1 /* SPDX-License-Identifier: GPL-2.0 OR MIT */
2
3 /*
4 * Copyright (c) 2006-2009 VMware, Inc., Palo Alto, CA., USA
5 * Copyright 2020 Advanced Micro Devices, Inc.
6 *
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
13 *
14 * The above copyright notice and this permission notice shall be included in
15 * all copies or substantial portions of the Software.
16 *
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
21 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
22 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
23 * OTHER DEALINGS IN THE SOFTWARE.
24 *
25 * Authors: Christian König
26 */
27
28 #define pr_fmt(fmt) "[TTM DEVICE] " fmt
29
30 #include <linux/debugfs.h>
31 #include <linux/export.h>
32 #include <linux/mm.h>
33
34 #include <drm/ttm/ttm_bo.h>
35 #include <drm/ttm/ttm_device.h>
36 #include <drm/ttm/ttm_tt.h>
37 #include <drm/ttm/ttm_placement.h>
38
39 #include "ttm_module.h"
40 #include "ttm_bo_internal.h"
41
42 /*
43 * ttm_global_mutex - protecting the global state
44 */
45 static DEFINE_MUTEX(ttm_global_mutex);
46 static unsigned ttm_glob_use_count;
47 struct ttm_global ttm_glob;
48 EXPORT_SYMBOL(ttm_glob);
49
50 struct dentry *ttm_debugfs_root;
51
ttm_global_release(void)52 static void ttm_global_release(void)
53 {
54 struct ttm_global *glob = &ttm_glob;
55
56 mutex_lock(&ttm_global_mutex);
57 if (--ttm_glob_use_count > 0)
58 goto out;
59
60 ttm_pool_mgr_fini();
61 debugfs_remove(ttm_debugfs_root);
62
63 __free_page(glob->dummy_read_page);
64 memset(glob, 0, sizeof(*glob));
65 out:
66 mutex_unlock(&ttm_global_mutex);
67 }
68
ttm_global_init(void)69 static int ttm_global_init(void)
70 {
71 struct ttm_global *glob = &ttm_glob;
72 unsigned long num_pages, num_dma32;
73 struct sysinfo si;
74 int ret = 0;
75
76 mutex_lock(&ttm_global_mutex);
77 if (++ttm_glob_use_count > 1)
78 goto out;
79
80 si_meminfo(&si);
81
82 ttm_debugfs_root = debugfs_create_dir("ttm", NULL);
83 if (IS_ERR(ttm_debugfs_root)) {
84 ttm_debugfs_root = NULL;
85 }
86
87 /* Limit the number of pages in the pool to about 50% of the total
88 * system memory.
89 */
90 num_pages = ((u64)si.totalram * si.mem_unit) >> PAGE_SHIFT;
91 num_pages /= 2;
92
93 /* But for DMA32 we limit ourself to only use 2GiB maximum. */
94 num_dma32 = (u64)(si.totalram - si.totalhigh) * si.mem_unit
95 >> PAGE_SHIFT;
96 num_dma32 = min(num_dma32, 2UL << (30 - PAGE_SHIFT));
97
98 ttm_pool_mgr_init(num_pages);
99 ttm_tt_mgr_init(num_pages, num_dma32);
100
101 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32 |
102 __GFP_NOWARN);
103
104 /* Retry without GFP_DMA32 for platforms DMA32 is not available */
105 if (unlikely(glob->dummy_read_page == NULL)) {
106 glob->dummy_read_page = alloc_page(__GFP_ZERO);
107 if (unlikely(glob->dummy_read_page == NULL)) {
108 ret = -ENOMEM;
109 goto out;
110 }
111 pr_warn("Using GFP_DMA32 fallback for dummy_read_page\n");
112 }
113
114 INIT_LIST_HEAD(&glob->device_list);
115 atomic_set(&glob->bo_count, 0);
116
117 debugfs_create_atomic_t("buffer_objects", 0444, ttm_debugfs_root,
118 &glob->bo_count);
119 out:
120 if (ret && ttm_debugfs_root)
121 debugfs_remove(ttm_debugfs_root);
122 if (ret)
123 --ttm_glob_use_count;
124 mutex_unlock(&ttm_global_mutex);
125 return ret;
126 }
127
128 /**
129 * ttm_device_prepare_hibernation - move GTT BOs to shmem for hibernation.
