mm/cma.c

   1 /*
   2  * Contiguous Memory Allocator
   3  *
   4  * Copyright (c) 2010-2011 by Samsung Electronics.
   5  * Copyright IBM Corporation, 2013
   6  * Copyright LG Electronics Inc., 2014
   7  * Written by:
   8  *      Marek Szyprowski <m.szyprowski@samsung.com>
   9  *      Michal Nazarewicz <mina86@mina86.com>
  10  *      Aneesh Kumar K.V <aneesh.kumar@linux.vnet.ibm.com>
  11  *      Joonsoo Kim <iamjoonsoo.kim@lge.com>
  12  *
  13  * This program is free software; you can redistribute it and/or
  14  * modify it under the terms of the GNU General Public License as
  15  * published by the Free Software Foundation; either version 2 of the
  16  * License or (at your optional) any later version of the license.
  17  */
  18
  19 #define pr_fmt(fmt) "cma: " fmt
  20
  21 #ifdef CONFIG_CMA_DEBUG
  22 #ifndef DEBUG
  23 #  define DEBUG
  24 #endif
  25 #endif
  26
  27 #include <linux/memblock.h>
  28 #include <linux/err.h>
  29 #include <linux/mm.h>
  30 #include <linux/mutex.h>
  31 #include <linux/sizes.h>
  32 #include <linux/slab.h>
  33 #include <linux/log2.h>
  34 #include <linux/cma.h>
  35 #include <linux/highmem.h>
  36 #include <linux/io.h>
  37
  38 struct cma {
  39         unsigned long   base_pfn;
  40         unsigned long   count;
  41         unsigned long   *bitmap;
  42         unsigned int order_per_bit; /* Order of pages represented by one bit */
  43         struct mutex    lock;
  44 };
  45
  46 static struct cma cma_areas[MAX_CMA_AREAS];
  47 static unsigned cma_area_count;
  48 static DEFINE_MUTEX(cma_mutex);
  49
  50 phys_addr_t cma_get_base(struct cma *cma)
  51 {
  52         return PFN_PHYS(cma->base_pfn);
  53 }
  54
  55 unsigned long cma_get_size(struct cma *cma)
  56 {
  57         return cma->count << PAGE_SHIFT;
  58 }
  59
  60 static unsigned long cma_bitmap_aligned_mask(struct cma *cma, int align_order)
  61 {
  62         if (align_order <= cma->order_per_bit)
  63                 return 0;
  64         return (1UL << (align_order - cma->order_per_bit)) - 1;
  65 }
  66
  67 static unsigned long cma_bitmap_aligned_offset(struct cma *cma, int align_order)
  68 {
  69         unsigned int alignment;
  70
  71         if (align_order <= cma->order_per_bit)
  72                 return 0;
  73         alignment = 1UL << (align_order - cma->order_per_bit);
  74         return ALIGN(cma->base_pfn, alignment) -
  75                 (cma->base_pfn >> cma->order_per_bit);
  76 }
  77
  78 static unsigned long cma_bitmap_maxno(struct cma *cma)
  79 {
  80         return cma->count >> cma->order_per_bit;
  81 }
  82
  83 static unsigned long cma_bitmap_pages_to_bits(struct cma *cma,
  84                                                 unsigned long pages)
  85 {
  86         return ALIGN(pages, 1UL << cma->order_per_bit) >> cma->order_per_bit;
  87 }
  88
  89 static void cma_clear_bitmap(struct cma *cma, unsigned long pfn, int count)
  90 {
  91         unsigned long bitmap_no, bitmap_count;
  92
  93         bitmap_no = (pfn - cma->base_pfn) >> cma->order_per_bit;
  94         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
  95
  96         mutex_lock(&cma->lock);
  97         bitmap_clear(cma->bitmap, bitmap_no, bitmap_count);
  98         mutex_unlock(&cma->lock);
  99 }
 100
 101 static int __init cma_activate_area(struct cma *cma)
 102 {
 103         int bitmap_size = BITS_TO_LONGS(cma_bitmap_maxno(cma)) * sizeof(long);
 104         unsigned long base_pfn = cma->base_pfn, pfn = base_pfn;
 105         unsigned i = cma->count >> pageblock_order;
 106         struct zone *zone;
 107
 108         cma->bitmap = kzalloc(bitmap_size, GFP_KERNEL);
 109
 110         if (!cma->bitmap)
 111                 return -ENOMEM;
 112
 113         WARN_ON_ONCE(!pfn_valid(pfn));
 114         zone = page_zone(pfn_to_page(pfn));
 115
 116         do {
 117                 unsigned j;
 118
 119                 base_pfn = pfn;
 120                 for (j = pageblock_nr_pages; j; --j, pfn++) {
 121                         WARN_ON_ONCE(!pfn_valid(pfn));
 122                         /*
 123                          * alloc_contig_range requires the pfn range
 124                          * specified to be in the same zone. Make this
 125                          * simple by forcing the entire CMA resv range
 126                          * to be in the same zone.
