mm/cma.c

   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
  34 struct cma cma_areas[MAX_CMA_AREAS];
  35 unsigned int cma_area_count;
  36 static DEFINE_MUTEX(cma_mutex);
  37
  38 phys_addr_t cma_get_base(const struct cma *cma)
  39 {
  40         return PFN_PHYS(cma->base_pfn);
  41 }
  42
  43 unsigned long cma_get_size(const struct cma *cma)
  44 {
  45         return cma->count << PAGE_SHIFT;
  46 }
  47
  48 const char *cma_get_name(const struct cma *cma)
  49 {
  50         return cma->name;
  51 }
  52
  53 static 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  */
  65 static 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
  72 static 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
  78 static 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
  92 static 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
 127 not_in_zone:
 128         bitmap_free(cma->bitmap);
 129 out_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 }
 139
 140 static int __init cma_init_reserved_areas(void)
 141 {
 142         int i;
 143
 144         for (i = 0; i < cma_area_count; i++)
 145                 cma_activate_area(&cma_areas[i]);
 146
 147         return 0;
 148 }
 149 core_initcall(cma_init_reserved_areas);
 150
 151 void __init cma_reserve_pages_on_error(struct cma *cma)
 152 {
 153         cma->reserve_pages_on_error = true;
 154 }
 155
 156 /**
 157  * cma_init_reserved_mem() - create custom contiguous area from reserved memory
 158  * @base: Base address of the reserved area
 159  * @size: Size of the reserved area (in bytes),
 160  * @order_per_bit: Order of pages represented by one bit on bitmap.
 161  * @name: The name of the area. If this parameter is NULL, the name of
 162  *        the area will be set to "cmaN", where N is a running counter of
 163  *        used areas.
 164  * @res_cma: Pointer to store the created cma region.
 165  *
 166  * This function creates custom contiguous area from already reserved memory.
 167  */
 168 int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
 169                                  unsigned int order_per_bit,
 170                                  const char *name,
 171                                  struct cma **res_cma)
 172 {
 173         struct cma *cma;
 174
 175         /* Sanity checks */
 176         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 177                 pr_err("Not enough slots for CMA reserved regions!\n");
 178                 return -ENOSPC;
 179         }
 180
 181         if (!size || !memblock_is_region_reserved(base, size))
 182                 return -EINVAL;
 183
 184         /*
 185          * CMA uses CMA_MIN_ALIGNMENT_BYTES as alignment requirement which
 186          * needs pageblock_order to be initialized. Let's enforce it.
 187          */
 188         if (!pageblock_order) {
 189                 pr_err("pageblock_order not yet initialized. Called during early boot?\n");
 190                 return -EINVAL;
 191         }
 192
 193         /* ensure minimal alignment required by mm core */
 194         if (!IS_ALIGNED(base | size, CMA_MIN_ALIGNMENT_BYTES))
 195                 return -EINVAL;
 196
 197         /*
 198          * Each reserved area must be initialised later, when more kernel
 199          * subsystems (like slab allocator) are available.
 200          */
 201         cma = &cma_areas[cma_area_count];
 202
 203         if (name)
 204                 snprintf(cma->name, CMA_MAX_NAME, name);
 205         else
 206                 snprintf(cma->name, CMA_MAX_NAME,  "cma%d\n", cma_area_count);
 207
 208         cma->base_pfn = PFN_DOWN(base);
 209         cma->count = size >> PAGE_SHIFT;
 210         cma->order_per_bit = order_per_bit;
 211         *res_cma = cma;
 212         cma_area_count++;
 213         totalcma_pages += cma->count;
 214
 215         return 0;
 216 }
 217
 218 /**
 219  * cma_declare_contiguous_nid() - reserve custom contiguous area
 220  * @base: Base address of the reserved area optional, use 0 for any
 221  * @size: Size of the reserved area (in bytes),
 222  * @limit: End address of the reserved memory (optional, 0 for any).
 223  * @alignment: Alignment for the CMA area, should be power of 2 or zero
 224  * @order_per_bit: Order of pages represented by one bit on bitmap.
 225  * @fixed: hint about where to place the reserved area
 226  * @name: The name of the area. See function cma_init_reserved_mem()
 227  * @res_cma: Pointer to store the created cma region.
 228  * @nid: nid of the free area to find, %NUMA_NO_NODE for any node
 229  *
 230  * This function reserves memory from early allocator. It should be
 231  * called by arch specific code once the early allocator (memblock or bootmem)
 232  * has been activated and all other subsystems have already allocated/reserved
 233  * memory. This function allows to create custom reserved areas.
 234  *
 235  * If @fixed is true, reserve contiguous area at exactly @base.  If false,
 236  * reserve in range from @base to @limit.
