xref: /linux/drivers/gpu/drm/ttm/ttm_tt.c (revision 02680c23d7b3febe45ea3d4f9818c2b2dc89020a)
1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */
2 /**************************************************************************
3  *
4  * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
5  * All Rights Reserved.
6  *
7  * Permission is hereby granted, free of charge, to any person obtaining a
8  * copy of this software and associated documentation files (the
9  * "Software"), to deal in the Software without restriction, including
10  * without limitation the rights to use, copy, modify, merge, publish,
11  * distribute, sub license, and/or sell copies of the Software, and to
12  * permit persons to whom the Software is furnished to do so, subject to
13  * the following conditions:
14  *
15  * The above copyright notice and this permission notice (including the
16  * next paragraph) shall be included in all copies or substantial portions
17  * of the Software.
18  *
19  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
22  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
23  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
24  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
25  * USE OR OTHER DEALINGS IN THE SOFTWARE.
26  *
27  **************************************************************************/
28 /*
29  * Authors: Thomas Hellstrom <thellstrom-at-vmware-dot-com>
30  */
31 
32 #define pr_fmt(fmt) "[TTM] " fmt
33 
34 #include <linux/sched.h>
35 #include <linux/pagemap.h>
36 #include <linux/shmem_fs.h>
37 #include <linux/file.h>
38 #include <drm/drm_cache.h>
39 #include <drm/ttm/ttm_bo_driver.h>
40 
41 #include "ttm_module.h"
42 
43 static unsigned long ttm_pages_limit;
44 
45 MODULE_PARM_DESC(pages_limit, "Limit for the allocated pages");
46 module_param_named(pages_limit, ttm_pages_limit, ulong, 0644);
47 
48 static unsigned long ttm_dma32_pages_limit;
49 
50 MODULE_PARM_DESC(dma32_pages_limit, "Limit for the allocated DMA32 pages");
51 module_param_named(dma32_pages_limit, ttm_dma32_pages_limit, ulong, 0644);
52 
53 static atomic_long_t ttm_pages_allocated;
54 static atomic_long_t ttm_dma32_pages_allocated;
55 
56 /*
57  * Allocates a ttm structure for the given BO.
58  */
59 int ttm_tt_create(struct ttm_buffer_object *bo, bool zero_alloc)
60 {
61 	struct ttm_device *bdev = bo->bdev;
62 	uint32_t page_flags = 0;
63 
64 	dma_resv_assert_held(bo->base.resv);
65 
66 	if (bo->ttm)
67 		return 0;
68 
69 	switch (bo->type) {
70 	case ttm_bo_type_device:
71 		if (zero_alloc)
72 			page_flags |= TTM_PAGE_FLAG_ZERO_ALLOC;
73 		break;
74 	case ttm_bo_type_kernel:
75 		break;
76 	case ttm_bo_type_sg:
77 		page_flags |= TTM_PAGE_FLAG_SG;
78 		break;
79 	default:
80 		pr_err("Illegal buffer object type\n");
81 		return -EINVAL;
82 	}
83 
84 	bo->ttm = bdev->funcs->ttm_tt_create(bo, page_flags);
85 	if (unlikely(bo->ttm == NULL))
86 		return -ENOMEM;
87 
88 	return 0;
89 }
90 
91 /*
92  * Allocates storage for pointers to the pages that back the ttm.
