xref: /openbmc/linux/drivers/gpu/drm/ttm/ttm_device.c (revision 75020f2d)
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 DEFINE_MUTEX(ttm_global_mutex);
43 unsigned ttm_glob_use_count;
44 struct ttm_global ttm_glob;
45 EXPORT_SYMBOL(ttm_glob);
46 
47 static void ttm_global_release(void)
48 {
49 	struct ttm_global *glob = &ttm_glob;
50 
51 	mutex_lock(&ttm_global_mutex);
52 	if (--ttm_glob_use_count > 0)
53 		goto out;
54 
55 	ttm_pool_mgr_fini();
56 
57 	__free_page(glob->dummy_read_page);
58 	memset(glob, 0, sizeof(*glob));
59 out:
60 	mutex_unlock(&ttm_global_mutex);
61 }
62 
63 static int ttm_global_init(void)
64 {
65 	struct ttm_global *glob = &ttm_glob;
66 	unsigned long num_pages, num_dma32;
67 	struct sysinfo si;
68 	int ret = 0;
69 
70 	mutex_lock(&ttm_global_mutex);
71 	if (++ttm_glob_use_count > 1)
72 		goto out;
73 
74 	si_meminfo(&si);
75 
76 	/* Limit the number of pages in the pool to about 50% of the total
77 	 * system memory.
78 	 */
79 	num_pages = ((u64)si.totalram * si.mem_unit) >> PAGE_SHIFT;
80 	num_pages /= 2;
81 
82 	/* But for DMA32 we limit ourself to only use 2GiB maximum. */
83 	num_dma32 = (u64)(si.totalram - si.totalhigh) * si.mem_unit
84 		>> PAGE_SHIFT;
85 	num_dma32 = min(num_dma32, 2UL << (30 - PAGE_SHIFT));
86 
87 	ttm_pool_mgr_init(num_pages);
88 	ttm_tt_mgr_init(num_pages, num_dma32);
89 
90 	glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32);
91 
92 	if (unlikely(glob->dummy_read_page == NULL)) {
93 		ret = -ENOMEM;
94 		goto out;
95 	}
96 
97 	INIT_LIST_HEAD(&glob->device_list);
98 	atomic_set(&glob->bo_count, 0);
99 
100 	debugfs_create_atomic_t("buffer_objects", 0444, ttm_debugfs_root,
101 				&glob->bo_count);
102 out:
103 	mutex_unlock(&ttm_global_mutex);
104 	return ret;
105 }
106 
107 /**
108  * A buffer object shrink method that tries to swap out the first
109  * buffer object on the global::swap_lru list.
110  */
111 int ttm_global_swapout(struct ttm_operation_ctx *ctx, gfp_t gfp_flags)
112 {
113 	struct ttm_global *glob = &ttm_glob;
114 	struct ttm_device *bdev;
115 	int ret = 0;
116 
117 	mutex_lock(&ttm_global_mutex);
118 	list_for_each_entry(bdev, &glob->device_list, device_list) {
119 		ret = ttm_device_swapout(bdev, ctx, gfp_flags);
120 		if (ret > 0) {
121 			list_move_tail(&bdev->device_list, &glob->device_list);
122 			break;
123 		}
124 	}
125 	mutex_unlock(&ttm_global_mutex);
126 	return ret;
127 }
128 EXPORT_SYMBOL(ttm_global_swapout);
129 
130 int ttm_device_swapout(struct ttm_device *bdev, struct ttm_operation_ctx *ctx,
131 		       gfp_t gfp_flags)
132 {
133 	struct ttm_resource_manager *man;
134 	struct ttm_buffer_object *bo;
135 	unsigned i, j;
136 	int ret;
137 
138 	spin_lock(&bdev->lru_lock);
139 	for (i = TTM_PL_SYSTEM; i < TTM_NUM_MEM_TYPES; ++i) {
140 		man = ttm_manager_type(bdev, i);
141 		if (!man || !man->use_tt)
142 			continue;
143 
144 		for (j = 0; j < TTM_MAX_BO_PRIORITY; ++j) {
145 			list_for_each_entry(bo, &man->lru[j], lru) {
146 				uint32_t num_pages = PFN_UP(bo->base.size);
147 
148 				ret = ttm_bo_swapout(bo, ctx, gfp_flags);
149 				/* ttm_bo_swapout has dropped the lru_lock */
150 				if (!ret)
151 					return num_pages;
152 				if (ret != -EBUSY)
153 					return ret;
154 			}
155 		}
156 	}
157 	spin_unlock(&bdev->lru_lock);
158 	return 0;
159 }
160 EXPORT_SYMBOL(ttm_device_swapout);
161 
162 static void ttm_init_sysman(struct ttm_device *bdev)
163 {
164 	struct ttm_resource_manager *man = &bdev->sysman;
165 
166 	/*
167 	 * Initialize the system memory buffer type.
