1// SPDX-License-Identifier: GPL-2.0-or-later 2/* 3 * Contiguous Memory Allocator 4 * 5 * Copyright (c) 2010-2011 by Samsung Electronics. 6 * Copyright IBM Corporation, 2013 7 * Copyright LG Electronics Inc., 2014 8 * Written by: 9 * Marek Szyprowski <m.szyprowski@samsung.com> 10 * Michal Nazarewicz <mina86@mina86.com> 11 * Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com> 12 * Joonsoo Kim <iamjoonsoo.kim@lge.com> 13 */ 14 15#define pr_fmt(fmt) "cma: " fmt 16 17#define CREATE_TRACE_POINTS 18 19#include <linux/memblock.h> 20#include <linux/err.h> 21#include <linux/mm.h> 22#include <linux/sizes.h> 23#include <linux/slab.h> 24#include <linux/log2.h> 25#include <linux/cma.h> 26#include <linux/highmem.h> 27#include <linux/io.h> 28#include <linux/kmemleak.h> 29#include <trace/events/cma.h> 30 31#include "internal.h" 32#include "cma.h" 33 34struct cma cma_areas[MAX_CMA_AREAS]; 35unsigned cma_area_count; 36static DEFINE_MUTEX(cma_mutex); 37 38phys_addr_t cma_get_base(const struct cma *cma) 39{ 40 return PFN_PHYS(cma->base_pfn); 41} 42 43unsigned long cma_get_size(const struct cma *cma) 44{ 45 return cma->count << PAGE_SHIFT; 46} 47 48const char *cma_get_name(const struct cma *cma) 49{ 50 return cma->name; 51} 52 53static unsigned long cma_bitmap_aligned_mask(const struct cma *cma, 54 unsigned int align_order) 55{ 56 if (align_order <= cma->order_per_bit) 57 return 0; 58 return (1UL << (align_order - cma->order_per_bit)) - 1; 59} 60 61/* 62 * Find the offset of the base PFN from the specified align_order. 63 * The value returned is represented in order_per_bits. 64 */ 65static unsigned long cma_bitmap_aligned_offset(const struct cma *cma, 66 unsigned int align_order) 67{ 68 return (cma->base_pfn & ((1UL << align_order) - 1)) 69 >> cma->order_per_bit; 70} 71 72static unsigned long cma_bitmap_pages_to_bits(const struct cma *cma, 73 unsigned long pages) 74{ 75 return ALIGN(pages, 1UL << cma->order_per_bit) >> cma->order_per_bit; 76} 77 78static void cma_clear_bitmap(struct cma *cma, unsigned long pfn, 79 unsigned long count) 80{ 81 unsigned long bitmap_no, bitmap_count; 82 unsigned long flags; 83 84 bitmap_no = (pfn - cma->base_pfn) >> cma->order_per_bit; 85 bitmap_count = cma_bitmap_pages_to_bits(cma, count); 86 87 spin_lock_irqsave(&cma->lock, flags); 88 bitmap_clear(cma->bitmap, bitmap_no, bitmap_count); 89 spin_unlock_irqrestore(&cma->lock, flags); 90} 91 92static void __init cma_activate_area(struct cma *cma) 93{ 94 unsigned long base_pfn = cma->base_pfn, pfn; 95 struct zone *zone; 96 97 cma->bitmap = bitmap_zalloc(cma_bitmap_maxno(cma), GFP_KERNEL); 98 if (!cma->bitmap) 99 goto out_error; 100 101 /* 102 * alloc_contig_range() requires the pfn range specified to be in the 103 * same zone. Simplify by forcing the entire CMA resv range to be in the 104 * same zone. 105 */ 106 WARN_ON_ONCE(!pfn_valid(base_pfn)); 107 zone = page_zone(pfn_to_page(base_pfn)); 108 for (pfn = base_pfn + 1; pfn < base_pfn + cma->count; pfn++) { 109 WARN_ON_ONCE(!pfn_valid(pfn)); 110 if (page_zone(pfn_to_page(pfn)) != zone) 111 goto not_in_zone; 112 } 113 114 for (pfn = base_pfn; pfn < base_pfn + cma->count; 115 pfn += pageblock_nr_pages) 116 init_cma_reserved_pageblock(pfn_to_page(pfn)); 117 118 spin_lock_init(&cma->lock); 119 120#ifdef CONFIG_CMA_DEBUGFS 121 INIT_HLIST_HEAD(&cma->mem_head); 122 spin_lock_init(&cma->mem_head_lock); 123#endif 124 125 return; 126 127not_in_zone: 128 bitmap_free(cma->bitmap); 