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 #include "amdgpu_vm.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 * DOC: mem_info_vram_total 37 * 38 * The amdgpu driver provides a sysfs API for reporting current total VRAM 39 * available on the device 40 * The file mem_info_vram_total is used for this and returns the total 41 * amount of VRAM in bytes 42 */ 43 static ssize_t amdgpu_mem_info_vram_total_show(struct device *dev, 44 struct device_attribute *attr, char *buf) 45 { 46 struct drm_device *ddev = dev_get_drvdata(dev); 47 struct amdgpu_device *adev = ddev->dev_private; 48 49 return snprintf(buf, PAGE_SIZE, "%llu\n", adev->gmc.real_vram_size); 50 } 51 52 /** 53 * DOC: mem_info_vis_vram_total 54 * 55 * The amdgpu driver provides a sysfs API for reporting current total 56 * visible VRAM available on the device 57 * The file mem_info_vis_vram_total is used for this and returns the total 58 * amount of visible VRAM in bytes 59 */ 60 static ssize_t amdgpu_mem_info_vis_vram_total_show(struct device *dev, 61 struct device_attribute *attr, char *buf) 62 { 63 struct drm_device *ddev = dev_get_drvdata(dev); 64 struct amdgpu_device *adev = ddev->dev_private; 65 66 return snprintf(buf, PAGE_SIZE, "%llu\n", adev->gmc.visible_vram_size); 67 } 68 69 /** 70 * DOC: mem_info_vram_used 71 * 72 * The amdgpu driver provides a sysfs API for reporting current total VRAM 73 * available on the device 74 * The file mem_info_vram_used is used for this and returns the total 75 * amount of currently used VRAM in bytes 76 */ 77 static ssize_t amdgpu_mem_info_vram_used_show(struct device *dev, 78 struct device_attribute *attr, char *buf) 79 { 80 struct drm_device *ddev = dev_get_drvdata(dev); 81 struct amdgpu_device *adev = ddev->dev_private; 82 83 return snprintf(buf, PAGE_SIZE, "%llu\n", 84 amdgpu_vram_mgr_usage(&adev->mman.bdev.man[TTM_PL_VRAM])); 85 } 86 87 /** 88 * DOC: mem_info_vis_vram_used 89 * 90 * The amdgpu driver provides a sysfs API for reporting current total of 91 * used visible VRAM 92 * The file mem_info_vis_vram_used is used for this and returns the total 93 * amount of currently used visible VRAM in bytes 94 */ 95 static ssize_t amdgpu_mem_info_vis_vram_used_show(struct device *dev, 96 struct device_attribute *attr, char *buf) 97 { 98 struct drm_device *ddev = dev_get_drvdata(dev); 99 struct amdgpu_device *adev = ddev->dev_private; 100 101 return snprintf(buf, PAGE_SIZE, "%llu\n", 102 amdgpu_vram_mgr_vis_usage(&adev->mman.bdev.man[TTM_PL_VRAM])); 103 } 104 105 static DEVICE_ATTR(mem_info_vram_total, S_IRUGO, 106 amdgpu_mem_info_vram_total_show, NULL); 107 static DEVICE_ATTR(mem_info_vis_vram_total, S_IRUGO, 108 amdgpu_mem_info_vis_vram_total_show,NULL); 109 static DEVICE_ATTR(mem_info_vram_used, S_IRUGO, 110 amdgpu_mem_info_vram_used_show, NULL); 111 static DEVICE_ATTR(mem_info_vis_vram_used, S_IRUGO, 112 amdgpu_mem_info_vis_vram_used_show, NULL); 113 114 /** 115 * amdgpu_vram_mgr_init - init VRAM manager and DRM MM 116 * 117 * @man: TTM memory type manager 118 * @p_size: maximum size of VRAM 119 * 120 * Allocate and initialize the VRAM manager. 