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 drm_mm_takedown(&mgr->mm); 72 spin_unlock(&mgr->lock); 73 kfree(mgr); 74 man->priv = NULL; 75 return 0; 76 } 77 78 /** 79 * amdgpu_vram_mgr_vis_size - Calculate visible node size 80 * 81 * @adev: amdgpu device structure 82 * @node: MM node structure 83 * 84 * Calculate how many bytes of the MM node are inside visible VRAM 85 */ 86 static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev, 87 struct drm_mm_node *node) 88 { 89 uint64_t start = node->start << PAGE_SHIFT; 90 uint64_t end = (node->size + node->start) << PAGE_SHIFT; 91 92 if (start >= adev->gmc.visible_vram_size) 93 return 0; 94 95 return (end > adev->gmc.visible_vram_size ? 96 adev->gmc.visible_vram_size : end) - start; 97 } 98 99 /** 100 * amdgpu_vram_mgr_bo_invisible_size - CPU invisible BO size 101 * 102 * @bo: &amdgpu_bo buffer object (must be in VRAM) 103 * 104 * Returns: 105 * How much of the given &amdgpu_bo buffer object lies in CPU invisible VRAM. 106 */ 107 u64 amdgpu_vram_mgr_bo_invisible_size(struct amdgpu_bo *bo) 108 { 109 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 110 struct ttm_mem_reg *mem = &bo->tbo.mem; 111 struct drm_mm_node *nodes = mem->mm_node; 112 unsigned pages = mem->num_pages; 113 u64 usage = 0; 114 115 if (adev->gmc.visible_vram_size == adev->gmc.real_vram_size) 116 return 0; 117 118 if (mem->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT) 119 return amdgpu_bo_size(bo); 120 121 while (nodes && pages) { 122 usage += nodes->size << PAGE_SHIFT; 123 usage -= amdgpu_vram_mgr_vis_size(adev, nodes); 124 pages -= nodes->size; 125 ++nodes; 126 } 127 128 return usage; 129 } 130 131 /** 132 * amdgpu_vram_mgr_new - allocate new ranges 133 * 134 * @man: TTM memory type manager 135 * @tbo: TTM BO we need this range for 136 * @place: placement flags and restrictions 137 * @mem: the resulting mem object 138 * 139 * Allocate VRAM for the given BO. 140 */ 141 static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man, 142 struct ttm_buffer_object *tbo, 143 const struct ttm_place *place, 144 struct ttm_mem_reg *mem) 145 { 146 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev); 147 struct amdgpu_vram_mgr *mgr = man->priv; 148 struct drm_mm *mm = &mgr->mm; 149 struct drm_mm_node *nodes; 150 enum drm_mm_insert_mode mode; 151 unsigned long lpfn, num_nodes, pages_per_node, pages_left; 152 uint64_t usage = 0, vis_usage = 0; 153 unsigned i; 154 int r; 155 156 lpfn = place->lpfn; 157 if (!lpfn) 158 lpfn = man->size; 159 160 if (place->flags & TTM_PL_FLAG_CONTIGUOUS || 161 amdgpu_vram_page_split == -1) { 162 pages_per_node = ~0ul; 163 num_nodes = 1; 164 } else { 165 pages_per_node = max((uint32_t)amdgpu_vram_page_split, 166 mem->page_alignment); 167 num_nodes = DIV_ROUND_UP(mem->num_pages, pages_per_node); 168 } 169 170 nodes = kvmalloc_array(num_nodes, sizeof(*nodes), 171 GFP_KERNEL | __GFP_ZERO); 172 if (!nodes) 173 return -ENOMEM; 174 175 mode = DRM_MM_INSERT_BEST; 176 if (place->flags & TTM_PL_FLAG_TOPDOWN) 177 mode = DRM_MM_INSERT_HIGH; 178 179 mem->start = 0; 180 pages_left = mem->num_pages; 181 182 spin_lock(&mgr->lock); 183 for (i = 0; i < num_nodes; ++i) { 184 unsigned long pages = min(pages_left, pages_per_node); 185 uint32_t alignment = mem->page_alignment; 186 unsigned long start; 187 188 if (pages == pages_per_node) 189 alignment = pages_per_node; 190 191 r = drm_mm_insert_node_in_range(mm, &nodes[i], 192 pages, alignment, 0, 193 place->fpfn, lpfn, 194 mode); 195 if (unlikely(r)) 196 goto error; 197 198 usage += nodes[i].size << PAGE_SHIFT; 199 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]); 200 201 /* Calculate a virtual BO start address to easily check if 202 * everything is CPU accessible. 