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 #define CREATE_TRACE_POINTS
  27
  28 #include <linux/memblock.h>
  29 #include <linux/err.h>
  30 #include <linux/mm.h>
  31 #include <linux/mutex.h>
  32 #include <linux/sizes.h>
  33 #include <linux/slab.h>
  34 #include <linux/log2.h>
  35 #include <linux/cma.h>
  36 #include <linux/highmem.h>
  37 #include <linux/io.h>
  38 #include <trace/events/cma.h>
  39
  40 #include "cma.h"
  41
  42 struct cma cma_areas[MAX_CMA_AREAS];
  43 unsigned cma_area_count;
  44 static DEFINE_MUTEX(cma_mutex);
  45
  46 phys_addr_t cma_get_base(const struct cma *cma)
  47 {
  48         return PFN_PHYS(cma->base_pfn);
  49 }
  50
  51 unsigned long cma_get_size(const struct cma *cma)
  52 {
  53         return cma->count << PAGE_SHIFT;
  54 }
  55
  56 static unsigned long cma_bitmap_aligned_mask(const struct cma *cma,
  57                                              int align_order)
  58 {
  59         if (align_order <= cma->order_per_bit)
  60                 return 0;
  61         return (1UL << (align_order - cma->order_per_bit)) - 1;
  62 }
  63
  64 /*
  65  * Find a PFN aligned to the specified order and return an offset represented in
  66  * order_per_bits.
  67  */
  68 static unsigned long cma_bitmap_aligned_offset(const struct cma *cma,
  69                                                int align_order)
  70 {
  71         if (align_order <= cma->order_per_bit)
  72                 return 0;
  73
  74         return (ALIGN(cma->base_pfn, (1UL << align_order))
  75                 - cma->base_pfn) >> cma->order_per_bit;
  76 }
  77
  78 static unsigned long cma_bitmap_pages_to_bits(const struct cma *cma,
  79                                               unsigned long pages)
  80 {
  81         return ALIGN(pages, 1UL << cma->order_per_bit) >> cma->order_per_bit;
  82 }
  83
  84 static void cma_clear_bitmap(struct cma *cma, unsigned long pfn,
  85                              unsigned int count)
  86 {
  87         unsigned long bitmap_no, bitmap_count;
  88
  89         bitmap_no = (pfn - cma->base_pfn) >> cma->order_per_bit;
  90         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
  91
  92         mutex_lock(&cma->lock);
  93         bitmap_clear(cma->bitmap, bitmap_no, bitmap_count);
  94         mutex_unlock(&cma->lock);
  95 }
  96
  97 static int __init cma_activate_area(struct cma *cma)
  98 {
  99         int bitmap_size = BITS_TO_LONGS(cma_bitmap_maxno(cma)) * sizeof(long);
 100         unsigned long base_pfn = cma->base_pfn, pfn = base_pfn;
 101         unsigned i = cma->count >> pageblock_order;
 102         struct zone *zone;
 103
 104         cma->bitmap = kzalloc(bitmap_size, GFP_KERNEL);
 105
 106         if (!cma->bitmap)
 107                 return -ENOMEM;
 108
 109         WARN_ON_ONCE(!pfn_valid(pfn));
 110         zone = page_zone(pfn_to_page(pfn));
 111
 112         do {
 113                 unsigned j;
 114
 115                 base_pfn = pfn;
 116                 for (j = pageblock_nr_pages; j; --j, pfn++) {
 117                         WARN_ON_ONCE(!pfn_valid(pfn));
 118                         /*
 119                          * alloc_contig_range requires the pfn range
 120                          * specified to be in the same zone. Make this
 121                          * simple by forcing the entire CMA resv range
 122                          * to be in the same zone.
