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                                              unsigned 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 the offset of the base PFN from the specified align_order.
  66  * The value returned is represented in order_per_bits.
  67  */
  68 static unsigned long cma_bitmap_aligned_offset(const struct cma *cma,
  69                                                unsigned int align_order)
  70 {
  71         return (cma->base_pfn & ((1UL << align_order) - 1))
  72                 >> cma->order_per_bit;
  73 }
  74
  75 static unsigned long cma_bitmap_pages_to_bits(const struct cma *cma,
  76                                               unsigned long pages)
  77 {
  78         return ALIGN(pages, 1UL << cma->order_per_bit) >> cma->order_per_bit;
  79 }
  80
  81 static void cma_clear_bitmap(struct cma *cma, unsigned long pfn,
  82                              unsigned int count)
  83 {
  84         unsigned long bitmap_no, bitmap_count;
  85
  86         bitmap_no = (pfn - cma->base_pfn) >> cma->order_per_bit;
  87         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
  88
  89         mutex_lock(&cma->lock);
  90         bitmap_clear(cma->bitmap, bitmap_no, bitmap_count);
  91         mutex_unlock(&cma->lock);
  92 }
  93
  94 static int __init cma_activate_area(struct cma *cma)
  95 {
  96         int bitmap_size = BITS_TO_LONGS(cma_bitmap_maxno(cma)) * sizeof(long);
  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 = kzalloc(bitmap_size, GFP_KERNEL);
 102
 103         if (!cma->bitmap)
 104                 return -ENOMEM;
 105
 106         WARN_ON_ONCE(!pfn_valid(pfn));
 107         zone = page_zone(pfn_to_page(pfn));
 108
 109         do {
 110                 unsigned j;
 111
 112                 base_pfn = pfn;
 113                 for (j = pageblock_nr_pages; j; --j, pfn++) {
 114                         WARN_ON_ONCE(!pfn_valid(pfn));
 115                         /*
 116                          * alloc_contig_range requires the pfn range
 117                          * specified to be in the same zone. Make this
 118                          * simple by forcing the entire CMA resv range
 119                          * to be in the same zone.
 120                          */
 121                         if (page_zone(pfn_to_page(pfn)) != zone)
 122                                 goto err;
 123                 }
 124                 init_cma_reserved_pageblock(pfn_to_page(base_pfn));
 125         } while (--i);
 126
 127         mutex_init(&cma->lock);
 128
 129 #ifdef CONFIG_CMA_DEBUGFS
 130         INIT_HLIST_HEAD(&cma->mem_head);
 131         spin_lock_init(&cma->mem_head_lock);
 132 #endif
 133
 134         return 0;
 135
 136 err:
 137         kfree(cma->bitmap);
 138         cma->count = 0;
 139         return -EINVAL;
 140 }
 141
 142 static int __init cma_init_reserved_areas(void)
 143 {
 144         int i;
 145
 146         for (i = 0; i < cma_area_count; i++) {
 147                 int ret = cma_activate_area(&cma_areas[i]);
 148
 149                 if (ret)
 150                         return ret;
 151         }
 152
 153         return 0;
 154 }
 155 core_initcall(cma_init_reserved_areas);
 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  * @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                                  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 required by mm core */
 183         alignment = PAGE_SIZE <<
 184                         max_t(unsigned long, MAX_ORDER - 1, pageblock_order);
 185
 186         /* alignment should be aligned with order_per_bit */
 187         if (!IS_ALIGNED(alignment >> PAGE_SHIFT, 1 << order_per_bit))
 188                 return -EINVAL;
 189
 190         if (ALIGN(base, alignment) != base || ALIGN(size, alignment) != size)
 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         cma->base_pfn = PFN_DOWN(base);
 199         cma->count = size >> PAGE_SHIFT;
 200         cma->order_per_bit = order_per_bit;
 201         *res_cma = cma;
 202         cma_area_count++;
 203         totalcma_pages += (size / PAGE_SIZE);
 204
 205         return 0;
 206 }
 207
 208 /**
 209  * cma_declare_contiguous() - reserve custom contiguous area
 210  * @base: Base address of the reserved area optional, use 0 for any
 211  * @size: Size of the reserved area (in bytes),
 212  * @limit: End address of the reserved memory (optional, 0 for any).
 213  * @alignment: Alignment for the CMA area, should be power of 2 or zero
 214  * @order_per_bit: Order of pages represented by one bit on bitmap.
 215  * @fixed: hint about where to place the reserved area
 216  * @res_cma: Pointer to store the created cma region.
 217  *
 218  * This function reserves memory from early allocator. It should be
 219  * called by arch specific code once the early allocator (memblock or bootmem)
 220  * has been activated and all other subsystems have already allocated/reserved
 221  * memory. This function allows to create custom reserved areas.