 237  */
 238 int __init cma_declare_contiguous_nid(phys_addr_t base,
 239                         phys_addr_t size, phys_addr_t limit,
 240                         phys_addr_t alignment, unsigned int order_per_bit,
 241                         bool fixed, const char *name, struct cma **res_cma,
 242                         int nid)
 243 {
 244         phys_addr_t memblock_end = memblock_end_of_DRAM();
 245         phys_addr_t highmem_start;
 246         int ret;
 247
 248         /*
 249          * We can't use __pa(high_memory) directly, since high_memory
 250          * isn't a valid direct map VA, and DEBUG_VIRTUAL will (validly)
 251          * complain. Find the boundary by adding one to the last valid
 252          * address.
 253          */
 254         highmem_start = __pa(high_memory - 1) + 1;
 255         pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
 256                 __func__, &size, &base, &limit, &alignment);
 257
 258         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 259                 pr_err("Not enough slots for CMA reserved regions!\n");
 260                 return -ENOSPC;
 261         }
 262
 263         if (!size)
 264                 return -EINVAL;
 265
 266         if (alignment && !is_power_of_2(alignment))
 267                 return -EINVAL;
 268
 269         if (!IS_ENABLED(CONFIG_NUMA))
 270                 nid = NUMA_NO_NODE;
 271
 272         /* Sanitise input arguments. */
 273         alignment = max_t(phys_addr_t, alignment, CMA_MIN_ALIGNMENT_BYTES);
 274         if (fixed && base & (alignment - 1)) {
 275                 ret = -EINVAL;
 276                 pr_err("Region at %pa must be aligned to %pa bytes\n",
 277                         &base, &alignment);
 278                 goto err;
 279         }
 280         base = ALIGN(base, alignment);
 281         size = ALIGN(size, alignment);
 282         limit &= ~(alignment - 1);
 283
 284         if (!base)
 285                 fixed = false;
 286
 287         /* size should be aligned with order_per_bit */
 288         if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit))
 289                 return -EINVAL;
 290
 291         /*
 292          * If allocating at a fixed base the request region must not cross the
 293          * low/high memory boundary.
 294          */
 295         if (fixed && base < highmem_start && base + size > highmem_start) {
 296                 ret = -EINVAL;
 297                 pr_err("Region at %pa defined on low/high memory boundary (%pa)\n",
 298                         &base, &highmem_start);
 299                 goto err;
 300         }
 301
 302         /*
 303          * If the limit is unspecified or above the memblock end, its effective
 304          * value will be the memblock end. Set it explicitly to simplify further
 305          * checks.
 306          */
 307         if (limit == 0 || limit > memblock_end)
 308                 limit = memblock_end;
 309
 310         if (base + size > limit) {
 311                 ret = -EINVAL;
 312                 pr_err("Size (%pa) of region at %pa exceeds limit (%pa)\n",
 313                         &size, &base, &limit);
 314                 goto err;
 315         }
 316
 317         /* Reserve memory */
 318         if (fixed) {
 319                 if (memblock_is_region_reserved(base, size) ||
 320                     memblock_reserve(base, size) < 0) {
 321                         ret = -EBUSY;
 322                         goto err;
 323                 }
 324         } else {
 325                 phys_addr_t addr = 0;
 326
 327                 /*
 328                  * If there is enough memory, try a bottom-up allocation first.
 329                  * It will place the new cma area close to the start of the node
 330                  * and guarantee that the compaction is moving pages out of the
 331                  * cma area and not into it.
 332                  * Avoid using first 4GB to not interfere with constrained zones
 333                  * like DMA/DMA32.
 334                  */
 335 #ifdef CONFIG_PHYS_ADDR_T_64BIT
 336                 if (!memblock_bottom_up() && memblock_end >= SZ_4G + size) {
 337                         memblock_set_bottom_up(true);
 338                         addr = memblock_alloc_range_nid(size, alignment, SZ_4G,
 339                                                         limit, nid, true);
 340                         memblock_set_bottom_up(false);
 341                 }
 342 #endif
 343
 344                 /*
 345                  * All pages in the reserved area must come from the same zone.
 346                  * If the requested region crosses the low/high memory boundary,
 347                  * try allocating from high memory first and fall back to low
 348                  * memory in case of failure.