129out_error: 130 /* Expose all pages to the buddy, they are useless for CMA. */ 131 if (!cma->reserve_pages_on_error) { 132 for (pfn = base_pfn; pfn < base_pfn + cma->count; pfn++) 133 free_reserved_page(pfn_to_page(pfn)); 134 } 135 totalcma_pages -= cma->count; 136 cma->count = 0; 137 pr_err("CMA area %s could not be activated\n", cma->name); 138 return; 139} 140 141static int __init cma_init_reserved_areas(void) 142{ 143 int i; 144 145 for (i = 0; i < cma_area_count; i++) 146 cma_activate_area(&cma_areas[i]); 147 148 return 0; 149} 150core_initcall(cma_init_reserved_areas); 151 152void __init cma_reserve_pages_on_error(struct cma *cma) 153{ 154 cma->reserve_pages_on_error = true; 155} 156 157/** 158 * cma_init_reserved_mem() - create custom contiguous area from reserved memory 159 * @base: Base address of the reserved area 160 * @size: Size of the reserved area (in bytes), 161 * @order_per_bit: Order of pages represented by one bit on bitmap. 162 * @name: The name of the area. If this parameter is NULL, the name of 163 * the area will be set to "cmaN", where N is a running counter of 164 * used areas. 165 * @res_cma: Pointer to store the created cma region. 166 * 167 * This function creates custom contiguous area from already reserved memory. 168 */ 169int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size, 170 unsigned int order_per_bit, 171 const char *name, 172 struct cma **res_cma) 173{ 174 struct cma *cma; 175 176 /* Sanity checks */ 177 if (cma_area_count == ARRAY_SIZE(cma_areas)) { 178 pr_err("Not enough slots for CMA reserved regions!\n"); 179 return -ENOSPC; 180 } 181 182 if (!size || !memblock_is_region_reserved(base, size)) 183 return -EINVAL; 184 185 /* alignment should be aligned with order_per_bit */ 186 if (!IS_ALIGNED(CMA_MIN_ALIGNMENT_PAGES, 1 << order_per_bit)) 187 return -EINVAL; 188 189 /* ensure minimal alignment required by mm core */ 190 if (!IS_ALIGNED(base | size, CMA_MIN_ALIGNMENT_BYTES)) 191 return -EINVAL; 192 193 /* 194 * Each reserved area must be initialised later, when more kernel 195 * subsystems (like slab allocator) are available. 196 */ 197 cma = &cma_areas[cma_area_count]; 198 199 if (name) 200 snprintf(cma->name, CMA_MAX_NAME, name); 201 else 202 snprintf(cma->name, CMA_MAX_NAME, "cma%d\n", cma_area_count); 203 204 cma->base_pfn = PFN_DOWN(base); 205 cma->count = size >> PAGE_SHIFT; 206 cma->order_per_bit = order_per_bit; 207 *res_cma = cma; 208 cma_area_count++; 209 totalcma_pages += (size / PAGE_SIZE); 210 211 return 0; 212} 213 214/** 215 * cma_declare_contiguous_nid() - reserve custom contiguous area 216 * @base: Base address of the reserved area optional, use 0 for any 217 * @size: Size of the reserved area (in bytes), 218 * @limit: End address of the reserved memory (optional, 0 for any). 219 * @alignment: Alignment for the CMA area, should be power of 2 or zero 220 * @order_per_bit: Order of pages represented by one bit on bitmap. 221 * @fixed: hint about where to place the reserved area 222 * @name: The name of the area. See function cma_init_reserved_mem() 223 * @res_cma: Pointer to store the created cma region. 224 * @nid: nid of the free area to find, %NUMA_NO_NODE for any node 225 * 226 * This function reserves memory from early allocator. It should be 227 * called by arch specific code once the early allocator (memblock or bootmem) 228 * has been activated and all other subsystems have already allocated/reserved 229 * memory. This function allows to create custom reserved areas. 