1 /* 2 * Copyright 2016 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 * 22 * Authors: Christian König 23 */ 24 25 #include <drm/drmP.h> 26 #include "amdgpu.h" 27 28 struct amdgpu_vram_mgr { 29 struct drm_mm mm; 30 spinlock_t lock; 31 atomic64_t usage; 32 atomic64_t vis_usage; 33 }; 34 35 /** 36 * amdgpu_vram_mgr_init - init VRAM manager and DRM MM 37 * 38 * @man: TTM memory type manager 39 * @p_size: maximum size of VRAM 40 * 41 * Allocate and initialize the VRAM manager. 42 */ 43 static int amdgpu_vram_mgr_init(struct ttm_mem_type_manager *man, 44 unsigned long p_size) 45 { 46 struct amdgpu_vram_mgr *mgr; 47 48 mgr = kzalloc(sizeof(*mgr), GFP_KERNEL); 49 if (!mgr) 50 return -ENOMEM; 51 52 drm_mm_init(&mgr->mm, 0, p_size); 53 spin_lock_init(&mgr->lock); 54 man->priv = mgr; 55 return 0; 56 } 57 58 /** 59 * amdgpu_vram_mgr_fini - free and destroy VRAM manager 60 * 61 * @man: TTM memory type manager 62 * 63 * Destroy and free the VRAM manager, returns -EBUSY if ranges are still 64 * allocated inside it. 65 */ 66 static int amdgpu_vram_mgr_fini(struct ttm_mem_type_manager *man) 67 { 68 struct amdgpu_vram_mgr *mgr = man->priv; 69 70 spin_lock(&mgr->lock); 71 if (!drm_mm_clean(&mgr->mm)) { 72 spin_unlock(&mgr->lock); 73 return -EBUSY; 74 } 75 76 drm_mm_takedown(&mgr->mm); 77 spin_unlock(&mgr->lock); 78 kfree(mgr); 79 man->priv = NULL; 80 return 0; 81 } 82 83 /** 84 * amdgpu_vram_mgr_vis_size - Calculate visible node size 85 * 86 * @adev: amdgpu device structure 87 * @node: MM node structure 88 * 89 * Calculate how many bytes of the MM node are inside visible VRAM 90 */ 91 static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev, 92 struct drm_mm_node *node) 93 { 94 uint64_t start = node->start << PAGE_SHIFT; 95 uint64_t end = (node->size + node->start) << PAGE_SHIFT; 96 97 if (start >= adev->mc.visible_vram_size) 98 return 0; 99 100 return (end > adev->mc.visible_vram_size ? 101 adev->mc.visible_vram_size : end) - start; 102 } 103 104 /** 105 * amdgpu_vram_mgr_new - allocate new ranges 106 * 107 * @man: TTM memory type manager 108 * @tbo: TTM BO we need this range for 109 * @place: placement flags and restrictions 110 * @mem: the resulting mem object 111 * 112 * Allocate VRAM for the given BO. 113 */ 114 static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man, 115 struct ttm_buffer_object *tbo, 116 const struct ttm_place *place, 117 struct ttm_mem_reg *mem) 118 { 119 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev); 120 struct amdgpu_vram_mgr *mgr = man->priv; 121 struct drm_mm *mm = &mgr->mm; 122 struct drm_mm_node *nodes; 123 enum drm_mm_insert_mode mode; 124 unsigned long lpfn, num_nodes, pages_per_node, pages_left; 125 uint64_t usage = 0, vis_usage = 0; 126 unsigned i; 127 int r; 128 129 lpfn = place->lpfn; 130 if (!lpfn) 131 lpfn = man->size; 132 133 if (place->flags & TTM_PL_FLAG_CONTIGUOUS || 134 amdgpu_vram_page_split == -1) { 135 pages_per_node = ~0ul; 136 num_nodes = 1; 137 } else { 138 pages_per_node = max((uint32_t)amdgpu_vram_page_split, 139 mem->page_alignment); 140 num_nodes = DIV_ROUND_UP(mem->num_pages, pages_per_node); 141 } 142 143 nodes = kcalloc(num_nodes, sizeof(*nodes), GFP_KERNEL); 144 if (!nodes) 145 return -ENOMEM; 146 147 mode = DRM_MM_INSERT_BEST; 148 if (place->flags & TTM_PL_FLAG_TOPDOWN) 149 mode = DRM_MM_INSERT_HIGH; 150 151 mem->start = 0; 152 pages_left = mem->num_pages; 153 154 spin_lock(&mgr->lock); 155 for (i = 0; i < num_nodes; ++i) { 156 unsigned long pages = min(pages_left, pages_per_node); 157 uint32_t alignment = mem->page_alignment; 158 unsigned long start; 159 160 if (pages == pages_per_node) 161 alignment = pages_per_node; 162 163 r = drm_mm_insert_node_in_range(mm, &nodes[i], 164 pages, alignment, 0, 165 place->fpfn, lpfn, 166 mode); 167 if (unlikely(r)) 168 goto error; 169 170 usage += nodes[i].size << PAGE_SHIFT; 171 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]); 172 173 /* Calculate a virtual BO start address to easily check if 174 * everything is CPU accessible. 