 349                  */
 350                 if (!addr && base < highmem_start && limit > highmem_start) {
 351                         addr = memblock_alloc_range_nid(size, alignment,
 352                                         highmem_start, limit, nid, true);
 353                         limit = highmem_start;
 354                 }
 355
 356                 if (!addr) {
 357                         addr = memblock_alloc_range_nid(size, alignment, base,
 358                                         limit, nid, true);
 359                         if (!addr) {
 360                                 ret = -ENOMEM;
 361                                 goto err;
 362                         }
 363                 }
 364
 365                 /*
 366                  * kmemleak scans/reads tracked objects for pointers to other
 367                  * objects but this address isn't mapped and accessible
 368                  */
 369                 kmemleak_ignore_phys(addr);
 370                 base = addr;
 371         }
 372
 373         ret = cma_init_reserved_mem(base, size, order_per_bit, name, res_cma);
 374         if (ret)
 375                 goto free_mem;
 376
 377         pr_info("Reserved %ld MiB at %pa on node %d\n", (unsigned long)size / SZ_1M,
 378                 &base, nid);
 379         return 0;
 380
 381 free_mem:
 382         memblock_phys_free(base, size);
 383 err:
 384         pr_err("Failed to reserve %ld MiB on node %d\n", (unsigned long)size / SZ_1M,
 385                nid);
 386         return ret;
 387 }
 388
 389 static void cma_debug_show_areas(struct cma *cma)
 390 {
 391         unsigned long next_zero_bit, next_set_bit, nr_zero;
 392         unsigned long start = 0;
 393         unsigned long nr_part, nr_total = 0;
 394         unsigned long nbits = cma_bitmap_maxno(cma);
 395
 396         spin_lock_irq(&cma->lock);
 397         pr_info("number of available pages: ");
 398         for (;;) {
 399                 next_zero_bit = find_next_zero_bit(cma->bitmap, nbits, start);
 400                 if (next_zero_bit >= nbits)
 401                         break;
 402                 next_set_bit = find_next_bit(cma->bitmap, nbits, next_zero_bit);
 403                 nr_zero = next_set_bit - next_zero_bit;
 404                 nr_part = nr_zero << cma->order_per_bit;
 405                 pr_cont("%s%lu@%lu", nr_total ? "+" : "", nr_part,
 406                         next_zero_bit);
 407                 nr_total += nr_part;
 408                 start = next_zero_bit + nr_zero;
 409         }
 410         pr_cont("=> %lu free of %lu total pages\n", nr_total, cma->count);
 411         spin_unlock_irq(&cma->lock);
 412 }
 413
 414 static struct page *__cma_alloc(struct cma *cma, unsigned long count,
 415                                 unsigned int align, gfp_t gfp)
 416 {
 417         unsigned long mask, offset;
 418         unsigned long pfn = -1;
 419         unsigned long start = 0;
 420         unsigned long bitmap_maxno, bitmap_no, bitmap_count;
 421         unsigned long i;
 422         struct page *page = NULL;
 423         int ret = -ENOMEM;
 424         const char *name = cma ? cma->name : NULL;
 425
 426         trace_cma_alloc_start(name, count, align);
 427
 428         if (!cma || !cma->count || !cma->bitmap)
 429                 return page;
 430
 431         pr_debug("%s(cma %p, name: %s, count %lu, align %d)\n", __func__,
 432                 (void *)cma, cma->name, count, align);
 433
 434         if (!count)
 435                 return page;
 436
 437         mask = cma_bitmap_aligned_mask(cma, align);
 438         offset = cma_bitmap_aligned_offset(cma, align);
 439         bitmap_maxno = cma_bitmap_maxno(cma);
 440         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
 441
 442         if (bitmap_count > bitmap_maxno)
 443                 return page;
 444
 445         for (;;) {
 446                 spin_lock_irq(&cma->lock);
 447                 bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
 448                                 bitmap_maxno, start, bitmap_count, mask,
 449                                 offset);
 450                 if (bitmap_no >= bitmap_maxno) {
 451                         spin_unlock_irq(&cma->lock);
 452                         break;
 453                 }
 454                 bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
 455                 /*
 456                  * It's safe to drop the lock here. We've marked this region for
 457                  * our exclusive use. If the migration fails we will take the
 458                  * lock again and unmark it.
 459                  */
 460                 spin_unlock_irq(&cma->lock);
 461
 462                 pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
 463                 mutex_lock(&cma_mutex);
 464                 ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA, gfp);
 465                 mutex_unlock(&cma_mutex);
 466                 if (ret == 0) {
 467                         page = pfn_to_page(pfn);
 468                         break;
 469                 }
 470
 471                 cma_clear_bitmap(cma, pfn, count);
 472                 if (ret != -EBUSY)
 473                         break;
 474
 475                 pr_debug("%s(): memory range at pfn 0x%lx %p is busy, retrying\n",
 476                          __func__, pfn, pfn_to_page(pfn));
 477
 478                 trace_cma_alloc_busy_retry(cma->name, pfn, pfn_to_page(pfn),
 479                                            count, align);
 480                 /* try again with a bit different memory target */
 481                 start = bitmap_no + mask + 1;
 482         }
 483
 484         /*
 485          * CMA can allocate multiple page blocks, which results in different
 486          * blocks being marked with different tags. Reset the tags to ignore
 487          * those page blocks.
