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 "amdgpu.h" 26 27 struct amdgpu_gtt_mgr { 28 struct drm_mm mm; 29 spinlock_t lock; 30 atomic64_t available; 31 }; 32 33 struct amdgpu_gtt_node { 34 struct drm_mm_node node; 35 struct ttm_buffer_object *tbo; 36 }; 37 38 /** 39 * DOC: mem_info_gtt_total 40 * 41 * The amdgpu driver provides a sysfs API for reporting current total size of 42 * the GTT. 43 * The file mem_info_gtt_total is used for this, and returns the total size of 44 * the GTT block, in bytes 45 */ 46 static ssize_t amdgpu_mem_info_gtt_total_show(struct device *dev, 47 struct device_attribute *attr, char *buf) 48 { 49 struct drm_device *ddev = dev_get_drvdata(dev); 50 struct amdgpu_device *adev = ddev->dev_private; 51 52 return snprintf(buf, PAGE_SIZE, "%llu\n", 53 (adev->mman.bdev.man[TTM_PL_TT].size) * PAGE_SIZE); 54 } 55 56 /** 57 * DOC: mem_info_gtt_used 58 * 59 * The amdgpu driver provides a sysfs API for reporting current total amount of 60 * used GTT. 61 * The file mem_info_gtt_used is used for this, and returns the current used 62 * size of the GTT block, in bytes 63 */ 64 static ssize_t amdgpu_mem_info_gtt_used_show(struct device *dev, 65 struct device_attribute *attr, char *buf) 66 { 67 struct drm_device *ddev = dev_get_drvdata(dev); 68 struct amdgpu_device *adev = ddev->dev_private; 69 70 return snprintf(buf, PAGE_SIZE, "%llu\n", 71 amdgpu_gtt_mgr_usage(&adev->mman.bdev.man[TTM_PL_TT])); 72 } 73 74 static DEVICE_ATTR(mem_info_gtt_total, S_IRUGO, 75 amdgpu_mem_info_gtt_total_show, NULL); 76 static DEVICE_ATTR(mem_info_gtt_used, S_IRUGO, 77 amdgpu_mem_info_gtt_used_show, NULL); 78 79 /** 80 * amdgpu_gtt_mgr_init - init GTT manager and DRM MM 81 * 82 * @man: TTM memory type manager 83 * @p_size: maximum size of GTT 84 * 85 * Allocate and initialize the GTT manager. 86 */ 87 static int amdgpu_gtt_mgr_init(struct ttm_mem_type_manager *man, 88 unsigned long p_size) 89 { 90 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev); 91 struct amdgpu_gtt_mgr *mgr; 92 uint64_t start, size; 93 int ret; 94 95 mgr = kzalloc(sizeof(*mgr), GFP_KERNEL); 96 if (!mgr) 97 return -ENOMEM; 98 99 start = AMDGPU_GTT_MAX_TRANSFER_SIZE * AMDGPU_GTT_NUM_TRANSFER_WINDOWS; 100 size = (adev->gmc.gart_size >> PAGE_SHIFT) - start; 101 drm_mm_init(&mgr->mm, start, size); 102 spin_lock_init(&mgr->lock); 103 atomic64_set(&mgr->available, p_size); 104 man->priv = mgr; 105 106 ret = device_create_file(adev->dev, &dev_attr_mem_info_gtt_total); 107 if (ret) { 108 DRM_ERROR("Failed to create device file mem_info_gtt_total\n"); 109 return ret; 110 } 111 ret = device_create_file(adev->dev, &dev_attr_mem_info_gtt_used); 112 if (ret) { 113 DRM_ERROR("Failed to create device file mem_info_gtt_used\n"); 114 return ret; 115 } 116 117 return 0; 118 } 119 120 /** 121 * amdgpu_gtt_mgr_fini - free and destroy GTT manager 122 * 123 * @man: TTM memory type manager 124 * 125 * Destroy and free the GTT manager, returns -EBUSY if ranges are still 126 * allocated inside it. 127 */ 128 static int