230 * 231 * If @fixed is true, reserve contiguous area at exactly @base. If false, 232 * reserve in range from @base to @limit. 233 */ 234int __init cma_declare_contiguous_nid(phys_addr_t base, 235 phys_addr_t size, phys_addr_t limit, 236 phys_addr_t alignment, unsigned int order_per_bit, 237 bool fixed, const char *name, struct cma **res_cma, 238 int nid) 239{ 240 phys_addr_t memblock_end = memblock_end_of_DRAM(); 241 phys_addr_t highmem_start; 242 int ret; 243 244 /* 245 * We can't use __pa(high_memory) directly, since high_memory 246 * isn't a valid direct map VA, and DEBUG_VIRTUAL will (validly) 247 * complain. Find the boundary by adding one to the last valid 248 * address. 249 */ 250 highmem_start = __pa(high_memory - 1) + 1; 251 pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n", 252 __func__, &size, &base, &limit, &alignment); 253 254 if (cma_area_count == ARRAY_SIZE(cma_areas)) { 255 pr_err("Not enough slots for CMA reserved regions!\n"); 256 return -ENOSPC; 257 } 258 259 if (!size) 260 return -EINVAL; 261 262 if (alignment && !is_power_of_2(alignment)) 263 return -EINVAL; 264 265 if (!IS_ENABLED(CONFIG_NUMA)) 266 nid = NUMA_NO_NODE; 267 268 /* Sanitise input arguments. */ 269 alignment = max_t(phys_addr_t, alignment, CMA_MIN_ALIGNMENT_BYTES); 270 if (fixed && base & (alignment - 1)) { 271 ret = -EINVAL; 272 pr_err("Region at %pa must be aligned to %pa bytes\n", 273 &base, &alignment); 274 goto err; 275 } 276 base = ALIGN(base, alignment); 277 size = ALIGN(size, alignment); 278 limit &= ~(alignment - 1); 279 280 if (!base) 281 fixed = false; 282 283 /* size should be aligned with order_per_bit */ 284 if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit)) 285 return -EINVAL; 286 287 /* 288 * If allocating at a fixed base the request region must not cross the 289 * low/high memory boundary. 290 */ 291 if (fixed && base < highmem_start && base + size > highmem_start) { 292 ret = -EINVAL; 293 pr_err("Region at %pa defined on low/high memory boundary (%pa)\n", 294 &base, &highmem_start); 295 goto err; 296 } 297 298 /* 299 * If the limit is unspecified or above the memblock end, its effective 300 * value will be the memblock end. Set it explicitly to simplify further 301 * checks. 302 */ 303 if (limit == 0 || limit > memblock_end) 304 limit = memblock_end; 305 306 if (base + size > limit) { 307 ret = -EINVAL; 308 pr_err("Size (%pa) of region at %pa exceeds limit (%pa)\n", 309 &size, &base, &limit); 310 goto err; 311 } 312 313 /* Reserve memory */ 314 if (fixed) { 315 if (memblock_is_region_reserved(base, size) || 316 memblock_reserve(base, size) < 0) { 317 ret = -EBUSY; 318 goto err; 319 } 320 } else { 321 phys_addr_t addr = 0; 322 323 /* 324 * If there is enough memory, try a bottom-up allocation first. 325 * It will place the new cma area close to the start of the node 326 * and guarantee that the compaction is moving pages out of the 327 * cma area and not into it. 328 * Avoid using first 4GB to not interfere with constrained zones 329 * like DMA/DMA32. 330 */ 331#ifdef CONFIG_PHYS_ADDR_T_64BIT 332 if (!memblock_bottom_up() && memblock_end >= SZ_4G + size) { 333 memblock_set_bottom_up(true); 334 addr = memblock_alloc_range_nid(size, alignment, SZ_4G, 335 limit, nid, true); 336 memblock_set_bottom_up(false); 337 } 338#endif 339 340 /* 341 * All pages in the reserved area must come from the same zone. 342 * If the requested region crosses the low/high memory boundary, 343 * try allocating from high memory first and fall back to low 344 * memory in case of failure. 