 127                          */
 128                         if (page_zone(pfn_to_page(pfn)) != zone)
 129                                 goto err;
 130                 }
 131                 init_cma_reserved_pageblock(pfn_to_page(base_pfn));
 132         } while (--i);
 133
 134         mutex_init(&cma->lock);
 135         return 0;
 136
 137 err:
 138         kfree(cma->bitmap);
 139         cma->count = 0;
 140         return -EINVAL;
 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                 int ret = cma_activate_area(&cma_areas[i]);
 149
 150                 if (ret)
 151                         return ret;
 152         }
 153
 154         return 0;
 155 }
 156 core_initcall(cma_init_reserved_areas);
 157
 158 /**
 159  * cma_init_reserved_mem() - create custom contiguous area from reserved memory
 160  * @base: Base address of the reserved area
 161  * @size: Size of the reserved area (in bytes),
 162  * @order_per_bit: Order of pages represented by one bit on bitmap.
 163  * @res_cma: Pointer to store the created cma region.
 164  *
 165  * This function creates custom contiguous area from already reserved memory.
 166  */
 167 int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
 168                                  int order_per_bit, struct cma **res_cma)
 169 {
 170         struct cma *cma;
 171         phys_addr_t alignment;
 172
 173         /* Sanity checks */
 174         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 175                 pr_err("Not enough slots for CMA reserved regions!\n");
 176                 return -ENOSPC;
 177         }
 178
 179         if (!size || !memblock_is_region_reserved(base, size))
 180                 return -EINVAL;
 181
 182         /* ensure minimal alignment requied by mm core */
 183         alignment = PAGE_SIZE << max(MAX_ORDER - 1, pageblock_order);
 184
 185         /* alignment should be aligned with order_per_bit */
 186         if (!IS_ALIGNED(alignment >> PAGE_SHIFT, 1 << order_per_bit))
 187                 return -EINVAL;
 188
 189         if (ALIGN(base, alignment) != base || ALIGN(size, alignment) != size)
 190                 return -EINVAL;
 191
 192         /*
 193          * Each reserved area must be initialised later, when more kernel
 194          * subsystems (like slab allocator) are available.
 195          */
 196         cma = &cma_areas[cma_area_count];
 197         cma->base_pfn = PFN_DOWN(base);
 198         cma->count = size >> PAGE_SHIFT;
 199         cma->order_per_bit = order_per_bit;
 200         *res_cma = cma;
 201         cma_area_count++;
 202         totalcma_pages += (size / PAGE_SIZE);
 203
 204         return 0;
 205 }
 206
 207 /**
 208  * cma_declare_contiguous() - reserve custom contiguous area
 209  * @base: Base address of the reserved area optional, use 0 for any
 210  * @size: Size of the reserved area (in bytes),
 211  * @limit: End address of the reserved memory (optional, 0 for any).
 212  * @alignment: Alignment for the CMA area, should be power of 2 or zero
 213  * @order_per_bit: Order of pages represented by one bit on bitmap.
 214  * @fixed: hint about where to place the reserved area
 215  * @res_cma: Pointer to store the created cma region.
