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_allocation.h> 35 #include <drm/ttm/ttm_bo.h> 36 #include <drm/ttm/ttm_device.h> 37 #include <drm/ttm/ttm_tt.h> 38 #include <drm/ttm/ttm_placement.h> 39 40 #include "ttm_module.h" 41 #include "ttm_bo_internal.h" 42 43 /* 44 * ttm_global_mutex - protecting the global state 45 */ 46 static DEFINE_MUTEX(ttm_global_mutex); 47 static unsigned ttm_glob_use_count; 48 struct ttm_global ttm_glob; 49 EXPORT_SYMBOL(ttm_glob); 50 51 struct dentry *ttm_debugfs_root; 52 53 static void ttm_global_release(void) 54 { 55 struct ttm_global *glob = &ttm_glob; 56 57 mutex_lock(&ttm_global_mutex); 58 if (--ttm_glob_use_count > 0) 59 goto out; 60 61 ttm_pool_mgr_fini(); 62 debugfs_remove(ttm_debugfs_root); 63 64 __free_page(glob->dummy_read_page); 65 memset(glob, 0, sizeof(*glob)); 66 out: 67 mutex_unlock(&ttm_global_mutex); 68 } 69 70 static int ttm_global_init(void) 71 { 72 struct ttm_global *glob = &ttm_glob; 73 unsigned long num_pages, num_dma32; 74 struct sysinfo si; 75 int ret = 0; 76 77 mutex_lock(&ttm_global_mutex); 78 if (++ttm_glob_use_count > 1) 79 goto out; 80 81 si_meminfo(&si); 82 83 ttm_debugfs_root = debugfs_create_dir("ttm", NULL); 84 if (IS_ERR(ttm_debugfs_root)) { 85 ttm_debugfs_root = NULL; 86 } 87 88 /* Limit the number of pages in the pool to about 50% of the total 89 * system memory. 90 */ 91 num_pages = ((u64)si.totalram * si.mem_unit) >> PAGE_SHIFT; 92 num_pages /= 2; 93 94 /* But for DMA32 we limit ourself to only use 2GiB maximum. */ 95 num_dma32 = (u64)(si.totalram - si.totalhigh) * si.mem_unit 96 >> PAGE_SHIFT; 97 num_dma32 = min(num_dma32, 2UL << (30 - PAGE_SHIFT)); 98 99 ttm_pool_mgr_init(num_pages); 100 ttm_tt_mgr_init(num_pages, num_dma32); 101 102 glob->dummy_read_page = alloc_page(__GFP_ZERO | GFP_DMA32 | 103 __GFP_NOWARN); 104 105 /* Retry without GFP_DMA32 for platforms DMA32 is not available */ 106 if (unlikely(glob->dummy_read_page == NULL)) { 107 glob->dummy_read_page = alloc_page(__GFP_ZERO); 108 if (unlikely(glob->dummy_read_page == NULL)) { 109 ret = -ENOMEM; 110 goto out; 111 } 112 pr_warn("Using GFP_DMA32 fallback for dummy_read_page\n"); 113 } 114 115 INIT_LIST_HEAD(&glob->device_list); 116 atomic_set(&glob->bo_count, 0); 117 118 debugfs_create_atomic_t("buffer_objects", 0444, ttm_debugfs_root, 119 &glob->bo_count); 120 out: 121 if (ret && ttm_debugfs_root) 122 debugfs_remove(ttm_debugfs_root); 123 if (ret) 124 --ttm_glob_use_count; 125 mutex_unlock(&ttm_global_mutex); 126 return ret; 127 } 128 129 /** 130 * ttm_device_prepare_hibernation - move GTT BOs to shmem for hibernation. 131 * 132 * @bdev: A pointer to a struct ttm_device to prepare hibernation for. 133 * 134 * Return: 0 on success, negative number on failure. 135 */ 136 int ttm_device_prepare_hibernation(struct ttm_device *bdev) 137 { 138 struct ttm_operation_ctx ctx = { 139 .interruptible = false, 140 .no_wait_gpu = false, 141 }; 142 int ret; 143 144 do { 145 ret = ttm_device_swapout(bdev, &ctx, GFP_KERNEL); 146 } while (ret > 0); 147 return ret; 148 } 149 EXPORT_SYMBOL(ttm_device_prepare_hibernation); 150 151 /* 152 * A buffer object shrink method that tries to swap out the first 153 * buffer object on the global::swap_lru list. 154 */ 155 int ttm_global_swapout(struct ttm_operation_ctx *ctx, gfp_t gfp_flags) 156 { 157 struct ttm_global *glob = &ttm_glob; 158 struct ttm_device *bdev; 159 int ret = 0; 160 161 mutex_lock(&ttm_global_mutex); 162 list_for_each_entry(bdev, &glob->device_list, device_list) { 163 ret = ttm_device_swapout(bdev, ctx, gfp_flags); 164 if (ret > 0) { 165 list_move_tail(&bdev->device_list, &glob->device_list); 166 break; 167 } 168 } 169 