 123                          */
 124                         if (page_zone(pfn_to_page(pfn)) != zone)
 125                                 goto err;
 126                 }
 127                 init_cma_reserved_pageblock(pfn_to_page(base_pfn));
 128         } while (--i);
 129
 130         mutex_init(&cma->lock);
 131
 132 #ifdef CONFIG_CMA_DEBUGFS
 133         INIT_HLIST_HEAD(&cma->mem_head);
 134         spin_lock_init(&cma->mem_head_lock);
 135 #endif
 136
 137         return 0;
 138
 139 err:
 140         kfree(cma->bitmap);
 141         cma->count = 0;
 142         return -EINVAL;
 143 }
 144
 145 static int __init cma_init_reserved_areas(void)
 146 {
 147         int i;
 148
 149         for (i = 0; i < cma_area_count; i++) {
 150                 int ret = cma_activate_area(&cma_areas[i]);
 151
 152                 if (ret)
 153                         return ret;
 154         }
 155
 156         return 0;
 157 }
 158 core_initcall(cma_init_reserved_areas);
 159
 160 /**
 161  * cma_init_reserved_mem() - create custom contiguous area from reserved memory
 162  * @base: Base address of the reserved area
 163  * @size: Size of the reserved area (in bytes),
 164  * @order_per_bit: Order of pages represented by one bit on bitmap.
 165  * @res_cma: Pointer to store the created cma region.
 166  *
 167  * This function creates custom contiguous area from already reserved memory.
 168  */
 169 int __init cma_init_reserved_mem(phys_addr_t base, phys_addr_t size,
 170                                  unsigned int order_per_bit,
 171                                  struct cma **res_cma)
 172 {
 173         struct cma *cma;
 174         phys_addr_t alignment;
 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         /* ensure minimal alignment required by mm core */
 186         alignment = PAGE_SIZE <<
 187                         max_t(unsigned long, MAX_ORDER - 1, pageblock_order);
 188
 189         /* alignment should be aligned with order_per_bit */
 190         if (!IS_ALIGNED(alignment >> PAGE_SHIFT, 1 << order_per_bit))
 191                 return -EINVAL;
 192
 193         if (ALIGN(base, alignment) != base || ALIGN(size, alignment) != size)
 194                 return -EINVAL;
 195
 196         /*
 197          * Each reserved area must be initialised later, when more kernel
 198          * subsystems (like slab allocator) are available.
 199          */
 200         cma = &cma_areas[cma_area_count];
 201         cma->base_pfn = PFN_DOWN(base);
 202         cma->count = size >> PAGE_SHIFT;
 203         cma->order_per_bit = order_per_bit;
 204         *res_cma = cma;
 205         cma_area_count++;
 206         totalcma_pages += (size / PAGE_SIZE);
 207
 208         return 0;
 209 }
 210
 211 /**
 212  * cma_declare_contiguous() - reserve custom contiguous area
 213  * @base: Base address of the reserved area optional, use 0 for any
 214  * @size: Size of the reserved area (in bytes),
 215  * @limit: End address of the reserved memory (optional, 0 for any).
 216  * @alignment: Alignment for the CMA area, should be power of 2 or zero
 217  * @order_per_bit: Order of pages represented by one bit on bitmap.
 218  * @fixed: hint about where to place the reserved area
 219  * @res_cma: Pointer to store the created cma region.
 220  *
 221  * This function reserves memory from early allocator. It should be
 222  * called by arch specific code once the early allocator (memblock or bootmem)
 223  * has been activated and all other subsystems have already allocated/reserved
 224  * memory. This function allows to create custom reserved areas.
 225  *
 226  * If @fixed is true, reserve contiguous area at exactly @base.  If false,
 227  * reserve in range from @base to @limit.
 228  */
 229 int __init cma_declare_contiguous(phys_addr_t base,
 230                         phys_addr_t size, phys_addr_t limit,
 231                         phys_addr_t alignment, unsigned int order_per_bit,
 232                         bool fixed, struct cma **res_cma)
 233 {
 234         phys_addr_t memblock_end = memblock_end_of_DRAM();
 235         phys_addr_t highmem_start;
 236         int ret = 0;
 237
 238 #ifdef CONFIG_X86
 239         /*
 240          * high_memory isn't direct mapped memory so retrieving its physical
 241          * address isn't appropriate.  But it would be useful to check the
 242          * physical address of the highmem boundary so it's justifiable to get
 243          * the physical address from it.  On x86 there is a validation check for
 244          * this case, so the following workaround is needed to avoid it.
