1/* $NetBSD: amdgpu_vram_mgr.c,v 1.4 2021/12/19 12:31:45 riastradh Exp $ */ 2 3/* 4 * Copyright 2016 Advanced Micro Devices, Inc. 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the "Software"), 8 * to deal in the Software without restriction, including without limitation 9 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 10 * and/or sell copies of the Software, and to permit persons to whom the 11 * Software is furnished to do so, subject to the following conditions: 12 * 13 * The above copyright notice and this permission notice shall be included in 14 * all copies or substantial portions of the Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 20 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 21 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 22 * OTHER DEALINGS IN THE SOFTWARE. 23 * 24 * Authors: Christian K��nig 25 */ 26 27#include <sys/cdefs.h> 28__KERNEL_RCSID(0, "$NetBSD: amdgpu_vram_mgr.c,v 1.4 2021/12/19 12:31:45 riastradh Exp $"); 29 30#include "amdgpu.h" 31#include "amdgpu_vm.h" 32#include "amdgpu_atomfirmware.h" 33#include "atom.h" 34 35struct amdgpu_vram_mgr { 36 struct drm_mm mm; 37 spinlock_t lock; 38 atomic64_t usage; 39 atomic64_t vis_usage; 40}; 41 42#ifndef __NetBSD__ /* XXX amdgpu sysfs */ 43 44/** 45 * DOC: mem_info_vram_total 46 * 47 * The amdgpu driver provides a sysfs API for reporting current total VRAM 48 * available on the device 49 * The file mem_info_vram_total is used for this and returns the total 50 * amount of VRAM in bytes 51 */ 52static ssize_t amdgpu_mem_info_vram_total_show(struct device *dev, 53 struct device_attribute *attr, char *buf) 54{ 55 struct drm_device *ddev = dev_get_drvdata(dev); 56 struct amdgpu_device *adev = ddev->dev_private; 57 58 return snprintf(buf, PAGE_SIZE, "%llu\n", adev->gmc.real_vram_size); 59} 60 61/** 62 * DOC: mem_info_vis_vram_total 63 * 64 * The amdgpu driver provides a sysfs API for reporting current total 65 * visible VRAM available on the device 66 * The file mem_info_vis_vram_total is used for this and returns the total 67 * amount of visible VRAM in bytes 68 */ 69static ssize_t amdgpu_mem_info_vis_vram_total_show(struct device *dev, 70 struct device_attribute *attr, char *buf) 71{ 72 struct drm_device *ddev = dev_get_drvdata(dev); 73 struct amdgpu_device *adev = ddev->dev_private; 74 75 return snprintf(buf, PAGE_SIZE, "%llu\n", adev->gmc.visible_vram_size); 76} 77 78/** 79 * DOC: mem_info_vram_used 80 * 81 * The amdgpu driver provides a sysfs API for reporting current total VRAM 82 * available on the device 83 * The file mem_info_vram_used is used for this and returns the total 84 * amount of currently used VRAM in bytes 85 */ 86static ssize_t amdgpu_mem_info_vram_used_show(struct device *dev, 87 struct device_attribute *attr, char *buf) 88{ 89 struct drm_device *ddev = dev_get_drvdata(dev); 90 struct amdgpu_device *adev = ddev->dev_private; 91 92 return snprintf(buf, PAGE_SIZE, "%llu\n", 93 amdgpu_vram_mgr_usage(&adev->mman.bdev.man[TTM_PL_VRAM])); 94} 95 96/** 97 * DOC: mem_info_vis_vram_used 98 * 99 * The amdgpu driver provides a sysfs API for reporting current total of 100 * used visible VRAM 101 * The file