93  */
94 static int ttm_tt_alloc_page_directory(struct ttm_tt *ttm)
95 {
96 	ttm->pages = kvmalloc_array(ttm->num_pages, sizeof(void*),
97 			GFP_KERNEL | __GFP_ZERO);
98 	if (!ttm->pages)
99 		return -ENOMEM;
100 	return 0;
101 }
102 
103 static int ttm_dma_tt_alloc_page_directory(struct ttm_tt *ttm)
104 {
105 	ttm->pages = kvmalloc_array(ttm->num_pages,
106 				    sizeof(*ttm->pages) +
107 				    sizeof(*ttm->dma_address),
108 				    GFP_KERNEL | __GFP_ZERO);
109 	if (!ttm->pages)
110 		return -ENOMEM;
111 
112 	ttm->dma_address = (void *)(ttm->pages + ttm->num_pages);
113 	return 0;
114 }
115 
116 static int ttm_sg_tt_alloc_page_directory(struct ttm_tt *ttm)
117 {
118 	ttm->dma_address = kvmalloc_array(ttm->num_pages,
119 					  sizeof(*ttm->dma_address),
120 					  GFP_KERNEL | __GFP_ZERO);
121 	if (!ttm->dma_address)
122 		return -ENOMEM;
123 	return 0;
124 }
125 
126 void ttm_tt_destroy_common(struct ttm_device *bdev, struct ttm_tt *ttm)
127 {
128 	ttm_tt_unpopulate(bdev, ttm);
129 
130 	if (ttm->swap_storage)
131 		fput(ttm->swap_storage);
132 
133 	ttm->swap_storage = NULL;
134 }
135 EXPORT_SYMBOL(ttm_tt_destroy_common);
136 
137 void ttm_tt_destroy(struct ttm_device *bdev, struct ttm_tt *ttm)
138 {
139 	bdev->funcs->ttm_tt_destroy(bdev, ttm);
140 }
141 
142 static void ttm_tt_init_fields(struct ttm_tt *ttm,
143 			       struct ttm_buffer_object *bo,
144 			       uint32_t page_flags,
145 			       enum ttm_caching caching)
146 {
147 	ttm->num_pages = PAGE_ALIGN(bo->base.size) >> PAGE_SHIFT;
148 	ttm->caching = ttm_cached;
149 	ttm->page_flags = page_flags;
150 	ttm->dma_address = NULL;
151 	ttm->swap_storage = NULL;
152 	ttm->sg = bo->sg;
153 	ttm->caching = caching;
154 }
155 
156 int ttm_tt_init(struct ttm_tt *ttm, struct ttm_buffer_object *bo,
157 		uint32_t page_flags, enum ttm_caching caching)
158 {
159 	ttm_tt_init_fields(ttm, bo, page_flags, caching);
160 
161 	if (ttm_tt_alloc_page_directory(ttm)) {
162 		pr_err("Failed allocating page table\n");
163 		return -ENOMEM;
164 	}
165 	return 0;
166 }
167 EXPORT_SYMBOL(ttm_tt_init);
168 
169 void ttm_tt_fini(struct ttm_tt *ttm)
170 {
171 	if (ttm->pages)
172 		kvfree(ttm->pages);
173 	else
174 		kvfree(ttm->dma_address);
175 	ttm->pages = NULL;
176 	ttm->dma_address = NULL;
177 }
178 EXPORT_SYMBOL(ttm_tt_fini);
179 
180 int ttm_sg_tt_init(struct ttm_tt *ttm, struct ttm_buffer_object *bo,
181 		   uint32_t page_flags, enum ttm_caching caching)
182 {
183 	int ret;
184 
185 	ttm_tt_init_fields(ttm, bo, page_flags, caching);
186 
187 	if (page_flags & TTM_PAGE_FLAG_SG)
188 		ret = ttm_sg_tt_alloc_page_directory(ttm);
189 	else
190 		ret = ttm_dma_tt_alloc_page_directory(ttm);
191 	if (ret) {
192 		pr_err("Failed allocating page table\n");
193 		return -ENOMEM;
194 	}
195 	return 0;
196 }
197 EXPORT_SYMBOL(ttm_sg_tt_init);
198 
199 int ttm_tt_swapin(struct ttm_tt *ttm)
200 {
201 	struct address_space *swap_space;
202 	struct file *swap_storage;
203 	struct page *from_page;
204 	struct page *to_page;
205 	gfp_t gfp_mask;
206 	int i, ret;
207 
208 	swap_storage = ttm->swap_storage;
209 	BUG_ON(swap_storage == NULL);
210 
211 	swap_space = swap_storage->f_mapping;
212 	gfp_mask = mapping_gfp_mask(swap_space);
213 
214 	for (i = 0; i < ttm->num_pages; ++i) {
215 		from_page = shmem_read_mapping_page_gfp(swap_space, i,
216 							gfp_mask);
217 		if (IS_ERR(from_page)) {
218 			ret = PTR_ERR(from_page);
219 			goto out_err;
220 		}
221 		to_page = ttm->pages[i];
222 		if (unlikely(to_page == NULL)) {
223 			ret = -ENOMEM;
224 			goto out_err;
225 		}
226 
227 		copy_highpage(to_page, from_page);
228 		put_page(from_page);
229 	}
230 
231 	fput(swap_storage);
232 	ttm->swap_storage = NULL;
233 	ttm->page_flags &= ~TTM_PAGE_FLAG_SWAPPED;
234 
235 	return 0;
236 
237 out_err:
238 	return ret;
239 }
240 
241 /**
242  * ttm_tt_swapout - swap out tt object
243  *
244  * @bdev: TTM device structure.
