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/mm.h>
31 
32 #include <drm/ttm/ttm_device.h>
33 #include <drm/ttm/ttm_tt.h>
34 #include <drm/ttm/ttm_placement.h>
35 #include <drm/ttm/ttm_bo_api.h>
36 
37 #include "ttm_module.h"
38 
39 /*
40  * ttm_global_mutex - protecting the global state
41  */
42 static DEFINE_MUTEX(ttm_global_mutex);
43 static unsigned ttm_glob_use_count;
44 struct ttm_global ttm_glob;
45 EXPORT_SYMBOL(ttm_glob);
46 
47 struct dentry *ttm_debugfs_root;
48 
ttm_global_release(void)49 static void ttm_global_release(void)
50 {
51 	struct ttm_global *glob = &ttm_glob;
52 
53 	mutex_lock(&ttm_global_mutex);
54 	if (--ttm_glob_use_count > 0)
55 		goto out;
56 
57 	ttm_pool_mgr_fini();
58 	debugfs_remove(ttm_debugfs_root);
59 
60 	__free_page(glob->dummy_read_page);
61 	memset(glob, 0, sizeof(*glob));
62 out:
63 	mutex_unlock(&ttm_global_mutex);
64 }
65 
ttm_global_init(void)66 static int ttm_global_init(void)
67 {
68 	struct ttm_global *glob = &ttm_glob;
69 	unsigned long num_pages, num_dma32;
70 	struct sysinfo si;
71 	int ret = 0;
72 
73 	mutex_lock(&ttm_global_mutex);
74 	if (++ttm_glob_use_count > 1)
75 		goto out;
76 
77 	si_meminfo(&si);
78 
79 	ttm_debugfs_root = debugfs_create_dir("ttm", NULL);
80 	if (IS_ERR(ttm_debugfs_root)) {
81 		ttm_debugfs_root = NULL;
82 	}
83 
84 	/* Limit the number of pages in the pool to about 50% of the total
85 	 * system memory.
86 	 */
87 	num_pages = ((u64)si.totalram * si.mem_unit) >> PAGE_SHIFT;
88 	num_pages /= 2;
89 
90 	/* But for DMA32 we limit ourself to only use 2GiB maximum. */
91 	num_dma32 = (u64)(si.totalram - si.totalhigh) * si.mem_unit
92 		>> PAGE_SHIFT;
93 	num_dma32 = min(num_dma32, 2UL << (30 - PAGE_SHIFT));
94 
95 	ttm_pool_mgr_init(num_pages);
96 	ttm_tt_mgr_init(num_pages, num_dma32);
97 
98 	glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
99 
100 	if (unlikely(glob->dummy_read_page == NULL)) {
101 		ret = -ENOMEM;
102 		goto out;
103 	}
104 
105 	INIT_LIST_HEAD(&glob->device_list);
106 	atomic_set(&glob->bo_count, 0);
107 
108 	debugfs_create_atomic_t("buffer_objects", 0444, ttm_debugfs_root,
109 				&glob->bo_count);
110 out:
111 	if (ret && ttm_debugfs_root)
112 		debugfs_remove(ttm_debugfs_root);
113 	if (ret)
114 		--ttm_glob_use_count;
115 	mutex_unlock(&ttm_global_mutex);
116 	return ret;
117 }
118 
119 /*
120  * A buffer object shrink method that tries to swap out the first
121  * buffer object on the global::swap_lru list.
122  */
ttm_global_swapout(struct ttm_operation_ctx * ctx,gfp_t gfp_flags)123 int ttm_global_swapout(struct ttm_operation_ctx *ctx, gfp_t gfp_flags)
124 {
125 	struct ttm_global *glob = &ttm_glob;
126 	struct ttm_device *bdev;
127 	int ret = 0;
128 
129 	mutex_lock(&ttm_global_mutex);
130 	list_for_each_entry(bdev, &glob->device_list, device_list) {
131 		ret = ttm_device_swapout(bdev, ctx, gfp_flags);
132 		if (ret > 0) {
133 			list_move_tail(&bdev->device_list, &glob->device_list);
134 			break;
135 		}
136 	}
137 	mutex_unlock(&ttm_global_mutex);
138 	return ret;
139 }
140 EXPORT_SYMBOL(ttm_global_swapout);
141 
ttm_device_swapout(struct ttm_device * bdev,struct ttm_operation_ctx * ctx,gfp_t gfp_flags)142 int ttm_device_swapout(struct ttm_device *bdev, struct ttm_operation_ctx *ctx,
143 		       gfp_t gfp_flags)
144 {
145 	struct ttm_resource_manager *man;
146 	struct ttm_buffer_object *bo;
147 	unsigned i, j;
148 	int ret;
149 
150 	spin_lock(&bdev->lru_lock);
151 	for (i = TTM_PL_SYSTEM; i < TTM_NUM_MEM_TYPES; ++i) {
152 		man = ttm_manager_type(bdev, i);
153 		if (!man || !man->use_tt)
154 			continue;
155 
156 		for (j = 0; j < TTM_MAX_BO_PRIORITY; ++j) {
157 			list_for_each_entry(bo, &man->lru[j], lru) {
158 				uint32_t num_pages = PFN_UP(bo->base.size);
159 
160 				ret = ttm_bo_swapout(bo, ctx, gfp_flags);
161 				/* ttm_bo_swapout has dropped the lru_lock */
162 				if (!ret)
163 					return num_pages;
164 				if (ret != -EBUSY)
165 					return ret;
166 			}
167 		}
168 	}
169 	spin_unlock(&bdev->lru_lock);
170 	return 0;
171 }
172 EXPORT_SYMBOL(ttm_device_swapout);
173 
ttm_device_delayed_workqueue(struct work_struct * work)174 static void ttm_device_delayed_workqueue(struct work_struct *work)
175 {
176 	struct ttm_device *bdev =
177 		container_of(work, struct ttm_device, wq.work);
178 
179 	if (!ttm_bo_delayed_delete(bdev, false))
180 		schedule_delayed_work(&bdev->wq,
181 				      ((HZ / 100) < 1) ? 1 : HZ / 100);
182 }
183 
184 /**
185  * ttm_device_init
186  *
187  * @bdev: A pointer to a struct ttm_device to initialize.