345 */ 346 if (!addr && base < highmem_start && limit > highmem_start) { 347 addr = memblock_alloc_range_nid(size, alignment, 348 highmem_start, limit, nid, true); 349 limit = highmem_start; 350 } 351 352 if (!addr) { 353 addr = memblock_alloc_range_nid(size, alignment, base, 354 limit, nid, true); 355 if (!addr) { 356 ret = -ENOMEM; 357 goto err; 358 } 359 } 360 361 /* 362 * kmemleak scans/reads tracked objects for pointers to other 363 * objects but this address isn't mapped and accessible 364 */ 365 kmemleak_ignore_phys(addr); 366 base = addr; 367 } 368 369 ret = cma_init_reserved_mem(base, size, order_per_bit, name, res_cma); 370 if (ret) 371 goto free_mem; 372 373 pr_info("Reserved %ld MiB at %pa on node %d\n", (unsigned long)size / SZ_1M, 374 &base, nid); 375 return 0; 376 377free_mem: 378 memblock_phys_free(base, size); 379err: 380 pr_err("Failed to reserve %ld MiB on node %d\n", (unsigned long)size / SZ_1M, 381 nid); 382 return ret; 383} 384 385static void cma_debug_show_areas(struct cma *cma) 386{ 387 unsigned long next_zero_bit, next_set_bit, nr_zero; 388 unsigned long start = 0; 389 unsigned long nr_part, nr_total = 0; 390 unsigned long nbits = cma_bitmap_maxno(cma); 391 392 spin_lock_irq(&cma->lock); 393 pr_info("number of available pages: "); 394 for (;;) { 395 next_zero_bit = find_next_zero_bit(cma->bitmap, nbits, start); 396 if (next_zero_bit >= nbits) 397 break; 398 next_set_bit = find_next_bit(cma->bitmap, nbits, next_zero_bit); 399 nr_zero = next_set_bit - next_zero_bit; 400 nr_part = nr_zero << cma->order_per_bit; 401 pr_cont("%s%lu@%lu", nr_total ? "+" : "", nr_part, 402 next_zero_bit); 403 nr_total += nr_part; 404 start = next_zero_bit + nr_zero; 405 } 406 pr_cont("=> %lu free of %lu total pages\n", nr_total, cma->count); 407 spin_unlock_irq(&cma->lock); 408} 409 410/** 411 * cma_alloc() - allocate pages from contiguous area 412 * @cma: Contiguous memory region for which the allocation is performed. 413 * @count: Requested number of pages. 414 * @align: Requested alignment of pages (in PAGE_SIZE order). 415 * @no_warn: Avoid printing message about failed allocation 416 * 417 * This function allocates part of contiguous memory on specific 418 * contiguous memory area. 419 */ 420struct page *cma_alloc(struct cma *cma, unsigned long count, 421 unsigned int align, bool no_warn) 422{ 423 unsigned long mask, offset; 424 unsigned long pfn = -1; 425 unsigned long start = 0; 426 unsigned long bitmap_maxno, bitmap_no, bitmap_count; 427 unsigned long i; 428 struct page *page = NULL; 429 int ret = -ENOMEM; 430 const char *name = cma ? cma->name : NULL; 431 432 trace_cma_alloc_start(name, count, align); 433 434 if (!cma || !cma->count || !cma->bitmap) 435 return page; 436 437 pr_debug("%s(cma %p, name: %s, count %lu, align %d)\n", __func__, 438 (void *)cma, cma->name, count, align); 439 440 if (!count) 441 return page; 442 443 mask = cma_bitmap_aligned_mask(cma, align); 444 offset = cma_bitmap_aligned_offset(cma, align); 445 bitmap_maxno = cma_bitmap_maxno(cma); 446 bitmap_count = cma_bitmap_pages_to_bits(cma, count); 447 448 if (bitmap_count > bitmap_maxno) 449 return page; 450 451 for (;;) { 452 spin_lock_irq(&cma->lock); 453 bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap, 454 bitmap_maxno, start, bitmap_count, mask, 455 