safe(block, on, objects, link) {
276 		i915_buddy_free(mm, block);
277 		cond_resched();
278 	}
279 	INIT_LIST_HEAD(objects);
280 }
281 
282 /*
283  * Allocate power-of-two block. The order value here translates to:
284  *
285  *   0 = 2^0 * mm->chunk_size
286  *   1 = 2^1 * mm->chunk_size
287  *   2 = 2^2 * mm->chunk_size
288  *   ...
289  */
290 struct i915_buddy_block *
i915_buddy_alloc(struct i915_buddy_mm * mm,unsigned int order)291 i915_buddy_alloc(struct i915_buddy_mm *mm, unsigned int order)
292 {
293 	struct i915_buddy_block *block = NULL;
294 	unsigned int i;
295 	int err;
296 
297 	for (i = order; i <= mm->max_order; ++i) {
298 		block = list_first_entry_or_null(&mm->free_list[i],
299 						 struct i915_buddy_block,
300 						 link);
301 		if (block)
302 			break;
303 	}
304 
305 	if (!block)
306 		return ERR_PTR(-ENOSPC);
307 
308 	GEM_BUG_ON(!i915_buddy_block_is_free(block));
309 
310 	while (i != order) {
311 		err = split_block(mm, block);
312 		if (unlikely(err))
313 			goto out_free;
314 
315 		/* Go low */
316 		block = block->left;
317 		i--;
318 	}
319 
320 	mark_allocated(block);
321 	kmemleak_update_trace(block);
322 	return block;
323 
324 out_free:
325 	__i915_buddy_free(mm, block);
326 	return ERR_PTR(err);
327 }
328 
overlaps(u64 s1,u64 e1,u64 s2,u64 e2)329 static inline bool overlaps(u64 s1, u64 e1, u64 s2, u64 e2)
330 {
331 	return s1 <= e2 && e1 >= s2;
332 }
333 
contains(u64 s1,u64 e1,u64 s2,u64 e2)334 static inline bool contains(u64 s1, u64 e1, u64 s2, u64 e2)
335 {
336 	return s1 <= s2 && e1 >= e2;
337 }
338 
339 /*
340  * Allocate range. Note that it's safe to chain together multiple alloc_ranges
341  * with the same blocks list.
342  *
343  * Intended for pre-allocating portions of the address space, for example to
344  * reserve a block for the initial framebuffer or similar, hence the expectation
345  * here is that i915_buddy_alloc() is still the main vehicle for
346  * allocations, so if that's not the case then the drm_mm range allocator is
347  * probably a much better fit, and so you should probably go use that instead.
348  */
i915_buddy_alloc_range(struct i915_buddy_mm * mm,struct list_head * blocks,u64 start,u64 size)349 int i915_buddy_alloc_range(struct i915_buddy_mm *mm,
350 			   struct list_head *blocks,
351 			   u64 start, u64 size)
352 {
353 	struct i915_buddy_block *block;
354 	struct i915_buddy_block *buddy;
355 	struct list_head allocated = LIST_HEAD_INIT(allocated);
356 	struct list_head dfs = LIST_HEAD_INIT(dfs);
357 	u64 end;
358 	int err;
359 	int i;
360 
361 	if (size < mm->chunk_size)
362 		return -EINVAL;
363 
364 	if (!IS_ALIGNED(size | start, mm->chunk_size))
365 		return -EINVAL;
366 
367 	if (range_overflows(start, size, mm->size))
368 		return -EINVAL;
369 
370 	for (i = 0; i < mm->n_roots; ++i)
371 		list_add_tail(&mm->roots[i]->tmp_link, &dfs);
372 
373 	end = start + size - 1;
374 
375 	do {
376 		u64 block_start;
377 		u64 block_end;
378 
379 		block = list_first_entry_or_null(&dfs,
380 						 struct i915_buddy_block,
381 						 tmp_link);
382 		if (!block)
383 			break;
384 
385 		list_del(&block->tmp_link);
386 
387 		block_start = i915_buddy_block_offset(block);
388 		block_end = block_start + i915_buddy_block_size(mm, block) - 1;
389 
390 		if (!overlaps(start, end, block_start, block_end))
391 			continue;
392 
393 		if (i915_buddy_block_is_allocated(block)) {
394 			err = -ENOSPC;
395 			goto err_free;
396 		}
397 
398 		if (contains(start, end, block_start, block_end)) {
399 			if (!i915_buddy_block_is_free(block)) {
400 				err = -ENOSPC;
401 				goto err_free;
402 			}
403 
404 			mark_allocated(block);
405 			list_add_tail(&block->link, &allocated);
406 			continue;
407 		}
408 
409 		if (!i915_buddy_block_is_split(block)) {
410 			err = split_block(mm, block);
411 			if (unlikely(err))
412 				goto err_undo;
413 		}
414 
415 		list_add(&block->right->tmp_link, &dfs);
416 		list_add(&block->left->tmp_link, &dfs);
417 	} while (1);
418 
419 	list_splice_tail(&allocated, blocks);
420 	return 0;
421 
422 err_undo:
423 	/*
424 	 * We really don't want to leave around a bunch of split blocks, since
425 	 * bigger is better, so make sure we merge everything back before we
426 	 * free the allocated blocks.
427 	 */
428 	buddy = get_buddy(block);
429 	if (buddy &&
430 	    (i915_buddy_block_is_free(block) &&
431 	     i915_buddy_block_is_free(buddy)))
432 		__i915_buddy_free(mm, block);
433 
434 err_free:
435 	i915_buddy_free_list(mm, &allocated);
436 	return err;
437 }
438 
439 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
440 #include "selftests/i915_buddy.c"
441 #endif
442