188  * @funcs: Function table for the device.
189  * @dev: The core kernel device pointer for DMA mappings and allocations.
190  * @mapping: The address space to use for this bo.
191  * @vma_manager: A pointer to a vma manager.
192  * @use_dma_alloc: If coherent DMA allocation API should be used.
193  * @use_dma32: If we should use GFP_DMA32 for device memory allocations.
194  *
195  * Initializes a struct ttm_device:
196  * Returns:
197  * !0: Failure.
198  */
ttm_device_init(struct ttm_device * bdev,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)199 int ttm_device_init(struct ttm_device *bdev, struct ttm_device_funcs *funcs,
200 		    struct device *dev, struct address_space *mapping,
201 		    struct drm_vma_offset_manager *vma_manager,
202 		    bool use_dma_alloc, bool use_dma32)
203 {
204 	struct ttm_global *glob = &ttm_glob;
205 	int ret;
206 
207 	if (WARN_ON(vma_manager == NULL))
208 		return -EINVAL;
209 
210 	ret = ttm_global_init();
211 	if (ret)
212 		return ret;
213 
214 	bdev->funcs = funcs;
215 
216 	ttm_sys_man_init(bdev);
217 	ttm_pool_init(&bdev->pool, dev, use_dma_alloc, use_dma32);
218 
219 	bdev->vma_manager = vma_manager;
220 	INIT_DELAYED_WORK(&bdev->wq, ttm_device_delayed_workqueue);
221 	spin_lock_init(&bdev->lru_lock);
222 	INIT_LIST_HEAD(&bdev->ddestroy);
223 	INIT_LIST_HEAD(&bdev->pinned);
224 	bdev->dev_mapping = mapping;
225 	mutex_lock(&ttm_global_mutex);
226 	list_add_tail(&bdev->device_list, &glob->device_list);
227 	mutex_unlock(&ttm_global_mutex);
228 
229 	return 0;
230 }
231 EXPORT_SYMBOL(ttm_device_init);
232 
ttm_device_fini(struct ttm_device * bdev)233 void ttm_device_fini(struct ttm_device *bdev)
234 {
235 	struct ttm_resource_manager *man;
236 	unsigned i;
237 
238 	man = ttm_manager_type(bdev, TTM_PL_SYSTEM);
239 	ttm_resource_manager_set_used(man, false);
240 	ttm_set_driver_manager(bdev, TTM_PL_SYSTEM, NULL);
241 
242 	mutex_lock(&ttm_global_mutex);
243 	list_del(&bdev->device_list);
244 	mutex_unlock(&ttm_global_mutex);
245 
246 	cancel_delayed_work_sync(&bdev->wq);
247 
248 	if (ttm_bo_delayed_delete(bdev, true))
249 		pr_debug("Delayed destroy list was clean\n");
250 
251 	spin_lock(&bdev->lru_lock);
252 	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
253 		if (list_empty(&man->lru[0]))
254 			pr_debug("Swap list %d was clean\n", i);
255 	spin_unlock(&bdev->lru_lock);
256 
257 	ttm_pool_fini(&bdev->pool);
258 	ttm_global_release();
259 }
260 EXPORT_SYMBOL(ttm_device_fini);
261 
ttm_device_clear_dma_mappings(struct ttm_device * bdev)262 void ttm_device_clear_dma_mappings(struct ttm_device *bdev)
263 {
264 	struct ttm_resource_manager *man;
265 	struct ttm_buffer_object *bo;
266 	unsigned int i, j;
267 
268 	spin_lock(&bdev->lru_lock);
269 	while (!list_empty(&bdev->pinned)) {
270 		bo = list_first_entry(&bdev->pinned, struct ttm_buffer_object, lru);
271 		/* Take ref against racing releases once lru_lock is unlocked */
272 		if (ttm_bo_get_unless_zero(bo)) {
273 			list_del_init(&bo->lru);
274 			spin_unlock(&bdev->lru_lock);
275 
276 			if (bo->ttm)
277 				ttm_tt_unpopulate(bo->bdev, bo->ttm);
278 
279 			ttm_bo_put(bo);
280 			spin_lock(&bdev->lru_lock);
281 		}
282 	}
283 
284 	for (i = TTM_PL_SYSTEM; i < TTM_NUM_MEM_TYPES; ++i) {
285 		man = ttm_manager_type(bdev, i);
286 		if (!man || !man->use_tt)
287 			continue;
288 
289 		for (j = 0; j < TTM_MAX_BO_PRIORITY; ++j) {
290 			while (!list_empty(&man->lru[j])) {
291 				bo = list_first_entry(&man->lru[j], struct ttm_buffer_object, lru);
292 				if (ttm_bo_get_unless_zero(bo)) {
293 					list_del_init(&bo->lru);
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 			}
303 		}
304 	}
305 	spin_unlock(&bdev->lru_lock);
306 }
307 EXPORT_SYMBOL(ttm_device_clear_dma_mappings);
308