offset); 456 if (bitmap_no >= bitmap_maxno) { 457 spin_unlock_irq(&cma->lock); 458 break; 459 } 460 bitmap_set(cma->bitmap, bitmap_no, bitmap_count); 461 /* 462 * It's safe to drop the lock here. We've marked this region for 463 * our exclusive use. If the migration fails we will take the 464 * lock again and unmark it. 465 */ 466 spin_unlock_irq(&cma->lock); 467 468 pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit); 469 mutex_lock(&cma_mutex); 470 ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA, 471 GFP_KERNEL | (no_warn ? __GFP_NOWARN : 0)); 472 mutex_unlock(&cma_mutex); 473 if (ret == 0) { 474 page = pfn_to_page(pfn); 475 break; 476 } 477 478 cma_clear_bitmap(cma, pfn, count); 479 if (ret != -EBUSY) 480 break; 481 482 pr_debug("%s(): memory range at pfn 0x%lx %p is busy, retrying\n", 483 __func__, pfn, pfn_to_page(pfn)); 484 485 trace_cma_alloc_busy_retry(cma->name, pfn, pfn_to_page(pfn), 486 count, align); 487 /* try again with a bit different memory target */ 488 start = bitmap_no + mask + 1; 489 } 490 491 /* 492 * CMA can allocate multiple page blocks, which results in different 493 * blocks being marked with different tags. Reset the tags to ignore 494 * those page blocks. 495 */ 496 if (page) { 497 for (i = 0; i < count; i++) 498 page_kasan_tag_reset(nth_page(page, i)); 499 } 500 501 if (ret && !no_warn) { 502 pr_err_ratelimited("%s: %s: alloc failed, req-size: %lu pages, ret: %d\n", 503 __func__, cma->name, count, ret); 504 cma_debug_show_areas(cma); 505 } 506 507 pr_debug("%s(): returned %p\n", __func__, page); 508 trace_cma_alloc_finish(name, pfn, page, count, align, ret); 509 if (page) { 510 count_vm_event(CMA_ALLOC_SUCCESS); 511 cma_sysfs_account_success_pages(cma, count); 512 } else { 513 count_vm_event(CMA_ALLOC_FAIL); 514 cma_sysfs_account_fail_pages(cma, count); 515 } 516 517 return page; 518} 519 520bool cma_pages_valid(struct cma *cma, const struct page *pages, 521 unsigned long count) 522{ 523 unsigned long pfn; 524 525 if (!cma || !pages) 526 return false; 527 528 pfn = page_to_pfn(pages); 529 530 if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count) { 531 pr_debug("%s(page %p, count %lu)\n", __func__, 532 (void *)pages, count); 533 return false; 534 } 535 536 return true; 537} 538 539/** 540 * cma_release() - release allocated pages 541 * @cma: Contiguous memory region for which the allocation is performed. 542 * @pages: Allocated pages. 543 * @count: Number of allocated pages. 544 * 545 * This function releases memory allocated by cma_alloc(). 546 * It returns false when provided pages do not belong to contiguous area and 547 * true otherwise. 548 */ 549bool cma_release(struct cma *cma, const struct page *pages, 550 unsigned long count) 551{ 552 unsigned long pfn; 553 554 if (!cma_pages_valid(cma, pages, count)) 555 return false; 556 557 pr_debug("%s(page %p, count %lu)\n", __func__, (void *)pages, count); 558 559 pfn = page_to_pfn(pages); 560 561 VM_BUG_ON(pfn + count > cma->base_pfn + cma->count); 562 563 free_contig_range(pfn, count); 564 cma_clear_bitmap(cma, pfn, count); 565 cma_sysfs_account_release_pages(cma, count); 566 trace_cma_release(cma->name, pfn, pages, count); 567 568 return true; 569} 570 571int cma_for_each_area(int (*it)(struct cma *cma, void *data), void *data) 572{ 573 int i; 574 575 for (i = 0; i < cma_area_count; i++) { 576 int ret = it(&cma_areas[i], data); 577 578 if (ret) 579 return ret; 580 } 581 582 return 0; 583} 584