203 */ 204 start = nodes[i].start + nodes[i].size; 205 if (start > mem->num_pages) 206 start -= mem->num_pages; 207 else 208 start = 0; 209 mem->start = max(mem->start, start); 210 pages_left -= pages; 211 } 212 spin_unlock(&mgr->lock); 213 214 atomic64_add(usage, &mgr->usage); 215 atomic64_add(vis_usage, &mgr->vis_usage); 216 217 mem->mm_node = nodes; 218 219 return 0; 220 221 error: 222 while (i--) 223 drm_mm_remove_node(&nodes[i]); 224 spin_unlock(&mgr->lock); 225 226 kvfree(nodes); 227 return r == -ENOSPC ? 0 : r; 228 } 229 230 /** 231 * amdgpu_vram_mgr_del - free ranges 232 * 233 * @man: TTM memory type manager 234 * @tbo: TTM BO we need this range for 235 * @place: placement flags and restrictions 236 * @mem: TTM memory object 237 * 238 * Free the allocated VRAM again. 239 */ 240 static void amdgpu_vram_mgr_del(struct ttm_mem_type_manager *man, 241 struct ttm_mem_reg *mem) 242 { 243 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev); 244 struct amdgpu_vram_mgr *mgr = man->priv; 245 struct drm_mm_node *nodes = mem->mm_node; 246 uint64_t usage = 0, vis_usage = 0; 247 unsigned pages = mem->num_pages; 248 249 if (!mem->mm_node) 250 return; 251 252 spin_lock(&mgr->lock); 253 while (pages) { 254 pages -= nodes->size; 255 drm_mm_remove_node(nodes); 256 usage += nodes->size << PAGE_SHIFT; 257 vis_usage += amdgpu_vram_mgr_vis_size(adev, nodes); 258 ++nodes; 259 } 260 spin_unlock(&mgr->lock); 261 262 atomic64_sub(usage, &mgr->usage); 263 atomic64_sub(vis_usage, &mgr->vis_usage); 264 265 kvfree(mem->mm_node); 266 mem->mm_node = NULL; 267 } 268 269 /** 270 * amdgpu_vram_mgr_usage - how many bytes are used in this domain 271 * 272 * @man: TTM memory type manager 273 * 274 * Returns how many bytes are used in this domain. 275 */ 276 uint64_t amdgpu_vram_mgr_usage(struct ttm_mem_type_manager *man) 277 { 278 struct amdgpu_vram_mgr *mgr = man->priv; 279 280 return atomic64_read(&mgr->usage); 281 } 282 283 /** 284 * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part 285 * 286 * @man: TTM memory type manager 287 * 288 * Returns how many bytes are used in the visible part of VRAM 289 */ 290 uint64_t amdgpu_vram_mgr_vis_usage(struct ttm_mem_type_manager *man) 291 { 292 struct amdgpu_vram_mgr *mgr = man->priv; 293 294 return atomic64_read(&mgr->vis_usage); 295 } 296 297 /** 298 * amdgpu_vram_mgr_debug - dump VRAM table 299 * 300 * @man: TTM memory type manager 301 * @printer: DRM printer to use 302 * 303 * Dump the table content using printk. 304 */ 305 static void amdgpu_vram_mgr_debug(struct ttm_mem_type_manager *man, 306 struct drm_printer *printer) 307 { 308 struct amdgpu_vram_mgr *mgr = man->priv; 309 310 spin_lock(&mgr->lock); 311 drm_mm_print(&mgr->mm, printer); 312 spin_unlock(&mgr->lock); 313 314 drm_printf(printer, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n", 315 man->size, amdgpu_vram_mgr_usage(man) >> 20, 316 amdgpu_vram_mgr_vis_usage(man) >> 20); 317 } 318 319 const struct ttm_mem_type_manager_func amdgpu_vram_mgr_func = { 320 .init = amdgpu_vram_mgr_init, 321 .takedown = amdgpu_vram_mgr_fini, 322 .get_node = amdgpu_vram_mgr_new, 323 .put_node = amdgpu_vram_mgr_del, 324 .debug = amdgpu_vram_mgr_debug 325 }; 326