245  * @ttm: The struct ttm_tt.
246  * @gfp_flags: Flags to use for memory allocation.
247  *
248  * Swapout a TT object to a shmem_file, return number of pages swapped out or
249  * negative error code.
250  */
251 int ttm_tt_swapout(struct ttm_device *bdev, struct ttm_tt *ttm,
252 		   gfp_t gfp_flags)
253 {
254 	loff_t size = (loff_t)ttm->num_pages << PAGE_SHIFT;
255 	struct address_space *swap_space;
256 	struct file *swap_storage;
257 	struct page *from_page;
258 	struct page *to_page;
259 	int i, ret;
260 
261 	swap_storage = shmem_file_setup("ttm swap", size, 0);
262 	if (IS_ERR(swap_storage)) {
263 		pr_err("Failed allocating swap storage\n");
264 		return PTR_ERR(swap_storage);
265 	}
266 
267 	swap_space = swap_storage->f_mapping;
268 	gfp_flags &= mapping_gfp_mask(swap_space);
269 
270 	for (i = 0; i < ttm->num_pages; ++i) {
271 		from_page = ttm->pages[i];
272 		if (unlikely(from_page == NULL))
273 			continue;
274 
275 		to_page = shmem_read_mapping_page_gfp(swap_space, i, gfp_flags);
276 		if (IS_ERR(to_page)) {
277 			ret = PTR_ERR(to_page);
278 			goto out_err;
279 		}
280 		copy_highpage(to_page, from_page);
281 		set_page_dirty(to_page);
282 		mark_page_accessed(to_page);
283 		put_page(to_page);
284 	}
285 
286 	ttm_tt_unpopulate(bdev, ttm);
287 	ttm->swap_storage = swap_storage;
288 	ttm->page_flags |= TTM_PAGE_FLAG_SWAPPED;
289 
290 	return ttm->num_pages;
291 
292 out_err:
293 	fput(swap_storage);
294 
295 	return ret;
296 }
297 
298 static void ttm_tt_add_mapping(struct ttm_device *bdev, struct ttm_tt *ttm)
299 {
300 	pgoff_t i;
301 
302 	if (ttm->page_flags & TTM_PAGE_FLAG_SG)
303 		return;
304 
305 	for (i = 0; i < ttm->num_pages; ++i)
306 		ttm->pages[i]->mapping = bdev->dev_mapping;
307 }
308 
309 int ttm_tt_populate(struct ttm_device *bdev,
310 		    struct ttm_tt *ttm, struct ttm_operation_ctx *ctx)
311 {
312 	int ret;
313 
314 	if (!ttm)
315 		return -EINVAL;
316 
317 	if (ttm_tt_is_populated(ttm))
318 		return 0;
319 
320 	if (!(ttm->page_flags & TTM_PAGE_FLAG_SG)) {
321 		atomic_long_add(ttm->num_pages, &ttm_pages_allocated);
322 		if (bdev->pool.use_dma32)
323 			atomic_long_add(ttm->num_pages,
324 					&ttm_dma32_pages_allocated);
325 	}
326 
327 	while (atomic_long_read(&ttm_pages_allocated) > ttm_pages_limit ||
328 	       atomic_long_read(&ttm_dma32_pages_allocated) >
329 	       ttm_dma32_pages_limit) {
330 
331 		ret = ttm_global_swapout(ctx, GFP_KERNEL);
332 		if (ret == 0)
333 			break;
334 		if (ret < 0)
335 			goto error;
336 	}
337 
338 	if (bdev->funcs->ttm_tt_populate)
