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_gtt_mgr { 29 struct drm_mm mm; 30 spinlock_t lock; 31 atomic64_t available; 32 }; 33 34 struct amdgpu_gtt_node { 35 struct drm_mm_node node; 36 struct ttm_buffer_object *tbo; 37 }; 38 39 /** 40 * DOC: mem_info_gtt_total 41 * 42 * The amdgpu driver provides a sysfs API for reporting current total size of 43 * the GTT. 44 * The file mem_info_gtt_total is used for this, and returns the total size of 45 * the GTT block, in bytes 46 */ 47 static ssize_t amdgpu_mem_info_gtt_total_show(struct device *dev, 48 struct device_attribute *attr, char *buf) 49 { 50 struct drm_device *ddev = dev_get_drvdata(dev); 51 struct amdgpu_device *adev = ddev->dev_private; 52 53 return snprintf(buf, PAGE_SIZE, "%llu\n", 54 (adev->mman.bdev.man[TTM_PL_TT].size) * PAGE_SIZE); 55 } 56 57 /** 58 * DOC: mem_info_gtt_used 59 * 60 * The amdgpu driver provides a sysfs API for reporting current total amount of 61 * used GTT. 62 * The file mem_info_gtt_used is used for this, and returns the current used 63 * size of the GTT block, in bytes 64 */ 65 static ssize_t amdgpu_mem_info_gtt_used_show(struct device *dev, 66 struct device_attribute *attr, char *buf) 67 { 68 struct drm_device *ddev = dev_get_drvdata(dev); 69 struct amdgpu_device *adev = ddev->dev_private; 70 71 return snprintf(buf, PAGE_SIZE, "%llu\n", 72 amdgpu_gtt_mgr_usage(&adev->mman.bdev.man[TTM_PL_TT])); 73 } 74 75 static DEVICE_ATTR(mem_info_gtt_total, S_IRUGO, 76 amdgpu_mem_info_gtt_total_show, NULL); 77 static DEVICE_ATTR(mem_info_gtt_used, S_IRUGO, 78 amdgpu_mem_info_gtt_used_show, NULL); 79 80 /** 81 * amdgpu_gtt_mgr_init - init GTT manager and DRM MM 82 * 83 * @man: TTM memory type manager 84 * @p_size: maximum size of GTT 85 * 86 * Allocate and initialize the GTT manager. 87 */ 88 static int amdgpu_gtt_mgr_init(struct ttm_mem_type_manager *man, 89 unsigned long p_size) 90 { 91 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev); 92 struct amdgpu_gtt_mgr *mgr; 93 uint64_t start, size; 94 int ret; 95 96 mgr = kzalloc(sizeof(*mgr), GFP_KERNEL); 97 if (!mgr) 98 return -ENOMEM; 99 100 start = AMDGPU_GTT_MAX_TRANSFER_SIZE * AMDGPU_GTT_NUM_TRANSFER_WINDOWS; 101 size = (adev->gmc.gart_size >> PAGE_SHIFT) - start; 102 drm_mm_init(&mgr->mm, start, size); 103 spin_lock_init(&mgr->lock); 104 atomic64_set(&mgr->available, p_size); 105 man->priv = mgr; 106 107 ret = device_create_file(adev->dev, &dev_attr_mem_info_gtt_total); 108 if (ret) { 109 DRM_ERROR("Failed to create device file mem_info_gtt_total\n"); 110 return ret; 111 } 112 ret = device_create_file(adev->dev, &dev_attr_mem_info_gtt_used); 113 if (ret) { 114 DRM_ERROR("Failed to create device file mem_info_gtt_used\n"); 115 return ret; 116 } 117 118 return 0; 119 } 120 121 /** 122 * amdgpu_gtt_mgr_fini - free and destroy GTT manager 123 * 124 * @man: TTM memory type manager 125 * 126 * Destroy and free the GTT manager, returns -EBUSY if ranges are still 127 * allocated inside it. 128 */ 129 static int amdgpu_gtt_mgr_fini(struct ttm_mem_type_manager *man) 130 { 131 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev); 132 struct amdgpu_gtt_mgr *mgr = man->priv; 133 spin_lock(&mgr->lock); 134 drm_mm_takedown(&mgr->mm); 135 spin_unlock(&mgr->lock); 136 kfree(mgr); 137 man->priv = NULL; 138 139 device_remove_file(adev->dev, &dev_attr_mem_info_gtt_total); 140 device_remove_file(adev->dev, &dev_attr_mem_info_gtt_used); 141 142 return 0; 143 } 144 145 /** 146 * amdgpu_gtt_mgr_has_gart_addr - Check if mem has address space 147 * 148 * @mem: the mem object to check 149 * 150 * Check if a mem object has already address space allocated. 151 */ 152 bool amdgpu_gtt_mgr_has_gart_addr(struct ttm_mem_reg *mem) 153 { 154 struct amdgpu_gtt_node *node = mem->mm_node; 155 156 return (node->node.start != AMDGPU_BO_INVALID_OFFSET); 157 } 158 159 /** 160 * amdgpu_gtt_mgr_alloc - allocate new ranges 161 * 162 * @man: TTM memory type manager 163 * @tbo: TTM BO we need this range for 164 * @place: placement flags and restrictions 165 * @mem: the resulting mem object 166 * 167 * Allocate the address space for a node. 168 */ 169 static int amdgpu_gtt_mgr_alloc(struct ttm_mem_type_manager *man, 170 struct ttm_buffer_object *tbo, 171 const struct ttm_place *place, 172 struct ttm_mem_reg *mem) 173 { 174 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev); 175 struct amdgpu_gtt_mgr *mgr = man->priv; 176 struct amdgpu_gtt_node *node = mem->mm_node; 177 enum drm_mm_insert_mode mode; 178 unsigned long fpfn, lpfn; 179 int r; 180 181 if (amdgpu_gtt_mgr_has_gart_addr(mem)) 182 return 0; 183 184 if (place) 185 fpfn = place->fpfn; 186 else 187 fpfn = 0; 188 189 if (place && place->lpfn) 190 lpfn = place->lpfn; 191 else 192 lpfn = adev->gart.num_cpu_pages; 193 194 mode = DRM_MM_INSERT_BEST; 195 if (place && place->flags & TTM_PL_FLAG_TOPDOWN) 196 mode = DRM_MM_INSERT_HIGH; 197 198 spin_lock(&mgr->lock); 199 r = drm_mm_insert_node_in_range(&mgr->mm, &node->node, mem->num_pages, 200 mem->page_alignment, 0, fpfn, lpfn, 201 mode); 202 spin_unlock(&mgr->lock); 203 204 if (!r) 205 mem->start = node->node.start; 206 207 return r; 208 } 209 210 /** 211 * amdgpu_gtt_mgr_new - allocate a new node 212 * 213 * @man: TTM memory type manager 214 * @tbo: TTM BO we need this range for 215 * @place: placement flags and restrictions 216 * @mem: the resulting mem object 217 * 218 * Dummy, allocate the node but no space for it yet. 219 */ 220 static int amdgpu_gtt_mgr_new(struct ttm_mem_type_manager *man, 221 struct ttm_buffer_object *tbo, 222 const struct ttm_place *place, 223 struct ttm_mem_reg *mem) 224 { 225 struct amdgpu_gtt_mgr *mgr = man->priv; 226 struct amdgpu_gtt_node *node; 227 int r; 228 229 spin_lock(&mgr->lock); 230 if ((&tbo->mem == mem || tbo->mem.mem_type != TTM_PL_TT) && 231 atomic64_read(&mgr->available) < mem->num_pages) { 232 spin_unlock(&mgr->lock); 233 return 0; 234 } 235 atomic64_sub(mem->num_pages, &mgr->available); 236 spin_unlock(&mgr->lock); 237 238 node = kzalloc(sizeof(*node), GFP_KERNEL); 239 if (!node) { 240 r = -ENOMEM; 241 goto err_out; 242 } 243 