amdgpu_gtt_mgr_fini(struct ttm_mem_type_manager *man) 129 { 130 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev); 131 struct amdgpu_gtt_mgr *mgr = man->priv; 132 spin_lock(&mgr->lock); 133 drm_mm_takedown(&mgr->mm); 134 spin_unlock(&mgr->lock); 135 kfree(mgr); 136 man->priv = NULL; 137 138 device_remove_file(adev->dev, &dev_attr_mem_info_gtt_total); 139 device_remove_file(adev->dev, &dev_attr_mem_info_gtt_used); 140 141 return 0; 142 } 143 144 /** 145 * amdgpu_gtt_mgr_has_gart_addr - Check if mem has address space 146 * 147 * @mem: the mem object to check 148 * 149 * Check if a mem object has already address space allocated. 150 */ 151 bool amdgpu_gtt_mgr_has_gart_addr(struct ttm_mem_reg *mem) 152 { 153 struct amdgpu_gtt_node *node = mem->mm_node; 154 155 return (node->node.start != AMDGPU_BO_INVALID_OFFSET); 156 } 157 158 /** 159 * amdgpu_gtt_mgr_alloc - allocate new ranges 160 * 161 * @man: TTM memory type manager 162 * @tbo: TTM BO we need this range for 163 * @place: placement flags and restrictions 164 * @mem: the resulting mem object 165 * 166 * Allocate the address space for a node. 167 */ 168 static int amdgpu_gtt_mgr_alloc(struct ttm_mem_type_manager *man, 169 struct ttm_buffer_object *tbo, 170 const struct ttm_place *place, 171 struct ttm_mem_reg *mem) 172 { 173 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev); 174 struct amdgpu_gtt_mgr *mgr = man->priv; 175 struct amdgpu_gtt_node *node = mem->mm_node; 176 enum drm_mm_insert_mode mode; 177 unsigned long fpfn, lpfn; 178 int r; 179 180 if (amdgpu_gtt_mgr_has_gart_addr(mem)) 181 return 0; 182 183 if (place) 184 fpfn = place->fpfn; 185 else 186 fpfn = 0; 187 188 if (place && place->lpfn) 189 lpfn = place->lpfn; 190 else 191 lpfn = adev->gart.num_cpu_pages; 192 193 mode = DRM_MM_INSERT_BEST; 194 if (place && place->flags & TTM_PL_FLAG_TOPDOWN) 195 mode = DRM_MM_INSERT_HIGH; 196 197 spin_lock(&mgr->lock); 198 r = drm_mm_insert_node_in_range(&mgr->mm, &node->node, mem->num_pages, 199 mem->page_alignment, 0, fpfn, lpfn, 200 mode); 201 spin_unlock(&mgr->lock); 202 203 if (!r) 204 mem->start = node->node.start; 205 206 return r; 207 } 208 209 /** 210 * amdgpu_gtt_mgr_new - allocate a new node 211 * 212 * @man: TTM memory type manager 213 * @tbo: TTM BO we need this range for 214 * @place: placement flags and restrictions 215 * @mem: the resulting mem object 216 * 217 * Dummy, allocate the node but no space for it yet. 218 */ 219 static int amdgpu_gtt_mgr_new(struct ttm_mem_type_manager *man, 220 struct ttm_buffer_object *tbo, 221 const struct ttm_place *place, 222 struct ttm_mem_reg *mem) 223 { 224 struct amdgpu_gtt_mgr *mgr = man->priv; 225 struct amdgpu_gtt_node *node; 226 int r; 227 228 spin_lock(&mgr->lock); 229 if ((&tbo->mem == mem || tbo->mem.mem_type != TTM_PL_TT) && 230 atomic64_read(&mgr->available) < mem->num_pages) { 231 spin_unlock(&mgr->lock); 232 return 0; 233 } 234 atomic64_sub(mem->num_pages, &mgr->available); 235 spin_unlock(&mgr->lock); 236 237 node = kzalloc(sizeof(*node), GFP_KERNEL); 