339 		ret = bdev->funcs->ttm_tt_populate(bdev, ttm, ctx);
340 	else
341 		ret = ttm_pool_alloc(&bdev->pool, ttm, ctx);
342 	if (ret)
343 		goto error;
344 
345 	ttm_tt_add_mapping(bdev, ttm);
346 	ttm->page_flags |= TTM_PAGE_FLAG_PRIV_POPULATED;
347 	if (unlikely(ttm->page_flags & TTM_PAGE_FLAG_SWAPPED)) {
348 		ret = ttm_tt_swapin(ttm);
349 		if (unlikely(ret != 0)) {
350 			ttm_tt_unpopulate(bdev, ttm);
351 			return ret;
352 		}
353 	}
354 
355 	return 0;
356 
357 error:
358 	if (!(ttm->page_flags & TTM_PAGE_FLAG_SG)) {
359 		atomic_long_sub(ttm->num_pages, &ttm_pages_allocated);
360 		if (bdev->pool.use_dma32)
361 			atomic_long_sub(ttm->num_pages,
362 					&ttm_dma32_pages_allocated);
363 	}
364 	return ret;
365 }
366 EXPORT_SYMBOL(ttm_tt_populate);
367 
368 static void ttm_tt_clear_mapping(struct ttm_tt *ttm)
369 {
370 	pgoff_t i;
371 	struct page **page = ttm->pages;
372 
373 	if (ttm->page_flags & TTM_PAGE_FLAG_SG)
374 		return;
375 
376 	for (i = 0; i < ttm->num_pages; ++i) {
377 		(*page)->mapping = NULL;
378 		(*page++)->index = 0;
379 	}
380 }
381 
382 void ttm_tt_unpopulate(struct ttm_device *bdev, struct ttm_tt *ttm)
383 {
384 	if (!ttm_tt_is_populated(ttm))
385 		return;
386 
387 	ttm_tt_clear_mapping(ttm);
388 	if (bdev->funcs->ttm_tt_unpopulate)
389 		bdev->funcs->ttm_tt_unpopulate(bdev, ttm);
390 	else
391 		ttm_pool_free(&bdev->pool, ttm);
392 
393 	if (!(ttm->page_flags & TTM_PAGE_FLAG_SG)) {
394 		atomic_long_sub(ttm->num_pages, &ttm_pages_allocated);
395 		if (bdev->pool.use_dma32)
396 			atomic_long_sub(ttm->num_pages,
397 					&ttm_dma32_pages_allocated);
398 	}
399 
400 	ttm->page_flags &= ~TTM_PAGE_FLAG_PRIV_POPULATED;
401 }
402 
403 #ifdef CONFIG_DEBUG_FS
404 
405 /* Test the shrinker functions and dump the result */
406 static int ttm_tt_debugfs_shrink_show(struct seq_file *m, void *data)
407 {
408 	struct ttm_operation_ctx ctx = { false, false };
409 
410 	seq_printf(m, "%d\n", ttm_global_swapout(&ctx, GFP_KERNEL));
411 	return 0;
412 }
413 DEFINE_SHOW_ATTRIBUTE(ttm_tt_debugfs_shrink);
414 
415 #endif
416 
417 
418 /**
419  * ttm_tt_mgr_init - register with the MM shrinker
420  *
421  * Register with the MM shrinker for swapping out BOs.
422  */
423 void ttm_tt_mgr_init(unsigned long num_pages, unsigned long num_dma32_pages)
424 {
425 #ifdef CONFIG_DEBUG_FS
426 	debugfs_create_file("tt_shrink", 0400, ttm_debugfs_root, NULL,
427 			    &ttm_tt_debugfs_shrink_fops);
428 #endif
429 
430 	if (!ttm_pages_limit)
431 		ttm_pages_limit = num_pages;
432 
433 	if (!ttm_dma32_pages_limit)
434 		ttm_dma32_pages_limit = num_dma32_pages;
435 }
436