 245          */
 246         highmem_start = __pa_nodebug(high_memory);
 247 #else
 248         highmem_start = __pa(high_memory);
 249 #endif
 250         pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
 251                 __func__, &size, &base, &limit, &alignment);
 252
 253         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 254                 pr_err("Not enough slots for CMA reserved regions!\n");
 255                 return -ENOSPC;
 256         }
 257
 258         if (!size)
 259                 return -EINVAL;
 260
 261         if (alignment && !is_power_of_2(alignment))
 262                 return -EINVAL;
 263
 264         /*
 265          * Sanitise input arguments.
 266          * Pages both ends in CMA area could be merged into adjacent unmovable
 267          * migratetype page by page allocator's buddy algorithm. In the case,
 268          * you couldn't get a contiguous memory, which is not what we want.
 269          */
 270         alignment = max(alignment,  (phys_addr_t)PAGE_SIZE <<
 271                           max_t(unsigned long, MAX_ORDER - 1, pageblock_order));
 272         base = ALIGN(base, alignment);
 273         size = ALIGN(size, alignment);
 274         limit &= ~(alignment - 1);
 275
 276         if (!base)
 277                 fixed = false;
 278
 279         /* size should be aligned with order_per_bit */
 280         if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit))
 281                 return -EINVAL;
 282
 283         /*
 284          * If allocating at a fixed base the request region must not cross the
 285          * low/high memory boundary.
 286          */
 287         if (fixed && base < highmem_start && base + size > highmem_start) {
 288                 ret = -EINVAL;
 289                 pr_err("Region at %pa defined on low/high memory boundary (%pa)\n",
 290                         &base, &highmem_start);
 291                 goto err;
 292         }
 293
 294         /*
 295          * If the limit is unspecified or above the memblock end, its effective
 296          * value will be the memblock end. Set it explicitly to simplify further
 297          * checks.
 298          */
 299         if (limit == 0 || limit > memblock_end)
 300                 limit = memblock_end;
 301
 302         /* Reserve memory */
 303         if (fixed) {
 304                 if (memblock_is_region_reserved(base, size) ||
 305                     memblock_reserve(base, size) < 0) {
 306                         ret = -EBUSY;
 307                         goto err;
 308                 }
 309         } else {
 310                 phys_addr_t addr = 0;
 311
 312                 /*
 313                  * All pages in the reserved area must come from the same zone.
 314                  * If the requested region crosses the low/high memory boundary,
 315                  * try allocating from high memory first and fall back to low
 316                  * memory in case of failure.
 317                  */
 318                 if (base < highmem_start && limit > highmem_start) {
 319                         addr = memblock_alloc_range(size, alignment,
 320                                                     highmem_start, limit,
 321                                                     MEMBLOCK_NONE);
 322                         limit = highmem_start;
 323                 }
 324
 325                 if (!addr) {
 326                         addr = memblock_alloc_range(size, alignment, base,
 327                                                     limit,
 328                                                     MEMBLOCK_NONE);
 329                         if (!addr) {
 330                                 ret = -ENOMEM;
 331                                 goto err;
 332                         }
 333                 }
 334
 335                 /*
 336                  * kmemleak scans/reads tracked objects for pointers to other
 337                  * objects but this address isn't mapped and accessible
 338                  */
 339                 kmemleak_ignore(phys_to_virt(addr));
 340                 base = addr;
 341         }
 342
 343         ret = cma_init_reserved_mem(base, size, order_per_bit, res_cma);
 344         if (ret)
 345                 goto err;
 346
 347         pr_info("Reserved %ld MiB at %pa\n", (unsigned long)size / SZ_1M,
 348                 &base);
 349         return 0;
 350
 351 err:
 352         pr_err("Failed to reserve %ld MiB\n", (unsigned long)size / SZ_1M);
 353         return ret;
 354 }
 355
 356 /**
 357  * cma_alloc() - allocate pages from contiguous area
 358  * @cma:   Contiguous memory region for which the allocation is performed.