130 *
131 * @bdev: A pointer to a struct ttm_device to prepare hibernation for.
132 *
133 * Return: 0 on success, negative number on failure.
134 */
ttm_device_prepare_hibernation(struct ttm_device * bdev)135 int ttm_device_prepare_hibernation(struct ttm_device *bdev)
136 {
137 struct ttm_operation_ctx ctx = {
138 .interruptible = false,
139 .no_wait_gpu = false,
140 };
141 int ret;
142
143 do {
144 ret = ttm_device_swapout(bdev, &ctx, GFP_KERNEL);
145 } while (ret > 0);
146 return ret;
147 }
148 EXPORT_SYMBOL(ttm_device_prepare_hibernation);
149
150 /*
151 * A buffer object shrink method that tries to swap out the first
152 * buffer object on the global::swap_lru list.
153 */
ttm_global_swapout(struct ttm_operation_ctx * ctx,gfp_t gfp_flags)154 int ttm_global_swapout(struct ttm_operation_ctx *ctx, gfp_t gfp_flags)
155 {
156 struct ttm_global *glob = &ttm_glob;
157 struct ttm_device *bdev;
158 int ret = 0;
159
160 mutex_lock(&ttm_global_mutex);
161 list_for_each_entry(bdev, &glob->device_list, device_list) {
162 ret = ttm_device_swapout(bdev, ctx, gfp_flags);
163 if (ret > 0) {
164 list_move_tail(&bdev->device_list, &glob->device_list);
165 break;
166 }
167 }
168 mutex_unlock(&ttm_global_mutex);
169 return ret;
170 }
171
ttm_device_swapout(struct ttm_device * bdev,struct ttm_operation_ctx * ctx,gfp_t gfp_flags)172 int ttm_device_swapout(struct ttm_device *bdev, struct ttm_operation_ctx *ctx,
173 gfp_t gfp_flags)
174 {
175 struct ttm_resource_manager *man;
176 unsigned i;
177 s64 lret;
178
179 for (i = TTM_PL_SYSTEM; i < TTM_NUM_MEM_TYPES; ++i) {
180 man = ttm_manager_type(bdev, i);
181 if (!man || !man->use_tt)
182 continue;
183
184 lret = ttm_bo_swapout(bdev, ctx, man, gfp_flags, 1);
185 /* Can be both positive (num_pages) and negative (error) */
186 if (lret)
187 return lret;
188 }
189 return 0;
190 }
191 EXPORT_SYMBOL(ttm_device_swapout);
192
193 /**
194 * ttm_device_init
195 *
196 * @bdev: A pointer to a struct ttm_device to initialize.
197 * @funcs: Function table for the device.
198 * @dev: The core kernel device pointer for DMA mappings and allocations.
199 * @mapping: The address space to use for this bo.
200 * @vma_manager: A pointer to a vma manager.
201 * @use_dma_alloc: If coherent DMA allocation API should be used.
202 * @use_dma32: If we should use GFP_DMA32 for device memory allocations.
203 *
204 * Initializes a struct ttm_device:
205 * Returns:
206 * !0: Failure.