 222  *
 223  * If @fixed is true, reserve contiguous area at exactly @base.  If false,
 224  * reserve in range from @base to @limit.
 225  */
 226 int __init cma_declare_contiguous(phys_addr_t base,
 227                         phys_addr_t size, phys_addr_t limit,
 228                         phys_addr_t alignment, unsigned int order_per_bit,
 229                         bool fixed, struct cma **res_cma)
 230 {
 231         phys_addr_t memblock_end = memblock_end_of_DRAM();
 232         phys_addr_t highmem_start;
 233         int ret = 0;
 234
 235 #ifdef CONFIG_X86
 236         /*
 237          * high_memory isn't direct mapped memory so retrieving its physical
 238          * address isn't appropriate.  But it would be useful to check the
 239          * physical address of the highmem boundary so it's justifiable to get
 240          * the physical address from it.  On x86 there is a validation check for
 241          * this case, so the following workaround is needed to avoid it.
 242          */
 243         highmem_start = __pa_nodebug(high_memory);
 244 #else
 245         highmem_start = __pa(high_memory);
 246 #endif
 247         pr_debug("%s(size %pa, base %pa, limit %pa alignment %pa)\n",
 248                 __func__, &size, &base, &limit, &alignment);
 249
 250         if (cma_area_count == ARRAY_SIZE(cma_areas)) {
 251                 pr_err("Not enough slots for CMA reserved regions!\n");
 252                 return -ENOSPC;
 253         }
 254
 255         if (!size)
 256                 return -EINVAL;
 257
 258         if (alignment && !is_power_of_2(alignment))
 259                 return -EINVAL;
 260
 261         /*
 262          * Sanitise input arguments.
 263          * Pages both ends in CMA area could be merged into adjacent unmovable
 264          * migratetype page by page allocator's buddy algorithm. In the case,
 265          * you couldn't get a contiguous memory, which is not what we want.
 266          */
 267         alignment = max(alignment,  (phys_addr_t)PAGE_SIZE <<
 268                           max_t(unsigned long, MAX_ORDER - 1, pageblock_order));
 269         base = ALIGN(base, alignment);
 270         size = ALIGN(size, alignment);
 271         limit &= ~(alignment - 1);
 272
 273         if (!base)
 274                 fixed = false;
 275
 276         /* size should be aligned with order_per_bit */
 277         if (!IS_ALIGNED(size >> PAGE_SHIFT, 1 << order_per_bit))
 278                 return -EINVAL;
 279
 280         /*
 281          * If allocating at a fixed base the request region must not cross the
 282          * low/high memory boundary.
 283          */
 284         if (fixed && base < highmem_start && base + size > highmem_start) {
 285                 ret = -EINVAL;
 286                 pr_err("Region at %pa defined on low/high memory boundary (%pa)\n",
 287                         &base, &highmem_start);
 288                 goto err;
 289         }
 290
 291         /*
 292          * If the limit is unspecified or above the memblock end, its effective
 293          * value will be the memblock end. Set it explicitly to simplify further
 294          * checks.
 295          */
 296         if (limit == 0 || limit > memblock_end)
 297                 limit = memblock_end;
 298
 299         /* Reserve memory */
 300         if (fixed) {
 301                 if (memblock_is_region_reserved(base, size) ||
 302                     memblock_reserve(base, size) < 0) {
 303                         ret = -EBUSY;
 304                         goto err;
 305                 }
 306         } else {
 307                 phys_addr_t addr = 0;
 308
 309                 /*
 310                  * All pages in the reserved area must come from the same zone.
 311                  * If the requested region crosses the low/high memory boundary,
 312                  * try allocating from high memory first and fall back to low
 313                  * memory in case of failure.
 314                  */
 315                 if (base < highmem_start && limit > highmem_start) {
 316                         addr = memblock_alloc_range(size, alignment,
 317                                                     highmem_start, limit,
 318                                                     MEMBLOCK_NONE);
 319                         limit = highmem_start;
 320                 }
 321
 322                 if (!addr) {
 323                         addr = memblock_alloc_range(size, alignment, base,
 324                                                     limit,
 325                                                     MEMBLOCK_NONE);
 326                         if (!addr) {
 327                                 ret = -ENOMEM;
 328                                 goto err;
 329                         }
 330                 }
 331
 332                 /*
 333                  * kmemleak scans/reads tracked objects for pointers to other
 334                  * objects but this address isn't mapped and accessible
 335                  */
 336                 kmemleak_ignore_phys(addr);
 337                 base = addr;
 338         }
 339
 340         ret = cma_init_reserved_mem(base, size, order_per_bit, res_cma);
 341         if (ret)
 342                 goto free_mem;
 343
 344         pr_info("Reserved %ld MiB at %pa\n", (unsigned long)size / SZ_1M,
 345                 &base);
 346         return 0;
 347
 348 free_mem:
 349         memblock_free(base, size);
 350 err:
 351         pr_err("Failed to reserve %ld MiB\n", (unsigned long)size / SZ_1M);
 352         return ret;
 353 }
 354
 355 /**
 356  * cma_alloc() - allocate pages from contiguous area
 357  * @cma:   Contiguous memory region for which the allocation is performed.