mutex_unlock(&ttm_global_mutex); 170 return ret; 171 } 172 173 int ttm_device_swapout(struct ttm_device *bdev, struct ttm_operation_ctx *ctx, 174 gfp_t gfp_flags) 175 { 176 struct ttm_resource_manager *man; 177 unsigned i; 178 s64 lret; 179 180 for (i = TTM_PL_SYSTEM; i < TTM_NUM_MEM_TYPES; ++i) { 181 man = ttm_manager_type(bdev, i); 182 if (!man || !man->use_tt) 183 continue; 184 185 lret = ttm_bo_swapout(bdev, ctx, man, gfp_flags, 1); 186 /* Can be both positive (num_pages) and negative (error) */ 187 if (lret) 188 return lret; 189 } 190 return 0; 191 } 192 EXPORT_SYMBOL(ttm_device_swapout); 193 194 /** 195 * ttm_device_init 196 * 197 * @bdev: A pointer to a struct ttm_device to initialize. 198 * @funcs: Function table for the device. 199 * @dev: The core kernel device pointer for DMA mappings and allocations. 200 * @mapping: The address space to use for this bo. 201 * @vma_manager: A pointer to a vma manager. 202 * @alloc_flags: TTM_ALLOCATION_ flags. 203 * 204 * Initializes a struct ttm_device: 205 * Returns: 206 * !0: Failure. 207 */ 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 unsigned int alloc_flags) 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->alloc_flags = alloc_flags; 231 bdev->funcs = funcs; 232 233 ttm_sys_man_init(bdev); 234 235 if (dev) 236 nid = dev_to_node(dev); 237 else 238 nid = NUMA_NO_NODE; 239 240 ttm_pool_init(&bdev->pool, dev, nid, alloc_flags); 241 242 bdev->vma_manager = vma_manager; 243 spin_lock_init(&bdev->lru_lock); 244 INIT_LIST_HEAD(&bdev->unevictable); 245 bdev->dev_mapping = mapping; 246 mutex_lock(&ttm_global_mutex); 247 list_add_tail(&bdev->device_list, &glob->device_list); 248 mutex_unlock(&ttm_global_mutex); 249 250 return 0; 251 } 252 EXPORT_SYMBOL(ttm_device_init); 253 254 void ttm_device_fini(struct ttm_device *bdev) 255 { 256 struct ttm_resource_manager *man; 257 unsigned i; 258 259 mutex_lock(&ttm_global_mutex); 260 list_del(&bdev->device_list); 261 mutex_unlock(&ttm_global_mutex); 262 263 drain_workqueue(bdev->wq); 264 destroy_workqueue(bdev->wq); 265 266 man = ttm_manager_type(bdev, TTM_PL_SYSTEM); 267 ttm_resource_manager_set_used(man, false); 268 ttm_set_driver_manager(bdev, TTM_PL_SYSTEM, NULL); 269 270 spin_lock(&bdev->lru_lock); 271 for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) 272 if (list_empty(&man->lru[0])) 273 pr_debug("Swap list %d was clean\n", i); 274 spin_unlock(&bdev->lru_lock); 275 276 ttm_pool_fini(&bdev->pool); 277 ttm_global_release(); 278 } 279 EXPORT_SYMBOL(ttm_device_fini); 280 281 static void ttm_device_clear_lru_dma_mappings(struct ttm_device *bdev, 282 struct list_head *list) 283 { 284 struct ttm_resource *res; 285 286 spin_lock(&bdev->lru_lock); 287 while ((res = ttm_lru_first_res_or_null(list))) { 288 struct ttm_buffer_object *bo = res->bo; 289 290 /* Take ref against racing releases once lru_lock is unlocked */ 291 if (!ttm_bo_get_unless_zero(bo)) 292 continue; 293 294 list_del_init(&bo->resource->lru.link); 295 spin_unlock(&bdev->lru_lock); 296 297 if (bo->ttm) 298 ttm_tt_unpopulate(bo->bdev, bo->ttm); 299 300 ttm_bo_put(bo); 301 spin_lock(&bdev->lru_lock); 302 } 303 spin_unlock(&bdev->lru_lock); 304 } 305 306 void ttm_device_clear_dma_mappings(struct ttm_device *bdev) 307 { 308 struct ttm_resource_manager *man; 309 unsigned int i, j; 310 311 ttm_device_clear_lru_dma_mappings(bdev, &bdev->unevictable); 312 313 for (i = TTM_PL_SYSTEM; i < TTM_NUM_MEM_TYPES; ++i) { 314 man = ttm_manager_type(bdev, i); 315 if (!man || !man->use_tt) 316 continue; 317 318 for (j = 0; j < TTM_MAX_BO_PRIORITY; ++j) 319 ttm_device_clear_lru_dma_mappings(bdev, &man->lru[j]); 320 } 321 } 322 EXPORT_SYMBOL(ttm_device_clear_dma_mappings); 323