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 #ifdef CONFIG_CMA_DEBUG
  18 #ifndef DEBUG
  19 #  define DEBUG
  20 #endif
  21 #endif
  22 #define CREATE_TRACE_POINTS
  23
  24 #include <linux/memblock.h>
  25 #include <linux/err.h>
  26 #include <linux/mm.h>
  27 #include <linux/mutex.h>
  28 #include <linux/sizes.h>
  29 #include <linux/slab.h>
  30 #include <linux/log2.h>
  31 #include <linux/cma.h>
  32 #include <linux/highmem.h>
  33 #include <linux/io.h>
  34 #include <linux/kmemleak.h>
  35 #include <trace/events/cma.h>
  36
  37 #include "cma.h"
  38
  39 struct cma cma_areas[MAX_CMA_AREAS];
  40 unsigned cma_area_count;
  41
  42 phys_addr_t cma_get_base(const struct cma *cma)
  43 {
  44         return PFN_PHYS(cma->base_pfn);
  45 }
  46
  47 unsigned long cma_get_size(const struct cma *cma)
  48 {
  49         return cma->count << PAGE_SHIFT;
  50 }
  51
  52 const char *cma_get_name(const struct cma *cma)
  53 {
  54         return cma->name;
  55 }
  56
  57 static unsigned long cma_bitmap_aligned_mask(const struct cma *cma,
  58                                              unsigned int align_order)
  59 {
  60         if (align_order <= cma->order_per_bit)
  61                 return 0;
  62         return (1UL << (align_order - cma->order_per_bit)) - 1;
  63 }
  64
  65 /*
  66  * Find the offset of the base PFN from the specified align_order.
  67  * The value returned is represented in order_per_bits.
  68  */
  69 static unsigned long cma_bitmap_aligned_offset(const struct cma *cma,
  70                                                unsigned int align_order)
  71 {
  72         return (cma->base_pfn & ((1UL << align_order) - 1))
  73                 >> cma->order_per_bit;
  74 }
  75
  76 static unsigned long cma_bitmap_pages_to_bits(const struct cma *cma,
  77                                               unsigned long pages)
  78 {
  79         return ALIGN(pages, 1UL << cma->order_per_bit) >> cma->order_per_bit;
  80 }
  81
  82 static void cma_clear_bitmap(struct cma *cma, unsigned long pfn,
  83                              unsigned int count)
  84 {
  85         unsigned long bitmap_no, bitmap_count;
  86
  87         bitmap_no = (pfn - cma->base_pfn) >> cma->order_per_bit;
  88         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
  89
  90         mutex_lock(&cma->lock);
  91         bitmap_clear(cma->bitmap, bitmap_no, bitmap_count);
  92         mutex_unlock(&cma->lock);
  93 }
  94
  95 static void __init cma_activate_area(struct cma *cma)
  96 {
  97         unsigned long base_pfn = cma->base_pfn, pfn = base_pfn;
  98         unsigned i = cma->count >> pageblock_order;
  99         struct zone *zone;
 100
 101         cma->bitmap = bitmap_zalloc(cma_bitmap_maxno(cma), GFP_KERNEL);
 102         if (!cma->bitmap)
 103                 goto out_error;
 104
 105         WARN_ON_ONCE(!pfn_valid(pfn));
 106         zone = page_zone(pfn_to_page(pfn));
 107
 108         do {
 109                 unsigned j;
 110
 111                 base_pfn = pfn;
 112                 for (j = pageblock_nr_pages; j; --j, pfn++) {
 113                         WARN_ON_ONCE(!pfn_valid(pfn));
 114                         /*
 115                          * alloc_contig_range requires the pfn range
 116                          * specified to be in the same zone. Make this
 117                          * simple by forcing the entire CMA resv range
 118                          * to be in the same zone.