mem_info_vis_vram_used is used for this and returns the total 102 * amount of currently used visible VRAM in bytes 103 */ 104static ssize_t amdgpu_mem_info_vis_vram_used_show(struct device *dev, 105 struct device_attribute *attr, char *buf) 106{ 107 struct drm_device *ddev = dev_get_drvdata(dev); 108 struct amdgpu_device *adev = ddev->dev_private; 109 110 return snprintf(buf, PAGE_SIZE, "%llu\n", 111 amdgpu_vram_mgr_vis_usage(&adev->mman.bdev.man[TTM_PL_VRAM])); 112} 113 114static ssize_t amdgpu_mem_info_vram_vendor(struct device *dev, 115 struct device_attribute *attr, 116 char *buf) 117{ 118 struct drm_device *ddev = dev_get_drvdata(dev); 119 struct amdgpu_device *adev = ddev->dev_private; 120 121 switch (adev->gmc.vram_vendor) { 122 case SAMSUNG: 123 return snprintf(buf, PAGE_SIZE, "samsung\n"); 124 case INFINEON: 125 return snprintf(buf, PAGE_SIZE, "infineon\n"); 126 case ELPIDA: 127 return snprintf(buf, PAGE_SIZE, "elpida\n"); 128 case ETRON: 129 return snprintf(buf, PAGE_SIZE, "etron\n"); 130 case NANYA: 131 return snprintf(buf, PAGE_SIZE, "nanya\n"); 132 case HYNIX: 133 return snprintf(buf, PAGE_SIZE, "hynix\n"); 134 case MOSEL: 135 return snprintf(buf, PAGE_SIZE, "mosel\n"); 136 case WINBOND: 137 return snprintf(buf, PAGE_SIZE, "winbond\n"); 138 case ESMT: 139 return snprintf(buf, PAGE_SIZE, "esmt\n"); 140 case MICRON: 141 return snprintf(buf, PAGE_SIZE, "micron\n"); 142 default: 143 return snprintf(buf, PAGE_SIZE, "unknown\n"); 144 } 145} 146 147static DEVICE_ATTR(mem_info_vram_total, S_IRUGO, 148 amdgpu_mem_info_vram_total_show, NULL); 149static DEVICE_ATTR(mem_info_vis_vram_total, S_IRUGO, 150 amdgpu_mem_info_vis_vram_total_show,NULL); 151static DEVICE_ATTR(mem_info_vram_used, S_IRUGO, 152 amdgpu_mem_info_vram_used_show, NULL); 153static DEVICE_ATTR(mem_info_vis_vram_used, S_IRUGO, 154 amdgpu_mem_info_vis_vram_used_show, NULL); 155static DEVICE_ATTR(mem_info_vram_vendor, S_IRUGO, 156 amdgpu_mem_info_vram_vendor, NULL); 157 158#endif /* __NetBSD__ */ 159 160/** 161 * amdgpu_vram_mgr_init - init VRAM manager and DRM MM 162 * 163 * @man: TTM memory type manager 164 * @p_size: maximum size of VRAM 165 * 166 * Allocate and initialize the VRAM manager. 167 */ 168static int amdgpu_vram_mgr_init(struct ttm_mem_type_manager *man, 169 unsigned long p_size) 170{ 171 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev); 172 struct amdgpu_vram_mgr *mgr; 173 int ret; 174 175 mgr = kzalloc(sizeof(*mgr), GFP_KERNEL); 176 if (!mgr) 177 return -ENOMEM; 178 179 drm_mm_init(&mgr->mm, 0, p_size); 180 spin_lock_init(&mgr->lock); 181 man->priv = mgr; 182 183#ifdef __NetBSD__ /* XXX amdgpu sysfs */ 184 __USE(adev); 185 __USE(ret); 186#else 187 /* Add the two VRAM-related sysfs files */ 188 ret = device_create_file(adev->dev, &dev_attr_mem_info_vram_total); 189 if (ret) { 190 DRM_ERROR("Failed to create device file mem_info_vram_total\n"); 191 return ret; 192 } 193 ret = device_create_file(adev->dev, &dev_attr_mem_info_vis_vram_total); 194 if (ret) { 195 DRM_ERROR("Failed to create device file mem_info_vis_vram_total\n"); 196 return ret; 197 } 198 ret = device_create_file(adev->dev, &dev_attr_mem_info_vram_used); 199 if (ret) { 