238 if (!node) { 239 r = -ENOMEM; 240 goto err_out; 241 } 242 243 node->node.start = AMDGPU_BO_INVALID_OFFSET; 244 node->node.size = mem->num_pages; 245 node->tbo = tbo; 246 mem->mm_node = node; 247 248 if (place->fpfn || place->lpfn || place->flags & TTM_PL_FLAG_TOPDOWN) { 249 r = amdgpu_gtt_mgr_alloc(man, tbo, place, mem); 250 if (unlikely(r)) { 251 kfree(node); 252 mem->mm_node = NULL; 253 r = 0; 254 goto err_out; 255 } 256 } else { 257 mem->start = node->node.start; 258 } 259 260 return 0; 261 err_out: 262 atomic64_add(mem->num_pages, &mgr->available); 263 264 return r; 265 } 266 267 /** 268 * amdgpu_gtt_mgr_del - free ranges 269 * 270 * @man: TTM memory type manager 271 * @mem: TTM memory object 272 * 273 * Free the allocated GTT again. 274 */ 275 static void amdgpu_gtt_mgr_del(struct ttm_mem_type_manager *man, 276 struct ttm_mem_reg *mem) 277 { 278 struct amdgpu_gtt_mgr *mgr = man->priv; 279 struct amdgpu_gtt_node *node = mem->mm_node; 280 281 if (!node) 282 return; 283 284 spin_lock(&mgr->lock); 285 if (node->node.start != AMDGPU_BO_INVALID_OFFSET) 286 drm_mm_remove_node(&node->node); 287 spin_unlock(&mgr->lock); 288 atomic64_add(mem->num_pages, &mgr->available); 289 290 kfree(node); 291 mem->mm_node = NULL; 292 } 293 294 /** 295 * amdgpu_gtt_mgr_usage - return usage of GTT domain 296 * 297 * @man: TTM memory type manager 298 * 299 * Return how many bytes are used in the GTT domain 300 */ 301 uint64_t amdgpu_gtt_mgr_usage(struct ttm_mem_type_manager *man) 302 { 303 struct amdgpu_gtt_mgr *mgr = man->priv; 304 s64 result = man->size - atomic64_read(&mgr->available); 305 306 return (result > 0 ? result : 0) * PAGE_SIZE; 307 } 308 309 int amdgpu_gtt_mgr_recover(struct ttm_mem_type_manager *man) 310 { 311 struct amdgpu_gtt_mgr *mgr = man->priv; 312 struct amdgpu_gtt_node *node; 313 struct drm_mm_node *mm_node; 314 int r = 0; 315 316 spin_lock(&mgr->lock); 317 drm_mm_for_each_node(mm_node, &mgr->mm) { 318 node = container_of(mm_node, struct amdgpu_gtt_node, node); 319 r = amdgpu_ttm_recover_gart(node->tbo); 320 if (r) 321 break; 322 } 323 spin_unlock(&mgr->lock); 324 325 return r; 326 } 327 328 /** 329 * amdgpu_gtt_mgr_debug - dump VRAM table 330 * 331 * @man: TTM memory type manager 332 * @printer: DRM printer to use 333 * 334 * Dump the table content using printk. 335 */ 336 static void amdgpu_gtt_mgr_debug(struct ttm_mem_type_manager *man, 337 struct drm_printer *printer) 338 { 339 struct amdgpu_gtt_mgr *mgr = man->priv; 340 341 spin_lock(&mgr->lock); 342 drm_mm_print(&mgr->mm, printer); 343 spin_unlock(&mgr->lock); 344 345 drm_printf(printer, "man size:%llu pages, gtt available:%lld pages, usage:%lluMB\n", 346 man->size, (u64)atomic64_read(&mgr->available), 347 amdgpu_gtt_mgr_usage(man) >> 20); 348 } 349 350 const struct ttm_mem_type_manager_func amdgpu_gtt_mgr_func = { 351 .init = amdgpu_gtt_mgr_init, 352 .takedown = amdgpu_gtt_mgr_fini, 353 .get_node = amdgpu_gtt_mgr_new, 354 .put_node = amdgpu_gtt_mgr_del, 355 .debug = amdgpu_gtt_mgr_debug 356 }; 357