xref: /openbmc/linux/drivers/gpu/drm/ttm/ttm_device.c (revision 9659281c)
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 
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 
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 		ret = PTR_ERR(ttm_debugfs_root);
82 		ttm_debugfs_root = NULL;
83 		goto out;
84 	}
85 
86 	/* Limit the number of pages in the pool to about 50% of the total
87 	 * system memory.
88 	 */
89 	num_pages = ((u64)si.totalram * si.mem_unit) >> PAGE_SHIFT;
90 	num_pages /= 2;
91 
92 	/* But for DMA32 we limit ourself to only use 2GiB maximum. */
93 	num_dma32 = (u64)(si.totalram - si.totalhigh) * si.mem_unit
94 		>> PAGE_SHIFT;
95 	num_dma32 = min(num_dma32, 2UL << (30 - PAGE_SHIFT));
96 
97 	ttm_pool_mgr_init(num_pages);
98 	ttm_tt_mgr_init(num_pages, num_dma32);
99 
100 	glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
101 
102 	if (unlikely(glob->dummy_read_page == NULL)) {
103 		ret = -ENOMEM;
104 		goto out;
105 	}
106 
107 	INIT_LIST_HEAD(&glob->device_list);
108 	atomic_set(&glob->bo_count, 0);
109 
110 	debugfs_create_atomic_t("buffer_objects", 0444, ttm_debugfs_root,
111 				&glob->bo_count);
112 out:
113 	if (ret && ttm_debugfs_root)
114 		debugfs_remove(ttm_debugfs_root);
115 	if (ret)
116 		--ttm_glob_use_count;
117 	mutex_unlock(&ttm_global_mutex);
118 	return ret;
119 }
120 
121 /*
122  * A buffer object shrink method that tries to swap out the first
123  * buffer object on the global::swap_lru list.
124  */
125 int ttm_global_swapout(struct ttm_operation_ctx *ctx, gfp_t gfp_flags)
126 {
127 	struct ttm_global *glob = &ttm_glob;
128 	struct ttm_device *bdev;
129 	int ret = 0;
130 
131 	mutex_lock(&ttm_global_mutex);
132 	list_for_each_entry(bdev, &glob->device_list, device_list) {
133 		ret = ttm_device_swapout(bdev, ctx, gfp_flags);
134 		if (ret > 0) {
135 			list_move_tail(&bdev->device_list, &glob->device_list);
136 			break;
137 		}
138 	}
139 	mutex_unlock(&ttm_global_mutex);
140 	return ret;
141 }
142 EXPORT_SYMBOL(ttm_global_swapout);
143 
144 int ttm_device_swapout(struct ttm_device *bdev, struct ttm_operation_ctx *ctx,
145 		       gfp_t gfp_flags)
146 {
147 	struct ttm_resource_manager *man;
148 	struct ttm_buffer_object *bo;
149 	unsigned i, j;
150 	int ret;
151 
152 	spin_lock(&bdev->lru_lock);
153 	for (i = TTM_PL_SYSTEM; i < TTM_NUM_MEM_TYPES; ++i) {
154 		man = ttm_manager_type(bdev, i);
155 		if (!man || !man->use_tt)
156 			continue;
157 
158 		for (j = 0; j < TTM_MAX_BO_PRIORITY; ++j) {
159 			list_for_each_entry(bo, &man->lru[j], lru) {
160 				uint32_t num_pages = PFN_UP(bo->base.size);
161 
162 				ret = ttm_bo_swapout(bo, ctx, gfp_flags);
163 				/* ttm_bo_swapout has dropped the lru_lock */
164 				if (!ret)
165 					return num_pages;
166 				if (ret != -EBUSY)
167 					return ret;
168 			}
169 		}
170 	}
171 	spin_unlock(&bdev->lru_lock);
172 	return 0;
173 }
174 EXPORT_SYMBOL(ttm_device_swapout);
175 
176 static void ttm_device_delayed_workqueue(struct work_struct *work)
177 {
178 	struct ttm_device *bdev =
179 		container_of(work, struct ttm_device, wq.work);
180 
181 	if (!ttm_bo_delayed_delete(bdev, false))
182 		schedule_delayed_work(&bdev->wq,
183 				      ((HZ / 100) < 1) ? 1 : HZ / 100);
184 }
185 
186 /**
187  * ttm_device_init
188  *
189  * @bdev: A pointer to a struct ttm_device to initialize.
190  * @funcs: Function table for the device.
191  * @dev: The core kernel device pointer for DMA mappings and allocations.
192  * @mapping: The address space to use for this bo.
193  * @vma_manager: A pointer to a vma manager.
194  * @use_dma_alloc: If coherent DMA allocation API should be used.
195  * @use_dma32: If we should use GFP_DMA32 for device memory allocations.
196  *
197  * Initializes a struct ttm_device:
198  * Returns:
199  * !0: Failure.
200  */
201 int ttm_device_init(struct ttm_device *bdev, struct ttm_device_funcs *funcs,
202 		    struct device *dev, struct address_space *mapping,
203 		    struct drm_vma_offset_manager *vma_manager,
204 		    bool use_dma_alloc, bool use_dma32)
205 {
206 	struct ttm_global *glob = &ttm_glob;
207 	int ret;
208 
209 	if (WARN_ON(vma_manager == NULL))
210 		return -EINVAL;
211 
212 	ret = ttm_global_init();
213 	if (ret)
214 		return ret;
215 
216 	bdev->funcs = funcs;
217 
218 	ttm_sys_man_init(bdev);
219 	ttm_pool_init(&bdev->pool, dev, use_dma_alloc, use_dma32);
220 
221 	bdev->vma_manager = vma_manager;
222 	INIT_DELAYED_WORK(&bdev->wq, ttm_device_delayed_workqueue);
223 	spin_lock_init(&bdev->lru_lock);
224 	INIT_LIST_HEAD(&bdev->ddestroy);
225 	bdev->dev_mapping = mapping;
226 	mutex_lock(&ttm_global_mutex);
227 	list_add_tail(&bdev->device_list, &glob->device_list);
228 	mutex_unlock(&ttm_global_mutex);
229 
230 	return 0;
231 }
232 EXPORT_SYMBOL(ttm_device_init);
233 
234 void ttm_device_fini(struct ttm_device *bdev)
235 {
236 	struct ttm_resource_manager *man;
237 	unsigned i;
238 
239 	man = ttm_manager_type(bdev, TTM_PL_SYSTEM);
240 	ttm_resource_manager_set_used(man, false);
241 	ttm_set_driver_manager(bdev, TTM_PL_SYSTEM, NULL);
242 
243 	mutex_lock(&ttm_global_mutex);
244 	list_del(&bdev->device_list);
245 	mutex_unlock(&ttm_global_mutex);
246 
247 	cancel_delayed_work_sync(&bdev->wq);
248 
249 	if (ttm_bo_delayed_delete(bdev, true))
250 		pr_debug("Delayed destroy list was clean\n");
251 
252 	spin_lock(&bdev->lru_lock);
253 	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
254 		if (list_empty(&man->lru[0]))
255 			pr_debug("Swap list %d was clean\n", i);
256 	spin_unlock(&bdev->lru_lock);
257 
258 	ttm_pool_fini(&bdev->pool);
259 	ttm_global_release();
260 }
261 EXPORT_SYMBOL(ttm_device_fini);
262