168 	 * Other types need to be driver / IOCTL initialized.
169 	 */
170 	man->use_tt = true;
171 
172 	ttm_resource_manager_init(man, 0);
173 	ttm_set_driver_manager(bdev, TTM_PL_SYSTEM, man);
174 	ttm_resource_manager_set_used(man, true);
175 }
176 
177 static void ttm_device_delayed_workqueue(struct work_struct *work)
178 {
179 	struct ttm_device *bdev =
180 		container_of(work, struct ttm_device, wq.work);
181 
182 	if (!ttm_bo_delayed_delete(bdev, false))
183 		schedule_delayed_work(&bdev->wq,
184 				      ((HZ / 100) < 1) ? 1 : HZ / 100);
185 }
186 
187 /**
188  * ttm_device_init
189  *
190  * @bdev: A pointer to a struct ttm_device to initialize.
191  * @funcs: Function table for the device.
192  * @dev: The core kernel device pointer for DMA mappings and allocations.
193  * @mapping: The address space to use for this bo.
194  * @vma_manager: A pointer to a vma manager.
195  * @use_dma_alloc: If coherent DMA allocation API should be used.
196  * @use_dma32: If we should use GFP_DMA32 for device memory allocations.
197  *
198  * Initializes a struct ttm_device:
199  * Returns:
200  * !0: Failure.
201  */
202 int ttm_device_init(struct ttm_device *bdev, struct ttm_device_funcs *funcs,
203 		    struct device *dev, struct address_space *mapping,
204 		    struct drm_vma_offset_manager *vma_manager,
205 		    bool use_dma_alloc, bool use_dma32)
206 {
207 	struct ttm_global *glob = &ttm_glob;
208 	int ret;
209 
210 	if (WARN_ON(vma_manager == NULL))
211 		return -EINVAL;
212 
213 	ret = ttm_global_init();
214 	if (ret)
215 		return ret;
216 
217 	bdev->funcs = funcs;
218 
219 	ttm_init_sysman(bdev);
220 	ttm_pool_init(&bdev->pool, dev, use_dma_alloc, use_dma32);
221 
222 	bdev->vma_manager = vma_manager;
223 	INIT_DELAYED_WORK(&bdev->wq, ttm_device_delayed_workqueue);
224 	spin_lock_init(&bdev->lru_lock);
225 	INIT_LIST_HEAD(&bdev->ddestroy);
226 	bdev->dev_mapping = mapping;
227 	mutex_lock(&ttm_global_mutex);
228 	list_add_tail(&bdev->device_list, &glob->device_list);
229 	mutex_unlock(&ttm_global_mutex);
230 
231 	return 0;
232 }
233 EXPORT_SYMBOL(ttm_device_init);
234 
235 void ttm_device_fini(struct ttm_device *bdev)
236 {
237 	struct ttm_resource_manager *man;
238 	unsigned i;
239 
240 	man = ttm_manager_type(bdev, TTM_PL_SYSTEM);
241 	ttm_resource_manager_set_used(man, false);
242 	ttm_set_driver_manager(bdev, TTM_PL_SYSTEM, NULL);
243 
244 	mutex_lock(&ttm_global_mutex);
245 	list_del(&bdev->device_list);
246 	mutex_unlock(&ttm_global_mutex);
247 
248 	cancel_delayed_work_sync(&bdev->wq);
249 
250 	if (ttm_bo_delayed_delete(bdev, true))
251 		pr_debug("Delayed destroy list was clean\n");
252 
253 	spin_lock(&bdev->lru_lock);
254 	for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i)
255 		if (list_empty(&man->lru[0]))
256 			pr_debug("Swap list %d was clean\n", i);
257 	spin_unlock(&bdev->lru_lock);
258 
259 	ttm_pool_fini(&bdev->pool);
260 	ttm_global_release();
261 }
262 EXPORT_SYMBOL(ttm_device_fini);
263