 119                          */
 120                         if (page_zone(pfn_to_page(pfn)) != zone)
 121                                 goto not_in_zone;
 122                 }
 123                 init_cma_reserved_pageblock(pfn_to_page(base_pfn));
 124         } while (--i);
 125
 126         mutex_init(&cma->lock);
 127
 128 #ifdef CONFIG_CMA_DEBUGFS
 129         INIT_HLIST_HEAD(&cma->mem_head);
 130         spin_lock_init(&cma->mem_head_lock);
 131 #endif
 132
 133         return;
 134
 135 not_in_zone:
 136         bitmap_free(cma->bitmap);
 137 out_error:
 138         cma->count = 0;
 139         pr_err("CMA area %s could not be activated\n", cma->name);
 140         return;
 141 }
 142
 143 static int __init cma_init_reserved_areas(void)
 144 {
 145         int i;
 146
 147         for (i = 0; i < cma_area_count; i++)
 148                 cma_activate_area(&cma_areas[i]);
 149
 150         return 0;
 151 }
 152 core_initcall(cma_init_reserved_areas);
 153
 154 /**
 155  * cma_init_reserved_mem() - create custom contiguous area from reserved memory
 156  * @base: Base address of the reserved area
 157  * @size: Size of the reserved area (in bytes),
 158  * @order_per_bit: Order of pages represented by one bit on bitmap.
 159  * @name: The name of the area. If this parameter is NULL, the name of
 160  *        the area will be set to "cmaN", where N is a running counter of
 161  *        used areas.
 162  * @res_cma: Pointer to store the created cma region.
 163  *
 164  * This function creates custom contiguous area from already reserved memory.
 165  */
 166 int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
 167                                  unsigned int order_per_bit,
 168                                  const char *name,
 169                                  struct cma **res_cma)
 170 {
 171         struct cma *cma;
 172         phys_addr_t alignment;
 173
 174         /* Sanity checks */
 175         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 176                 pr_err("Not enough slots for CMA reserved regions!\n");
 177                 return -ENOSPC;
 178         }
 179
 180         if (!size || !memblock_is_region_reserved(base, size))
 181                 return -EINVAL;
 182
 183         /* ensure minimal alignment required by mm core */
 184         alignment = PAGE_SIZE <<
 185                         max_t(unsigned long, MAX_ORDER - 1, pageblock_order);
 186
 187         /* alignment should be aligned with order_per_bit */
 188         if (!IS_ALIGNED(alignment >> PAGE_SHIFT, 1 << order_per_bit))
 189                 return -EINVAL;
 190
 191         if (ALIGN(base, alignment) != base || ALIGN(size, alignment) != size)
 192                 return -EINVAL;
 193
 194         /*
 195          * Each reserved area must be initialised later, when more kernel
 196          * subsystems (like slab allocator) are available.
 197          */
 198         cma = &cma_areas[cma_area_count];
 199
 200         if (name)
 201                 snprintf(cma->name, CMA_MAX_NAME, name);
 202         else
 203                 snprintf(cma->name, CMA_MAX_NAME,  "cma%d\n", cma_area_count);
 204
 205         cma->base_pfn = PFN_DOWN(base);
 206         cma->count = size >> PAGE_SHIFT;
 207         cma->order_per_bit = order_per_bit;
 208         *res_cma = cma;
 209         cma_area_count++;
 210         totalcma_pages += (size / PAGE_SIZE);
 211
 212         return 0;
 213 }
 214
 215 /**
 216  * cma_declare_contiguous_nid() - reserve custom contiguous area
 217  * @base: Base address of the reserved area optional, use 0 for any
 218  * @size: Size of the reserved area (in bytes),
 219  * @limit: End address of the reserved memory (optional, 0 for any).
 220  * @alignment: Alignment for the CMA area, should be power of 2 or zero
 221  * @order_per_bit: Order of pages represented by one bit on bitmap.
 222  * @fixed: hint about where to place the reserved area
 223  * @name: The name of the area. See function cma_init_reserved_mem()
 224  * @res_cma: Pointer to store the created cma region.
 225  * @nid: nid of the free area to find, %NUMA_NO_NODE for any node
 226  *
 227  * This function reserves memory from early allocator. It should be
 228  * called by arch specific code once the early allocator (memblock or bootmem)
 229  * has been activated and all other subsystems have already allocated/reserved
 230  * memory. This function allows to create custom reserved areas.
 231  *
 232  * If @fixed is true, reserve contiguous area at exactly @base.  If false,
 233  * reserve in range from @base to @limit.
 234  */
 235 int __init cma_declare_contiguous_nid(phys_addr_t base,
 236                         phys_addr_t size, phys_addr_t limit,
 237                         phys_addr_t alignment, unsigned int order_per_bit,
 238                         bool fixed, const char *name, struct cma **res_cma,
 239                         int nid)
 240 {
 241         phys_addr_t memblock_end = memblock_end_of_DRAM();
 242         phys_addr_t highmem_start;
 243         int ret = 0;
 244
 245         /*
 246          * We can't use __pa(high_memory) directly, since high_memory
 247          * isn't a valid direct map VA, and DEBUG_VIRTUAL will (validly)
 248          * complain. Find the boundary by adding one to the last valid
 249          * address.