200 DRM_ERROR("Failed to create device file mem_info_vram_used\n"); 201 return ret; 202 } 203 ret = device_create_file(adev->dev, &dev_attr_mem_info_vis_vram_used); 204 if (ret) { 205 DRM_ERROR("Failed to create device file mem_info_vis_vram_used\n"); 206 return ret; 207 } 208 ret = device_create_file(adev->dev, &dev_attr_mem_info_vram_vendor); 209 if (ret) { 210 DRM_ERROR("Failed to create device file mem_info_vram_vendor\n"); 211 return ret; 212 } 213#endif /* __NetBSD__ */ 214 215 return 0; 216} 217 218/** 219 * amdgpu_vram_mgr_fini - free and destroy VRAM manager 220 * 221 * @man: TTM memory type manager 222 * 223 * Destroy and free the VRAM manager, returns -EBUSY if ranges are still 224 * allocated inside it. 225 */ 226static int amdgpu_vram_mgr_fini(struct ttm_mem_type_manager *man) 227{ 228 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev); 229 struct amdgpu_vram_mgr *mgr = man->priv; 230 231 spin_lock(&mgr->lock); 232 drm_mm_takedown(&mgr->mm); 233 spin_unlock(&mgr->lock); 234 spin_lock_destroy(&mgr->lock); 235 kfree(mgr); 236 man->priv = NULL; 237#ifdef __NetBSD__ /* XXX amdgpu sysfs */ 238 __USE(adev); 239#else 240 device_remove_file(adev->dev, &dev_attr_mem_info_vram_total); 241 device_remove_file(adev->dev, &dev_attr_mem_info_vis_vram_total); 242 device_remove_file(adev->dev, &dev_attr_mem_info_vram_used); 243 device_remove_file(adev->dev, &dev_attr_mem_info_vis_vram_used); 244 device_remove_file(adev->dev, &dev_attr_mem_info_vram_vendor); 245#endif 246 return 0; 247} 248 249/** 250 * amdgpu_vram_mgr_vis_size - Calculate visible node size 251 * 252 * @adev: amdgpu device structure 253 * @node: MM node structure 254 * 255 * Calculate how many bytes of the MM node are inside visible VRAM 256 */ 257static u64 amdgpu_vram_mgr_vis_size(struct amdgpu_device *adev, 258 struct drm_mm_node *node) 259{ 260 uint64_t start = node->start << PAGE_SHIFT; 261 uint64_t end = (node->size + node->start) << PAGE_SHIFT; 262 263 if (start >= adev->gmc.visible_vram_size) 264 return 0; 265 266 return (end > adev->gmc.visible_vram_size ? 267 adev->gmc.visible_vram_size : end) - start; 268} 269 270/** 271 * amdgpu_vram_mgr_bo_visible_size - CPU visible BO size 272 * 273 * @bo: &amdgpu_bo buffer object (must be in VRAM) 274 * 275 * Returns: 276 * How much of the given &amdgpu_bo buffer object lies in CPU visible VRAM. 277 */ 278u64 amdgpu_vram_mgr_bo_visible_size(struct amdgpu_bo *bo) 279{ 280 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); 281 struct ttm_mem_reg *mem = &bo->tbo.mem; 282 struct drm_mm_node *nodes = mem->mm_node; 283 unsigned pages = mem->num_pages; 284 u64 usage; 285 286 if (amdgpu_gmc_vram_full_visible(&adev->gmc)) 287 return amdgpu_bo_size(bo); 288 289 if (mem->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT) 290 return 0; 291 292 for (usage = 0; nodes && pages; pages -= nodes->size, nodes++) 293 usage += amdgpu_vram_mgr_vis_size(adev, nodes); 294 295 return usage; 296} 297 298/** 299 * amdgpu_vram_mgr_virt_start - update virtual start address 300 * 301 * @mem: ttm_mem_reg to update 302 * @node: just allocated node 303 * 304 * Calculate a virtual BO start address to easily check if everything is CPU 305 * accessible. 