 358  * @count: Requested number of pages.
 359  * @align: Requested alignment of pages (in PAGE_SIZE order).
 360  *
 361  * This function allocates part of contiguous memory on specific
 362  * contiguous memory area.
 363  */
 364 struct page *cma_alloc(struct cma *cma, size_t count, unsigned int align)
 365 {
 366         unsigned long mask, offset;
 367         unsigned long pfn = -1;
 368         unsigned long start = 0;
 369         unsigned long bitmap_maxno, bitmap_no, bitmap_count;
 370         struct page *page = NULL;
 371         int ret;
 372
 373         if (!cma || !cma->count)
 374                 return NULL;
 375
 376         pr_debug("%s(cma %p, count %zu, align %d)\n", __func__, (void *)cma,
 377                  count, align);
 378
 379         if (!count)
 380                 return NULL;
 381
 382         mask = cma_bitmap_aligned_mask(cma, align);
 383         offset = cma_bitmap_aligned_offset(cma, align);
 384         bitmap_maxno = cma_bitmap_maxno(cma);
 385         bitmap_count = cma_bitmap_pages_to_bits(cma, count);
 386
 387         if (bitmap_count > bitmap_maxno)
 388                 return NULL;
 389
 390         for (;;) {
 391                 mutex_lock(&cma->lock);
 392                 bitmap_no = bitmap_find_next_zero_area_off(cma->bitmap,
 393                                 bitmap_maxno, start, bitmap_count, mask,
 394                                 offset);
 395                 if (bitmap_no >= bitmap_maxno) {
 396                         mutex_unlock(&cma->lock);
 397                         break;
 398                 }
 399                 bitmap_set(cma->bitmap, bitmap_no, bitmap_count);
 400                 /*
 401                  * It's safe to drop the lock here. We've marked this region for
 402                  * our exclusive use. If the migration fails we will take the
 403                  * lock again and unmark it.
 404                  */
 405                 mutex_unlock(&cma->lock);
 406
 407                 pfn = cma->base_pfn + (bitmap_no << cma->order_per_bit);
 408                 mutex_lock(&cma_mutex);
 409                 ret = alloc_contig_range(pfn, pfn + count, MIGRATE_CMA);
 410                 mutex_unlock(&cma_mutex);
 411                 if (ret == 0) {
 412                         page = pfn_to_page(pfn);
 413                         break;
 414                 }
 415
 416                 cma_clear_bitmap(cma, pfn, count);
 417                 if (ret != -EBUSY)
 418                         break;
 419
 420                 pr_debug("%s(): memory range at %p is busy, retrying\n",
 421                          __func__, pfn_to_page(pfn));
 422                 /* try again with a bit different memory target */
 423                 start = bitmap_no + mask + 1;
 424         }
 425
 426         trace_cma_alloc(pfn, page, count, align);
 427
 428         pr_debug("%s(): returned %p\n", __func__, page);
 429         return page;
 430 }
 431
 432 /**
 433  * cma_release() - release allocated pages
 434  * @cma:   Contiguous memory region for which the allocation is performed.
 435  * @pages: Allocated pages.
 436  * @count: Number of allocated pages.
 437  *
 438  * This function releases memory allocated by alloc_cma().
 439  * It returns false when provided pages do not belong to contiguous area and
 440  * true otherwise.
 441  */
 442 bool cma_release(struct cma *cma, const struct page *pages, unsigned int count)
 443 {
 444         unsigned long pfn;
 445
 446         if (!cma || !pages)
 447                 return false;
 448
 449         pr_debug("%s(page %p)\n", __func__, (void *)pages);
 450
 451         pfn = page_to_pfn(pages);
 452
 453         if (pfn < cma->base_pfn || pfn >= cma->base_pfn + cma->count)
 454                 return false;
 455
 456         VM_BUG_ON(pfn + count > cma->base_pfn + cma->count);
 457
 458         free_contig_range(pfn, count);
 459         cma_clear_bitmap(cma, pfn, count);
 460         trace_cma_release(pfn, pages, count);
 461
 462         return true;
 463 }