 216  *
 217  * This function reserves memory from early allocator. It should be
 218  * called by arch specific code once the early allocator (memblock or bootmem)
 219  * has been activated and all other subsystems have already allocated/reserved
 220  * memory. This function allows to create custom reserved areas.
 221  *
 222  * If @fixed is true, reserve contiguous area at exactly @base.  If false,
 223  * reserve in range from @base to @limit.
 224  */
 225 int __init cma_declare_contiguous(phys_addr_t base,
 226                         phys_addr_t size, phys_addr_t limit,
 227                         phys_addr_t alignment, unsigned int order_per_bit,
 228                         bool fixed, struct cma **res_cma)
 229 {
 230         phys_addr_t memblock_end = memblock_end_of_DRAM();
 231         phys_addr_t highmem_start;
 232         int ret = 0;
 233
 234 #ifdef CONFIG_X86
 235         /*
 236          * high_memory isn't direct mapped memory so retrieving its physical
 237          * address isn't appropriate.  But it would be useful to check the
 238          * physical address of the highmem boundary so it's justfiable to get
 239          * the physical address from it.  On x86 there is a validation check for
 240          * this case, so the following workaround is needed to avoid it.
 241          */
 242         highmem_start = __pa_nodebug(high_memory);
 243 #else
 244         highmem_start = __pa(high_memory);
 245 #endif
 246         pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
 247                 __func__, &size, &base, &limit, &alignment);
 248
 249         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 250                 pr_err("Not enough slots for CMA reserved regions!\n");
 251                 return -ENOSPC;
 252         }
 253
 254         if (!size)
 255                 return -EINVAL;
 256
 257         if (alignment && !is_power_of_2(alignment))
 258                 return -EINVAL;
 259
 260         /*
 261          * Sanitise input arguments.
 262          * Pages both ends in CMA area could be merged into adjacent unmovable
 263          * migratetype page by page allocator's buddy algorithm. In the case,
 264          * you couldn't get a contiguous memory, which is not what we want.
 265          */
 266         alignment = max(alignment,
 267                 (phys_addr_t)PAGE_SIZE << max(MAX_ORDER - 1, pageblock_order));
 268         base = ALIGN(base, alignment);
 269         size = ALIGN(size, alignment);
 270         limit &= ~(alignment - 1);
 271
 272         if (!base)
 273                 fixed = false;
 274
 275         /* size should be aligned with order_per_bit */
 276         if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit))
 277                 return -EINVAL;
 278
 279         /*
 280          * If allocating at a fixed base the request region must not cross the
 281          * low/high memory boundary.
 282          */
 283         if (fixed && base < highmem_start && base + size > highmem_start) {
 284                 ret = -EINVAL;
 285                 pr_err("Region at %pa defined on low/high memory boundary (%pa)\n",
 286                         &base, &highmem_start);
 287                 goto err;
 288         }
 289
 290         /*
 291          * If the limit is unspecified or above the memblock end, its effective
 292          * value will be the memblock end. Set it explicitly to simplify further
 293          * checks.
 294          */
 295         if (limit == 0 || limit > memblock_end)
 296                 limit = memblock_end;
 297
 298         /* Reserve memory */
 299         if (fixed) {
 300                 if (memblock_is_region_reserved(base, size) ||
 301                     memblock_reserve(base, size) < 0) {
 302                         ret = -EBUSY;
 303                         goto err;
 304                 }
 305         } else {
 306                 phys_addr_t addr = 0;
 307
 308                 /*
 309                  * All pages in the reserved area must come from the same zone.
 310                  * If the requested region crosses the low/high memory boundary,
 311                  * try allocating from high memory first and fall back to low
 312                  * memory in case of failure.