207 */
ttm_device_init(struct ttm_device * bdev,const struct ttm_device_funcs * funcs,struct device * dev,struct address_space * mapping,struct drm_vma_offset_manager * vma_manager,bool use_dma_alloc,bool use_dma32)208 int ttm_device_init(struct ttm_device *bdev, const struct ttm_device_funcs *funcs,
209 struct device *dev, struct address_space *mapping,
210 struct drm_vma_offset_manager *vma_manager,
211 bool use_dma_alloc, bool use_dma32)
212 {
213 struct ttm_global *glob = &ttm_glob;
214 int ret, nid;
215
216 if (WARN_ON(vma_manager == NULL))
217 return -EINVAL;
218
219 ret = ttm_global_init();
220 if (ret)
221 return ret;
222
223 bdev->wq = alloc_workqueue("ttm",
224 WQ_MEM_RECLAIM | WQ_HIGHPRI | WQ_UNBOUND, 16);
225 if (!bdev->wq) {
226 ttm_global_release();
227 return -ENOMEM;
228 }
229
230 bdev->funcs = funcs;
231
232 ttm_sys_man_init(bdev);
233
234 if (dev)
235 nid = dev_to_node(dev);
236 else
237 nid = NUMA_NO_NODE;
238
239 ttm_pool_init(&bdev->pool, dev, nid, use_dma_alloc, use_dma32);
240
241 bdev->vma_manager = vma_manager;
242 spin_lock_init(&bdev->lru_lock);
243 INIT_LIST_HEAD(&bdev->unevictable);
244 bdev->dev_mapping = mapping;
245 mutex_lock(&ttm_global_mutex);
246 list_add_tail(&bdev->device_list, &glob->device_list);
247 mutex_unlock(&ttm_global_mutex);
248
249 return 0;
250 }
251 EXPORT_SYMBOL(ttm_device_init);
252
ttm_device_fini(struct ttm_device * bdev)253 void ttm_device_fini(struct ttm_device *bdev)
254 {
255 struct ttm_resource_manager *man;
256 unsigned i;
257
258 mutex_lock(&ttm_global_mutex);
259 list_del(&bdev->device_list);
260 mutex_unlock(&ttm_global_mutex);
261
262 drain_workqueue(bdev->wq);
263 destroy_workqueue(bdev->wq);
264
265 man = ttm_manager_type(bdev, TTM_PL_SYSTEM);
266 ttm_resource_manager_set_used(man, false);
267 ttm_set_driver_manager(bdev, TTM_PL_SYSTEM, NULL);
268
269 spin_lock(&bdev->lru_lock);
270 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
271 if (list_empty(&man->lru[0]))
272 pr_debug("Swap list %d was clean\n", i);
273 spin_unlock(&bdev->lru_lock);
274
275 ttm_pool_fini(&bdev->pool);
276 ttm_global_release();
277 }
278 EXPORT_SYMBOL(ttm_device_fini);
279
ttm_device_clear_lru_dma_mappings(struct ttm_device * bdev,struct list_head * list)280 static void ttm_device_clear_lru_dma_mappings(struct ttm_device *bdev,
281 struct list_head *list)
282 {
283 struct ttm_resource *res;
284
285 spin_lock(&bdev->lru_lock);
286 while ((res = ttm_lru_first_res_or_null(list))) {
287 struct ttm_buffer_object *bo = res->bo;
288
289 /* Take ref against racing releases once lru_lock is unlocked */
290 if (!ttm_bo_get_unless_zero(bo))
291 continue;
292
293 list_del_init(&bo->resource->lru.link);
294 spin_unlock(&bdev->lru_lock);
295
296 if (bo->ttm)
297 ttm_tt_unpopulate(bo->bdev, bo->ttm);
298
299 ttm_bo_put(bo);
300 spin_lock(&bdev->lru_lock);
301 }
302 spin_unlock(&bdev->lru_lock);
303 }
304
ttm_device_clear_dma_mappings(struct ttm_device * bdev)305 void ttm_device_clear_dma_mappings(struct ttm_device *bdev)
306 {
307 struct ttm_resource_manager *man;
308 unsigned int i, j;
309
310 ttm_device_clear_lru_dma_mappings(bdev, &bdev->unevictable);
311
312 for (i = TTM_PL_SYSTEM; i < TTM_NUM_MEM_TYPES; ++i) {
313 man = ttm_manager_type(bdev, i);
314 if (!man || !man->use_tt)
315 continue;
316
317 for (j = 0; j < TTM_MAX_BO_PRIORITY; ++j)
318 ttm_device_clear_lru_dma_mappings(bdev, &man->lru[j]);
319 }
320 }
321 EXPORT_SYMBOL(ttm_device_clear_dma_mappings);
322