121 */ 122 static int amdgpu_vram_mgr_init(struct ttm_mem_type_manager *man, 123 unsigned long p_size) 124 { 125 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev); 126 struct amdgpu_vram_mgr *mgr; 127 int ret; 128 129 mgr = kzalloc(sizeof(*mgr), GFP_KERNEL); 130 if (!mgr) 131 return -ENOMEM; 132 133 drm_mm_init(&mgr->mm, 0, p_size); 134 spin_lock_init(&mgr->lock); 135 man->priv = mgr; 136 137 /* Add the two VRAM-related sysfs files */ 138 ret = device_create_file(adev->dev, &dev_attr_mem_info_vram_total); 139 if (ret) { 140 DRM_ERROR("Failed to create device file mem_info_vram_total\n"); 141 return ret; 142 } 143 ret = device_create_file(adev->dev, &dev_attr_mem_info_vis_vram_total); 144 if (ret) { 145 DRM_ERROR("Failed to create device file mem_info_vis_vram_total\n"); 146 return ret; 147 } 148 ret = device_create_file(adev->dev, &dev_attr_mem_info_vram_used); 149 if (ret) { 150 DRM_ERROR("Failed to create device file mem_info_vram_used\n"); 151 return ret; 152 } 153 ret = device_create_file(adev->dev, &dev_attr_mem_info_vis_vram_used); 154 if (ret) { 155 DRM_ERROR("Failed to create device file mem_info_vis_vram_used\n"); 156 return ret; 157 } 158 159 return 0; 160 } 161 162 /** 163 * amdgpu_vram_mgr_fini - free and destroy VRAM manager 164 * 165 * @man: TTM memory type manager 166 * 167 * Destroy and free the VRAM manager, returns -EBUSY if ranges are still 168 * allocated inside it. 169 */ 170 static int amdgpu_vram_mgr_fini(struct ttm_mem_type_manager *man) 171 { 172 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev); 173 struct amdgpu_vram_mgr *mgr = man->priv; 174 175 spin_lock(&mgr->lock); 176 drm_mm_takedown(&mgr->mm); 177 spin_unlock(&mgr->lock); 178 kfree(mgr); 179 man->priv = NULL; 180 device_remove_file(adev->dev, &dev_attr_mem_info_vram_total); 181 device_remove_file(adev->dev, &dev_attr_mem_info_vis_vram_total); 182 device_remove_file(adev->dev, &dev_attr_mem_info_vram_used); 183 device_remove_file(adev->dev, &dev_attr_mem_info_vis_vram_used); 184 return 0; 185 } 186 187 /** 188 * amdgpu_vram_mgr_vis_size - Calculate visible node size 189 * 190 * @adev: amdgpu device structure 191 * @node: MM node structure 192 * 193 * Calculate how many bytes of the MM node are inside visible VRAM 194 */ 195 static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev, 196 struct drm_mm_node *node) 197 { 198 uint64_t start = node->start << PAGE_SHIFT; 199 uint64_t end = (node->size + node->start) << PAGE_SHIFT; 200 201 if (start >= adev->gmc.visible_vram_size) 202 return 0; 203 204 return (end > adev->gmc.visible_vram_size ? 205 adev->gmc.visible_vram_size : end) - start; 206 } 207 208 /** 209 * amdgpu_vram_mgr_bo_visible_size - CPU visible BO size 210 * 211 * @bo: &amdgpu_bo buffer object (must be in VRAM) 212 * 213 * Returns: 214 * How much of the given &amdgpu_bo buffer object lies in CPU visible VRAM. 215 */ 216 u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo) 217 { 218 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 219 struct ttm_mem_reg *mem = &bo->tbo.mem; 220 struct drm_mm_node *nodes = mem->mm_node; 221 unsigned pages = mem->num_pages; 222 u64 usage; 223 224 if (amdgpu_gmc_vram_full_visible(&adev->gmc)) 225 return amdgpu_bo_size(bo); 226 227 if (mem->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT) 228 return 0; 229 230 for (usage = 0; nodes && pages; pages -= nodes->size, nodes++) 231 usage += amdgpu_vram_mgr_vis_size(adev, nodes); 232 233 return usage; 234 } 235 236 /** 237 * amdgpu_vram_mgr_virt_start - update virtual start address 238 * 239 * @mem: ttm_mem_reg to update 240 * @node: just allocated node 241 * 242 * Calculate a virtual BO start address to easily check if everything is CPU 243 * accessible. 