 313                  */
 314                 if (base < highmem_start && limit > highmem_start) {
 315                         addr = memblock_alloc_range(size, alignment,
 316                                                     highmem_start, limit);
 317                         limit = highmem_start;
 318                 }
 319
 320                 if (!addr) {
 321                         addr = memblock_alloc_range(size, alignment, base,
 322                                                     limit);
 323                         if (!addr) {
 324                                 ret = -ENOMEM;
 325                                 goto err;
 326                         }
 327                 }
 328
 329                 /*
 330                  * kmemleak scans/reads tracked objects for pointers to other
 331                  * objects but this address isn't mapped and accessible
 332                  */
 333                 kmemleak_ignore(phys_to_virt(addr));
 334                 base = addr;
 335         }
 336
 337         ret = cma_init_reserved_mem(base, size, order_per_bit, res_cma);
 338         if (ret)
 339                 goto err;
 340
 341         pr_info("Reserved %ld MiB at %pa\n", (unsigned long)size / SZ_1M,
 342                 &base);
 343         return 0;
 344
 345 err:
 346         pr_err("Failed to reserve %ld MiB\n", (unsigned long)size / SZ_1M);
 347         return ret;
 348 }
 349
 350 /**
 351  * cma_alloc() - allocate pages from contiguous area
 352  * @cma:   Contiguous memory region for which the allocation is performed.
 353  * @count: Requested number of pages.
 354  * @align: Requested alignment of pages (in PAGE_SIZE order).
 355  *
 356  * This function allocates part of contiguous memory on specific
 357  * contiguous memory area.
 358  */
 359 struct page *cma_alloc(struct cma *cma, int count, unsigned int align)
 360 {
 361         unsigned long mask, offset, pfn, start = 0;
 362         unsigned long bitmap_maxno, bitmap_no, bitmap_count;
 363         struct page *page = NULL;
 364         int ret;
 365
 366         if (!cma || !cma->count)
 367                 return NULL;
 368
 369         pr_debug("%s(cma %p, count %d, align %d)\n", __func__, (void *)cma,
 370                  count, align);
 371
 372         if (!count)
 373                 return NULL;
 374
 375         mask = cma_bitmap_aligned_mask(cma, align);
 376         offset = cma_bitmap_aligned_offset(cma, align);
 377         bitmap_maxno = cma_bitmap_maxno(cma);
 378         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
 379
 380         for (;;) {
 381                 mutex_lock(&cma->lock);
 382                 bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
 383                                 bitmap_maxno, start, bitmap_count, mask,
 384                                 offset);
 385                 if (bitmap_no >= bitmap_maxno) {
 386                         mutex_unlock(&cma->lock);
 387                         break;
 388                 }
 389                 bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
 390                 /*
 391                  * It's safe to drop the lock here. We've marked this region for
 392                  * our exclusive use. If the migration fails we will take the
 393                  * lock again and unmark it.
 394                  */
 395                 mutex_unlock(&cma->lock);
 396
 397                 pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
 398                 mutex_lock(&cma_mutex);
 399                 ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA);
 400                 mutex_unlock(&cma_mutex);
 401                 if (ret == 0) {
 402                         page = pfn_to_page(pfn);
 403                         break;
 404                 }
 405
 406                 cma_clear_bitmap(cma, pfn, count);
 407                 if (ret != -EBUSY)
 408                         break;
 409
 410                 pr_debug("%s(): memory range at %p is busy, retrying\n",
 411                          __func__, pfn_to_page(pfn));
 412                 /* try again with a bit different memory target */
 413                 start = bitmap_no + mask + 1;
 414         }
 415
 416         pr_debug("%s(): returned %p\n", __func__, page);
 417         return page;
 418 }
 419
 420 /**
 421  * cma_release() - release allocated pages
 422  * @cma:   Contiguous memory region for which the allocation is performed.
 423  * @pages: Allocated pages.
 424  * @count: Number of allocated pages.
 425  *
 426  * This function releases memory allocated by alloc_cma().
 427  * It returns false when provided pages do not belong to contiguous area and
 428  * true otherwise.
 429  */
 430 bool cma_release(struct cma *cma, struct page *pages, int count)
 431 {
 432         unsigned long pfn;
 433
 434         if (!cma || !pages)
 435                 return false;
 436
 437         pr_debug("%s(page %p)\n", __func__, (void *)pages);
 438
 439         pfn = page_to_pfn(pages);
 440
 441         if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count)
 442                 return false;
 443
 444         VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
 445
 446         free_contig_range(pfn, count);
 447         cma_clear_bitmap(cma, pfn, count);
 448
 449         return true;
 450 }