244 node->node.start = AMDGPU_BO_INVALID_OFFSET; 245 node->node.size = mem->num_pages; 246 node->tbo = tbo; 247 mem->mm_node = node; 248 249 if (place->fpfn || place->lpfn || place->flags & TTM_PL_FLAG_TOPDOWN) { 250 r = amdgpu_gtt_mgr_alloc(man, tbo, place, mem); 251 if (unlikely(r)) { 252 kfree(node); 253 mem->mm_node = NULL; 254 r = 0; 255 goto err_out; 256 } 257 } else { 258 mem->start = node->node.start; 259 } 260 261 return 0; 262 err_out: 263 atomic64_add(mem->num_pages, &mgr->available); 264 265 return r; 266 } 267 268 /** 269 * amdgpu_gtt_mgr_del - free ranges 270 * 271 * @man: TTM memory type manager 272 * @tbo: TTM BO we need this range for 273 * @place: placement flags and restrictions 274 * @mem: TTM memory object 275 * 276 * Free the allocated GTT again. 277 */ 278 static void amdgpu_gtt_mgr_del(struct ttm_mem_type_manager *man, 279 struct ttm_mem_reg *mem) 280 { 281 struct amdgpu_gtt_mgr *mgr = man->priv; 282 struct amdgpu_gtt_node *node = mem->mm_node; 283 284 if (!node) 285 return; 286 287 spin_lock(&mgr->lock); 288 if (node->node.start != AMDGPU_BO_INVALID_OFFSET) 289 drm_mm_remove_node(&node->node); 290 spin_unlock(&mgr->lock); 291 atomic64_add(mem->num_pages, &mgr->available); 292 293 kfree(node); 294 mem->mm_node = NULL; 295 } 296 297 /** 298 * amdgpu_gtt_mgr_usage - return usage of GTT domain 299 * 300 * @man: TTM memory type manager 301 * 302 * Return how many bytes are used in the GTT domain 303 */ 304 uint64_t amdgpu_gtt_mgr_usage(struct ttm_mem_type_manager *man) 305 { 306 struct amdgpu_gtt_mgr *mgr = man->priv; 307 s64 result = man->size - atomic64_read(&mgr->available); 308 309 return (result > 0 ? result : 0) * PAGE_SIZE; 310 } 311 312 int amdgpu_gtt_mgr_recover(struct ttm_mem_type_manager *man) 313 { 314 struct amdgpu_gtt_mgr *mgr = man->priv; 315 struct amdgpu_gtt_node *node; 316 struct drm_mm_node *mm_node; 317 int r = 0; 318 319 spin_lock(&mgr->lock); 320 drm_mm_for_each_node(mm_node, &mgr->mm) { 321 node = container_of(mm_node, struct amdgpu_gtt_node, node); 322 r = amdgpu_ttm_recover_gart(node->tbo); 323 if (r) 324 break; 325 } 326 spin_unlock(&mgr->lock); 327 328 return r; 329 } 330 331 /** 332 * amdgpu_gtt_mgr_debug - dump VRAM table 333 * 334 * @man: TTM memory type manager 335 * @printer: DRM printer to use 336 * 337 * Dump the table content using printk. 338 */ 339 static void amdgpu_gtt_mgr_debug(struct ttm_mem_type_manager *man, 340 struct drm_printer *printer) 341 { 342 struct amdgpu_gtt_mgr *mgr = man->priv; 343 344 spin_lock(&mgr->lock); 345 drm_mm_print(&mgr->mm, printer); 346 spin_unlock(&mgr->lock); 347 348 drm_printf(printer, "man size:%llu pages, gtt available:%lld pages, usage:%lluMB\n", 349 man->size, (u64)atomic64_read(&mgr->available), 350 amdgpu_gtt_mgr_usage(man) >> 20); 351 } 352 353 const struct ttm_mem_type_manager_func amdgpu_gtt_mgr_func = { 354 .init = amdgpu_gtt_mgr_init, 355 .takedown = amdgpu_gtt_mgr_fini, 356 .get_node = amdgpu_gtt_mgr_new, 357 .put_node = amdgpu_gtt_mgr_del, 358 .debug = amdgpu_gtt_mgr_debug 359 }; 360