306 */ 307static void amdgpu_vram_mgr_virt_start(struct ttm_mem_reg *mem, 308 struct drm_mm_node *node) 309{ 310 unsigned long start; 311 312 start = node->start + node->size; 313 if (start > mem->num_pages) 314 start -= mem->num_pages; 315 else 316 start = 0; 317 mem->start = max(mem->start, start); 318} 319 320/** 321 * amdgpu_vram_mgr_new - allocate new ranges 322 * 323 * @man: TTM memory type manager 324 * @tbo: TTM BO we need this range for 325 * @place: placement flags and restrictions 326 * @mem: the resulting mem object 327 * 328 * Allocate VRAM for the given BO. 329 */ 330static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man, 331 struct ttm_buffer_object *tbo, 332 const struct ttm_place *place, 333 struct ttm_mem_reg *mem) 334{ 335 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev); 336 struct amdgpu_vram_mgr *mgr = man->priv; 337 struct drm_mm *mm = &mgr->mm; 338 struct drm_mm_node *nodes; 339 enum drm_mm_insert_mode mode; 340 unsigned long lpfn, num_nodes, pages_per_node, pages_left; 341 uint64_t vis_usage = 0, mem_bytes, max_bytes; 342 unsigned i; 343 int r; 344 345 lpfn = place->lpfn; 346 if (!lpfn) 347 lpfn = man->size; 348 349 max_bytes = adev->gmc.mc_vram_size; 350 if (tbo->type != ttm_bo_type_kernel) 351 max_bytes -= AMDGPU_VM_RESERVED_VRAM; 352 353 /* bail out quickly if there's likely not enough VRAM for this BO */ 354 mem_bytes = (u64)mem->num_pages << PAGE_SHIFT; 355 if (atomic64_add_return(mem_bytes, &mgr->usage) > max_bytes) { 356 atomic64_sub(mem_bytes, &mgr->usage); 357 mem->mm_node = NULL; 358 return 0; 359 } 360 361 if (place->flags & TTM_PL_FLAG_CONTIGUOUS) { 362 pages_per_node = ~0ul; 363 num_nodes = 1; 364 } else { 365#ifdef CONFIG_TRANSPARENT_HUGEPAGE 366 pages_per_node = HPAGE_PMD_NR; 367#else 368 /* default to 2MB */ 369 pages_per_node = (2UL << (20UL - PAGE_SHIFT)); 370#endif 371 pages_per_node = max((uint32_t)pages_per_node, mem->page_alignment); 372 num_nodes = DIV_ROUND_UP(mem->num_pages, pages_per_node); 373 } 374 375 nodes = kvmalloc_array((uint32_t)num_nodes, sizeof(*nodes), 376 GFP_KERNEL | __GFP_ZERO); 377 if (!nodes) { 378 atomic64_sub(mem_bytes, &mgr->usage); 379 return -ENOMEM; 380 } 381 382 mode = DRM_MM_INSERT_BEST; 383 if (place->flags & TTM_PL_FLAG_TOPDOWN) 384 mode = DRM_MM_INSERT_HIGH; 385 386 mem->start = 0; 387 pages_left = mem->num_pages; 388 389 spin_lock(&mgr->lock); 390 for (i = 0; pages_left >= pages_per_node; ++i) { 391 unsigned long pages = rounddown_pow_of_two(pages_left); 392 393 r = drm_mm_insert_node_in_range(mm, &nodes[i], pages, 394 pages_per_node, 0, 395 place->fpfn, lpfn, 396 mode); 397 if (unlikely(r)) 398 break; 399 400 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]); 401 amdgpu_vram_mgr_virt_start(mem, &nodes[i]); 402 pages_left -= pages; 403 } 404 405 for (; pages_left; ++i) { 406 unsigned long pages = min(pages_left, pages_per_node); 407 uint32_t alignment = mem->page_alignment; 408 409 if (pages == pages_per_node) 410 alignment = pages_per_node; 411 412 r = drm_mm_insert_node_in_range(mm, &nodes[i], 413 pages, alignment, 0, 414 place->fpfn, lpfn, 415 mode); 416 if (unlikely(r)) 417 goto