175 */ 176 start = nodes[i].start + nodes[i].size; 177 if (start > mem->num_pages) 178 start -= mem->num_pages; 179 else 180 start = 0; 181 mem->start = max(mem->start, start); 182 pages_left -= pages; 183 } 184 spin_unlock(&mgr->lock); 185 186 atomic64_add(usage, &mgr->usage); 187 atomic64_add(vis_usage, &mgr->vis_usage); 188 189 mem->mm_node = nodes; 190 191 return 0; 192 193 error: 194 while (i--) 195 drm_mm_remove_node(&nodes[i]); 196 spin_unlock(&mgr->lock); 197 198 kfree(nodes); 199 return r == -ENOSPC ? 0 : r; 200 } 201 202 /** 203 * amdgpu_vram_mgr_del - free ranges 204 * 205 * @man: TTM memory type manager 206 * @tbo: TTM BO we need this range for 207 * @place: placement flags and restrictions 208 * @mem: TTM memory object 209 * 210 * Free the allocated VRAM again. 211 */ 212 static void amdgpu_vram_mgr_del(struct ttm_mem_type_manager *man, 213 struct ttm_mem_reg *mem) 214 { 215 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev); 216 struct amdgpu_vram_mgr *mgr = man->priv; 217 struct drm_mm_node *nodes = mem->mm_node; 218 uint64_t usage = 0, vis_usage = 0; 219 unsigned pages = mem->num_pages; 220 221 if (!mem->mm_node) 222 return; 223 224 spin_lock(&mgr->lock); 225 while (pages) { 226 pages -= nodes->size; 227 drm_mm_remove_node(nodes); 228 usage += nodes->size << PAGE_SHIFT; 229 vis_usage += amdgpu_vram_mgr_vis_size(adev, nodes); 230 ++nodes; 231 } 232 spin_unlock(&mgr->lock); 233 234 atomic64_sub(usage, &mgr->usage); 235 atomic64_sub(vis_usage, &mgr->vis_usage); 236 237 kfree(mem->mm_node); 238 mem->mm_node = NULL; 239 } 240 241 /** 242 * amdgpu_vram_mgr_usage - how many bytes are used in this domain 243 * 244 * @man: TTM memory type manager 245 * 246 * Returns how many bytes are used in this domain. 247 */ 248 uint64_t amdgpu_vram_mgr_usage(struct ttm_mem_type_manager *man) 249 { 250 struct amdgpu_vram_mgr *mgr = man->priv; 251 252 return atomic64_read(&mgr->usage); 253 } 254 255 /** 256 * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part 257 * 258 * @man: TTM memory type manager 259 * 260 * Returns how many bytes are used in the visible part of VRAM 261 */ 262 uint64_t amdgpu_vram_mgr_vis_usage(struct ttm_mem_type_manager *man) 263 { 264 struct amdgpu_vram_mgr *mgr = man->priv; 265 266 return atomic64_read(&mgr->vis_usage); 267 } 268 269 /** 270 * amdgpu_vram_mgr_debug - dump VRAM table 271 * 272 * @man: TTM memory type manager 273 * @printer: DRM printer to use 274 * 275 * Dump the table content using printk. 276 */ 277 static void amdgpu_vram_mgr_debug(struct ttm_mem_type_manager *man, 278 struct drm_printer *printer) 279 { 280 struct amdgpu_vram_mgr *mgr = man->priv; 281 282 spin_lock(&mgr->lock); 283 drm_mm_print(&mgr->mm, printer); 284 spin_unlock(&mgr->lock); 285 286 drm_printf(printer, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n", 287 man->size, amdgpu_vram_mgr_usage(man) >> 20, 288 amdgpu_vram_mgr_vis_usage(man) >> 20); 289 } 290 291 const struct ttm_mem_type_manager_func amdgpu_vram_mgr_func = { 292 .init = amdgpu_vram_mgr_init, 293 .takedown = amdgpu_vram_mgr_fini, 294 .get_node = amdgpu_vram_mgr_new, 295 .put_node = amdgpu_vram_mgr_del, 296 .debug = amdgpu_vram_mgr_debug 297 }; 298