 250          */
 251         highmem_start = __pa(high_memory - 1) + 1;
 252         pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
 253                 __func__, &size, &base, &limit, &alignment);
 254
 255         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 256                 pr_err("Not enough slots for CMA reserved regions!\n");
 257                 return -ENOSPC;
 258         }
 259
 260         if (!size)
 261                 return -EINVAL;
 262
 263         if (alignment && !is_power_of_2(alignment))
 264                 return -EINVAL;
 265
 266         /*
 267          * Sanitise input arguments.
 268          * Pages both ends in CMA area could be merged into adjacent unmovable
 269          * migratetype page by page allocator's buddy algorithm. In the case,
 270          * you couldn't get a contiguous memory, which is not what we want.
 271          */
 272         alignment = max(alignment,  (phys_addr_t)PAGE_SIZE <<
 273                           max_t(unsigned long, MAX_ORDER - 1, pageblock_order));
 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                  * All pages in the reserved area must come from the same zone.
 329                  * If the requested region crosses the low/high memory boundary,
 330                  * try allocating from high memory first and fall back to low
 331                  * memory in case of failure.
 332                  */
 333                 if (base < highmem_start && limit > highmem_start) {
 334                         addr = memblock_alloc_range_nid(size, alignment,
 335                                         highmem_start, limit, nid, true);
 336                         limit = highmem_start;
 337                 }
 338
 339                 if (!addr) {
 340                         addr = memblock_alloc_range_nid(size, alignment, base,
 341                                         limit, nid, true);
 342                         if (!addr) {
 343                                 ret = -ENOMEM;
 344                                 goto err;
 345                         }
 346                 }
 347
 348                 /*
 349                  * kmemleak scans/reads tracked objects for pointers to other
 350                  * objects but this address isn't mapped and accessible
 351                  */
 352                 kmemleak_ignore_phys(addr);
 353                 base = addr;
 354         }
 355
 356         ret = cma_init_reserved_mem(base, size, order_per_bit, name, res_cma);
 357         if (ret)
 358                 goto free_mem;
 359
 360         pr_info("Reserved %ld MiB at %pa\n", (unsigned long)size / SZ_1M,
 361                 &base);
 362         return 0;
 363
 364 free_mem:
 365         memblock_free(base, size);
 366 err:
 367         pr_err("Failed to reserve %ld MiB\n", (unsigned long)size / SZ_1M);
 368         return ret;
 369 }
 370
 371 #ifdef CONFIG_CMA_DEBUG
 372 static void cma_debug_show_areas(struct cma *cma)
 373 {
 374         unsigned long next_zero_bit, next_set_bit, nr_zero;
 375         unsigned long start = 0;
 376         unsigned long nr_part, nr_total = 0;
 377         unsigned long nbits = cma_bitmap_maxno(cma);
 378
 379         mutex_lock(&cma->lock);
 380         pr_info("number of available pages: ");
 381         for (;;) {
 382                 next_zero_bit = find_next_zero_bit(cma->bitmap, nbits, start);
 383                 if (next_zero_bit >= nbits)
 384                         break;
 385                 next_set_bit = find_next_bit(cma->bitmap, nbits, next_zero_bit);
 386                 nr_zero = next_set_bit - next_zero_bit;
 387                 nr_part = nr_zero << cma->order_per_bit;
 388                 pr_cont("%s%lu@%lu", nr_total ? "+" : "", nr_part,
 389                         next_zero_bit);
 390                 nr_total += nr_part;
 391                 start = next_zero_bit + nr_zero;
 392         }
 393         pr_cont("=> %lu free of %lu total pages\n", nr_total, cma->count);
 394         mutex_unlock(&cma->lock);
 395 }
 396 #else
 397 static inline void cma_debug_show_areas(struct cma *cma) { }
 398 #endif
 399
 400 /**
 401  * cma_alloc() - allocate pages from contiguous area
 402  * @cma:   Contiguous memory region for which the allocation is performed.
 403  * @count: Requested number of pages.
 404  * @align: Requested alignment of pages (in PAGE_SIZE order).
 405  * @no_warn: Avoid printing message about failed allocation
 406  *
 407  * This function allocates part of contiguous memory on specific
 408  * contiguous memory area.