244 */ 245 static void amdgpu_vram_mgr_virt_start(struct ttm_mem_reg *mem, 246 struct drm_mm_node *node) 247 { 248 unsigned long start; 249 250 start = node->start + node->size; 251 if (start > mem->num_pages) 252 start -= mem->num_pages; 253 else 254 start = 0; 255 mem->start = max(mem->start, start); 256 } 257 258 /** 259 * amdgpu_vram_mgr_new - allocate new ranges 260 * 261 * @man: TTM memory type manager 262 * @tbo: TTM BO we need this range for 263 * @place: placement flags and restrictions 264 * @mem: the resulting mem object 265 * 266 * Allocate VRAM for the given BO. 267 */ 268 static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man, 269 struct ttm_buffer_object *tbo, 270 const struct ttm_place *place, 271 struct ttm_mem_reg *mem) 272 { 273 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev); 274 struct amdgpu_vram_mgr *mgr = man->priv; 275 struct drm_mm *mm = &mgr->mm; 276 struct drm_mm_node *nodes; 277 enum drm_mm_insert_mode mode; 278 unsigned long lpfn, num_nodes, pages_per_node, pages_left; 279 uint64_t vis_usage = 0, mem_bytes, max_bytes; 280 unsigned i; 281 int r; 282 283 lpfn = place->lpfn; 284 if (!lpfn) 285 lpfn = man->size; 286 287 max_bytes = adev->gmc.mc_vram_size; 288 if (tbo->type != ttm_bo_type_kernel) 289 max_bytes -= AMDGPU_VM_RESERVED_VRAM; 290 291 /* bail out quickly if there's likely not enough VRAM for this BO */ 292 mem_bytes = (u64)mem->num_pages << PAGE_SHIFT; 293 if (atomic64_add_return(mem_bytes, &mgr->usage) > max_bytes) { 294 atomic64_sub(mem_bytes, &mgr->usage); 295 mem->mm_node = NULL; 296 return 0; 297 } 298 299 if (place->flags & TTM_PL_FLAG_CONTIGUOUS) { 300 pages_per_node = ~0ul; 301 num_nodes = 1; 302 } else { 303 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 304 pages_per_node = HPAGE_PMD_NR; 305 #else 306 /* default to 2MB */ 307 pages_per_node = (2UL << (20UL - PAGE_SHIFT)); 308 #endif 309 pages_per_node = max((uint32_t)pages_per_node, mem->page_alignment); 310 num_nodes = DIV_ROUND_UP(mem->num_pages, pages_per_node); 311 } 312 313 nodes = kvmalloc_array((uint32_t)num_nodes, sizeof(*nodes), 314 GFP_KERNEL | __GFP_ZERO); 315 if (!nodes) { 316 atomic64_sub(mem_bytes, &mgr->usage); 317 return -ENOMEM; 318 } 319 320 mode = DRM_MM_INSERT_BEST; 321 if (place->flags & TTM_PL_FLAG_TOPDOWN) 322 mode = DRM_MM_INSERT_HIGH; 323 324 mem->start = 0; 325 pages_left = mem->num_pages; 326 327 spin_lock(&mgr->lock); 328 for (i = 0; pages_left >= pages_per_node; ++i) { 329 unsigned long pages = rounddown_pow_of_two(pages_left); 330 331 r = drm_mm_insert_node_in_range(mm, &nodes[i], pages, 332 pages_per_node, 0, 333 place->fpfn, lpfn, 334 mode); 335 if (unlikely(r)) 336 break; 337 338 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]); 339 amdgpu_vram_mgr_virt_start(mem, &nodes[i]); 340 pages_left -= pages; 341 } 342 343 for (; pages_left; ++i) { 344 unsigned long pages = min(pages_left, pages_per_node); 345 uint32_t alignment = mem->page_alignment; 346 347 if (pages == pages_per_node) 348 alignment = pages_per_node; 349 350 r = drm_mm_insert_node_in_range(mm, &nodes[i], 351 pages, alignment, 0, 352 place->fpfn, lpfn, 353 mode); 354 if (unlikely(r)) 355 goto error; 356 357 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]); 358 amdgpu_vram_mgr_virt_start(mem, &nodes[i]); 359 pages_left -= pages; 360 } 361 spin_unlock(&mgr->lock); 362 363 atomic64_add(vis_usage, &mgr->vis_usage); 364 365 mem->mm_node = nodes; 366 367 return 0; 368 369 error: 370 while (i--) 371 drm_mm_remove_node(&nodes[i]); 372 spin_unlock(&mgr->lock); 373 atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage); 374 375 kvfree(nodes); 376 return r == -ENOSPC ? 0 : r; 377 } 378 379 /** 380 * amdgpu_vram_mgr_del - free ranges 381 * 382 * @man: TTM memory type manager 383 * @tbo: TTM BO we need this range for 384 * @place: placement flags and restrictions 385 * @mem: TTM memory object 386 * 387 * Free the allocated VRAM again. 388 */ 389 static void amdgpu_vram_mgr_del(struct ttm_mem_type_manager *man, 390 struct ttm_mem_reg *mem) 391 { 392 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev); 393 struct amdgpu_vram_mgr *mgr = man->priv; 394 struct drm_mm_node *nodes = mem->mm_node; 395 uint64_t usage = 0, vis_usage = 0; 396 unsigned pages = mem->num_pages; 397 398 if (!mem->mm_node) 399 return; 400 401 spin_lock(&mgr->lock); 402 while (pages) { 403 pages -= nodes->size; 404 drm_mm_remove_node(nodes); 405 usage += nodes->size << PAGE_SHIFT; 406 vis_usage += amdgpu_vram_mgr_vis_size(adev, nodes); 407 ++nodes; 408 } 409 spin_unlock(&mgr->lock); 410 411 atomic64_sub(usage, &mgr->usage); 412 atomic64_sub(vis_usage, &mgr->vis_usage); 413 414 kvfree(mem->mm_node); 415 mem->mm_node = NULL; 416 } 417 418 /** 419 * amdgpu_vram_mgr_usage - how many bytes are used in this domain 420 * 421 * @man: TTM memory type manager 422 * 423 * Returns how many bytes are used in this domain. 424 */ 425 uint64_t amdgpu_vram_mgr_usage(struct ttm_mem_type_manager *man) 426 { 427 struct amdgpu_vram_mgr *mgr = man->priv; 428 429 return atomic64_read(&mgr->usage); 430 } 431 432 /** 433 * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part 434 * 435 * @man: TTM memory type manager 436 * 437 * Returns how many bytes are used in the visible part of VRAM 438 */ 439 uint64_t amdgpu_vram_mgr_vis_usage(struct ttm_mem_type_manager *man) 440 { 441 struct amdgpu_vram_mgr *mgr = man->priv; 442 443 return atomic64_read(&mgr->vis_usage); 444 } 445 446 /** 447 * amdgpu_vram_mgr_debug - dump VRAM table 448 * 449 * @man: TTM memory type manager 450 * @printer: DRM printer to use 451 * 452 * Dump the table content using printk. 453 */ 454 static void amdgpu_vram_mgr_debug(struct ttm_mem_type_manager *man, 455 struct drm_printer *printer) 456 { 457 struct amdgpu_vram_mgr *mgr = man->priv; 458 459 spin_lock(&mgr->lock); 460 drm_mm_print(&mgr->mm, printer); 461 spin_unlock(&mgr->lock); 462 463 drm_printf(printer, "man size:%llu pages, ram usage:%lluMB, vis usage:%lluMB\n", 464 man->size, amdgpu_vram_mgr_usage(man) >> 20, 465 amdgpu_vram_mgr_vis_usage(man) >> 20); 466 } 467 468 const struct ttm_mem_type_manager_func amdgpu_vram_mgr_func = { 469 .init = amdgpu_vram_mgr_init, 470 .takedown = amdgpu_vram_mgr_fini, 471 .get_node = amdgpu_vram_mgr_new, 472 .put_node = amdgpu_vram_mgr_del, 473 .debug = amdgpu_vram_mgr_debug 474 }; 475