 488          */
 489         if (page) {
 490                 for (i = 0; i < count; i++)
 491                         page_kasan_tag_reset(nth_page(page, i));
 492         }
 493
 494         if (ret && !(gfp & __GFP_NOWARN)) {
 495                 pr_err_ratelimited("%s: %s: alloc failed, req-size: %lu pages, ret: %d\n",
 496                                    __func__, cma->name, count, ret);
 497                 cma_debug_show_areas(cma);
 498         }
 499
 500         pr_debug("%s(): returned %p\n", __func__, page);
 501         trace_cma_alloc_finish(name, pfn, page, count, align, ret);
 502         if (page) {
 503                 count_vm_event(CMA_ALLOC_SUCCESS);
 504                 cma_sysfs_account_success_pages(cma, count);
 505         } else {
 506                 count_vm_event(CMA_ALLOC_FAIL);
 507                 cma_sysfs_account_fail_pages(cma, count);
 508         }
 509
 510         return page;
 511 }
 512
 513 /**
 514  * cma_alloc() - allocate pages from contiguous area
 515  * @cma:   Contiguous memory region for which the allocation is performed.
 516  * @count: Requested number of pages.
 517  * @align: Requested alignment of pages (in PAGE_SIZE order).
 518  * @no_warn: Avoid printing message about failed allocation
 519  *
 520  * This function allocates part of contiguous memory on specific
 521  * contiguous memory area.
 522  */
 523 struct page *cma_alloc(struct cma *cma, unsigned long count,
 524                        unsigned int align, bool no_warn)
 525 {
 526         return __cma_alloc(cma, count, align, GFP_KERNEL | (no_warn ? __GFP_NOWARN : 0));
 527 }
 528
 529 struct folio *cma_alloc_folio(struct cma *cma, int order, gfp_t gfp)
 530 {
 531         struct page *page;
 532
 533         if (WARN_ON(!order || !(gfp & __GFP_COMP)))
 534                 return NULL;
 535
 536         page = __cma_alloc(cma, 1 << order, order, gfp);
 537
 538         return page ? page_folio(page) : NULL;
 539 }
 540
 541 bool cma_pages_valid(struct cma *cma, const struct page *pages,
 542                      unsigned long count)
 543 {
 544         unsigned long pfn;
 545
 546         if (!cma || !pages)
 547                 return false;
 548
 549         pfn = page_to_pfn(pages);
 550
 551         if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count) {
 552                 pr_debug("%s(page %p, count %lu)\n", __func__,
 553                                                 (void *)pages, count);
 554                 return false;
 555         }
 556
 557         return true;
 558 }
 559
 560 /**
 561  * cma_release() - release allocated pages
 562  * @cma:   Contiguous memory region for which the allocation is performed.
 563  * @pages: Allocated pages.
 564  * @count: Number of allocated pages.
 565  *
 566  * This function releases memory allocated by cma_alloc().
 567  * It returns false when provided pages do not belong to contiguous area and
 568  * true otherwise.
 569  */
 570 bool cma_release(struct cma *cma, const struct page *pages,
 571                  unsigned long count)
 572 {
 573         unsigned long pfn;
 574
 575         if (!cma_pages_valid(cma, pages, count))
 576                 return false;
 577
 578         pr_debug("%s(page %p, count %lu)\n", __func__, (void *)pages, count);
 579
 580         pfn = page_to_pfn(pages);
 581
 582         VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
 583
 584         free_contig_range(pfn, count);
 585         cma_clear_bitmap(cma, pfn, count);
 586         cma_sysfs_account_release_pages(cma, count);
 587         trace_cma_release(cma->name, pfn, pages, count);
 588
 589         return true;
 590 }
 591
 592 bool cma_free_folio(struct cma *cma, const struct folio *folio)
 593 {
 594         if (WARN_ON(!folio_test_large(folio)))
 595                 return false;
 596
 597         return cma_release(cma, &folio->page, folio_nr_pages(folio));
 598 }
 599
 600 int cma_for_each_area(int (*it)(struct cma *cma, void *data), void *data)
 601 {
 602         int i;
 603
 604         for (i = 0; i < cma_area_count; i++) {
 605                 int ret = it(&cma_areas[i], data);
 606
 607                 if (ret)
 608                         return ret;
 609         }
 610
 611         return 0;
 612 }