 409  */
 410 struct page *cma_alloc(struct cma *cma, size_t count, unsigned int align,
 411                        bool no_warn)
 412 {
 413         unsigned long mask, offset;
 414         unsigned long pfn = -1;
 415         unsigned long start = 0;
 416         unsigned long bitmap_maxno, bitmap_no, bitmap_count;
 417         size_t i;
 418         struct page *page = NULL;
 419         int ret = -ENOMEM;
 420
 421         if (!cma || !cma->count || !cma->bitmap)
 422                 return NULL;
 423
 424         pr_debug("%s(cma %p, count %zu, align %d)\n", __func__, (void *)cma,
 425                  count, align);
 426
 427         if (!count)
 428                 return NULL;
 429
 430         mask = cma_bitmap_aligned_mask(cma, align);
 431         offset = cma_bitmap_aligned_offset(cma, align);
 432         bitmap_maxno = cma_bitmap_maxno(cma);
 433         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
 434
 435         if (bitmap_count > bitmap_maxno)
 436                 return NULL;
 437
 438         for (;;) {
 439                 mutex_lock(&cma->lock);
 440                 bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
 441                                 bitmap_maxno, start, bitmap_count, mask,
 442                                 offset);
 443                 if (bitmap_no >= bitmap_maxno) {
 444                         mutex_unlock(&cma->lock);
 445                         break;
 446                 }
 447                 bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
 448                 /*
 449                  * It's safe to drop the lock here. We've marked this region for
 450                  * our exclusive use. If the migration fails we will take the
 451                  * lock again and unmark it.
 452                  */
 453                 mutex_unlock(&cma->lock);
 454
 455                 pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
 456                 ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA,
 457                                      GFP_KERNEL | (no_warn ? __GFP_NOWARN : 0));
 458
 459                 if (ret == 0) {
 460                         page = pfn_to_page(pfn);
 461                         break;
 462                 }
 463
 464                 cma_clear_bitmap(cma, pfn, count);
 465                 if (ret != -EBUSY)
 466                         break;
 467
 468                 pr_debug("%s(): memory range at %p is busy, retrying\n",
 469                          __func__, pfn_to_page(pfn));
 470                 /* try again with a bit different memory target */
 471                 start = bitmap_no + mask + 1;
 472         }
 473
 474         trace_cma_alloc(pfn, page, count, align);
 475
 476         /*
 477          * CMA can allocate multiple page blocks, which results in different
 478          * blocks being marked with different tags. Reset the tags to ignore
 479          * those page blocks.
 480          */
 481         if (page) {
 482                 for (i = 0; i < count; i++)
 483                         page_kasan_tag_reset(page + i);
 484         }
 485
 486         if (ret && !no_warn) {
 487                 pr_err("%s: alloc failed, req-size: %zu pages, ret: %d\n",
 488                         __func__, count, ret);
 489                 cma_debug_show_areas(cma);
 490         }
 491
 492         pr_debug("%s(): returned %p\n", __func__, page);
 493         return page;
 494 }
 495
 496 /**
 497  * cma_release() - release allocated pages
 498  * @cma:   Contiguous memory region for which the allocation is performed.
 499  * @pages: Allocated pages.
 500  * @count: Number of allocated pages.
 501  *
 502  * This function releases memory allocated by cma_alloc().
 503  * It returns false when provided pages do not belong to contiguous area and
 504  * true otherwise.
 505  */
 506 bool cma_release(struct cma *cma, const struct page *pages, unsigned int count)
 507 {
 508         unsigned long pfn;
 509
 510         if (!cma || !pages)
 511                 return false;
 512
 513         pr_debug("%s(page %p, count %u)\n", __func__, (void *)pages, count);
 514
 515         pfn = page_to_pfn(pages);
 516
 517         if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count)
 518                 return false;
 519
 520         VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
 521
 522         free_contig_range(pfn, count);
 523         cma_clear_bitmap(cma, pfn, count);
 524         trace_cma_release(pfn, pages, count);
 525
 526         return true;
 527 }
 528
 529 int cma_for_each_area(int (*it)(struct cma *cma, void *data), void *data)
 530 {
 531         int i;
 532
 533         for (i = 0; i < cma_area_count; i++) {
 534                 int ret = it(&cma_areas[i], data);
 535
 536                 if (ret)
 537                         return ret;
 538         }
 539
 540         return 0;
 541 }