error; 418 419 vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]); 420 amdgpu_vram_mgr_virt_start(mem, &nodes[i]); 421 pages_left -= pages; 422 } 423 spin_unlock(&mgr->lock); 424 425 atomic64_add(vis_usage, &mgr->vis_usage); 426 427 mem->mm_node = nodes; 428 429 return 0; 430 431error: 432 while (i--) 433 drm_mm_remove_node(&nodes[i]); 434 spin_unlock(&mgr->lock); 435 atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage); 436 437 kvfree(nodes); 438 return r == -ENOSPC ? 0 : r; 439} 440 441/** 442 * amdgpu_vram_mgr_del - free ranges 443 * 444 * @man: TTM memory type manager 445 * @tbo: TTM BO we need this range for 446 * @place: placement flags and restrictions 447 * @mem: TTM memory object 448 * 449 * Free the allocated VRAM again. 450 */ 451static void amdgpu_vram_mgr_del(struct ttm_mem_type_manager *man, 452 struct ttm_mem_reg *mem) 453{ 454 struct amdgpu_device *adev = amdgpu_ttm_adev(man->bdev); 455 struct amdgpu_vram_mgr *mgr = man->priv; 456 struct drm_mm_node *nodes = mem->mm_node; 457 uint64_t usage = 0, vis_usage = 0; 458 unsigned pages = mem->num_pages; 459 460 if (!mem->mm_node) 461 return; 462 463 spin_lock(&mgr->lock); 464 while (pages) { 465 pages -= nodes->size; 466 drm_mm_remove_node(nodes); 467 usage += nodes->size << PAGE_SHIFT; 468 vis_usage += amdgpu_vram_mgr_vis_size(adev, nodes); 469 ++nodes; 470 } 471 spin_unlock(&mgr->lock); 472 473 atomic64_sub(usage, &mgr->usage); 474 atomic64_sub(vis_usage, &mgr->vis_usage); 475 476 kvfree(mem->mm_node); 477 mem->mm_node = NULL; 478} 479 480/** 481 * amdgpu_vram_mgr_usage - how many bytes are used in this domain 482 * 483 * @man: TTM memory type manager 484 * 485 * Returns how many bytes are used in this domain. 486 */ 487uint64_t amdgpu_vram_mgr_usage(struct ttm_mem_type_manager *man) 488{ 489 struct amdgpu_vram_mgr *mgr = man->priv; 490 491 return atomic64_read(&mgr->usage); 492} 493 494/** 495 * amdgpu_vram_mgr_vis_usage - how many bytes are used in the visible part 496 * 497 * @man: TTM memory type manager 498 * 499 * Returns how many bytes are used in the visible part of VRAM 500 */ 501uint64_t amdgpu_vram_mgr_vis_usage(struct ttm_mem_type_manager *man) 502{ 503 struct amdgpu_vram_mgr *mgr = man->priv; 504 505 return atomic64_read(&mgr->vis_usage); 506} 507 508/** 509 * amdgpu_vram_mgr_debug - dump VRAM table 510 * 511 * @man: TTM memory type manager 512 * @printer: DRM printer to use 513 * 514 * Dump the table content using printk. 515 */ 516static void amdgpu_vram_mgr_debug(struct ttm_mem_type_manager *man, 517 struct drm_printer *printer) 518{ 519 struct amdgpu_vram_mgr *mgr = man->priv; 520 521 spin_lock(&mgr->lock); 522 drm_mm_print(&mgr->mm, printer); 523 spin_unlock(&mgr->lock); 524 525 drm_printf(printer, "man size:%"PRIu64" pages, ram usage:%"PRIu64"MB, vis usage:%"PRIu64"MB\n", 526 man->size, amdgpu_vram_mgr_usage(man) >> 20, 527 amdgpu_vram_mgr_vis_usage(man) >> 20); 528} 529 530const struct ttm_mem_type_manager_func amdgpu_vram_mgr_func = { 531 .init = amdgpu_vram_mgr_init, 532 .takedown = amdgpu_vram_mgr_fini, 533 .get_node = amdgpu_vram_mgr_new, 534 .put_node = amdgpu_vram_mgr_del, 535 .debug = amdgpu_vram_mgr_debug 536}; 537