 359  * @count: Requested number of pages.
 360  * @align: Requested alignment of pages (in PAGE_SIZE order).
 361  *
 362  * This function allocates part of contiguous memory on specific
 363  * contiguous memory area.
 364  */
 365 struct page *cma_alloc(struct cma *cma, size_t count, unsigned int align)
 366 {
 367         unsigned long mask, offset;
 368         unsigned long pfn = -1;
 369         unsigned long start = 0;
 370         unsigned long bitmap_maxno, bitmap_no, bitmap_count;
 371         struct page *page = NULL;
 372         int ret;
 373
 374         if (!cma || !cma->count)
 375                 return NULL;
 376
 377         pr_debug("%s(cma %p, count %zu, align %d)\n", __func__, (void *)cma,
 378                  count, align);
 379
 380         if (!count)
 381                 return NULL;
 382
 383         mask = cma_bitmap_aligned_mask(cma, align);
 384         offset = cma_bitmap_aligned_offset(cma, align);
 385         bitmap_maxno = cma_bitmap_maxno(cma);
 386         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
 387
 388         for (;;) {
 389                 mutex_lock(&cma->lock);
 390                 bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
 391                                 bitmap_maxno, start, bitmap_count, mask,
 392                                 offset);
 393                 if (bitmap_no >= bitmap_maxno) {
 394                         mutex_unlock(&cma->lock);
 395                         break;
 396                 }
 397                 bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
 398                 /*
 399                  * It's safe to drop the lock here. We've marked this region for
 400                  * our exclusive use. If the migration fails we will take the
 401                  * lock again and unmark it.
 402                  */
 403                 mutex_unlock(&cma->lock);
 404
 405                 pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
 406                 mutex_lock(&cma_mutex);
 407                 ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA);
 408                 mutex_unlock(&cma_mutex);
 409                 if (ret == 0) {
 410                         page = pfn_to_page(pfn);
 411                         break;
 412                 }
 413
 414                 cma_clear_bitmap(cma, pfn, count);
 415                 if (ret != -EBUSY)
 416                         break;
 417
 418                 pr_debug("%s(): memory range at %p is busy, retrying\n",
 419                          __func__, pfn_to_page(pfn));
 420                 /* try again with a bit different memory target */
 421                 start = bitmap_no + mask + 1;
 422         }
 423
 424         trace_cma_alloc(pfn, page, count, align);
 425
 426         pr_debug("%s(): returned %p\n", __func__, page);
 427         return page;
 428 }
 429
 430 /**
 431  * cma_release() - release allocated pages
 432  * @cma:   Contiguous memory region for which the allocation is performed.
 433  * @pages: Allocated pages.
 434  * @count: Number of allocated pages.
 435  *
 436  * This function releases memory allocated by alloc_cma().
 437  * It returns false when provided pages do not belong to contiguous area and
 438  * true otherwise.
 439  */
 440 bool cma_release(struct cma *cma, const struct page *pages, unsigned int count)
 441 {
 442         unsigned long pfn;
 443
 444         if (!cma || !pages)
 445                 return false;
 446
 447         pr_debug("%s(page %p)\n", __func__, (void *)pages);
 448
 449         pfn = page_to_pfn(pages);
 450
 451         if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count)
 452                 return false;
 453
 454         VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
 455
 456         free_contig_range(pfn, count);
 457         cma_clear_bitmap(cma, pfn, count);
 458         trace_cma_release(pfn, pages, count);
 459
 460         return true;
 461 }